Authors: Zehua Xia, Yuyang Wu, Yiyun Xia, Cam-Tu Nguyen
Abstract: Multi-hop question answering (QA) often requires sequential retrieval (multi-hop retrieval), where each hop retrieves missing knowledge based on information from previous hops. To facilitate more effective retrieval, we aim to distill knowledge from a posterior retrieval, which has access to posterior information like an answer, into a prior retrieval used during inference when such information is unavailable. Unfortunately, current methods for knowledge distillation in one-time retrieval are ineffective for multi-hop QA due to two issues: 1) Posterior information is often defined as the response (i.e. the answer), which may not clearly connect to the query without intermediate retrieval; and 2) The large knowledge gap between prior and posterior retrievals makes existing distillation methods unstable, even resulting in performance loss. As such, we propose MoPo (Momentum Posterior Regularization) with two key innovations: 1) Posterior information of one hop is defined as a query-focus summary from the golden knowledge of the previous and current hops; 2) We develop an effective training strategy where the posterior retrieval is updated along with the prior retrieval via momentum moving average method, allowing smoother and effective distillation. Experiments on HotpotQA and StrategyQA demonstrate that MoPo outperforms existing baselines in both retrieval and downstream QA tasks.
Authors: James Begin, Namit Agrawal, Eshan Singh, Yicheng Fu, Sean O'Brien, Vasu Sharma, Kevin Zhu
Abstract: LLMs have demonstrated remarkable proficiency in understanding tasks but continue to struggle with long-context comprehension, particularly with content located in the middle of extensive inputs. This limitation, known as the Lost-in-the-Middle (LITM) problem, hinders models from fully processing and utilizing information across lengthy contexts. To address this issue, we introduce pause-tuning, a technique that redistributes attention to enhance comprehension of long-context inputs. Our approach involves fine-tuning language models on datasets with artificially inserted pause tokens, which serve to segment the input into smaller, more manageable parts. We evaluate pause-tuning against alternative approaches using the Needle-in-a-Haystack benchmark, where models must retrieve information embedded within contexts of up to 128K tokens. Experimental results demonstrate significant performance gains, with the LLaMA 3.2 3B Instruct model and the LLaMA 3.1 8B Instruct model improving by 10.61% and 3.57% respectively on average, suggesting that pause-tuning successfully enhances attention redistribution and improves long-context retention. The code and data are available at https://anonymous.4open.science/r/LITM-PauseTokens-7357.
URLs: https://anonymous.4open.science/r/LITM-PauseTokens-7357.
Authors: Shaharukh Khan, Ayush Tarun, Ali Faraz, Palash Kamble, Vivek Dahiya, Praveen Pokala, Ashish Kulkarni, Chandra Khatri, Abhinav Ravi, Shubham Agarwal
Abstract: In this work, we provide the system description of our submission as part of the English to Lowres Multimodal Translation Task at the Workshop on Asian Translation (WAT2024). We introduce Chitranuvad, a multimodal model that effectively integrates Multilingual LLM and a vision module for Multimodal Translation. Our method uses a ViT image encoder to extract visual representations as visual token embeddings which are projected to the LLM space by an adapter layer and generates translation in an autoregressive fashion. We participated in all the three tracks (Image Captioning, Text only and Multimodal translation tasks) for Indic languages (ie. English translation to Hindi, Bengali and Malyalam) and achieved SOTA results for Hindi in all of them on the Challenge set while remaining competitive for the other languages in the shared task.
Authors: Zicheng Cai, Yaohua Tang, Yutao Lai, Hua Wang, Zhi Chen, Hao Chen
Abstract: We introduce SEKI, a novel large language model (LLM)-based neural architecture search (NAS) method. Inspired by the chain-of-thought (CoT) paradigm in modern LLMs, SEKI operates in two key stages: self-evolution and knowledge distillation. In the self-evolution stage, LLMs initially lack sufficient reference examples, so we implement an iterative refinement mechanism that enhances architectures based on performance feedback. Over time, this process accumulates a repository of high-performance architectures. In the knowledge distillation stage, LLMs analyze common patterns among these architectures to generate new, optimized designs. Combining these two stages, SEKI greatly leverages the capacity of LLMs on NAS and without requiring any domain-specific data. Experimental results show that SEKI achieves state-of-the-art (SOTA) performance across various datasets and search spaces while requiring only 0.05 GPU-days, outperforming existing methods in both efficiency and accuracy. Furthermore, SEKI demonstrates strong generalization capabilities, achieving SOTA-competitive results across multiple tasks.
Authors: Mateusz Idziejczak, Vasyl Korzavatykh, Mateusz Stawicki, Andrii Chmutov, Marcin Korcz, Iwo B{\l}\k{a}dek, Dariusz Brzezinski
Abstract: The proliferation of large language models (LLMs) and autonomous AI agents has raised concerns about their potential for automated persuasion and social influence. While existing research has explored isolated instances of LLM-based manipulation, systematic evaluations of persuasion capabilities across different models remain limited. In this paper, we present an Among Us-inspired game framework for assessing LLM deception skills in a controlled environment. The proposed framework makes it possible to compare LLM models by game statistics, as well as quantify in-game manipulation according to 25 persuasion strategies from social psychology and rhetoric. Experiments between 8 popular language models of different types and sizes demonstrate that all tested models exhibit persuasive capabilities, successfully employing 22 of the 25 anticipated techniques. We also find that larger models do not provide any persuasion advantage over smaller models and that longer model outputs are negatively correlated with the number of games won. Our study provides insights into the deception capabilities of LLMs, as well as tools and data for fostering future research on the topic.
Authors: Jennifer Hu, Felix Sosa, Tomer Ullman
Abstract: Some things are impossible, but some things may be even more impossible than impossible. Levitating a feather using one's mind is impossible in our world, but fits into our intuitive theories of possible worlds, whereas levitating a feather using the number five cannot be conceived in any possible world ("inconceivable"). While prior work has examined the distinction between improbable and impossible events, there has been little empirical research on inconceivability. Here, we investigate whether people maintain a distinction between impossibility and inconceivability, and how such distinctions might be made. We find that people can readily distinguish the impossible from the inconceivable, using categorization studies similar to those used to investigate the differences between impossible and improbable (Experiment 1). However, this distinction is not explained by people's subjective ratings of event likelihood, which are near zero and indistinguishable between impossible and inconceivable event descriptions (Experiment 2). Finally, we ask whether the probabilities assigned to event descriptions by statistical language models (LMs) can be used to separate modal categories, and whether these probabilities align with people's ratings (Experiment 3). We find high-level similarities between people and LMs: both distinguish among impossible and inconceivable event descriptions, and LM-derived string probabilities predict people's ratings of event likelihood across modal categories. Our findings suggest that fine-grained knowledge about exceedingly rare events (i.e., the impossible and inconceivable) may be learned via statistical learning over linguistic forms, yet leave open the question of whether people represent the distinction between impossible and inconceivable as a difference not of degree, but of kind.
Authors: Tianyi Lorena Yan, Robin Jia
Abstract: To answer one-to-many factual queries (e.g., listing cities of a country), a language model (LM) must simultaneously recall knowledge and avoid repeating previous answers. How are these two subtasks implemented and integrated internally? Across multiple datasets and models, we identify a promote-then-suppress mechanism: the model first recalls all answers, and then suppresses previously generated ones. Specifically, LMs use both the subject and previous answer tokens to perform knowledge recall, with attention propagating subject information and MLPs promoting the answers. Then, attention attends to and suppresses previous answer tokens, while MLPs amplify the suppression signal. Our mechanism is corroborated by extensive experimental evidence: in addition to using early decoding and causal tracing, we analyze how components use different tokens by introducing both \emph{Token Lens}, which decodes aggregated attention updates from specified tokens, and a knockout method that analyzes changes in MLP outputs after removing attention to specified tokens. Overall, we provide new insights into how LMs' internal components interact with different input tokens to support complex factual recall. Code is available at https://github.com/Lorenayannnnn/how-lms-answer-one-to-many-factual-queries.
URLs: https://github.com/Lorenayannnnn/how-lms-answer-one-to-many-factual-queries.
Authors: Jun Hou, Lucy Lu Wang
Abstract: Explainable AI (XAI) techniques are necessary to help clinicians make sense of AI predictions and integrate predictions into their decision-making workflow. In this work, we conduct a survey study to understand clinician preference among different XAI techniques when they are used to interpret model predictions over text-based EHR data. We implement four XAI techniques (LIME, Attention-based span highlights, exemplar patient retrieval, and free-text rationales generated by LLMs) on an outcome prediction model that uses ICU admission notes to predict a patient's likelihood of experiencing in-hospital mortality. Using these XAI implementations, we design and conduct a survey study of 32 practicing clinicians, collecting their feedback and preferences on the four techniques. We synthesize our findings into a set of recommendations describing when each of the XAI techniques may be more appropriate, their potential limitations, as well as recommendations for improvement.
Authors: Jonathan Tonglet, Tinne Tuytelaars, Marie-Francine Moens, Iryna Gurevych
Abstract: We assess the vulnerability of multimodal large language models to misleading visualizations - charts that distort the underlying data using techniques such as truncated or inverted axes, leading readers to draw inaccurate conclusions that may support misinformation or conspiracy theories. Our analysis shows that these distortions severely harm multimodal large language models, reducing their question-answering accuracy to the level of the random baseline. To mitigate this vulnerability, we introduce six inference-time methods to improve performance of MLLMs on misleading visualizations while preserving their accuracy on non-misleading ones. The most effective approach involves (1) extracting the underlying data table and (2) using a text-only large language model to answer questions based on the table. This method improves performance on misleading visualizations by 15.4 to 19.6 percentage points.
Authors: Julius Broomfield, Kartik Sharma, Srijan Kumar
Abstract: Large language models (LLMs) have recently demonstrated remarkable advancements in embodying diverse personas, enhancing their effectiveness as conversational agents and virtual assistants. Consequently, LLMs have made significant strides in processing and integrating multimodal information. However, even though human personas can be expressed in both text and image, the extent to which the modality of a persona impacts the embodiment by the LLM remains largely unexplored. In this paper, we investigate how do different modalities influence the expressiveness of personas in multimodal LLMs. To this end, we create a novel modality-parallel dataset of 40 diverse personas varying in age, gender, occupation, and location. This consists of four modalities to equivalently represent a persona: image-only, text-only, a combination of image and small text, and typographical images, where text is visually stylized to convey persona-related attributes. We then create a systematic evaluation framework with 60 questions and corresponding metrics to assess how well LLMs embody each persona across its attributes and scenarios. Comprehensive experiments on $5$ multimodal LLMs show that personas represented by detailed text show more linguistic habits, while typographical images often show more consistency with the persona. Our results reveal that LLMs often overlook persona-specific details conveyed through images, highlighting underlying limitations and paving the way for future research to bridge this gap. We release the data and code at https://github.com/claws-lab/persona-modality .
Authors: Soumyabrata Chaudhuri, Pranav Purkar, Ritwik Raghav, Shubhojit Mallick, Manish Gupta, Abhik Jana, Shreya Ghosh
Abstract: Recent advancements in probing Large Language Models (LLMs) have explored their latent potential as personalized travel planning agents, yet existing benchmarks remain limited in real world applicability. Existing datasets, such as TravelPlanner and TravelPlanner+, suffer from semi synthetic data reliance, spatial inconsistencies, and a lack of key travel constraints, making them inadequate for practical itinerary generation. To address these gaps, we introduce TripCraft, a spatiotemporally coherent travel planning dataset that integrates real world constraints, including public transit schedules, event availability, diverse attraction categories, and user personas for enhanced personalization. To evaluate LLM generated plans beyond existing binary validation methods, we propose five continuous evaluation metrics, namely Temporal Meal Score, Temporal Attraction Score, Spatial Score, Ordering Score, and Persona Score which assess itinerary quality across multiple dimensions. Our parameter informed setting significantly enhances meal scheduling, improving the Temporal Meal Score from 61% to 80% in a 7 day scenario. TripCraft establishes a new benchmark for LLM driven personalized travel planning, offering a more realistic, constraint aware framework for itinerary generation. Dataset and Codebase will be made publicly available upon acceptance.
Authors: Emily Ross, Yuval Kansal, Jake Renzella, Alexandra Vassar, Andrew Taylor
Abstract: Large language models (LLMs) are increasingly being explored in higher education, yet their effectiveness as teaching agents remains underexamined. In this paper, we present the development of GuideLM, a fine-tuned LLM designed for programming education. GuideLM has been integrated into the Debugging C Compiler (DCC), an educational C compiler that leverages LLMs to generate pedagogically sound error explanations. Previously, DCC relied on off-the-shelf OpenAI models, which, while accurate, often over-assisted students by directly providing solutions despite contrary prompting. To address this, we employed supervised fine-tuning (SFT) on a dataset of 528 student-question/teacher-answer pairs, creating two models: GuideLM and GuideLM-mini, fine-tuned on ChatGPT-4o and 4o-mini, respectively. We conducted an expert analysis of 400 responses per model, comparing their pedagogical effectiveness against base OpenAI models. Our evaluation, grounded in constructivism and cognitive load theory, assessed factors such as conceptual scaffolding, clarity, and Socratic guidance. Results indicate that GuideLM and GuideLM-mini improve pedagogical performance, with an 8% increase in Socratic guidance and a 58% improvement in economy of words compared to GPT-4o. However, this refinement comes at the cost of a slight reduction in general accuracy. While further work is needed, our findings suggest that fine-tuning LLMs with targeted datasets is a promising approach for developing models better suited to educational contexts.
Authors: Achuth Chandrasekhar, Omid Barati Farimani, Olabode T. Ajenifujah, Janghoon Ock, Amir Barati Farimani
Abstract: This paper presents the development and application of a Large Language Model Retrieval-Augmented Generation (LLM-RAG) system tailored for nanotechnology research. The system leverages the capabilities of a sophisticated language model to serve as an intelligent research assistant, enhancing the efficiency and comprehensiveness of literature reviews in the nanotechnology domain. Central to this LLM-RAG system is its advanced query backend retrieval mechanism, which integrates data from multiple reputable sources. The system retrieves relevant literature by utilizing Google Scholar's advanced search, and scraping open-access papers from Elsevier, Springer Nature, and ACS Publications. This multifaceted approach ensures a broad and diverse collection of up-to-date scholarly articles and papers. The proposed system demonstrates significant potential in aiding researchers by providing a streamlined, accurate, and exhaustive literature retrieval process, thereby accelerating research advancements in nanotechnology. The effectiveness of the LLM-RAG system is validated through rigorous testing, illustrating its capability to significantly reduce the time and effort required for comprehensive literature reviews, while maintaining high accuracy, query relevance and outperforming standard, publicly available LLMS.
Authors: Botond Barta, Endre Hamerlik, Mil\'an Konor Nyist, Judit \'Acs
Abstract: We present HuAMR, the first Abstract Meaning Representation (AMR) dataset and a suite of large language model-based AMR parsers for Hungarian, targeting the scarcity of semantic resources for non-English languages. To create HuAMR, we employed Llama-3.1-70B to automatically generate silver-standard AMR annotations, which we then refined manually to ensure quality. Building on this dataset, we investigate how different model architectures - mT5 Large and Llama-3.2-1B - and fine-tuning strategies affect AMR parsing performance. While incorporating silver-standard AMRs from Llama-3.1-70B into the training data of smaller models does not consistently boost overall scores, our results show that these techniques effectively enhance parsing accuracy on Hungarian news data (the domain of HuAMR). We evaluate our parsers using Smatch scores and confirm the potential of HuAMR and our parsers for advancing semantic parsing research.
Authors: Hong Chen, Misha Teplitskiy, David Jurgens
Abstract: Academic citations are widely used for evaluating research and tracing knowledge flows. Such uses typically rely on raw citation counts and neglect variability in citation types. In particular, citations can vary in their fidelity as original knowledge from cited studies may be paraphrased, summarized, or reinterpreted, possibly wrongly, leading to variation in how much information changes from cited to citing paper. In this study, we introduce a computational pipeline to quantify citation fidelity at scale. Using full texts of papers, the pipeline identifies citations in citing papers and the corresponding claims in cited papers, and applies supervised models to measure fidelity at the sentence level. Analyzing a large-scale multi-disciplinary dataset of approximately 13 million citation sentence pairs, we find that citation fidelity is higher when authors cite papers that are 1) more recent and intellectually close, 2) more accessible, and 3) the first author has a lower H-index and the author team is medium-sized. Using a quasi-experiment, we establish the "telephone effect" - when citing papers have low fidelity to the original claim, future papers that cite the citing paper and the original have lower fidelity to the original. Our work reveals systematic differences in citation fidelity, underscoring the limitations of analyses that rely on citation quantity alone and the potential for distortion of evidence.
Authors: Juntai Cao, Xiang Zhang, Raymond Li, Chuyuan Li, Shafiq Joty, Giuseppe Carenini
Abstract: Recent advances in test-time scaling have shown promising results in improving Large Language Models (LLMs) performance through strategic computation allocation during inference. While this approach has demonstrated strong performance improvements in logical and mathematical reasoning tasks, its application to natural language generation (NLG), especially summarization, has yet to be explored. Multi-Document Summarization (MDS) is a challenging task that focuses on extracting and synthesizing useful information from multiple lengthy documents. Unlike reasoning tasks, MDS requires a more nuanced approach to prompt design and ensemble, as there is no "best" prompt to satisfy diverse summarization requirements. To address this, we propose a novel framework that leverages inference-time scaling for this task. Precisely, we take prompt ensemble approach by leveraging various prompt to first generate candidate summaries and then ensemble them with an aggregator to produce a refined summary. We also introduce two new evaluation metrics: Consistency-Aware Preference (CAP) score and LLM Atom-Content-Unit (ACU) score, to enhance LLM's contextual understanding while mitigating its positional bias. Extensive experiments demonstrate the effectiveness of our approach in improving summary quality while identifying and analyzing the scaling boundaries in summarization tasks.
Authors: Nguyen Xuan Thanh, Anh Duc Le, Quyen Tran, Thanh-Thien Le, Linh Ngo Van, Thien Huu Nguyen
Abstract: Few-shot Continual Relation Extraction is a crucial challenge for enabling AI systems to identify and adapt to evolving relationships in dynamic real-world domains. Traditional memory-based approaches often overfit to limited samples, failing to reinforce old knowledge, with the scarcity of data in few-shot scenarios further exacerbating these issues by hindering effective data augmentation in the latent space. In this paper, we propose a novel retrieval-based solution, starting with a large language model to generate descriptions for each relation. From these descriptions, we introduce a bi-encoder retrieval training paradigm to enrich both sample and class representation learning. Leveraging these enhanced representations, we design a retrieval-based prediction method where each sample "retrieves" the best fitting relation via a reciprocal rank fusion score that integrates both relation description vectors and class prototypes. Extensive experiments on multiple datasets demonstrate that our method significantly advances the state-of-the-art by maintaining robust performance across sequential tasks, effectively addressing catastrophic forgetting.
Authors: J\k{e}drzej Warczy\'nski, Mateusz Lango, Ondrej Dusek
Abstract: We introduce a simple approach that uses a large language model (LLM) to automatically implement a fully interpretable rule-based data-to-text system in pure Python. Experimental evaluation on the WebNLG dataset showed that such a constructed system produces text of better quality (according to the BLEU and BLEURT metrics) than the same LLM prompted to directly produce outputs, and produces fewer hallucinations than a BART language model fine-tuned on the same data. Furthermore, at runtime, the approach generates text in a fraction of the processing time required by neural approaches, using only a single CPU
Authors: Seungah Son, Andrez Saurez, Dongsoo Har
Abstract: While pre-trained language models excel at semantic understanding, they often struggle to capture nuanced affective information critical for affective recognition tasks. To address these limitations, we propose a novel framework for enhancing emotion-aware embeddings in transformer-based models. Our approach introduces a continuous valence-arousal labeling system to guide contrastive learning, which captures subtle and multi-dimensional emotional nuances more effectively. Furthermore, we employ a dynamic token perturbation mechanism, using gradient-based saliency to focus on sentiment-relevant tokens, improving model sensitivity to emotional cues. The experimental results demonstrate that the proposed framework outperforms existing methods, achieving up to 15.5% improvement in the emotion classification benchmark, highlighting the importance of employing continuous labels. This improvement demonstrates that the proposed framework is effective in affective representation learning and enables precise and contextually relevant emotional understanding.
Authors: Ayana Niwa, Masahiro Kaneko, Kentaro Inui
Abstract: Large language models (LLMs) can exhibit advanced reasoning yet still generate incorrect answers. We hypothesize that such errors frequently stem from spurious beliefs, propositions the model internally considers true but are incorrect. To address this, we propose a method to rectify the belief space by suppressing these spurious beliefs while simultaneously enhancing true ones, thereby enabling more reliable inferences. Our approach first identifies the beliefs that lead to incorrect or correct answers by prompting the model to generate textual explanations, using our Forward-Backward Beam Search (FBBS). We then apply unlearning to suppress the identified spurious beliefs and enhance the true ones, effectively rectifying the model's belief space. Empirical results on multiple QA datasets and LLMs show that our method corrects previously misanswered questions without harming overall model performance. Furthermore, our approach yields improved generalization on unseen data, suggesting that rectifying a model's belief space is a promising direction for mitigating errors and enhancing overall reliability.
Authors: Haitao Li, Yifan Chen, Yiran Hu, Qingyao Ai, Junjie Chen, Xiaoyu Yang, Jianhui Yang, Yueyue Wu, Zeyang Liu, Yiqun Liu
Abstract: Retrieval-augmented generation (RAG) has proven highly effective in improving large language models (LLMs) across various domains. However, there is no benchmark specifically designed to assess the effectiveness of RAG in the legal domain, which restricts progress in this area. To fill this gap, we propose LexRAG, the first benchmark to evaluate RAG systems for multi-turn legal consultations. LexRAG consists of 1,013 multi-turn dialogue samples and 17,228 candidate legal articles. Each sample is annotated by legal experts and consists of five rounds of progressive questioning. LexRAG includes two key tasks: (1) Conversational knowledge retrieval, requiring accurate retrieval of relevant legal articles based on multi-turn context. (2) Response generation, focusing on producing legally sound answers. To ensure reliable reproducibility, we develop LexiT, a legal RAG toolkit that provides a comprehensive implementation of RAG system components tailored for the legal domain. Additionally, we introduce an LLM-as-a-judge evaluation pipeline to enable detailed and effective assessment. Through experimental analysis of various LLMs and retrieval methods, we reveal the key limitations of existing RAG systems in handling legal consultation conversations. LexRAG establishes a new benchmark for the practical application of RAG systems in the legal domain, with its code and data available at https://github.com/CSHaitao/LexRAG.
Authors: Colleen Gilhuly, Haleh Shahzad
Abstract: Text summarizing is a critical Natural Language Processing (NLP) task with applications ranging from information retrieval to content generation. Large Language Models (LLMs) have shown remarkable promise in generating fluent abstractive summaries but they can produce hallucinated details not grounded in the source text. Regardless of the method of generating a summary, high quality automated evaluations remain an open area of investigation. This paper embarks on an exploration of text summarization with a diverse set of techniques, including TextRank, BART, Mistral-7B-Instruct, and OpenAI GPT-3.5-Turbo. The generated summaries are evaluated using traditional metrics such as the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) Score and Bidirectional Encoder Representations from Transformers (BERT) Score, as well as LLM-powered evaluation methods that directly assess a generated summary's consistency with the source text. We introduce a meta evaluation score which directly assesses the performance of the LLM evaluation system (prompt + model). We find that that all summarization models produce consistent summaries when tested on the XL-Sum dataset, exceeding the consistency of the reference summaries.
Authors: Radhika Kapoor, Sang T. Truong, Nick Haber, Maria Araceli Ruiz-Primo, Benjamin W. Domingue
Abstract: Prediction of item difficulty based on its text content is of substantial interest. In this paper, we focus on the related problem of recovering IRT-based difficulty when the data originally reported item p-value (percent correct responses). We model this item difficulty using a repository of reading passages and student data from US standardized tests from New York and Texas for grades 3-8 spanning the years 2017-23. This repository is annotated with meta-data on (1) linguistic features of the reading items, (2) test features of the passage, and (3) context features. A penalized regression prediction model with all these features can predict item difficulty with RMSE 0.52 compared to baseline RMSE of 0.92, and with a correlation of 0.77 between true and predicted difficulty. We supplement these features with embeddings from LLMs (ModernBERT, BERT, and LlAMA), which marginally improve item difficulty prediction. When models use only item linguistic features or LLM embeddings, prediction performance is similar, which suggests that only one of these feature categories may be required. This item difficulty prediction model can be used to filter and categorize reading items and will be made publicly available for use by other stakeholders.
Authors: Zixuan Gong, Jiaye Teng, Yong Liu
Abstract: Transformers may exhibit two-stage training dynamics during the real-world training process. For instance, when training GPT-2 on the Counterfact dataset, the answers progress from syntactically incorrect to syntactically correct to semantically correct. However, existing theoretical analyses hardly account for this two-stage phenomenon. In this paper, we theoretically demonstrate how such two-stage training dynamics occur in transformers. Specifically, we analyze the dynamics of transformers using feature learning techniques under in-context learning regimes, based on a disentangled two-type feature structure. Such disentanglement of feature structure is general in practice, e.g., natural languages contain syntax and semantics, and proteins contain primary and secondary structures. To our best known, this is the first rigorous result regarding a two-stage optimization process in transformers. Additionally, a corollary indicates that such a two-stage process is closely related to the spectral properties of the attention weights, which accords well with empirical findings.
Authors: Gibson Nkhata, Susan Gauch, Usman Anjum, Justin Zhan
Abstract: Sentiment Analysis (SA) is instrumental in understanding peoples viewpoints facilitating social media monitoring recognizing products and brands and gauging customer satisfaction. Consequently SA has evolved into an active research domain within Natural Language Processing (NLP). Many approaches outlined in the literature devise intricate frameworks aimed at achieving high accuracy, focusing exclusively on either binary sentiment classification or fine-grained sentiment classification. In this paper our objective is to fine-tune the pre-trained BERT model with Bidirectional LSTM (BiLSTM) to enhance both binary and fine-grained SA specifically for movie reviews. Our approach involves conducting sentiment classification for each review followed by computing the overall sentiment polarity across all reviews. We present our findings on binary classification as well as fine-grained classification utilizing benchmark datasets. Additionally we implement and assess two accuracy improvement techniques Synthetic Minority Oversampling Technique (SMOTE) and NLP Augmenter (NLPAUG) to bolster the models generalization in fine-grained sentiment classification. Finally a heuristic algorithm is employed to calculate the overall polarity of predicted reviews from the BERT+BiLSTM output vector. Our approach performs comparably with state-of-the-art (SOTA) techniques in both classifications. For instance in binary classification we achieve 97.67% accuracy surpassing the leading SOTA model NB-weighted-BON+dv-cosine by 0.27% on the renowned IMDb dataset. Conversely for five-class classification on SST-5 while the top SOTA model RoBERTa+large+Self-explaining attains 55.5% accuracy our model achieves 59.48% accuracy surpassing the BERT-large baseline by 3.6%.
Authors: Yingbing Huang, Deming Chen, Abhishek K. Umrawal
Abstract: While large language models (LLMs) have made significant strides in generating coherent and contextually relevant text, they often function as opaque black boxes, trained on vast unlabeled datasets with statistical objectives, lacking an interpretable framework for responsible control. In this paper, we introduce JAM (Just A Move), a novel framework that interprets and controls text generation by integrating cause-effect analysis within the latent space of LLMs. Based on our observations, we uncover the inherent causality in LLM generation, which is critical for producing responsible and realistic outputs. Moreover, we explore latent vectors as fundamental components in LLM architectures, aiming to understand and manipulate them for more effective and efficient controllable text generation. We evaluate our framework using a range of tools, including the HHH criteria, toxicity reduction benchmarks, and GPT-4 alignment measures. Our results show that JAM achieves up to a 22% improvement over previous Controllable Text Generation (CTG) methods across multiple quantitative metrics and human-centric evaluations. Furthermore, JAM demonstrates greater computational efficiency compared to other CTG methods. These results highlight the effectiveness and efficiency of JAM for responsible and realistic text generation, paving the way for more interpretable and controllable models.
Authors: Yifei Duan, Raphael Shang, Deng Liang, Yongqiang Cai
Abstract: Language models can be viewed as functions that embed text into Euclidean space, where the quality of the embedding vectors directly determines model performance, training such neural networks involves various uncertainties. This paper focuses on improving the performance of pre-trained language models in zero-shot settings through a simple and easily implementable method. We propose a novel backward attention mechanism to enhance contextual information encoding. Evaluated on the Chinese Massive Text Embedding Benchmark (C-MTEB), our approach achieves significant improvements across multiple tasks, providing valuable insights for advancing zero-shot learning capabilities.
Authors: Heejin Do, Sangwon Ryu, Gary Geunbae Lee
Abstract: Multi-trait automated essay scoring (AES) systems provide a fine-grained evaluation of an essay's diverse aspects. While they excel in scoring, prior systems fail to explain why specific trait scores are assigned. This lack of transparency leaves instructors and learners unconvinced of the AES outputs, hindering their practical use. To address this, we propose a self-explainable Rationale-Driven Multi-trait automated Essay scoring (RaDME) framework. RaDME leverages the reasoning capabilities of large language models (LLMs) by distilling them into a smaller yet effective scorer. This more manageable student model is optimized to sequentially generate a trait score followed by the corresponding rationale, thereby inherently learning to select a more justifiable score by considering the subsequent rationale during training. Our findings indicate that while LLMs underperform in direct AES tasks, they excel in rationale generation when provided with precise numerical scores. Thus, RaDME integrates the superior reasoning capacities of LLMs into the robust scoring accuracy of an optimized smaller model. Extensive experiments demonstrate that RaDME achieves both accurate and adequate reasoning while supporting high-quality multi-trait scoring, significantly enhancing the transparency of AES.
Authors: Jiaqi Bai, Hongcheng Guo, Zhongyuan Peng, Jian Yang, Zhoujun Li, Mohan Li, Zhihong Tian
Abstract: Large vision-language models show tremendous potential in understanding visual information through human languages. However, they are prone to suffer from object hallucination, i.e., the generated image descriptions contain objects that do not exist in the image. In this paper, we reveal that object hallucination can be attributed to overconfidence in irrelevant visual features when soft visual tokens map to the LLM's word embedding space. Specifically, by figuring out the semantic similarity between visual tokens and LLM's word embedding, we observe that the smoothness of similarity distribution strongly correlates with the emergence of object hallucinations. To mitigate hallucinations, we propose using the Variational Information Bottleneck (VIB) to alleviate overconfidence by introducing stochastic noise, facilitating the constraining of irrelevant information. Furthermore, we propose an entropy-based noise-controlling strategy to enable the injected noise to be adaptively constrained regarding the smoothness of the similarity distribution. We adapt the proposed AdaVIB across distinct model architectures. Experimental results demonstrate that the proposed AdaVIB mitigates object hallucinations by effectively alleviating the overconfidence in irrelevant visual features, with consistent improvements on two object hallucination benchmarks.
Authors: Yihong Tang, Kehai Chen, Xuefeng Bai, Zhengyu Niu, Bo Wang, Jie Liu, Min Zhang
Abstract: Large Language Models (LLMs) have made remarkable advances in role-playing dialogue agents, demonstrating their utility in character simulations. However, it remains challenging for these agents to balance character portrayal utility with content safety because this essential character simulation often comes with the risk of generating unsafe content. To address this issue, we first conduct a systematic exploration of the safety-utility trade-off across multiple LLMs. Our analysis reveals that risk scenarios created by villain characters and user queries (referred to as risk coupling) contribute to this trade-off. Building on this, we propose a novel Adaptive Dynamic Multi-Preference (ADMP) method, which dynamically adjusts safety-utility preferences based on the degree of risk coupling and guides the model to generate responses biased toward utility or safety. We further introduce Coupling Margin Sampling (CMS) into coupling detection to enhance the model's ability to handle high-risk scenarios. Experimental results demonstrate that our approach improves safety metrics while maintaining utility.
Authors: Yuko Nakagi, Keigo Tada, Sota Yoshino, Shinji Nishimoto, Yu Takagi
Abstract: Large language models (LLMs) often exhibit abrupt emergent behavior, whereby new abilities arise at certain points during their training. This phenomenon, commonly referred to as a ''phase transition'', remains poorly understood. In this study, we conduct an integrative analysis of such phase transitions by examining three interconnected perspectives: the similarity between LLMs and the human brain, the internal states of LLMs, and downstream task performance. We propose a novel interpretation for the learning dynamics of LLMs that vary in both training data and architecture, revealing that three phase transitions commonly emerge across these models during training: (1) alignment with the entire brain surges as LLMs begin adhering to task instructions Brain Alignment and Instruction Following, (2) unexpectedly, LLMs diverge from the brain during a period in which downstream task accuracy temporarily stagnates Brain Detachment and Stagnation, and (3) alignment with the brain reoccurs as LLMs become capable of solving the downstream tasks Brain Realignment and Consolidation. These findings illuminate the underlying mechanisms of phase transitions in LLMs, while opening new avenues for interdisciplinary research bridging AI and neuroscience.
Authors: Hyewon Jeon, Jay-Yoon Lee
Abstract: Automated fact-checking aims to assess the truthfulness of text based on relevant evidence, yet verifying complex claims requiring multi-hop reasoning remains a significant challenge. We propose GraphCheck, a novel framework that converts claims into entity-relationship graphs for comprehensive verification. By identifying relation between explicit entities and latent entities across multiple paths, GraphCheck enhances the adaptability and robustness of verification. Furthermore, we introduce DP-GraphCheck, a two-stage variant that improves performance by incorporating direct prompting as an initial filtering step. Experiments on the HOVER and EX-FEVER datasets show that our approach outperforms existing methods, particularly in multi-hop reasoning tasks. Furthermore, our two-stage framework generalizes well to other fact-checking pipelines, demonstrating its versatility.
Authors: Dawei Zhu, Xiyu Wei, Guangxiang Zhao, Wenhao Wu, Haosheng Zou, Junfeng Ran, Xun Wang, Lin Sun, Xiangzheng Zhang, Sujian Li
Abstract: Recent advances in Large Language Models (LLMs) have highlighted the challenge of handling long-context tasks, where models need to reason over extensive input contexts to aggregate target information. While Chain-of-Thought (CoT) prompting has shown promise for multi-step reasoning, its effectiveness for long-context scenarios remains underexplored. Through systematic investigation across diverse tasks, we demonstrate that CoT's benefits generalize across most long-context scenarios and amplify with increasing context length. Motivated by this critical observation, we propose LongRePS, a process-supervised framework that teaches models to generate high-quality reasoning paths for enhanced long-context performance. Our framework incorporates a self-sampling mechanism to bootstrap reasoning paths and a novel quality assessment protocol specifically designed for long-context scenarios. Experimental results on various long-context benchmarks demonstrate the effectiveness of our approach, achieving significant improvements over outcome supervision baselines on both in-domain tasks (+13.6/+3.8 points for LLaMA/Qwen on MuSiQue) and cross-domain generalization (+9.3/+8.1 points on average across diverse QA tasks). Our code, data and trained models are made public to facilitate future research.
Authors: Kuang-Da Wang, Teng-Ruei Chen, Yu Heng Hung, Shuoyang Ding, Yueh-Hua Wu, Yu-Chiang Frank Wang, Chao-Han Huck Yang, Wen-Chih Peng, Ping-Chun Hsieh
Abstract: Machine Translation (MT) has been predominantly designed for sentence-level translation using transformer-based architectures. While next-token prediction based Large Language Models (LLMs) demonstrate strong capabilities in long-text translation, non-extensive language models often suffer from omissions and semantic inconsistencies when processing paragraphs. Existing preference alignment methods improve sentence-level translation but fail to ensure coherence over extended contexts due to the myopic nature of next-token generation. We introduce Plan2Align, a test-time alignment framework that treats translation as a predictive planning problem, adapting Model Predictive Control to iteratively refine translation outputs. Experiments on WMT24 Discourse-Level Literary Translation show that Plan2Align significantly improves paragraph-level translation, achieving performance surpassing or on par with the existing training-time and test-time alignment methods on LLaMA-3.1 8B.
Authors: Zhaoyi Li, Gangwei Jiang, Chenwang Wu, Ying Wei, Defu Lian, Enhong Chen
Abstract: Despite the rising prevalence of neural language models, recent empirical evidence suggests their deficiency in compositional generalization. One of the current de-facto solutions to this problem is compositional data augmentation, which aims to introduce additional compositional inductive bias. However, existing handcrafted augmentation strategies offer limited improvement when systematic generalization of neural language models requires multi-grained compositional bias (i.e., not limited to either lexical or structural biases alone) or when training sentences have an imbalanced difficulty distribution. To address these challenges, we first propose a novel compositional augmentation strategy called Component Substitution (CompSub), which enables multi-grained composition of substantial substructures across the entire training set. Furthermore, we introduce the Learning Component Substitution (LCS) framework. This framework empowers the learning of component substitution probabilities in CompSub in an end-to-end manner by maximizing the loss of neural language models, thereby prioritizing challenging compositions with elusive concepts and novel contexts. We extend the key ideas of CompSub and LCS to the recently emerging in-context learning scenarios of pre-trained large language models (LLMs), proposing the LCS-ICL algorithm to enhance the few-shot compositional generalization of state-of-the-art (SOTA) LLMs. Theoretically, we provide insights into why applying our algorithms to language models can improve compositional generalization performance. Empirically, our results on four standard compositional generalization benchmarks(SCAN, COGS, GeoQuery, and COGS-QL) demonstrate the superiority of CompSub, LCS, and LCS-ICL, with improvements of up to 66.5%, 10.3%, 1.4%, and 8.8%, respectively.
Authors: Jonathan Drechsel, Anja Reusch, Steffen Herbold
Abstract: Mathematical formulas are a fundamental and widely used component in various scientific fields, serving as a universal language for expressing complex concepts and relationships. While state-of-the-art transformer models excel in processing and understanding natural language, they encounter challenges with mathematical notation, which involves a complex structure and diverse representations. This study focuses on the development of specialized training datasets to enhance the encoding of mathematical content. We introduce Math Mutator (MAMUT), a framework capable of generating equivalent and falsified versions of a given mathematical formula in LaTeX notation, effectively capturing the mathematical variety in notation of the same concept. Based on MAMUT, we have generated four large mathematical datasets containing diverse notation, which can be used to train language models with enhanced mathematical embeddings.
Authors: Yifan Duan, Yihong Tang, Xuefeng Bai, Kehai Chen, Juntao Li, Min Zhang
Abstract: Large language models (LLMs) excel in both closed tasks (including problem-solving, and code generation) and open tasks (including creative writing), yet existing explanations for their capabilities lack connections to real-world human intelligence. To fill this gap, this paper systematically investigates LLM intelligence through the lens of ``human simulation'', addressing three core questions: (1) How do personality traits affect problem-solving in closed tasks? (2) How do traits shape creativity in open tasks? (3) How does single-agent performance influence multi-agent collaboration? By assigning Big Five personality traits to LLM agents and evaluating their performance in single- and multi-agent settings, we reveal that specific traits significantly influence reasoning accuracy (closed tasks) and creative output (open tasks). Furthermore, multi-agent systems exhibit collective intelligence distinct from individual capabilities, driven by distinguishing combinations of personalities. We demonstrate that LLMs inherently simulate human behavior through next-token prediction, mirroring human language, decision-making, and collaborative dynamics.
Authors: Mohammad Rifqi Farhansyah, Iwan Darmawan, Adryan Kusumawardhana, Genta Indra Winata, Alham Fikri Aji, Derry Tanti Wijaya
Abstract: The Javanese language features a complex system of honorifics that vary according to the social status of the speaker, listener, and referent. Despite its cultural and linguistic significance, there has been limited progress in developing a comprehensive corpus to capture these variations for natural language processing (NLP) tasks. In this paper, we present Unggah-Ungguh, a carefully curated dataset designed to encapsulate the nuances of Unggah-Ungguh Basa, the Javanese speech etiquette framework that dictates the choice of words and phrases based on social hierarchy and context. Using Unggah-Ungguh, we assess the ability of language models (LMs) to process various levels of Javanese honorifics through classification and machine translation tasks. To further evaluate cross-lingual LMs, we conduct machine translation experiments between Javanese (at specific honorific levels) and Indonesian. Additionally, we explore whether LMs can generate contextually appropriate Javanese honorifics in conversation tasks, where the honorific usage should align with the social role and contextual cues. Our findings indicate that current LMs struggle with most honorific levels, exhibitinga bias toward certain honorific tiers.
Authors: Ana Ezquerro, Carlos G\'omez-Rodr\'iguez, David Vilares
Abstract: While LLMs excel in zero-shot tasks, their performance in linguistic challenges like syntactic parsing has been less scrutinized. This paper studies state-of-the-art open-weight LLMs on the task by comparing them to baselines that do not have access to the input sentence, including baselines that have not been used in this context such as random projective trees or optimal linear arrangements. The results show that most of the tested LLMs cannot outperform the best uninformed baselines, with only the newest and largest versions of LLaMA doing so for most languages, and still achieving rather low performance. Thus, accurate zero-shot syntactic parsing is not forthcoming with open LLMs.
Authors: Lei Yang, Renren Jin, Ling Shi, Jianxiang Peng, Yue Chen, Deyi Xiong
Abstract: With reasoning language models such as OpenAI-o3 and DeepSeek-R1 emerging, large language models (LLMs) have entered a new phase of development. However, existing benchmarks for coding evaluation are gradually inadequate to assess the capability of advanced LLMs in code reasoning. To bridge the gap for high-level code reasoning assessment, we propose ProBench to benchmark LLMs in competitive programming, drawing inspiration from the International Collegiate Programming Contest. ProBench collects a comprehensive set of competitive programming problems from Codeforces, Luogu, and Nowcoder platforms during the period from July to December 2024, obtaining real test results through online submissions to ensure the fairness and accuracy of the evaluation. We establish a unified problem attribute system, including difficulty grading and algorithm tagging. With carefully collected and annotated data in ProBench, we systematically assess 9 latest LLMs in competitive programming across multiple dimensions, including thought chain analysis, error type diagnosis, and reasoning depth evaluation. Experimental results show that QwQ-32B-Preview achieves the best score of 20.93 followed by DeepSeek-V3 with a score of 16.38, suggesting that models trained with specialized reasoning tasks significantly outperform general-purpose models (even larger than reasoning-oriented models) in programming. Further analysis also reveals key areas for programming capability enhancement, e.g., algorithm adaptability and reasoning sufficiency, providing important insights for the future development of reasoning models.
Authors: Matthias Orlikowski, Jiaxin Pei, Paul R\"ottger, Philipp Cimiano, David Jurgens, Dirk Hovy
Abstract: People naturally vary in their annotations for subjective questions and some of this variation is thought to be due to the person's sociodemographic characteristics. LLMs have also been used to label data, but recent work has shown that models perform poorly when prompted with sociodemographic attributes, suggesting limited inherent sociodemographic knowledge. Here, we ask whether LLMs can be trained to be accurate sociodemographic models of annotator variation. Using a curated dataset of five tasks with standardized sociodemographics, we show that models do improve in sociodemographic prompting when trained but that this performance gain is largely due to models learning annotator-specific behaviour rather than sociodemographic patterns. Across all tasks, our results suggest that models learn little meaningful connection between sociodemographics and annotation, raising doubts about the current use of LLMs for simulating sociodemographic variation and behaviour.
Authors: Luise Mehner, Lena Alicija Philine Fiedler, Sabine Ammon, Dorothea Kolossa
Abstract: The widespread application of Large Language Models (LLMs) involves ethical risks for users and societies. A prominent ethical risk of LLMs is the generation of unfair language output that reinforces or exacerbates harm for members of disadvantaged social groups through gender biases (Weidinger et al., 2022; Bender et al., 2021; Kotek et al., 2023). Hence, the evaluation of the fairness of LLM outputs with respect to such biases is a topic of rising interest. To advance research in this field, promote discourse on suitable normative bases and evaluation methodologies, and enhance the reproducibility of related studies, we propose a novel approach to database construction. This approach enables the assessment of gender-related biases in LLM-generated language beyond merely evaluating their degree of neutralization.
Authors: Ilya Koziev
Abstract: Generative poetry systems require effective tools for data engineering and automatic evaluation, particularly to assess how well a poem adheres to versification rules, such as the correct alternation of stressed and unstressed syllables and the presence of rhymes. In this work, we introduce the Russian Poetry Scansion Tool library designed for stress mark placement in Russian-language syllabo-tonic poetry, rhyme detection, and identification of defects of poeticness. Additionally, we release RIFMA -- a dataset of poem fragments spanning various genres and forms, annotated with stress marks. This dataset can be used to evaluate the capability of modern large language models to accurately place stress marks in poetic texts. The published resources provide valuable tools for researchers and practitioners in the field of creative generative AI, facilitating advancements in the development and evaluation of generative poetry systems.
Authors: Michael Dinzinger, Laura Caspari, Kanishka Ghosh Dastidar, Jelena Mitrovi\'c, Michael Granitzer
Abstract: We present WebFAQ, a large-scale collection of open-domain question answering datasets derived from FAQ-style schema.org annotations. In total, the data collection consists of 96 million natural question-answer (QA) pairs across 75 languages, including 47 million (49%) non-English samples. WebFAQ further serves as the foundation for 20 monolingual retrieval benchmarks with a total size of 11.2 million QA pairs (5.9 million non-English). These datasets are carefully curated through refined filtering and near-duplicate detection, yielding high-quality resources for training and evaluating multilingual dense retrieval models. To empirically confirm WebFAQ's efficacy, we use the collected QAs to fine-tune an in-domain pretrained XLM-RoBERTa model. Through this process of dataset-specific fine-tuning, the model achieves significant retrieval performance gains, which generalize - beyond WebFAQ - to other multilingual retrieval benchmarks evaluated in zero-shot setting. Last but not least, we utilize WebFAQ to construct a set of QA-aligned bilingual corpora spanning over 1000 language pairs using state-of-the-art bitext mining and automated LLM-assessed translation evaluation. Due to our advanced, automated method of bitext dataset generation, the resulting bilingual corpora demonstrate higher translation quality compared to similar datasets. WebFAQ and all associated resources are publicly available on GitHub and HuggingFace.
Authors: Weixiang Zhao, Yulin Hu, Yang Deng, Jiahe Guo, Xingyu Sui, Xinyang Han, An Zhang, Yanyan Zhao, Bing Qin, Tat-Seng Chua, Ting Liu
Abstract: Role-playing enables large language models (LLMs) to engage users in immersive and personalized interactions, but it also introduces significant safety risks. Existing role-play fine-tuning techniques improve role adaptability but may degrade safety performance, particularly for villainous characters. In this work, we conduct the first comprehensive assessment of role-play fine-tuning risks by training 95 role-specific LLMs using RoleBench. Our experiments reveal that role-play fine-tuning leads to a noticeable decline in safety performance, with safety risks varying based on character traits. To tackle this challenge, we propose Safety-Aware Role-Play Fine-Tuning (SaRFT), a novel method designed to balance role-playing capabilities and safety. Extensive experiments on LLaMA-3-8B-Instruct, Gemma-2-9B-it, and Qwen2.5-7B-Instruct demonstrate that SaRFT consistently outperforms state-of-the-art baselines under both LoRA and full-parameter fine-tuning settings. Our findings highlight the necessity of role-adaptive safety measures and provide insights into mitigating role-specific safety risks in role-playing LLMs.
Authors: Perla Al Almaoui, Pierrette Bouillon, Simon Hengchen
Abstract: In this era of rapid technological advancements, communication continues to evolve as new linguistic phenomena emerge. Among these is Arabizi, a hybrid form of Arabic that incorporates Latin characters and numbers to represent the spoken dialects of Arab communities. Arabizi is widely used on social media and allows people to communicate in an informal and dynamic way, but it poses significant challenges for machine translation due to its lack of formal structure and deeply embedded cultural nuances. This case study arises from a growing need to translate Arabizi for gisting purposes. It evaluates the capacity of different LLMs to decode and translate Arabizi, focusing on multiple Arabic dialects that have rarely been studied up until now. Using a combination of human evaluators and automatic metrics, this research project investigates the model's performance in translating Arabizi into both Modern Standard Arabic and English. Key questions explored include which dialects are translated most effectively and whether translations into English surpass those into Arabic.
Authors: Naman Bansal, Yash mahajan, Sanjeev Sinha, Santu Karmaker
Abstract: Sentence encoders play a pivotal role in various NLP tasks; hence, an accurate evaluation of their compositional properties is paramount. However, existing evaluation methods predominantly focus on goal task-specific performance. This leaves a significant gap in understanding how well sentence embeddings demonstrate fundamental compositional properties in a task-independent context. Leveraging classical set theory, we address this gap by proposing six criteria based on three core "set-like" compositions/operations: \textit{TextOverlap}, \textit{TextDifference}, and \textit{TextUnion}. We systematically evaluate $7$ classical and $9$ Large Language Model (LLM)-based sentence encoders to assess their alignment with these criteria. Our findings show that SBERT consistently demonstrates set-like compositional properties, surpassing even the latest LLMs. Additionally, we introduce a new dataset of ~$192$K samples designed to facilitate future benchmarking efforts on set-like compositionality of sentence embeddings.
Authors: Thanet Markchom, Tong Wu, Liting Huang, Huizhi Liang
Abstract: SemEval-2025 Task 1 focuses on ranking images based on their alignment with a given nominal compound that may carry idiomatic meaning in both English and Brazilian Portuguese. To address this challenge, this work uses generative large language models (LLMs) and multilingual CLIP models to enhance idiomatic compound representations. LLMs generate idiomatic meanings for potentially idiomatic compounds, enriching their semantic interpretation. These meanings are then encoded using multilingual CLIP models, serving as representations for image ranking. Contrastive learning and data augmentation techniques are applied to fine-tune these embeddings for improved performance. Experimental results show that multimodal representations extracted through this method outperformed those based solely on the original nominal compounds. The fine-tuning approach shows promising outcomes but is less effective than using embeddings without fine-tuning. The source code used in this paper is available at https://github.com/tongwu17/SemEval-2025-Task1-UoR-NCL.
URLs: https://github.com/tongwu17/SemEval-2025-Task1-UoR-NCL.
Authors: Xiusheng Huang, Jiaxiang Liu, Yequan Wang, Jun Zhao, Kang Liu
Abstract: Large scale language models have achieved superior performance in tasks related to natural language processing, however, it is still unclear how model parameters affect performance improvement. Previous studies assumed that individual knowledge is stored in local parameters, and the storage form of individual knowledge is dispersed parameters, parameter layers, or parameter chains, which are not unified. We found through fidelity and reliability evaluation experiments that individual knowledge cannot be localized. Afterwards, we constructed a dataset for decoupling experiments and discovered the potential for localizing data commonalities. To further reveal this phenomenon, this paper proposes a Commonality Neuron Localization (CNL) method, which successfully locates commonality neurons and achieves a neuron overlap rate of 96.42% on the GSM8K dataset. Finally, we have demonstrated through cross data experiments that commonality neurons are a collection of capability neurons that possess the capability to enhance performance. Our code is available at https://github.com/nlpkeg/Capability-Neuron-Localization.
URLs: https://github.com/nlpkeg/Capability-Neuron-Localization.
Authors: Fangxu Yu, Lai Jiang, Shenyi Huang, Zhen Wu, Xinyu Dai
Abstract: The ability to understand and predict the mental states of oneself and others, known as the Theory of Mind (ToM), is crucial for effective social interactions. Recent research has emerged to evaluate whether Large Language Models (LLMs) exhibit a form of ToM. Although recent studies have evaluated ToM in LLMs, existing benchmarks focus predominantly on physical perception with principles guided by the Sally-Anne test in synthetic stories and conversations, failing to capture the complex psychological activities of mental states in real-life social interactions. To mitigate this gap, we propose PersuasiveToM, a benchmark designed to evaluate the ToM abilities of LLMs in persuasive dialogues. Our framework introduces two categories of questions: (1) ToM Reasoning, assessing the capacity of LLMs to track evolving mental states (e.g., desire shifts in persuadees), and (2) ToM Application, evaluating whether LLMs can take advantage of inferred mental states to select effective persuasion strategies (e.g., emphasize rarity) and evaluate the effectiveness of persuasion strategies. Experiments across eight state-of-the-art LLMs reveal that while models excel on multiple questions, they struggle to answer questions that need tracking the dynamics and shifts of mental states and understanding the mental states in the whole dialogue comprehensively. Our aim with PersuasiveToM is to allow an effective evaluation of the ToM reasoning ability of LLMs with more focus on complex psychological activities. Our code is available at https://github.com/Yu-Fangxu/PersuasiveToM.
Authors: Jos\'e I. Orlicki
Abstract: Recent advances in large language models (LLMs) have popularized the chain-of-thought (CoT) paradigm, in which models produce explicit reasoning steps in natural language. Although this approach improves interpretability and facilitates external auditing, it may not represent the most computationally efficient method for internal reasoning. In contrast, human cognition relies on implicit mental representations that recall past sensory and episodic information without requiring complete verbalization. In this paper, we propose a framework that integrates implicit mental representations into the internal reasoning processes of LLMs. Preliminary experiments indicate that incorporating an Implicit Memory Module (IMM) into a simple GPT model yields a reduction of between 35% and 57% in final training loss compared to a regular GPT baseline. The addition of an explicit interpretability channel (e.g., a chain-of-thought decoder) is straightforward to implement within this approach. We outline theoretical foundations, propose technical mechanisms to scale the memory module, and discuss how these ideas may lead to more efficient and robust reasoning, with optional future extensions for explicit auditability.
Authors: Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, Yulan He
Abstract: Chain-of-Thought (CoT) enhances Large Language Models (LLMs) by enabling step-by-step reasoning in natural language. However, the language space may be suboptimal for reasoning. While implicit CoT methods attempt to enable reasoning without explicit CoT tokens, they have consistently lagged behind explicit CoT method in task performance. We propose CODI (Continuous Chain-of-Thought via Self-Distillation), a novel framework that distills CoT into a continuous space, where a shared model acts as both teacher and student, jointly learning explicit and implicit CoT while aligning their hidden activation on the token generating the final answer. CODI is the first implicit CoT method to match explicit CoT's performance on GSM8k while achieving 3.1x compression, surpassing the previous state-of-the-art by 28.2% in accuracy. Furthermore, CODI demonstrates scalability, robustness, and generalizability to more complex CoT datasets. Additionally, CODI retains interpretability by decoding its continuous thoughts, making its reasoning process transparent. Our findings establish implicit CoT as not only a more efficient but a powerful alternative to explicit CoT.
Authors: Hansi Yang, Qi Zhang, Wei Jiang, Jianguo Li
Abstract: Large language models (LLMs) have shown impressive abilities in answering questions across various domains, but they often encounter hallucination issues on questions that require professional and up-to-date knowledge. To address this limitation, retrieval-augmented generation (RAG) techniques have been proposed, which retrieve relevant information from external sources to inform their responses. However, existing RAG methods typically focus on a single type of external data, such as vectorized text database or knowledge graphs, and cannot well handle real-world questions on semi-structured data containing both text and relational information. To bridge this gap, we introduce PASemiQA, a novel approach that jointly leverages text and relational information in semi-structured data to answer questions. PASemiQA first generates a plan to identify relevant text and relational information to answer the question in semi-structured data, and then uses an LLM agent to traverse the semi-structured data and extract necessary information. Our empirical results demonstrate the effectiveness of PASemiQA across different semi-structured datasets from various domains, showcasing its potential to improve the accuracy and reliability of question answering systems on semi-structured data.
Authors: Angelo Ziletti, Leonardo D'Ambrosi
Abstract: Clinical cohort definition is crucial for patient recruitment and observational studies, yet translating inclusion/exclusion criteria into SQL queries remains challenging and manual. We present an automated system utilizing large language models that combines criteria parsing, two-level retrieval augmented generation with specialized knowledge bases, medical concept standardization, and SQL generation to retrieve patient cohorts with patient funnels. The system achieves 0.75 F1-score in cohort identification on EHR data, effectively capturing complex temporal and logical relationships. These results demonstrate the feasibility of automated cohort generation for epidemiological research.
Authors: Sidney Wong, Benjamin Adams, Jonathan Dunn
Abstract: This chapter explores the efficacy of using social media data to examine changing linguistic behaviour of a place. We focus our investigation on Aotearoa New Zealand where official statistics from the census is the only source of language use data. We use published census data as the ground truth and the social media sub-corpus from the Corpus of Global Language Use as our alternative data source. We use place as the common denominator between the two data sources. We identify the language conditions of each tweet in the social media data set and validated our results with two language identification models. We then compare levels of linguistic diversity at national, regional, and local geographies. The results suggest that social media language data has the possibility to provide a rich source of spatial and temporal insights on the linguistic profile of a place. We show that social media is sensitive to demographic and sociopolitical changes within a language and at low-level regional and local geographies.
Authors: Omer Goldman, Uri Shaham, Dan Malkin, Sivan Eiger, Avinatan Hassidim, Yossi Matias, Joshua Maynez, Adi Mayrav Gilady, Jason Riesa, Shruti Rijhwani, Laura Rimell, Idan Szpektor, Reut Tsarfaty, Matan Eyal
Abstract: To achieve equitable performance across languages, multilingual large language models (LLMs) must be able to abstract knowledge beyond the language in which it was acquired. However, the current literature lacks reliable ways to measure LLMs' capability of cross-lingual knowledge transfer. To that end, we present ECLeKTic, a multilingual closed-book QA (CBQA) dataset that Evaluates Cross-Lingual Knowledge Transfer in a simple, black-box manner. We detected information with uneven coverage across languages by controlling for presence and absence of Wikipedia articles in 12 languages. We generated knowledge-seeking questions in a source language, for which the answer appears in a relevant Wikipedia article and translated them to all other 11 languages, for which the respective Wikipedias lack equivalent articles. Assuming that Wikipedia reflects the prominent knowledge in the LLM's training data, to solve ECLeKTic's CBQA task the model is required to transfer knowledge between languages. Experimenting with 8 LLMs, we show that SOTA models struggle to effectively share knowledge across, languages even if they can predict the answer well for queries in the same language the knowledge was acquired in.
Authors: Xiaomin Li, Zhou Yu, Ziji Zhang, Yingying Zhuang, Swair Shah, Anurag Beniwal
Abstract: Large language models (LLMs) have demonstrated remarkable performance across diverse tasks by encoding vast amounts of factual knowledge. However, they are still prone to hallucinations, generating incorrect or misleading information, often accompanied by high uncertainty. Existing methods for hallucination detection primarily focus on quantifying internal uncertainty, which arises from missing or conflicting knowledge within the model. However, hallucinations can also stem from external uncertainty, where ambiguous user queries lead to multiple possible interpretations. In this work, we introduce Semantic Volume, a novel mathematical measure for quantifying both external and internal uncertainty in LLMs. Our approach perturbs queries and responses, embeds them in a semantic space, and computes the determinant of the Gram matrix of the embedding vectors, capturing their dispersion as a measure of uncertainty. Our framework provides a generalizable and unsupervised uncertainty detection method without requiring white-box access to LLMs. We conduct extensive experiments on both external and internal uncertainty detection, demonstrating that our Semantic Volume method consistently outperforms existing baselines in both tasks. Additionally, we provide theoretical insights linking our measure to differential entropy, unifying and extending previous sampling-based uncertainty measures such as the semantic entropy. Semantic Volume is shown to be a robust and interpretable approach to improving the reliability of LLMs by systematically detecting uncertainty in both user queries and model responses.
Authors: Aleksandr Nesterov, Andrey Sakhovskiy, Ivan Sviridov, Airat Valiev, Vladimir Makharev, Petr Anokhin, Galina Zubkova, Elena Tutubalina
Abstract: This study investigates the feasibility of automating clinical coding in Russian, a language with limited biomedical resources. We present a new dataset for ICD coding, which includes diagnosis fields from electronic health records (EHRs) annotated with over 10,000 entities and more than 1,500 unique ICD codes. This dataset serves as a benchmark for several state-of-the-art models, including BERT, LLaMA with LoRA, and RAG, with additional experiments examining transfer learning across domains (from PubMed abstracts to medical diagnosis) and terminologies (from UMLS concepts to ICD codes). We then apply the best-performing model to label an in-house EHR dataset containing patient histories from 2017 to 2021. Our experiments, conducted on a carefully curated test set, demonstrate that training with the automated predicted codes leads to a significant improvement in accuracy compared to manually annotated data from physicians. We believe our findings offer valuable insights into the potential for automating clinical coding in resource-limited languages like Russian, which could enhance clinical efficiency and data accuracy in these contexts.
Authors: Rachel Wicks, Kartik Ravisankar, Xinchen Yang, Philipp Koehn, Matt Post
Abstract: Model ensembling is a technique to combine the predicted distributions of two or more models, often leading to improved robustness and performance. For ensembling in text generation, the next token's probability distribution is derived from a weighted sum of the distributions of each individual model. This requires the underlying models to share the same subword vocabulary, limiting the applicability of ensembling, since many open-sourced models have distinct vocabularies. In research settings, experimentation or upgrades to vocabularies may introduce multiple vocabulary sizes. This paper proposes an inference-time only algorithm that allows for ensembling models with different vocabularies, without the need to learn additional parameters or alter the underlying models. Instead, the algorithm ensures that tokens generated by the ensembled models \textit{agree} in their surface form. We apply this technique to combinations of traditional encoder-decoder models and decoder-only LLMs and evaluate on machine translation. In addition to expanding to model pairs that were previously incapable of token-level ensembling, our algorithm frequently improves translation performance over either model individually.
Authors: Dingyi Zhang, Deyu Zhou
Abstract: Persuasive dialogue plays a pivotal role in human communication, influencing various domains. Recent persuasive dialogue datasets often fail to align with real-world interpersonal interactions, leading to unfaithful representations. For instance, unrealistic scenarios may arise, such as when the persuadee explicitly instructs the persuader on which persuasion strategies to employ, with each of the persuadee's questions corresponding to a specific strategy for the persuader to follow. This issue can be attributed to a violation of the "Double Blind" condition, where critical information is fully shared between participants. In actual human interactions, however, key information such as the mental state of the persuadee and the persuasion strategies of the persuader is not directly accessible. The persuader must infer the persuadee's mental state using Theory of Mind capabilities and construct arguments that align with the persuadee's motivations. To address this gap, we introduce ToMMA, a novel multi-agent framework for dialogue generation that is guided by causal Theory of Mind. This framework ensures that information remains undisclosed between agents, preserving "double-blind" conditions, while causal ToM directs the persuader's reasoning, enhancing alignment with human-like persuasion dynamics. Consequently, we present CToMPersu, a multi-domain, multi-turn persuasive dialogue dataset that tackles both double-blind and logical coherence issues, demonstrating superior performance across multiple metrics and achieving better alignment with real human dialogues. Our dataset and prompts are available at https://github.com/DingyiZhang/ToMMA-CToMPersu .
Authors: Yihong Dong, Ge Li, Xue Jiang, Yongding Tao, Kechi Zhang, Hao Zhu, Huanyu Liu, Jiazheng Ding, Jia Li, Jinliang Deng, Hong Mei
Abstract: Periodicity, as one of the most important basic characteristics, lays the foundation for facilitating structured knowledge acquisition and systematic cognitive processes within human learning paradigms. However, the potential flaws of periodicity modeling in Transformer affect the learning efficiency and establishment of underlying principles from data for large language models (LLMs) built upon it. In this paper, we demonstrate that integrating effective periodicity modeling can improve the learning efficiency and performance of LLMs. We introduce FANformer, which integrates Fourier Analysis Network (FAN) into attention mechanism to achieve efficient periodicity modeling, by modifying the feature projection process of attention mechanism. Extensive experimental results on language modeling show that FANformer consistently outperforms Transformer when scaling up model size and training tokens, underscoring its superior learning efficiency. To further validate the effectiveness of FANformer, we pretrain a FANformer-1B on 1 trillion tokens. FANformer-1B exhibits marked improvements on downstream tasks compared to open-source LLMs with similar model parameters or training tokens. The results position FANformer as an effective and promising architecture for advancing LLMs.
Authors: Sibo Ma, Julian Nyarko
Abstract: We introduce a new method to identify emerging concepts in large text corpora. By analyzing changes in the heatmaps of the underlying embedding space, we are able to detect these concepts with high accuracy shortly after they originate, in turn outperforming common alternatives. We further demonstrate the utility of our approach by analyzing speeches in the U.S. Senate from 1941 to 2015. Our results suggest that the minority party is more active in introducing new concepts into the Senate discourse. We also identify specific concepts that closely correlate with the Senators' racial, ethnic, and gender identities. An implementation of our method is publicly available.
Authors: Komal Kumar, Tajamul Ashraf, Omkar Thawakar, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, Phillip H. S. Torr, Salman Khan, Fahad Shahbaz Khan
Abstract: Large Language Models (LLMs) have transformed the natural language processing landscape and brought to life diverse applications. Pretraining on vast web-scale data has laid the foundation for these models, yet the research community is now increasingly shifting focus toward post-training techniques to achieve further breakthroughs. While pretraining provides a broad linguistic foundation, post-training methods enable LLMs to refine their knowledge, improve reasoning, enhance factual accuracy, and align more effectively with user intents and ethical considerations. Fine-tuning, reinforcement learning, and test-time scaling have emerged as critical strategies for optimizing LLMs performance, ensuring robustness, and improving adaptability across various real-world tasks. This survey provides a systematic exploration of post-training methodologies, analyzing their role in refining LLMs beyond pretraining, addressing key challenges such as catastrophic forgetting, reward hacking, and inference-time trade-offs. We highlight emerging directions in model alignment, scalable adaptation, and inference-time reasoning, and outline future research directions. We also provide a public repository to continually track developments in this fast-evolving field: https://github.com/mbzuai-oryx/Awesome-LLM-Post-training.
URLs: https://github.com/mbzuai-oryx/Awesome-LLM-Post-training.
Authors: Jingtao Zhan, Jiahao Zhao, Jiayu Li, Yiqun Liu, Bo Zhang, Qingyao Ai, Jiaxin Mao, Hongning Wang, Min Zhang, Shaoping Ma
Abstract: How does intelligence emerge? We propose that intelligence is not a sudden gift or random occurrence, but rather a necessary trait for species to survive through Natural Selection. If a species passes the test of Natural Selection, it demonstrates the intelligence to survive in nature. Extending this perspective, we introduce Intelligence Test, a method to quantify the intelligence of any subject on any task. Like how species evolve by trial and error, Intelligence Test quantifies intelligence by the number of failed attempts before success. Fewer failures correspond to higher intelligence. When the expectation and variance of failure counts are both finite, it signals the achievement of an autonomous level of intelligence. Using Intelligence Test, we comprehensively evaluate existing AI systems. Our results show that while AI systems achieve a level of autonomy in simple tasks, they are still far from autonomous in more complex tasks, such as vision, search, recommendation, and language. While scaling model size might help, this would come at an astronomical cost. Projections suggest that achieving general autonomy would require unimaginable $10^{26}$ parameters. Even if Moore's Law continuously holds, such a parameter scale would take $70$ years. This staggering cost highlights the complexity of human tasks and the inadequacies of current AI. To further understand this phenomenon, we conduct a theoretical analysis. Our simulations suggest that human tasks possess a criticality property. As a result, autonomy requires a deep understanding of the task's underlying mechanisms. Current AI, however, does not fully grasp these mechanisms and instead relies on superficial mimicry, making it difficult to reach an autonomous level. We believe Intelligence Test can not only guide the future development of AI but also offer profound insights into the intelligence of humans ourselves.
Authors: Yiheng Liu, Xiaohui Gao, Haiyang Sun, Bao Ge, Tianming Liu, Junwei Han, Xintao Hu
Abstract: In recent years, the rapid advancement of large language models (LLMs) in natural language processing has sparked significant interest among researchers to understand their mechanisms and functional characteristics. Although existing studies have attempted to explain LLM functionalities by identifying and interpreting specific neurons, these efforts mostly focus on individual neuron contributions, neglecting the fact that human brain functions are realized through intricate interaction networks. Inspired by cognitive neuroscience research on functional brain networks (FBNs), this study introduces a novel approach to investigate whether similar functional networks exist within LLMs. We use methods similar to those in the field of functional neuroimaging analysis to locate and identify functional networks in LLM. Experimental results show that, similar to the human brain, LLMs contain functional networks that frequently recur during operation. Further analysis shows that these functional networks are crucial for LLM performance. Masking key functional networks significantly impairs the model's performance, while retaining just a subset of these networks is adequate to maintain effective operation. This research provides novel insights into the interpretation of LLMs and the lightweighting of LLMs for certain downstream tasks. Code is available at https://github.com/WhatAboutMyStar/LLM_ACTIVATION.
Authors: MohammadHossein Rezaei, Yicheng Fu, Phil Cuvin, Caleb Ziems, Yanzhe Zhang, Hao Zhu, Diyi Yang
Abstract: Human activity is moderated by norms. When performing actions in the real world, humans not only follow norms, but also consider the trade-off between different norms However, machines are often trained without explicit supervision on norm understanding and reasoning, especially when the norms are grounded in a physical and social context. To improve and evaluate the normative reasoning capability of vision-language models (VLMs), we present EgoNormia $\|\epsilon\|$, consisting of 1,853 ego-centric videos of human interactions, each of which has two related questions evaluating both the prediction and justification of normative actions. The normative actions encompass seven categories: safety, privacy, proxemics, politeness, cooperation, coordination/proactivity, and communication/legibility. To compile this dataset at scale, we propose a novel pipeline leveraging video sampling, automatic answer generation, filtering, and human validation. Our work demonstrates that current state-of-the-art vision-language models lack robust norm understanding, scoring a maximum of 45% on EgoNormia (versus a human bench of 92%). Our analysis of performance in each dimension highlights the significant risks of safety, privacy, and the lack of collaboration and communication capability when applied to real-world agents. We additionally show that through a retrieval-based generation method, it is possible to use EgoNomia to enhance normative reasoning in VLMs.
Authors: Jin Peng Zhou, Kaiwen Wang, Jonathan Chang, Zhaolin Gao, Nathan Kallus, Kilian Q. Weinberger, Kiant\'e Brantley, Wen Sun
Abstract: Reinforcement learning (RL) post-training is crucial for LLM alignment and reasoning, but existing policy-based methods, such as PPO and DPO, can fall short of fixing shortcuts inherited from pre-training. In this work, we introduce $Q\sharp$, a value-based algorithm for KL-regularized RL that guides the reference policy using the optimal regularized $Q$ function. We propose to learn the optimal $Q$ function using distributional RL on an aggregated online dataset. Unlike prior value-based baselines that guide the model using unregularized $Q$-values, our method is theoretically principled and provably learns the optimal policy for the KL-regularized RL problem. Empirically, $Q\sharp$ outperforms prior baselines in math reasoning benchmarks while maintaining a smaller KL divergence to the reference policy. Theoretically, we establish a reduction from KL-regularized RL to no-regret online learning, providing the first bounds for deterministic MDPs under only realizability. Thanks to distributional RL, our bounds are also variance-dependent and converge faster when the reference policy has small variance. In sum, our results highlight $Q\sharp$ as an effective approach for post-training LLMs, offering both improved performance and theoretical guarantees. The code can be found at https://github.com/jinpz/q_sharp.
Authors: Zhuohang Li, Chao Yan, Nicholas J. Jackson, Wendi Cui, Bo Li, Jiaxin Zhang, Bradley A. Malin
Abstract: Advancements in Large Vision-Language Models (LVLMs) have demonstrated promising performance in a variety of vision-language tasks involving image-conditioned free-form text generation. However, growing concerns about hallucinations in LVLMs, where the generated text is inconsistent with the visual context, are becoming a major impediment to deploying these models in applications that demand guaranteed reliability. In this paper, we introduce a framework to address this challenge, ConfLVLM, which is grounded on conformal prediction to achieve finite-sample distribution-free statistical guarantees on the factuality of LVLM output. This framework treats an LVLM as a hypothesis generator, where each generated text detail (or claim) is considered an individual hypothesis. It then applies a statistical hypothesis testing procedure to verify each claim using efficient heuristic uncertainty measures to filter out unreliable claims before returning any responses to users. We conduct extensive experiments covering three representative application domains, including general scene understanding, medical radiology report generation, and document understanding. Remarkably, ConfLVLM reduces the error rate of claims generated by LLaVa-1.5 for scene descriptions from 87.8\% to 10.0\% by filtering out erroneous claims with a 95.3\% true positive rate. Our results further demonstrate that ConfLVLM is highly flexible, and can be applied to any black-box LVLMs paired with any uncertainty measure for any image-conditioned free-form text generation task while providing a rigorous guarantee on controlling the risk of hallucination.
Authors: Mohammad Abu Tami, Mohammed Elhenawy, Huthaifa I. Ashqar
Abstract: Traffic safety remains a vital concern in contemporary urban settings, intensified by the increase of vehicles and the complicated nature of road networks. Traditional safety-critical event detection systems predominantly rely on sensor-based approaches and conventional machine learning algorithms, necessitating extensive data collection and complex training processes to adhere to traffic safety regulations. This paper introduces HazardNet, a small-scale Vision Language Model designed to enhance traffic safety by leveraging the reasoning capabilities of advanced language and vision models. We built HazardNet by fine-tuning the pre-trained Qwen2-VL-2B model, chosen for its superior performance among open-source alternatives and its compact size of two billion parameters. This helps to facilitate deployment on edge devices with efficient inference throughput. In addition, we present HazardQA, a novel Vision Question Answering (VQA) dataset constructed specifically for training HazardNet on real-world scenarios involving safety-critical events. Our experimental results show that the fine-tuned HazardNet outperformed the base model up to an 89% improvement in F1-Score and has comparable results with improvement in some cases reach up to 6% when compared to larger models, such as GPT-4o. These advancements underscore the potential of HazardNet in providing real-time, reliable traffic safety event detection, thereby contributing to reduced accidents and improved traffic management in urban environments. Both HazardNet model and the HazardQA dataset are available at https://huggingface.co/Tami3/HazardNet and https://huggingface.co/datasets/Tami3/HazardQA, respectively.
URLs: https://huggingface.co/Tami3/HazardNet, https://huggingface.co/datasets/Tami3/HazardQA,
Authors: Sari Masri, Huthaifa I. Ashqar, Mohammed Elhenawy
Abstract: Traffic control in unsignalized urban intersections presents significant challenges due to the complexity, frequent conflicts, and blind spots. This study explores the capability of leveraging Multimodal Large Language Models (MLLMs), such as GPT-4o, to provide logical and visual reasoning by directly using birds-eye-view videos of four-legged intersections. In this proposed method, GPT-4o acts as intelligent system to detect conflicts and provide explanations and recommendations for the drivers. The fine-tuned model achieved an accuracy of 77.14%, while the manual evaluation of the true predicted values of the fine-tuned GPT-4o showed significant achievements of 89.9% accuracy for model-generated explanations and 92.3% for the recommended next actions. These results highlight the feasibility of using MLLMs for real-time traffic management using videos as inputs, offering scalable and actionable insights into intersections traffic management and operation. Code used in this study is available at https://github.com/sarimasri3/Traffic-Intersection-Conflict-Detection-using-images.git.
URLs: https://github.com/sarimasri3/Traffic-Intersection-Conflict-Detection-using-images.git.
Authors: Kai Mei, Wujiang Xu, Shuhang Lin, Yongfeng Zhang
Abstract: As large language models (LLMs) are increasingly deployed as service endpoints in systems, the surge in query volume creates significant scheduling challenges. Existing scheduling frameworks mainly target at latency optimization while neglecting the capability of LLMs to serve different level of queries, which could lead to computational resource waste. This paper addresses this challenge by proposing a capability-cost coordinated scheduling framework, ECCOS, for multi-LLM serving, which explicitly constrains response quality and workload to optimize LLM inference cost. Specifically, it introduces the two-stage scheduling by designing a multi-objective predictor and a constrained optimizer. The predictor estimates both model capabilities and computational costs through training-based and retrieval-based approaches, while the optimizer determines cost-optimal assignments under quality and workload constraints. It also introduces QAServe, a dataset collected for sample-wise response quality and costs by zero-shot prompting different LLMs on knowledge QA and mathematical reasoning. Extensive experiments demonstrate that ECCOS improves success rates by 6.30% while reducing costs by 10.15% compared to existing methods, consuming less than 0.5% of LLM response time. The code is available at: https://github.com/agiresearch/ECCOS.
Authors: Saeif Alhazbi, Ahmed Mohamed Hussain, Gabriele Oligeri, Panos Papadimitratos
Abstract: As Large Language Models (LLMs) become increasingly integrated into many technological ecosystems across various domains and industries, identifying which model is deployed or being interacted with is critical for the security and trustworthiness of the systems. Current verification methods typically rely on analyzing the generated output to determine the source model. However, these techniques are susceptible to adversarial attacks, operate in a post-hoc manner, and may require access to the model weights to inject a verifiable fingerprint. In this paper, we propose a novel passive and non-invasive fingerprinting technique that operates in real-time and remains effective even under encrypted network traffic conditions. Our method leverages the intrinsic autoregressive generation nature of language models, which generate text one token at a time based on all previously generated tokens, creating a unique temporal pattern like a rhythm or heartbeat that persists even when the output is streamed over a network. We find that measuring the Inter-Token Times (ITTs)-time intervals between consecutive tokens-can identify different language models with high accuracy. We develop a Deep Learning (DL) pipeline to capture these timing patterns using network traffic analysis and evaluate it on 16 Small Language Models (SLMs) and 10 proprietary LLMs across different deployment scenarios, including local host machine (GPU/CPU), Local Area Network (LAN), Remote Network, and Virtual Private Network (VPN). The experimental results confirm that our proposed technique is effective and maintains high accuracy even when tested in different network conditions. This work opens a new avenue for model identification in real-world scenarios and contributes to more secure and trustworthy language model deployment.
Authors: Saman Khamesian, Asiful Arefeen, Stephanie M. Carpenter, Hassan Ghasemzadeh
Abstract: Maintaining a balanced diet is essential for overall health, yet many individuals struggle with meal planning due to nutritional complexity, time constraints, and lack of dietary knowledge. Personalized food recommendations can help address these challenges by tailoring meal plans to individual preferences, habits, and dietary restrictions. However, existing dietary recommendation systems often lack adaptability, fail to consider real-world constraints such as food ingredient availability, and require extensive user input, making them impractical for sustainable and scalable daily use. To address these limitations, we introduce NutriGen, a framework based on large language models (LLM) designed to generate personalized meal plans that align with user-defined dietary preferences and constraints. By building a personalized nutrition database and leveraging prompt engineering, our approach enables LLMs to incorporate reliable nutritional references like the USDA nutrition database while maintaining flexibility and ease-of-use. We demonstrate that LLMs have strong potential in generating accurate and user-friendly food recommendations, addressing key limitations in existing dietary recommendation systems by providing structured, practical, and scalable meal plans. Our evaluation shows that Llama 3.1 8B and GPT-3.5 Turbo achieve the lowest percentage errors of 1.55\% and 3.68\%, respectively, producing meal plans that closely align with user-defined caloric targets while minimizing deviation and improving precision. Additionally, we compared the performance of DeepSeek V3 against several established models to evaluate its potential in personalized nutrition planning.
Authors: Yingqi Gao, Zhiling Luo
Abstract: In the context of the Text-to-SQL task, table and column descriptions are crucial for bridging the gap between natural language and database schema. This report proposes a method for automatically generating effective database descriptions when explicit descriptions are unavailable. The proposed method employs a dual-process approach: a coarse-to-fine process, followed by a fine-to-coarse process. The coarse-to-fine approach leverages the inherent knowledge of LLM to guide the understanding process from databases to tables and finally to columns. This approach provides a holistic understanding of the database structure and ensures contextual alignment. Conversely, the fine-to-coarse approach starts at the column level, offering a more accurate and nuanced understanding when stepping back to the table level. Experimental results on the Bird benchmark indicate that using descriptions generated by the proposed improves SQL generation accuracy by 0.93\% compared to not using descriptions, and achieves 37\% of human-level performance. The source code is publicly available at https://github.com/XGenerationLab/XiYan-DBDescGen.
Authors: Yuqi Wu, Guangya Wan, Jingjing Li, Shengming Zhao, Lingfeng Ma, Tianyi Ye, Ion Pop, Yanbo Zhang, Jie Chen
Abstract: Most LLM-driven conversational AI systems operate reactively, responding to user prompts without guiding the interaction. Most LLM-driven conversational AI systems operate reactively, responding to user prompts without guiding the interaction. However, many real-world applications-such as psychiatric diagnosis, consulting, and interviews-require AI to take a proactive role, asking the right questions and steering conversations toward specific objectives. Using mental health differential diagnosis as an application context, we introduce ProAI, a goal-oriented, proactive conversational AI framework. ProAI integrates structured knowledge-guided memory, multi-agent proactive reasoning, and a multi-faceted evaluation strategy, enabling LLMs to engage in clinician-style diagnostic reasoning rather than simple response generation. Through simulated patient interactions, user experience assessment, and professional clinical validation, we demonstrate that ProAI achieves up to 83.3% accuracy in mental disorder differential diagnosis while maintaining professional and empathetic interaction standards. These results highlight the potential for more reliable, adaptive, and goal-driven AI diagnostic assistants, advancing LLMs beyond reactive dialogue systems.
Authors: Xiwen Liang, Min Lin, Weiqi Ruan, Rongtao Xu, Yuecheng Liu, Jiaqi Chen, Bingqian Lin, Yuzheng Zhuang, Xiaodan Liang
Abstract: Existing methods for vision-language task planning excel in short-horizon tasks but often fall short in complex, long-horizon planning within dynamic environments. These challenges primarily arise from the difficulty of effectively training models to produce high-quality reasoning processes for long-horizon tasks. To address this, we propose Structured Preference Optimization (SPO), which aims to enhance reasoning and action selection in long-horizon task planning through structured preference evaluation and optimized training strategies. Specifically, SPO introduces: 1) Preference-Based Scoring and Optimization, which systematically evaluates reasoning chains based on task relevance, visual grounding, and historical consistency; and 2) Curriculum-Guided Training, where the model progressively adapts from simple to complex tasks, improving its generalization ability in long-horizon scenarios and enhancing reasoning robustness. To advance research in vision-language long-horizon task planning, we introduce ExtendaBench, a comprehensive benchmark covering 1,509 tasks across VirtualHome and Habitat 2.0, categorized into ultra-short, short, medium, and long tasks. Experimental results demonstrate that SPO significantly improves reasoning quality and final decision accuracy, outperforming prior methods on long-horizon tasks and underscoring the effectiveness of preference-driven optimization in vision-language task planning. Specifically, SPO achieves a +5.98% GCR and +4.68% SR improvement in VirtualHome and a +3.30% GCR and +2.11% SR improvement in Habitat over the best-performing baselines.
Authors: Shiwali Mohan, Aaron H. Mininger, James R. Kirk, John E. Laird
Abstract: We present an approach for acquiring grounded representations of words from mixed-initiative, situated interactions with a human instructor. The work focuses on the acquisition of diverse types of knowledge including perceptual, semantic, and procedural knowledge along with learning grounded meanings. Interactive learning allows the agent to control its learning by requesting instructions about unknown concepts, making learning efficient. Our approach has been instantiated in Soar and has been evaluated on a table-top robotic arm capable of manipulating small objects.
Authors: Seyed Pouyan Mousavi Davoudi, Alireza Shafiee Fard, Alireza Amiri-Margavi
Abstract: We present a collaborative framework where multiple large language models, namely GPT-4-0125-preview, Meta-LLaMA-3-70B-Instruct, Claude-3-Opus, and Gemini-1.5-Flash, work together to generate and respond to complex PhD-level probability questions in the absence of definitive ground truth. This study explores how inter-model consensus enhances response reliability and serves as a proxy for assessing the quality of generated questions. To quantify agreement and consistency, we employ statistical methods including chi-square tests, Fleiss' Kappa, and confidence interval analysis, measuring both response precision and question clarity. Our findings highlight that Claude and Gemini generate well-structured and less ambiguous questions, leading to higher inter-model agreement. This is reflected in their narrower confidence intervals and stronger alignment with answering models. Conversely, LLaMA demonstrates increased variability and lower reliability in question formulation, as indicated by broader confidence intervals and reduced consensus rates. These results suggest that multi-model collaboration not only enhances the reliability of responses but also provides a valuable framework for assessing and improving question quality in the absence of explicit ground truth. This research offers meaningful insights into optimizing AI-driven reasoning through collaborative large-language model interactions.
Authors: Xunhao Lai, Jianqiao Lu, Yao Luo, Yiyuan Ma, Xun Zhou
Abstract: Large language models (LLMs) encounter computational challenges during long-sequence inference, especially in the attention pre-filling phase, where the complexity grows quadratically with the prompt length. Previous efforts to mitigate these challenges have relied on fixed sparse attention patterns or identifying sparse attention patterns based on limited cases. However, these methods lacked the flexibility to efficiently adapt to varying input demands. In this paper, we introduce FlexPrefill, a Flexible sparse Pre-filling mechanism that dynamically adjusts sparse attention patterns and computational budget in real-time to meet the specific requirements of each input and attention head. The flexibility of our method is demonstrated through two key innovations: 1) Query-Aware Sparse Pattern Determination: By measuring Jensen-Shannon divergence, this component adaptively switches between query-specific diverse attention patterns and predefined attention patterns. 2) Cumulative-Attention Based Index Selection: This component dynamically selects query-key indexes to be computed based on different attention patterns, ensuring the sum of attention scores meets a predefined threshold. FlexPrefill adaptively optimizes the sparse pattern and sparse ratio of each attention head based on the prompt, enhancing efficiency in long-sequence inference tasks. Experimental results show significant improvements in both speed and accuracy over prior methods, providing a more flexible and efficient solution for LLM inference.
Authors: Qiao Yan, Yuchen Yuan, Xiaowei Hu, Yihan Wang, Jiaqi Xu, Jinpeng Li, Chi-Wing Fu, Pheng-Ann Heng
Abstract: The increasing use of vision-language models (VLMs) in healthcare applications presents great challenges related to hallucinations, in which the models may generate seemingly plausible results that are in fact incorrect. Such hallucinations can jeopardize clinical decision making, potentially harming the diagnosis and treatments. In this work, we propose MedHallTune, a large-scale benchmark designed specifically to evaluate and mitigate hallucinations in medical VLMs. Comprising over 100,000 images and 1,000,000 instruction pairs, MedHallTune includes both hallucination and non-hallucination samples, each with ground-truth annotations. We conduct a comprehensive evaluation of current medical and general VLMs using MedHallTune, assessing their performance across key metrics, including clinical accuracy, relevance, detail level, and risk level. The experimental results show that fine-tuning with MedHallTune successfully improves the ability of several existing models to manage hallucinations and boost their zero-shot performance on downstream visual-question-answering (VQA) tasks, making them more reliable for practical medical applications. Our work contributes to the development of more trustworthy VLMs. Codes and dataset will be available at \href{https://github.com/russellyq/MedHallTune}{MedHallTune}.
Authors: Xiao Wang, Jingyun Hua, Weihong Lin, Yuanxing Zhang, Fuzheng Zhang, Jianlong Wu, Di Zhang, Liqiang Nie
Abstract: Recent Multi-modal Large Language Models (MLLMs) have made great progress in video understanding. However, their performance on videos involving human actions is still limited by the lack of high-quality data. To address this, we introduce a two-stage data annotation pipeline. First, we design strategies to accumulate videos featuring clear human actions from the Internet. Second, videos are annotated in a standardized caption format that uses human attributes to distinguish individuals and chronologically details their actions and interactions. Through this pipeline, we curate two datasets, namely HAICTrain and HAICBench. \textbf{HAICTrain} comprises 126K video-caption pairs generated by Gemini-Pro and verified for training purposes. Meanwhile, \textbf{HAICBench} includes 500 manually annotated video-caption pairs and 1,400 QA pairs, for a comprehensive evaluation of human action understanding. Experimental results demonstrate that training with HAICTrain not only significantly enhances human understanding abilities across 4 benchmarks, but can also improve text-to-video generation results. Both the HAICTrain and HAICBench are released at https://huggingface.co/datasets/KuaishouHAIC/HAIC.
Authors: Xujie Yuan, Yongxu Liu, Shimin Di, Shiwen Wu, Libin Zheng, Rui Meng, Xiaofang Zhou, Lei Chen, Jian Yin
Abstract: The integration of Knowledge Graphs (KGs) into the Retrieval Augmented Generation (RAG) framework has attracted significant interest, with early studies showing promise in mitigating hallucinations and improving model accuracy. However, a systematic understanding and comparative analysis of the rapidly emerging KG-RAG methods are still lacking. This paper seeks to lay the foundation for systematically answering the question of when and how to use KG-RAG by analyzing their performance in various application scenarios associated with different technical configurations. After outlining the mind map using KG-RAG framework and summarizing its popular pipeline, we conduct a pilot empirical study of KG-RAG works to reimplement and evaluate 6 KG-RAG methods across 7 datasets in diverse scenarios, analyzing the impact of 9 KG-RAG configurations in combination with 17 LLMs. Our results underscore the critical role of appropriate application conditions and optimal configurations of KG-RAG components.
Authors: Maxime M\'eloux, Silviu Maniu, Fran\c{c}ois Portet, Maxime Peyrard
Abstract: As AI systems are used in high-stakes applications, ensuring interpretability is crucial. Mechanistic Interpretability (MI) aims to reverse-engineer neural networks by extracting human-understandable algorithms to explain their behavior. This work examines a key question: for a given behavior, and under MI's criteria, does a unique explanation exist? Drawing on identifiability in statistics, where parameters are uniquely inferred under specific assumptions, we explore the identifiability of MI explanations. We identify two main MI strategies: (1) "where-then-what," which isolates a circuit replicating model behavior before interpreting it, and (2) "what-then-where," which starts with candidate algorithms and searches for neural activation subspaces implementing them, using causal alignment. We test both strategies on Boolean functions and small multi-layer perceptrons, fully enumerating candidate explanations. Our experiments reveal systematic non-identifiability: multiple circuits can replicate behavior, a circuit can have multiple interpretations, several algorithms can align with the network, and one algorithm can align with different subspaces. Is uniqueness necessary? A pragmatic approach may require only predictive and manipulability standards. If uniqueness is essential for understanding, stricter criteria may be needed. We also reference the inner interpretability framework, which validates explanations through multiple criteria. This work contributes to defining explanation standards in AI.
Authors: Qiyuan Li, Haijiang Liu, Caicai Guo, Deyu Chen, Meng Wang, Feng Gao, Jinguang Gu
Abstract: Clinical knowledge is the collection of information learned from studies on the causes, prognosis, diagnosis, and treatment of diseases. This type of knowledge can improve curing performances, and promote physical health. With the emergence of large language models (LLMs), medical artificial intelligence (medical AI), which aims to apply academic medical AI systems to real-world medical scenarios, has entered a new age of development, resulting in excellent works such as DoctorGPT and Pangu-Drug from academic and industrial researches. However, the field lacks a comprehensive compendium and comparison of building medical AI systems from academia and industry. Therefore, this survey focuses on the building paradigms of medical AI systems including the use of clinical databases, datasets, training pipelines, integrating medical knowledge graphs, system applications, and evaluation systems. We hope that this survey can help relevant practical researchers understand the current performance of academic models in various fields of healthcare, as well as the potential problems and future directions for implementing these scientific achievements.
Authors: Abdelrahman Abdallah, Bhawna Piryani, Jonas Wallat, Avishek Anand, Adam Jatowt
Abstract: Temporal awareness is crucial in many information retrieval tasks, particularly in scenarios where the relevance of documents depends on their alignment with the query's temporal context. Traditional retrieval methods such as BM25 and Dense Passage Retrieval (DPR) excel at capturing lexical and semantic relevance but fall short in addressing time-sensitive queries. To bridge this gap, we introduce the temporal retrieval model that integrates explicit temporal signals by incorporating query timestamps and document dates into the representation space. Our approach ensures that retrieved passages are not only topically relevant but also temporally aligned with user intent. We evaluate our approach on two large-scale benchmark datasets, ArchivalQA and ChroniclingAmericaQA, achieving substantial performance gains over standard retrieval baselines. In particular, our model improves Top-1 retrieval accuracy by 6.63% and NDCG@10 by 3.79% on ArchivalQA, while yielding a 9.56% boost in Top-1 retrieval accuracy and 4.68% in NDCG@10 on ChroniclingAmericaQA. Additionally, we introduce a time-sensitive negative sampling strategy, which refines the model's ability to distinguish between temporally relevant and irrelevant documents during training. Our findings highlight the importance of explicitly modeling time in retrieval systems and set a new standard for handling temporally grounded queries.
Authors: Jennifer Hu, Felix Sosa, Tomer Ullman
Abstract: The question of whether large language models (LLMs) possess Theory of Mind (ToM) -- often defined as the ability to reason about others' mental states -- has sparked significant scientific and public interest. However, the evidence as to whether LLMs possess ToM is mixed, and the recent growth in evaluations has not resulted in a convergence. Here, we take inspiration from cognitive science to re-evaluate the state of ToM evaluation in LLMs. We argue that a major reason for the disagreement on whether LLMs have ToM is a lack of clarity on whether models should be expected to match human behaviors, or the computations underlying those behaviors. We also highlight ways in which current evaluations may be deviating from "pure" measurements of ToM abilities, which also contributes to the confusion. We conclude by discussing several directions for future research, including the relationship between ToM and pragmatic communication, which could advance our understanding of artificial systems as well as human cognition.
Authors: Mattia Birti, Francesco Osborne, Andrea Maurino
Abstract: The integration of Environmental, Social, and Governance (ESG) factors into corporate decision-making is a fundamental aspect of sustainable finance. However, ensuring that business practices align with evolving regulatory frameworks remains a persistent challenge. AI-driven solutions for automatically assessing the alignment of sustainability reports and non-financial disclosures with specific ESG activities could greatly support this process. Yet, this task remains complex due to the limitations of general-purpose Large Language Models (LLMs) in domain-specific contexts and the scarcity of structured, high-quality datasets. In this paper, we investigate the ability of current-generation LLMs to identify text related to environmental activities. Furthermore, we demonstrate that their performance can be significantly enhanced through fine-tuning on a combination of original and synthetically generated data. To this end, we introduce ESG-Activities, a benchmark dataset containing 1,325 labelled text segments classified according to the EU ESG taxonomy. Our experimental results show that fine-tuning on ESG-Activities significantly enhances classification accuracy, with open models such as Llama 7B and Gemma 7B outperforming large proprietary solutions in specific configurations. These findings have important implications for financial analysts, policymakers, and AI researchers seeking to enhance ESG transparency and compliance through advanced natural language processing techniques.
Authors: Daniil Filienko, Mahek Nizar, Javier Roberti, Denise Galdamez, Haroon Jakher, Sarah Iribarren, Weichao Yuwen, Martine De Cock
Abstract: Tuberculosis (TB) is the leading cause of death from an infectious disease globally, with the highest burden in low- and middle-income countries. In these regions, limited healthcare access and high patient-to-provider ratios impede effective patient support, communication, and treatment completion. To bridge this gap, we propose integrating a specialized Large Language Model into an efficacious digital adherence technology to augment interactive communication with treatment supporters. This AI-powered approach, operating within a human-in-the-loop framework, aims to enhance patient engagement and improve TB treatment outcomes.
Authors: Hongzhan Lin, Ziyang Luo, Bo Wang, Ruichao Yang, Jing Ma
Abstract: The exponential growth of social media has profoundly transformed how information is created, disseminated, and absorbed, exceeding any precedent in the digital age. Regrettably, this explosion has also spawned a significant increase in the online abuse of memes. Evaluating the negative impact of memes is notably challenging, owing to their often subtle and implicit meanings, which are not directly conveyed through the overt text and image. In light of this, large multimodal models (LMMs) have emerged as a focal point of interest due to their remarkable capabilities in handling diverse multimodal tasks. In response to this development, our paper aims to thoroughly examine the capacity of various LMMs (e.g., GPT-4o) to discern and respond to the nuanced aspects of social abuse manifested in memes. We introduce the comprehensive meme benchmark, GOAT-Bench, comprising over 6K varied memes encapsulating themes such as implicit hate speech, sexism, and cyberbullying, etc. Utilizing GOAT-Bench, we delve into the ability of LMMs to accurately assess hatefulness, misogyny, offensiveness, sarcasm, and harmful content. Our extensive experiments across a range of LMMs reveal that current models still exhibit a deficiency in safety awareness, showing insensitivity to various forms of implicit abuse. We posit that this shortfall represents a critical impediment to the realization of safe artificial intelligence. The GOAT-Bench and accompanying resources are publicly accessible at https://goatlmm.github.io/, contributing to ongoing research in this vital field.
Authors: Jiahuan Yan, Jintai Chen, Chaowen Hu, Bo Zheng, Yaojun Hu, Jimeng Sun, Jian Wu
Abstract: Recent development in large language models (LLMs) has demonstrated impressive domain proficiency on unstructured textual or multi-modal tasks. However, despite with intrinsic world knowledge, their application on structured tabular data prediction still lags behind, primarily due to the numerical insensitivity and modality discrepancy that brings a gap between LLM reasoning and statistical tabular learning. Unlike textual or vision data (e.g., electronic clinical notes or medical imaging data), tabular data is often presented in heterogeneous numerical values (e.g., CBC reports). This ubiquitous data format requires intensive expert annotation, and its numerical nature limits LLMs' capability to effectively transfer untapped domain expertise. In this paper, we propose SERSAL, a general self-prompting method by synergy learning with small models to enhance LLM tabular prediction in an unsupervised manner. Specifically, SERSAL utilizes the LLM's prior outcomes as original soft noisy annotations, which are dynamically leveraged to teach a better small student model. Reversely, the outcomes from the trained small model are used to teach the LLM to further refine its real capability. This process can be repeatedly applied to gradually distill refined knowledge for continuous progress. Comprehensive experiments on widely used medical domain tabular datasets show that, without access to gold labels, applying SERSAL to OpenAI GPT reasoning process attains substantial improvement compared to linguistic prompting methods, which serves as an orthogonal direction for tabular LLM, and increasing prompting bonus is observed as more powerful LLMs appear.
Authors: Seanie Lee, Minsu Kim, Lynn Cherif, David Dobre, Juho Lee, Sung Ju Hwang, Kenji Kawaguchi, Gauthier Gidel, Yoshua Bengio, Nikolay Malkin, Moksh Jain
Abstract: Red-teaming, or identifying prompts that elicit harmful responses, is a critical step in ensuring the safe and responsible deployment of large language models (LLMs). Developing effective protection against many modes of attack prompts requires discovering diverse attacks. Automated red-teaming typically uses reinforcement learning to fine-tune an attacker language model to generate prompts that elicit undesirable responses from a target LLM, as measured, for example, by an auxiliary toxicity classifier. We show that even with explicit regularization to favor novelty and diversity, existing approaches suffer from mode collapse or fail to generate effective attacks. As a flexible and probabilistically principled alternative, we propose to use GFlowNet fine-tuning, followed by a secondary smoothing phase, to train the attacker model to generate diverse and effective attack prompts. We find that the attacks generated by our method are effective against a wide range of target LLMs, both with and without safety tuning, and transfer well between target LLMs. Finally, we demonstrate that models safety-tuned using a dataset of red-teaming prompts generated by our method are robust to attacks from other RL-based red-teaming approaches.
Authors: Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng
Abstract: Retrieval-augmented generation (RAG) utilizes retrieved texts to enhance large language models (LLMs). Studies show that while RAG provides valuable external information (benefit), it may also mislead LLMs (detriment) with noisy or incorrect retrieved texts. Although many existing methods attempt to preserve benefit and avoid detriment, they lack a theoretical explanation for RAG. The benefit and detriment in the next token prediction of RAG remain a black box that cannot be quantified or compared in an explainable manner, so existing methods are data-driven, need additional utility evaluators or post-hoc. This paper takes the first step towards providing a theory to explain and trade off the benefit and detriment in RAG. First, we model RAG as the fusion between distribution of LLMs knowledge and distribution of retrieved texts. Then, we formalize the trade-off between the value of external knowledge (benefit) and its potential risk of misleading LLMs (detriment) in next token prediction of RAG by distribution difference in this fusion. Finally, we prove that the actual effect of RAG on the token, which is the comparison between benefit and detriment, can be predicted without any training or accessing the utility of retrieval. Based on our theory, we propose a practical novel method, Tok-RAG, which achieves collaborative generation between the pure LLM and RAG at token level to preserve benefit and avoid detriment. Experiments in real-world tasks using LLMs such as OPT, LLaMA-2, and Mistral show the effectiveness of our method and support our theoretical findings.
Authors: Liliang Ren, Yang Liu, Yadong Lu, Yelong Shen, Chen Liang, Weizhu Chen
Abstract: Efficiently modeling sequences with infinite context length has long been a challenging problem. Previous approaches have either suffered from quadratic computational complexity or limited extrapolation ability in length generalization. In this work, we present Samba, a simple hybrid architecture that layer-wise combines Mamba, a selective State Space Model (SSM), with Sliding Window Attention (SWA). Samba selectively compresses a given sequence into recurrent hidden states while still maintaining the ability to precisely recall recent memories with the attention mechanism. We scale Samba up to 3.8B parameters with 3.2T training tokens and demonstrate that it significantly outperforms state-of-the-art models across a variety of benchmarks. Pretrained on sequences of 4K length, Samba shows improved perplexity in context lengths of up to 1M in zero-shot. When finetuned on 4K-length sequences, Samba efficiently extrapolates to a 256K context length with perfect memory recall on the Passkey Retrieval task, and exhibits superior retrieval extrapolation on the challenging Phonebook task compared to full-attention models. As a linear-time sequence model, Samba achieves a 3.73x higher throughput compared to Transformers with grouped-query attention for user prompts of 128K length, and a 3.64x speedup when generating 64K tokens with unlimited streaming. Our code for training on open source data is publicly available at https://github.com/microsoft/Samba.
Authors: Xiaoshuai Song, Muxi Diao, Guanting Dong, Zhengyang Wang, Yujia Fu, Runqi Qiao, Zhexu Wang, Dayuan Fu, Huangxuan Wu, Bin Liang, Weihao Zeng, Yejie Wang, Zhuoma GongQue, Jianing Yu, Qiuna Tan, Weiran Xu
Abstract: Large language models (LLMs) have demonstrated significant potential in advancing various fields of research and society. However, the current community of LLMs overly focuses on benchmarks for analyzing specific foundational skills (e.g. mathematics and code generation), neglecting an all-round evaluation of the computer science field. To bridge this gap, we introduce CS-Bench, the first multilingual (English, Chinese, French, German) benchmark dedicated to evaluating the performance of LLMs in computer science. CS-Bench comprises approximately 10K meticulously curated test samples, covering 26 subfields across 4 key areas of computer science, encompassing various task forms and divisions of knowledge and reasoning. Utilizing CS-Bench, we conduct a comprehensive evaluation of over 30 mainstream LLMs, revealing the relationship between CS performance and model scales. We also quantitatively analyze the reasons for failures in existing LLMs and highlight directions for improvements, including knowledge supplementation and CS-specific reasoning. Further cross-capability experiments show a high correlation between LLMs' capabilities in computer science and their abilities in mathematics and coding. Moreover, expert LLMs specialized in mathematics and coding also demonstrate strong performances in several CS subfields. Looking ahead, we envision CS-Bench serving as a cornerstone for LLM applications in the CS field and paving new avenues in assessing LLMs' diverse reasoning capabilities. The CS-Bench data and evaluation code are available at https://github.com/csbench/csbench.
Authors: Francisco Eiras, Aleksandar Petrov, Philip H. S. Torr, M. Pawan Kumar, Adel Bibi
Abstract: Recent research shows that fine-tuning on benign instruction-following data can inadvertently undo the safety alignment process and increase a model's propensity to comply with harmful queries. While instruction-following fine-tuning is important, task-specific fine-tuning - where models are trained on datasets with clear ground truth answers (e.g., multiple choice questions) - can enhance model performance on specialized downstream tasks. Understanding and mitigating safety risks in the task-specific setting remains distinct from the instruction-following context due to structural differences in the data. Our work demonstrates how malicious actors can subtly manipulate the structure of almost any task-specific dataset to foster significantly more dangerous model behaviors, while maintaining an appearance of innocuity and reasonable downstream task performance. To address this issue, we propose a novel mitigation strategy that mixes in safety data which mimics the task format and prompting style of the user data, showing this is significantly more effective and efficient than existing baselines at re-establishing safety alignment while maintaining similar task performance.
Authors: Wenkai Yang, Shiqi Shen, Guangyao Shen, Wei Yao, Yong Liu, Zhi Gong, Yankai Lin, Ji-Rong Wen
Abstract: Superalignment, where humans act as weak supervisors for superhuman models, has become a crucial problem with the rapid development of Large Language Models (LLMs). Recent work has preliminarily studied this problem by using weak models to supervise strong models, and discovered that weakly supervised strong students can consistently outperform weak teachers towards the alignment target, leading to a weak-to-strong generalization phenomenon. However, we are concerned that behind such a promising phenomenon, whether there exists an issue of weak-to-strong deception, where strong models deceive weak models by exhibiting well-aligned in areas known to weak models but producing misaligned behaviors in cases weak models do not know. We take an initial step towards exploring this security issue in a specific but realistic multi-objective alignment case, where there may be some alignment targets conflicting with each other (e.g., helpfulness v.s. harmlessness). We aim to explore whether, in such cases, strong models might deliberately make mistakes in areas known to them but unknown to weak models within one alignment dimension, in exchange for a higher reward in another dimension. Through extensive experiments in both the reward modeling and preference optimization scenarios, we find: (1) The weak-to-strong deception phenomenon exists across all settings. (2) The deception intensifies as the capability gap between weak and strong models increases. (3) Bootstrapping with an intermediate model can mitigate the deception to some extent, though its effectiveness remains limited. Our work highlights the urgent need to pay more attention to the true reliability of superalignment.
Authors: Jeffrey Willette, Heejun Lee, Youngwan Lee, Myeongjae Jeon, Sung Ju Hwang
Abstract: The transformer's context window is vital for tasks such as few-shot learning and conditional generation as it preserves previous tokens for active memory. However, as the context lengths increase, the computational costs grow quadratically, hindering the deployment of large language models (LLMs) in real-world, long sequence scenarios. Although some recent key-value caching (KV Cache) methods offer linear inference complexity, they naively manage the stored context, prematurely evicting tokens and losing valuable information. Moreover, they lack an optimized prefill/prompt stage strategy, resulting in higher latency than even quadratic attention for realistic context sizes. In response, we introduce a novel mechanism that leverages cascading sub-cache buffers to selectively retain the most relevant tokens, enabling the model to maintain longer context histories without increasing the cache size. Our approach outperforms linear caching baselines across key benchmarks, including streaming perplexity, question answering, book summarization, and passkey retrieval, where it retains better retrieval accuracy at 1M tokens after four doublings of the cache size of 65K. Additionally, our method reduces prefill stage latency by a factor of 6.8 when compared to flash attention on 1M tokens. These innovations not only enhance the computational efficiency of LLMs but also pave the way for their effective deployment in resource-constrained environments, enabling large-scale, real-time applications with significantly reduced latency.
Authors: Zilong Wang, Zifeng Wang, Long Le, Huaixiu Steven Zheng, Swaroop Mishra, Vincent Perot, Yuwei Zhang, Anush Mattapalli, Ankur Taly, Jingbo Shang, Chen-Yu Lee, Tomas Pfister
Abstract: Retrieval augmented generation (RAG) combines the generative abilities of large language models (LLMs) with external knowledge sources to provide more accurate and up-to-date responses. Recent RAG advancements focus on improving retrieval outcomes through iterative LLM refinement or self-critique capabilities acquired through additional instruction tuning of LLMs. In this work, we introduce Speculative RAG - a framework that leverages a larger generalist LM to efficiently verify multiple RAG drafts produced in parallel by a smaller, distilled specialist LM. Each draft is generated from a distinct subset of retrieved documents, offering diverse perspectives on the evidence while reducing input token counts per draft. This approach enhances comprehension of each subset and mitigates potential position bias over long context. Our method accelerates RAG by delegating drafting to the smaller specialist LM, with the larger generalist LM performing a single verification pass over the drafts. Extensive experiments demonstrate that Speculative RAG achieves state-of-the-art performance with reduced latency on TriviaQA, MuSiQue, PopQA, PubHealth, and ARC-Challenge benchmarks. It notably enhances accuracy by up to 12.97% while reducing latency by 50.83% compared to conventional RAG systems on PubHealth.
Authors: Xiang Lisa Li, Farzaan Kaiyom, Evan Zheran Liu, Yifan Mai, Percy Liang, Tatsunori Hashimoto
Abstract: We present AutoBencher, a declarative framework for automatic benchmark construction, and use it to scalably discover novel insights and vulnerabilities of existing language models. Concretely, given a few desiderata of benchmarks (e.g., question difficulty, topic salience), we operationalize each desideratum and cast benchmark creation as an optimization problem. Specifically, we experiment with two settings with different optimization objectives: (i) for capability evaluation, we declare the goal of finding a salient, difficult dataset that induces novel performance patterns; (ii) for safety evaluation, we declare the goal of finding a dataset of unsafe prompts that existing LMs fail to decline. To tackle this optimization problem, we use a language model to iteratively propose and refine dataset descriptions, which are then used to generate topic-specific questions and answers. These descriptions are optimized to improve the declared desiderata. We use AutoBencher (powered by GPT-4) to create datasets for math, multilinguality, knowledge, and safety. The scalability of AutoBencher allows it to test fine-grained categories and tail knowledge, creating datasets that elicit 22% more model errors (i.e., difficulty) than existing benchmarks. On the novelty ends, AutoBencher also helps identify specific gaps not captured by existing benchmarks: e.g., Gemini-Pro has knowledge gaps on Permian Extinction and Fordism while GPT-4o fails to decline harmful requests about cryptocurrency scams.
Authors: Kaushal Kumar Maurya, KV Aditya Srivatsa, Ekaterina Kochmar
Abstract: Large language models (LLMs) have been widely adopted due to their remarkable performance across various applications, driving the accelerated development of a large number of diverse models. However, these individual LLMs show limitations in generalization and performance on complex tasks due to inherent training biases, model size constraints, and the quality or diversity of pre-training datasets. A promising direction is to efficiently harness the diverse capabilities of LLMs to overcome these individual limitations. To address these limitations, we introduce a novel LLM selection algorithm called SelectLLM, which efficiently directs input queries to the most suitable subset of LLMs from a large pool, ensuring that the selected models collectively provide accurate responses. SelectLLM employs a multi-label classifier and policy based on the classifier's predictions and confidence scores in selecting an optimal, query-aware, and lightweight subset of LLMs. Our findings indicate that the proposed model outperforms existing ensemble-based baselines and achieves competitive performance with similarly sized top-performing LLMs while maintaining efficiency. Specifically, it achieves a huge reduction in inference latency on two challenging reasoning benchmarks: 13% on GSM8K and 70% on MMLU, compared to the top-performing baseline. Also, we establish a theoretical upper bound by an Oracle with LLMs and perform an in-depth linguistic analysis to understand the performance gap between the Oracle and SelectLLM.
Authors: Shiwen Ni, Xiangtao Kong, Chengming Li, Xiping Hu, Ruifeng Xu, Jia Zhu, Min Yang
Abstract: The success of Large Language Models (LLMs) relies heavily on the huge amount of pre-training data learned in the pre-training phase. The opacity of the pre-training process and the training data causes the results of many benchmark tests to become unreliable. If any model has been trained on a benchmark test set, it can seriously hinder the health of the field. In order to automate and efficiently test the capabilities of large language models, numerous mainstream benchmarks adopt a multiple-choice format. As the swapping of the contents of multiple-choice options does not affect the meaning of the question itself, we propose a simple and effective data leakage detection method based on this property. Specifically, we shuffle the contents of the options in the data to generate the corresponding derived data sets, and then detect data leakage based on the model's log probability distribution over the derived data sets. If there is a maximum and outlier in the set of log probabilities, it indicates that the data is leaked. Our method is able to work under gray-box conditions without access to model training data or weights, effectively identifying data leakage from benchmark test sets in model pre-training data, including both normal scenarios and complex scenarios where options may have been shuffled intentionally or unintentionally. Through experiments based on two LLMs and benchmark designs, we demonstrate the effectiveness of our method. In addition, we evaluate the degree of data leakage of 35 mainstream open-source LLMs on four benchmark datasets and give a ranking of the leaked LLMs for each benchmark, and we find that the Qwen family of LLMs has the highest degree of data leakage.
Authors: Andrea Silvi, Jonathan Thomas, Emil Carlsson, Devdatt Dubhashi, Moa Johansson
Abstract: It has previously been shown that by using reinforcement learning (RL), agents can derive simple approximate and exact-restricted numeral systems that are similar to human ones (Carlsson, 2021). However, it is a major challenge to show how more complex recursive numeral systems, similar to for example English, could arise via a simple learning mechanism such as RL. Here, we introduce an approach towards deriving a mechanistic explanation of the emergence of efficient recursive number systems. We consider pairs of agents learning how to communicate about numerical quantities through a meta-grammar that can be gradually modified throughout the interactions. %We find that the seminal meta-grammar of Hurford (Hurford, 1975) is not suitable for this application as its optimization results in systems that deviate from standard conventions observed within human numeral systems. We propose a simple modification which addresses this issue. Utilising a slightly modified version of the meta-grammar of Hurford, we demonstrate that our RL agents, shaped by the pressures for efficient communication, can effectively modify their lexicon towards Pareto-optimal configurations which are comparable to those observed within human numeral systems in terms of their efficiency.
Authors: Yanming Liu, Xinyue Peng, Jiannan Cao, Shi Bo, Yanxin Shen, Tianyu Du, Sheng Cheng, Xun Wang, Jianwei Yin, Xuhong Zhang
Abstract: Large language models (LLMs) have shown remarkable capabilities in natural language processing; however, they still face difficulties when tasked with understanding lengthy contexts and executing effective question answering. These challenges often arise due to the complexity and ambiguity present in longer texts. To enhance the performance of LLMs in such scenarios, we introduce the Long Question Coreference Adaptation (LQCA) method. This innovative framework focuses on coreference resolution tailored to long contexts, allowing the model to identify and manage references effectively. The LQCA method encompasses four key steps: resolving coreferences within sub-documents, computing the distances between mentions, defining a representative mention for coreference, and answering questions through mention replacement. By processing information systematically, the framework provides easier-to-handle partitions for LLMs, promoting better understanding. Experimental evaluations on a range of LLMs and datasets have yielded positive results, with a notable improvements on OpenAI-o1-mini and GPT-4o models, highlighting the effectiveness of leveraging coreference resolution to bridge context gaps in question answering. Our code is public at https://github.com/OceannTwT/LQCA.
Authors: Xiao Yu, Baolin Peng, Vineeth Vajipey, Hao Cheng, Michel Galley, Jianfeng Gao, Zhou Yu
Abstract: Autonomous agents have demonstrated significant potential in automating complex multistep decision-making tasks. However, even state-of-the-art vision-language models (VLMs), such as GPT-4o, still fall short of human-level performance, particularly in intricate web environments and long-horizon tasks. To address these limitations, we present ExACT, an approach to combine test-time search and self-learning to build o1-like models for agentic applications. We first introduce Reflective Monte Carlo Tree Search (R-MCTS), a novel test time algorithm designed to enhance AI agents' ability to explore decision space on the fly. R-MCTS extends traditional MCTS by 1) incorporating contrastive reflection, allowing agents to learn from past interactions and dynamically improve their search efficiency; and 2) using multi-agent debate for reliable state evaluation. Next, we introduce Exploratory Learning, a novel learning strategy to teach agents to search at inference time without relying on any external search algorithms. On the challenging VisualWebArena benchmark, our GPT-4o based R-MCTS agent achieves a 6% to 30% relative improvement across various tasks compared to the previous state-of-the-art. Additionally, we show that the knowledge and experience gained from test-time search can be effectively transferred back to GPT-4o via fine-tuning. After Exploratory Learning, GPT-4o 1) demonstrates the ability to explore the environment, evaluate a state, and backtrack to viable ones when it detects that the current state cannot lead to success, and 2) matches 87% of R-MCTS's performance while using significantly less compute. Notably, our work demonstrates the compute scaling properties in both training - data collection with R-MCTS - and testing time. These results suggest a promising research direction to enhance VLMs' capabilities for agentic applications via test-time search and self-learning.
Authors: Guy Kaplan, Matanel Oren, Yuval Reif, Roy Schwartz
Abstract: Natural language is composed of words, but modern large language models (LLMs) process sub-words as input. A natural question raised by this discrepancy is whether LLMs encode words internally, and if so how. We present evidence that LLMs engage in an intrinsic detokenization process, where sub-word sequences are combined into coherent whole-word representations at their last token. Our experiments show that this process primarily takes place within the early and middle layers of the model. We further demonstrate its robustness to arbitrary splits (e.g., "cats" to "ca" and "ts"), typos, and importantly-to out-of-vocabulary words: when feeding the last token internal representations of such words to the model as input, it can "understand" them as the complete word despite never seeing such representations as input during training. Our findings suggest that LLMs maintain a latent vocabulary beyond the tokenizer's scope. These insights provide a practical, finetuning-free application for expanding the vocabulary of pre-trained models. By enabling the addition of new vocabulary words, we reduce input length and inference iterations, which reduces both space and model latency, with little to no loss in model accuracy.
Authors: Guoxin Chen, Zhong Zhang, Xin Cong, Fangda Guo, Yesai Wu, Yankai Lin, Wenzheng Feng, Yasheng Wang
Abstract: Tool learning enables large language models (LLMs) to interact with external tools and APIs, greatly expanding the application scope of LLMs. However, due to the dynamic nature of external environments, these tools and APIs may become outdated over time, preventing LLMs from correctly invoking tools. Existing research primarily focuses on static environments and overlooks this issue, limiting the adaptability of LLMs in real-world applications. In this paper, we propose ToolEVO, a novel framework designed to enhance the adaptive and reflective capabilities of LLMs against tool variability. By leveraging Monte Carlo Tree Search, ToolEVO facilitates active exploration and interaction of LLMs within dynamic environments, allowing for autonomous self-reflection and self-updating of tool usage based on environmental feedback. Additionally, we introduce ToolQA-D, a benchmark specifically designed to evaluate the impact of tool variability. Extensive experiments demonstrate the effectiveness and stability of our approach, highlighting the importance of adaptability to tool variability for effective tool learning. Code: https://github.com/Chen-GX/ToolEVO
Authors: Kai Zhang, Congchao Wang, Liqian Peng, Alec Go, Xiaozhong Liu
Abstract: Large Language Models (LLMs) have gained significant attention in on-device applications due to their remarkable performance across real-world tasks. However, on-device LLMs often suffer from suboptimal performance due to hardware limitations. A promising solution to this challenge is cascading a weaker local (on-device) LLM with a more powerful server LLM. While existing research on LLM cascade primarily optimizes the performance-cost trade-off, real-world applications impose additional requirements, such as privacy preservation, which remain largely unaddressed. In this work, we move beyond existing confidence- and logit-based LLM cascade methods and propose $\mathbf{P^{3}Defer}$, a novel Chain-of-Thought (CoT)-enhanced \textbf{p}olicy learning framework for \textbf{p}rivacy-\textbf{p}reserved \textbf{defer}ral decision-making. Our approach effectively improves cascade efficiency while mitigating privacy risks. Extensive experiments on three benchmark datasets demonstrate the effectiveness and superiority of $\mathbf{P^{3}Defer}$ over existing methods.
Authors: Maximus Powers, Shaina Raza, Alex Chang, Umang Mavani, Harshitha Reddy Jonala, Ansh Tiwari, Hua Wei
Abstract: The detection of social bias in text is a critical challenge, particularly due to the limitations of binary classification methods. These methods often oversimplify nuanced biases, leading to high emotional impact when content is misclassified as either "biased" or "fair." To address these shortcomings, we propose a more nuanced framework that focuses on three key linguistic components underlying social bias: Generalizations, Unfairness, and Stereotypes (the GUS framework). The GUS framework employs a semi-automated approach to create a comprehensive synthetic dataset, which is then verified by humans to maintain ethical standards. This dataset enables robust multi-label token classification. Our methodology, which combines discriminative (encoder-only) models and generative (auto-regressive large language models), identifies biased entities in text. Through extensive experiments, we demonstrate that encoder-only models are effective for this complex task, often outperforming state-of-the-art methods, both in terms of macro and entity-wise F1-score and Hamming loss. These findings can guide the choice of model for different use cases, highlighting the GUS framework's effectiveness in capturing explicit and implicit biases across diverse contexts, and offering a pathway for future research and applications in various fields.
Authors: Jiachun Li, Pengfei Cao, Zhuoran Jin, Yubo Chen, Kang Liu, Jun Zhao
Abstract: Inductive reasoning is an essential capability for large language models (LLMs) to achieve higher intelligence, which requires the model to generalize rules from observed facts and then apply them to unseen examples. We present MIRAGE, a synthetic dataset that addresses the limitations of previous work, specifically the lack of comprehensive evaluation and flexible test data. In it, we evaluate LLMs' capabilities in both the inductive and deductive stages, allowing for flexible variation in input distribution, task scenario, and task difficulty to analyze the factors influencing LLMs' inductive reasoning. Based on these multi-faceted evaluations, we demonstrate that the LLM is a poor rule-based reasoner. In many cases, when conducting inductive reasoning, they do not rely on a correct rule to answer the unseen case. From the perspectives of different prompting methods, observation numbers, and task forms, models tend to consistently conduct correct deduction without correct inductive rules. Besides, we find that LLMs are good neighbor-based reasoners. In the inductive reasoning process, the model tends to focus on observed facts that are close to the current test example in feature space. By leveraging these similar examples, the model maintains strong inductive capabilities within a localized region, significantly improving its deductive performance.
Authors: Yein Park, Chanwoong Yoon, Jungwoo Park, Donghyeon Lee, Minbyul Jeong, Jaewoo Kang
Abstract: Large language models (LLMs) have brought significant changes to many aspects of our lives. However, assessing and ensuring their chronological knowledge remains challenging. Existing approaches fall short in addressing the temporal adaptability of knowledge, often relying on a fixed time-point view. To overcome this, we introduce ChroKnowBench, a benchmark dataset designed to evaluate chronologically accumulated knowledge across three key aspects: multiple domains, time dependency, temporal state. Our benchmark distinguishes between knowledge that evolves (e.g., personal history, scientific discoveries, amended laws) and knowledge that remain constant (e.g., mathematical truths, commonsense facts). Building on this benchmark, we present ChroKnowledge (Chronological Categorization of Knowledge), a novel sampling-based framework for evaluating LLMs' non-parametric chronological knowledge. Our evaluation led to the following observations: (1) The ability of eliciting temporal knowledge varies depending on the data format that model was trained on. (2) LLMs partially recall knowledge or show a cut-off at temporal boundaries rather than recalling all aspects of knowledge correctly. Thus, we apply our ChroKnowPrompt, an in-depth prompting to elicit chronological knowledge by traversing step-by-step through the surrounding time spans. We observe that it successfully recalls objects across both open-source and proprietary LLMs, demonstrating versatility, though it faces challenges with dynamic datasets and unstructured formats.
Authors: Xiongtao Zhou, Jie He, Lanyu Chen, Jingyu Li, Haojing Chen, V\'ictor Guti\'errez-Basulto, Jeff Z. Pan, Hanjie Chen
Abstract: Multimodal Chain of Thought (MCoT) is a popular prompting strategy for improving the performance of multimodal large language models (MLLMs) across a range of complex reasoning tasks. Despite its popularity, there is a notable absence of automated methods for evaluating the quality of reasoning steps in MCoT. To address this gap, we propose Multimodal Chain-of-Thought Evaluation (MiCEval), a framework designed to assess the correctness of reasoning chains by evaluating the quality of both the description and each reasoning step. The evaluation of the description component focuses on the accuracy of the image descriptions, while the reasoning step evaluates the quality of each step as it is conditionally generated based on the preceding steps. MiCEval is built upon a fine-grained dataset with annotations that rate each step according to correctness, relevance, and informativeness. Extensive experiments on four state-of-the-art MLLMs show that step-wise evaluations using MiCEval align more closely with human judgments compared to existing methods based on cosine similarity or fine-tuning approaches. MiCEval datasets and code can be found in https://github.com/alenai97/MiCEval.
Authors: Samuele Poppi, Zheng-Xin Yong, Yifei He, Bobbie Chern, Han Zhao, Aobo Yang, Jianfeng Chi
Abstract: Recent advancements in Large Language Models (LLMs) have sparked widespread concerns about their safety. Recent work demonstrates that safety alignment of LLMs can be easily removed by fine-tuning with a few adversarially chosen instruction-following examples, i.e., fine-tuning attacks. We take a further step to understand fine-tuning attacks in multilingual LLMs. We first discover cross-lingual generalization of fine-tuning attacks: using a few adversarially chosen instruction-following examples in one language, multilingual LLMs can also be easily compromised (e.g., multilingual LLMs fail to refuse harmful prompts in other languages). Motivated by this finding, we hypothesize that safety-related information is language-agnostic and propose a new method termed Safety Information Localization (SIL) to identify the safety-related information in the model parameter space. Through SIL, we validate this hypothesis and find that only changing 20% of weight parameters in fine-tuning attacks can break safety alignment across all languages. Furthermore, we provide evidence to the alternative pathways hypothesis for why freezing safety-related parameters does not prevent fine-tuning attacks, and we demonstrate that our attack vector can still jailbreak LLMs adapted to new languages.
Authors: Sangmin Bae, Adam Fisch, Hrayr Harutyunyan, Ziwei Ji, Seungyeon Kim, Tal Schuster
Abstract: Large language models (LLMs) are expensive to deploy. Parameter sharing offers a possible path towards reducing their size and cost, but its effectiveness in modern LLMs remains fairly limited. In this work, we revisit "layer tying" as form of parameter sharing in Transformers, and introduce novel methods for converting existing LLMs into smaller "Recursive Transformers" that share parameters across layers, with minimal loss of performance. Here, our Recursive Transformers are efficiently initialized from standard pretrained Transformers, but only use a single block of unique layers that is then repeated multiple times in a loop. We further improve performance by introducing Relaxed Recursive Transformers that add flexibility to the layer tying constraint via depth-wise low-rank adaptation (LoRA) modules, yet still preserve the compactness of the overall model. We show that our recursive models (e.g., recursive Gemma 1B) outperform both similar-sized vanilla pretrained models (such as TinyLlama 1.1B and Pythia 1B) and knowledge distillation baselines -- and can even recover most of the performance of the original "full-size" model (e.g., Gemma 2B with no shared parameters). Finally, we propose Continuous Depth-wise Batching, a promising new inference paradigm enabled by the Recursive Transformer when paired with early exiting. In a theoretical analysis, we show that this has the potential to lead to significant (2-3x) gains in inference throughput.
Authors: Minkai Xu, Tomas Geffner, Karsten Kreis, Weili Nie, Yilun Xu, Jure Leskovec, Stefano Ermon, Arash Vahdat
Abstract: Despite remarkable progress in autoregressive language models, alternative generative paradigms beyond left-to-right generation are still being actively explored. Discrete diffusion models, with the capacity for parallel generation, have recently emerged as a promising alternative. Unfortunately, these models still underperform the autoregressive counterparts, with the performance gap increasing when reducing the number of sampling steps. Our analysis reveals that this degradation is a consequence of an imperfect approximation used by diffusion models. In this work, we propose Energy-based Diffusion Language Model (EDLM), an energy-based model operating at the full sequence level for each diffusion step, introduced to improve the underlying approximation used by diffusion models. More specifically, we introduce an EBM in a residual form, and show that its parameters can be obtained by leveraging a pretrained autoregressive model or by finetuning a bidirectional transformer via noise contrastive estimation. We also propose an efficient generation algorithm via parallel important sampling. Comprehensive experiments on language modeling benchmarks show that our model can consistently outperform state-of-the-art diffusion models by a significant margin, and approaches autoregressive models' perplexity. We further show that, without any generation performance drop, our framework offers a 1.3$\times$ sampling speedup over existing diffusion models.
Authors: Chenghua Huang, Zhizhen Fan, Lu Wang, Fangkai Yang, Pu Zhao, Zeqi Lin, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan, Qi Zhang
Abstract: Reinforcement Learning from Human Feedback (RLHF) is a crucial technique for aligning language models with human preferences, playing a pivotal role in the success of conversational models like GPT-4, ChatGPT, and Llama 2. A core challenge in employing RLHF lies in training a reliable reward model (RM), which relies on high-quality labels typically provided by human experts or advanced AI system. These methods can be costly and may introduce biases that affect the language model's responses. As language models improve, human input may become less effective in further enhancing their performance. In this paper, we propose Self-Evolved Reward Learning (SER), a novel approach where the RM generates additional training data to iteratively improve itself. We conducted extensive experiments on multiple datasets such as HH-RLHF and UltraFeedback, using models like Mistral and Llama 3, and compare SER against various baselines. Our results demonstrate that even with limited human-annotated data, learning from self-feedback can robustly enhance RM performance, thereby boosting the capabilities of large language models (LLMs).
Authors: Fujie Zhang, Peiqi Yu, Biao Yi, Baolei Zhang, Tong Li, Zheli Liu
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across a variety of tasks in different domains. However, they sometimes generate responses that are logically coherent but factually incorrect or misleading, which is known as LLM hallucinations. Data-driven supervised methods train hallucination detectors by leveraging the internal states of LLMs, but detectors trained on specific domains often struggle to generalize well to other domains. In this paper, we aim to enhance the cross-domain performance of supervised detectors with only in-domain data. We propose a novel framework, prompt-guided internal states for hallucination detection of LLMs, namely PRISM. By utilizing appropriate prompts to guide changes to the structure related to text truthfulness in LLMs' internal states, we make this structure more salient and consistent across texts from different domains. We integrated our framework with existing hallucination detection methods and conducted experiments on datasets from different domains. The experimental results indicate that our framework significantly enhances the cross-domain generalization of existing hallucination detection methods.
Authors: Shu Wang, Lei Ji, Renxi Wang, Wenxiao Zhao, Haokun Liu, Yifan Hou, Ying Nian Wu
Abstract: Reasoning is a central capability of human intelligence. In recent years, with the advent of large-scale datasets, pretrained large language models have emerged with new capabilities, including reasoning. However, these models still struggle with long-term, complex reasoning tasks, such as playing chess. Based on the observation that expert chess players employ a dual approach combining long-term strategic play with short-term tactical play along with language explanation, we propose improving the reasoning capability of large language models in chess by integrating annotated strategy and tactic. Specifically, we collect a dataset named MATE, which consists of 1 million chess positions with candidate moves annotated by chess experts for strategy and tactics. We finetune the LLaMA-3-8B model and compare it against state-of-the-art commercial language models in the task of selecting better chess moves. Our experiments show that our models perform better than GPT, Claude, and Gemini models. We find that language explanations can enhance the reasoning capability of large language models.
Authors: Daniel P. Jeong, Pranav Mani, Saurabh Garg, Zachary C. Lipton, Michael Oberst
Abstract: Several recent works seek to adapt general-purpose large language models (LLMs) and vision-language models (VLMs) for medical applications through continued pretraining on publicly available biomedical corpora. These works typically claim that such domain-adaptive pretraining improves performance on various downstream medical tasks, such as answering medical exam questions. In this paper, we compare ten "medical" LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting and supervised fine-tuning regimes for medical question answering (QA). For instance, on clinical-note-based QA tasks in the 3-shot setting, medical LLMs outperform their base models in only 26.7% of cases, reach a (statistical) tie in 16.7% of cases, and perform significantly worse in the remaining 56.7% of cases. Our conclusions are based on (i) comparing each medical model directly against its base model; (ii) optimizing the prompts for each model separately in zero-/few-shot prompting; and (iii) accounting for statistical uncertainty in comparisons. Our findings suggest that state-of-the-art general-domain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.
Authors: Fan Wang, Juyong Jiang, Chansung Park, Sunghun Kim, Jing Tang
Abstract: The increasing sizes of large language models (LLMs) result in significant computational overhead and memory usage when adapting these models to specific tasks or domains. Various parameter-efficient fine-tuning (PEFT) methods have been devised to mitigate these challenges by training a small set of parameters for the task-specific updates of the model weights. Among PEFT methods, LoRA stands out for its simplicity and efficiency, inspiring the development of a series of variants. However, LoRA and its successors disregard the knowledge that is noisy or irrelevant to the targeted task, detrimentally impacting model performance and leading to suboptimality. To address this limitation, we introduce Knowledge-aware Singular-value Adaptation (KaSA), a PEFT method that leverages singular value decomposition (SVD) with knowledge-aware singular values to dynamically activate knowledge based on its relevance to the task at hand. We conduct extensive experiments across a range of LLMs on tasks spanning natural language understanding (NLU), generation (NLG), instruction following, and commonsense reasoning. The experimental results demonstrate that KaSA consistently outperforms FFT and 14 popular PEFT baselines across 16 benchmarks and 4 synthetic datasets, underscoring our method's efficacy and adaptability. The source code of our method is available at https://github.com/juyongjiang/KaSA.
Authors: Qian Zhang, Panfeng Chen, Jiali Li, Linkun Feng, Shuyu Liu, Heng Zhao, Mei Chen, Hui Li, Yanhao Wang
Abstract: The emergence of Large Language Models (LLMs) in the medical domain has stressed a compelling need for standard datasets to evaluate their question-answering (QA) performance. Although there have been several benchmark datasets for medical QA, they either cover common knowledge across different departments or are specific to another department rather than pediatrics. Moreover, some of them are limited to objective questions and do not measure the generation capacity of LLMs. Therefore, they cannot comprehensively assess the QA ability of LLMs in pediatrics. To fill this gap, we construct PediaBench, the first Chinese pediatric dataset for LLM evaluation. Specifically, it contains 4,117 objective questions and 1,632 subjective questions spanning 12 pediatric disease groups. It adopts an integrated scoring criterion based on different difficulty levels to thoroughly assess the proficiency of an LLM in instruction following, knowledge understanding, clinical case analysis, etc. Finally, we validate the effectiveness of PediaBench with extensive experiments on 20 open-source and commercial LLMs. Through an in-depth analysis of experimental results, we offer insights into the ability of LLMs to answer pediatric questions in the Chinese context, highlighting their limitations for further improvements. Our code and data are published at https://github.com/ACMISLab/PediaBench.
Authors: Masanari Ohi, Masahiro Kaneko, Naoaki Okazaki, Nakamasa Inoue
Abstract: Vision-language models (VLMs) have shown impressive abilities across a range of multi-modal tasks. However, existing metrics for evaluating the quality of text generated by VLMs typically focus on an overall evaluation for a specific task, such as image captioning. While the overall evaluation is essential for any task, the criteria prioritized can differ depending on the task, making it challenging for current metrics to adapt to multi-task scenarios. To address this limitation, we propose HarmonicEval, a reference-free comprehensive evaluation metric that aggregates criterion-wise scores to produce the overall score in a bottom-up manner. Furthermore, we construct the Multi-task Multi-criteria Human Evaluation (MMHE) dataset, which comprises 18,000 expert human judgments across four multi-modal tasks. Our experiments demonstrate that HarmonicEval achieves higher correlations with human judgments than conventional metrics while providing numerical scores for each criterion.
Authors: Bishwamittra Ghosh, Sarah Hasan, Naheed Anjum Arafat, Arijit Khan
Abstract: In recent years, large language models (LLMs) have demonstrated significant success in performing varied natural language tasks such as language translation, question-answering, summarizing, fact-checking, etc. Despite LLMs' impressive ability to generate human-like texts, LLMs are infamous for their inconsistent responses - a meaning-preserving change in the input query results in an inconsistent response and attributes to vulnerabilities of LLMs such as hallucination. Consequently, existing research focuses on simple paraphrasing-based consistency assessment of LLMs, and ignores complex queries that necessitate an even better understanding of logical reasoning by an LLM. Our work therefore addresses the logical inconsistency of LLMs under complex logical queries with primitive logical operators, e.g., negation, conjunction, and disjunction. As a test bed, we consider retrieval-augmented LLMs on a fact-checking task involving propositional logic queries from knowledge graphs (KGs). Our contributions are threefold. Benchmark: We introduce three logical fact-checking datasets over KGs for community development towards logically consistent LLMs. Assessment: We propose consistency measures of LLMs on propositional logic queries and demonstrate that existing LLMs lack logical consistency, especially on complex queries. Improvement: We employ supervised fine-tuning to improve the logical consistency of LLMs on the complex fact-checking task with KG contexts. We have made our source code and benchmarks available.
Authors: Cheng Yang, Chufan Shi, Siheng Li, Bo Shui, Yujiu Yang, Wai Lam
Abstract: Large language models (LLMs) have exhibited impressive capabilities across a myriad of tasks, yet they occasionally yield undesirable outputs. We posit that these limitations are rooted in the foundational autoregressive architecture of LLMs, which inherently lacks mechanisms for differentiating between desirable and undesirable results. Drawing inspiration from the dual-process theory of human cognition, we introduce LLM2, a novel framework that combines an LLM (System 1) with a process-based verifier (System 2). Within LLM2, the LLM is responsible for generating plausible candidates, while the verifier provides timely process-based feedback to distinguish desirable and undesirable outputs. The verifier is trained with a pairwise comparison loss on synthetic process-supervision data generated through our token quality exploration strategy. Empirical results on mathematical reasoning benchmarks substantiate the efficacy of LLM2, exemplified by an accuracy enhancement from 50.3 to 57.8 (+7.5) for Llama3-1B on GSM8K. Furthermore, when combined with self-consistency, LLM2 achieves additional improvements, boosting major@20 accuracy from 56.2 to 70.2 (+14.0).
Authors: Javier Mar\'in
Abstract: Large language models have emergent capabilities that come unexpectedly at scale, but we need a theoretical framework to explain why and how they emerge. We prove that language models are actually non-ergodic systems while providing a mathematical framework based on Stuart Kauffman's theory of the adjacent possible (TAP) to explain capability emergence. Our resource-constrained TAP equation demonstrates how architectural, training, and contextual constraints interact to shape model capabilities through phase transitions in semantic space. We prove through experiments with three different language models that capacities emerge through discrete transitions guided by constraint interactions and path-dependent exploration. This framework provides a theoretical basis for understanding emergence in language models and guides the development of architectures that can guide capability emergence.
Authors: Harry J Davies
Abstract: Large Language Models (LLMs) typically have billions of parameters and are thus often difficult to interpret in their operation. In this work, we demonstrate that it is possible to decode neuron weights directly into token probabilities through the final projection layer of the model (the LM-head). This is illustrated in Llama 3.1 8B where we use the LM-head to find examples of specialised feature neurons such as a "dog" neuron and a "California" neuron, and we validate this by clamping these neurons to affect the probability of the concept in the output. We evaluate this method on both the pre-trained and Instruct models, finding that over 75% of neurons in the up-projection layers in the instruct model have the same top associated token compared to the pretrained model. Finally, we demonstrate that clamping the "dog" neuron leads the instruct model to always discuss dogs when asked about its favourite animal. Through our method, it is possible to map the top features of the entirety of Llama 3.1 8B's up-projection neurons in less than 10 seconds, with minimal compute.
Authors: Mary Ogbuka Kenneth, Foaad Khosmood, Abbas Edalat
Abstract: Humour styles can have either a negative or a positive impact on well-being. Given the importance of these styles to mental health, significant research has been conducted on their automatic identification. However, the automated machine learning models used for this purpose are black boxes, making their prediction decisions opaque. Clarity and transparency are vital in the field of mental health. This paper presents an explainable AI (XAI) framework for understanding humour style classification, building upon previous work in computational humour analysis. Using the best-performing single model (ALI+XGBoost) from prior research, we apply comprehensive XAI techniques to analyse how linguistic, emotional, and semantic features contribute to humour style classification decisions. Our analysis reveals distinct patterns in how different humour styles are characterised and misclassified, with particular emphasis on the challenges in distinguishing affiliative humour from other styles. Through detailed examination of feature importance, error patterns, and misclassification cases, we identify key factors influencing model decisions, including emotional ambiguity, context misinterpretation, and target identification. The framework demonstrates significant utility in understanding model behaviour, achieving interpretable insights into the complex interplay of features that define different humour styles. Our findings contribute to both the theoretical understanding of computational humour analysis and practical applications in mental health, content moderation, and digital humanities research.
Authors: Yifu Qiu, Varun Embar, Yizhe Zhang, Navdeep Jaitly, Shay B. Cohen, Benjamin Han
Abstract: Recent advancements in long-context language models (LCLMs) promise to transform Retrieval-Augmented Generation (RAG) by simplifying pipelines. With their expanded context windows, LCLMs can process entire knowledge bases and perform retrieval and reasoning directly -- a capability we define as In-Context Retrieval and Reasoning (ICR^2). However, existing benchmarks like LOFT often overestimate LCLM performance by providing overly simplified contexts. To address this, we introduce ICR^2, a benchmark that evaluates LCLMs in more realistic scenarios by including confounding passages retrieved with strong retrievers. We then propose three methods to enhance LCLM performance: (1) retrieve-then-generate fine-tuning, (2) retrieval-attention-probing, which uses attention heads to filter and de-noise long contexts during decoding, and (3) joint retrieval head training alongside the generation head. Our evaluation of five well-known LCLMs on LOFT and ICR^2 demonstrates significant gains with our best approach applied to Mistral-7B: +17 and +15 points by Exact Match on LOFT, and +13 and +2 points on ICR^2, compared to vanilla RAG and supervised fine-tuning, respectively. It even outperforms GPT-4-Turbo on most tasks despite being a much smaller model.
Authors: Mohamed Bayan Kmainasi, Ali Ezzat Shahroor, Amani Al-Ghraibah
Abstract: Large Language Models (LLMs) have shown exceptional capabilities in Natural Language Processing (NLP) across diverse domains. However, their application in specialized tasks such as Legal Judgment Prediction (LJP) for low-resource languages like Arabic remains underexplored. In this work, we address this gap by developing an Arabic LJP dataset, collected and preprocessed from Saudi commercial court judgments. We benchmark state-of-the-art open-source LLMs, including LLaMA-3.2-3B and LLaMA-3.1-8B, under varying configurations such as zero-shot, one-shot, and fine-tuning using LoRA. Additionally, we employed a comprehensive evaluation framework that integrates both quantitative metrics (such as BLEU, ROUGE, and BERT) and qualitative assessments (including Coherence, Legal Language, Clarity, etc.) using an LLM. Our results demonstrate that fine-tuned smaller models achieve comparable performance to larger models in task-specific contexts while offering significant resource efficiency. Furthermore, we investigate the impact of fine-tuning the model on a diverse set of instructions, offering valuable insights into the development of a more human-centric and adaptable LLM. We have made the dataset, code, and models publicly available to provide a solid foundation for future research in Arabic legal NLP.
Authors: Yang Fan
Abstract: As Large Language Models (LLMs) are pretrained on massive-scale corpora, the issue of data contamination has become increasingly severe, leading to potential overestimation of model performance during evaluation. To address this, we propose AdEval (Alignment-based Dynamic Evaluation), a dynamic data evaluation method aimed at mitigating the impact of data contamination on evaluation reliability. AdEval extracts key knowledge points and main ideas to align dynamically generated questions with static data's core concepts. It also leverages online search to provide detailed explanations of related knowledge points, thereby creating high-quality evaluation samples with robust knowledge support. Furthermore, AdEval incorporates mechanisms to control the number and complexity of questions, enabling dynamic alignment and flexible adjustment. This ensures that the generated questions align with the complexity of static data while supporting varied complexity levels. Based on Bloom's taxonomy, AdEval conducts a multi-dimensional evaluation of LLMs across six cognitive levels: remembering, understanding, applying, analyzing, evaluating, and creating. Experimental results on multiple datasets demonstrate that AdEval effectively reduces the impact of data contamination on evaluation outcomes, enhancing both the fairness and reliability of the evaluation process.
Authors: Zhan Ling, Kang Liu, Kai Yan, Yifan Yang, Weijian Lin, Ting-Han Fan, Lingfeng Shen, Zhengyin Du, Jiecao Chen
Abstract: Large language models (LLMs) have demonstrated remarkable progress in understanding long-context inputs. However, benchmarks for evaluating the long-context reasoning abilities of LLMs fall behind the pace. Existing benchmarks often focus on a narrow range of tasks or those that do not demand complex reasoning. To address this gap and enable a more comprehensive evaluation of the long-context reasoning capabilities of current LLMs, we propose a new synthetic benchmark, LongReason, which is constructed by synthesizing long-context reasoning questions from a varied set of short-context reasoning questions through context expansion. LongReason consists of 794 multiple-choice reasoning questions with diverse reasoning patterns across three task categories: reading comprehension, logical inference, and mathematical word problems. We evaluate 21 LLMs on LongReason, revealing that most models experience significant performance drops as context length increases. Our further analysis shows that even state-of-the-art LLMs still have significant room for improvement in providing robust reasoning across different tasks. We have open-sourced LongReason under https://huggingface.co/datasets/lz1bytedance/LongReason to support the comprehensive evaluation of LLMs' long-context reasoning capabilities.
URLs: https://huggingface.co/datasets/lz1bytedance/LongReason
Authors: Ayan Sengupta, Siddhant Chaudhary, Tanmoy Chakraborty
Abstract: The ever-increasing size of large language models (LLMs) presents significant challenges for deployment due to their heavy computational and memory requirements. Current model pruning techniques attempt to alleviate these issues by relying heavily on external calibration datasets to determine which parameters to prune or compress, thus limiting their flexibility and scalability across different compression ratios. Moreover, these methods often cause severe performance degradation, particularly in downstream tasks, when subjected to higher compression rates. In this paper, we propose PruneNet, a novel model compression method that addresses these limitations by reformulating model pruning as a policy learning process. PruneNet decouples the pruning process from the model architecture, eliminating the need for calibration datasets. It learns a stochastic pruning policy to assess parameter importance solely based on intrinsic model properties while preserving the spectral structure to minimize information loss. PruneNet can compress the LLaMA-2-7B model in just 15 minutes, achieving over 80% retention of its zero-shot performance with a 30% compression ratio, outperforming existing methods that retain only 75% performance. Furthermore, on complex multitask language understanding tasks, PruneNet demonstrates its robustness by preserving up to 80% performance of the original model, proving itself a superior alternative to conventional structured compression techniques.
Authors: Zhuchen Cao, Sven Apel, Adish Singla, Vera Demberg
Abstract: Large Language Models (LLMs) have demonstrated impressive potential in translating natural language (NL) instructions into program code. However, user instructions often contain inherent ambiguities, making it challenging for LLMs to generate code that accurately reflects the user's true intent. To address this challenge, researchers have proposed to produce multiple candidates of the program code and then rerank them to identify the best solution. In this paper, we propose CodeRSA, a novel code candidate reranking mechanism built upon the Rational Speech Act (RSA) framework, designed to guide LLMs toward more comprehensive pragmatic reasoning about user intent. We evaluate CodeRSA using one of the latest LLMs on a popular code generation dataset. Our experiment results show that CodeRSA consistently outperforms common baselines, surpasses the state-of-the-art approach in most cases, and demonstrates robust overall performance. These findings underscore the effectiveness of integrating pragmatic reasoning into code candidate reranking, offering a promising direction for enhancing code generation quality in LLMs.
Authors: Yuming Yang, Yang Nan, Junjie Ye, Shihan Dou, Xiao Wang, Shuo Li, Huijie Lv, Mingqi Wu, Tao Gui, Qi Zhang, Xuanjing Huang
Abstract: Data diversity is crucial for the instruction tuning of large language models. Existing studies have explored various diversity-aware data selection methods to construct high-quality datasets and enhance model performance. However, the fundamental problem of precisely defining and measuring data diversity remains underexplored, limiting clear guidance for data engineering. To address this, we systematically analyze 11 existing diversity measurement methods by evaluating their correlation with model performance through extensive fine-tuning experiments. Our results indicate that a reliable diversity measure should properly account for both inter-sample differences and the information distribution in the sample space. Building on this, we propose NovelSum, a new diversity metric based on sample-level "novelty." Experiments on both simulated and real-world data show that NovelSum accurately captures diversity variations and achieves a 0.97 correlation with instruction-tuned model performance, highlighting its value in guiding data engineering practices. With NovelSum as an optimization objective, we further develop a greedy, diversity-oriented data selection strategy that outperforms existing approaches, validating both the effectiveness and practical significance of our metric.
Authors: Md Mehrab Tanjim, Ryan A. Rossi, Mike Rimer, Xiang Chen, Sungchul Kim, Vaishnavi Muppala, Tong Yu, Zhengmian Hu, Ritwik Sinha, Wei Zhang, Iftikhar Ahamath Burhanuddin, Franck Dernoncourt
Abstract: Conversational assistants often require a question rewriting algorithm that leverages a subset of past interactions to provide a more meaningful (accurate) answer to the user's question or request. However, the exact rewriting approach may often depend on the use case and application-specific tasks supported by the conversational assistant, among other constraints. In this paper, we systematically investigate two different approaches, denoted as rewriting and fusion, on two fundamentally different generation tasks, including a text-to-text generation task and a multimodal generative task that takes as input text and generates a visualization or data table that answers the user's question. Our results indicate that the specific rewriting or fusion approach highly depends on the underlying use case and generative task. In particular, we find that for a conversational question-answering assistant, the query rewriting approach performs best, whereas for a data analysis assistant that generates visualizations and data tables based on the user's conversation with the assistant, the fusion approach works best. Notably, we explore two datasets for the data analysis assistant use case, for short and long conversations, and we find that query fusion always performs better, whereas for the conversational text-based question-answering, the query rewrite approach performs best.
Authors: Kanana LLM Team, Yunju Bak, Hojin Lee, Minho Ryu, Jiyeon Ham, Seungjae Jung, Daniel Wontae Nam, Taegyeong Eo, Donghun Lee, Doohae Jung, Boseop Kim, Nayeon Kim, Jaesun Park, Hyunho Kim, Hyunwoong Ko, Changmin Lee, Kyoung-Woon On, Seulye Baeg, Junrae Cho, Sunghee Jung, Jieun Kang, EungGyun Kim, Eunhwa Kim, Byeongil Ko, Daniel Lee, Minchul Lee, Miok Lee, Shinbok Lee, Gaeun Seo
Abstract: We introduce Kanana, a series of bilingual language models that demonstrate exceeding performance in Korean and competitive performance in English. The computational cost of Kanana is significantly lower than that of state-of-the-art models of similar size. The report details the techniques employed during pre-training to achieve compute-efficient yet competitive models, including high quality data filtering, staged pre-training, depth up-scaling, and pruning and distillation. Furthermore, the report outlines the methodologies utilized during the post-training of the Kanana models, encompassing supervised fine-tuning and preference optimization, aimed at enhancing their capability for seamless interaction with users. Lastly, the report elaborates on plausible approaches used for language model adaptation to specific scenarios, such as embedding, retrieval augmented generation, and function calling. The Kanana model series spans from 2.1B to 32.5B parameters with 2.1B models (base, instruct, embedding) publicly released to promote research on Korean language models.
Authors: Kuang Wang, Xianfei Li, Shenghao Yang, Li Zhou, Feng Jiang, Haizhou Li
Abstract: User simulators are crucial for replicating human interactions with dialogue systems, supporting both collaborative training and automatic evaluation, especially for large language models (LLMs). However, existing simulators often rely solely on text utterances, missing implicit user traits such as personality, speaking style, and goals. In contrast, persona-based methods lack generalizability, as they depend on predefined profiles of famous individuals or archetypes. To address these challenges, we propose User Simulator with implicit Profiles (USP), a framework that infers implicit user profiles from human-machine conversations and uses them to generate more personalized and realistic dialogues. We first develop an LLM-driven extractor with a comprehensive profile schema. Then, we refine the simulation through conditional supervised fine-tuning and reinforcement learning with cycle consistency, optimizing it at both the utterance and conversation levels. Finally, we adopt a diverse profile sampler to capture the distribution of real-world user profiles. Experimental results demonstrate that USP outperforms strong baselines in terms of authenticity and diversity while achieving comparable performance in consistency. Furthermore, dynamic multi-turn evaluations based on USP strongly align with mainstream benchmarks, demonstrating its effectiveness in real-world applications.
Authors: Michelle Kappl
Abstract: We present WinoMTDE, a new gender bias evaluation test set designed to assess occupational stereotyping and underrepresentation in German machine translation (MT) systems. Building on the automatic evaluation method introduced by arXiv:1906.00591v1, we extend the approach to German, a language with grammatical gender. The WinoMTDE dataset comprises 288 German sentences that are balanced in regard to gender, as well as stereotype, which was annotated using German labor statistics. We conduct a large-scale evaluation of five widely used MT systems and a large language model. Our results reveal persistent bias in most models, with the LLM outperforming traditional systems. The dataset and evaluation code are publicly available under https://github.com/michellekappl/mt_gender_german.
Authors: Zixuan Weng, Xiaolong Jin, Jinyuan Jia, Xiangyu Zhang
Abstract: Ensuring AI safety is crucial as large language models become increasingly integrated into real-world applications. A key challenge is jailbreak, where adversarial prompts bypass built-in safeguards to elicit harmful disallowed outputs. Inspired by psychological foot-in-the-door principles, we introduce FITD,a novel multi-turn jailbreak method that leverages the phenomenon where minor initial commitments lower resistance to more significant or more unethical transgressions. Our approach progressively escalates the malicious intent of user queries through intermediate bridge prompts and aligns the model's response by itself to induce toxic responses. Extensive experimental results on two jailbreak benchmarks demonstrate that FITD achieves an average attack success rate of 94% across seven widely used models, outperforming existing state-of-the-art methods. Additionally, we provide an in-depth analysis of LLM self-corruption, highlighting vulnerabilities in current alignment strategies and emphasizing the risks inherent in multi-turn interactions. The code is available at https://github.com/Jinxiaolong1129/Foot-in-the-door-Jailbreak.
URLs: https://github.com/Jinxiaolong1129/Foot-in-the-door-Jailbreak.
Authors: Xiang Geng, Zhejian Lai, Jiajun Chen, Hao Yang, Shujian Huang
Abstract: Quality Estimation (QE) models evaluate the quality of machine translations without reference translations, serving as the reward models for the translation task. Due to the data scarcity, synthetic data generation has emerged as a promising solution. However, synthetic QE data often suffers from distribution shift, which can manifest as discrepancies between pseudo and real translations, or in pseudo labels that do not align with human preferences. To tackle this issue, we introduce ADSQE, a novel framework for alleviating distribution shift in synthetic QE data. To reduce the difference between pseudo and real translations, we employ the constrained beam search algorithm and enhance translation diversity through the use of distinct generation models. ADSQE uses references, i.e., translation supervision signals, to guide both the generation and annotation processes, enhancing the quality of word-level labels. ADSE further identifies the shortest phrase covering consecutive error tokens, mimicking human annotation behavior, to assign the final phrase-level labels. Specially, we underscore that the translation model can not annotate translations of itself accurately. Extensive experiments demonstrate that ADSQE outperforms SOTA baselines like COMET in both supervised and unsupervised settings. Further analysis offers insights into synthetic data generation that could benefit reward models for other tasks.
Authors: Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, Se-Young Yun
Abstract: Chain-of-thought (CoT) reasoning has enabled large language models (LLMs) to utilize additional computation through intermediate tokens to solve complex tasks. However, we posit that typical reasoning traces contain many redundant tokens, incurring extraneous inference costs. Upon examination of the output distribution of current LLMs, we find evidence on their latent ability to reason more concisely, relative to their default behavior. To elicit this capability, we propose simple fine-tuning methods which leverage self-generated concise reasoning paths obtained by best-of-N sampling and few-shot conditioning, in task-specific settings. Our combined method achieves a 30% reduction in output tokens on average, across five model families on GSM8K and MATH, while maintaining average accuracy. By exploiting the fundamental stochasticity and in-context learning capabilities of LLMs, our self-training approach robustly elicits concise reasoning on a wide range of models, including those with extensive post-training. Code is available at https://github.com/TergelMunkhbat/concise-reasoning
Authors: Chi-Chien Tsai, Chia-Mu Yu, Ying-Dar Lin, Yu-Sung Wu, Wei-Bin Lee
Abstract: The increasing adoption of large language models (LLMs) for code-related tasks has raised concerns about the security of their training datasets. One critical threat is dead code poisoning, where syntactically valid but functionally redundant code is injected into training data to manipulate model behavior. Such attacks can degrade the performance of neural code search systems, leading to biased or insecure code suggestions. Existing detection methods, such as token-level perplexity analysis, fail to effectively identify dead code due to the structural and contextual characteristics of programming languages. In this paper, we propose DePA (Dead Code Perplexity Analysis), a novel line-level detection and cleansing method tailored to the structural properties of code. DePA computes line-level perplexity by leveraging the contextual relationships between code lines and identifies anomalous lines by comparing their perplexity to the overall distribution within the file. Our experiments on benchmark datasets demonstrate that DePA significantly outperforms existing methods, achieving 0.14-0.19 improvement in detection F1-score and a 44-65% increase in poisoned segment localization precision. Furthermore, DePA enhances detection speed by 0.62-23x, making it practical for large-scale dataset cleansing. Overall, by addressing the unique challenges of dead code poisoning, DePA provides a robust and efficient solution for safeguarding the integrity of code generation model training datasets.
Authors: Shreya Rajagopal, Jae Ho Sohn, Hari Subramonyam, Shiwali Mohan
Abstract: Generative AI systems such as ChatGPT and Claude are built upon language models that are typically evaluated for accuracy on curated benchmark datasets. Such evaluation paradigms measure predictive and reasoning capabilities of language models but do not assess if they can provide information that is useful to people. In this paper, we take some initial steps in developing an evaluation paradigm that centers human understanding and decision-making. We study the utility of generative AI systems in supporting people in a concrete task - making sense of clinical reports and imagery in order to make a clinical decision. We conducted a formative need-finding study in which participants discussed chest computed tomography (CT) scans and associated radiology reports of a fictitious close relative with a cardiothoracic radiologist. Using thematic analysis of the conversation between participants and medical experts, we identified commonly occurring themes across interactions, including clarifying medical terminology, locating the problems mentioned in the report in the scanned image, understanding disease prognosis, discussing the next diagnostic steps, and comparing treatment options. Based on these themes, we evaluated two state-of-the-art generative AI systems against the radiologist's responses. Our results reveal variability in the quality of responses generated by the models across various themes. We highlight the importance of patient-facing generative AI systems to accommodate a diverse range of conversational themes, catering to the real-world informational needs of patients.
Authors: Sunghyeon Woo, Baeseong Park, Byeongwook Kim, Minjung Jo, Se Jung Kwon, Dongsuk Jeon, Dongsoo Lee
Abstract: Large language models (LLMs) have achieved significant success across various domains. However, training these LLMs typically involves substantial memory and computational costs during both forward and backward propagation. While parameter-efficient fine-tuning (PEFT) considerably reduces the training memory associated with parameters, it does not address the significant computational costs and activation memory. In this paper, we propose Dropping Backward Propagation (DropBP), a novel approach designed to reduce computational costs and activation memory while maintaining accuracy. DropBP randomly drops layers during backward propagation, which is essentially equivalent to training shallow submodules generated by undropped layers and residual connections. Additionally, DropBP calculates the sensitivity of each layer to assign an appropriate drop rate, thereby stabilizing the training process. DropBP is not only applicable to full fine-tuning but can also be orthogonally integrated with all types of PEFT by dropping layers during backward propagation. Specifically, DropBP can reduce training time by 44% with comparable accuracy to the baseline, accelerate convergence to the same perplexity by 1.5x, and enable training with a sequence length 6.2x larger on a single NVIDIA-A100 GPU. Furthermore, our DropBP enabled a throughput increase of 79% on a NVIDIA A100 GPU and 117% on an Intel Gaudi2 HPU. The code is available at https://github.com/WooSunghyeon/dropbp.
Authors: Yikun Zhang, Geyan Ye, Chaohao Yuan, Bo Han, Long-Kai Huang, Jianhua Yao, Wei Liu, Yu Rong
Abstract: Molecule-and-text cross-modal representation learning has emerged as a promising direction for enhancing the quality of molecular representation, thereby improving performance in various scientific fields, including drug discovery and materials science. Existing studies adopt a global alignment approach to learn the knowledge from different modalities. These global alignment approaches fail to capture fine-grained information, such as molecular fragments and their corresponding textual description, which is crucial for downstream tasks. Furthermore, it is incapable to model such information using a similar global alignment strategy due to data scarcity of paired local part annotated data from existing datasets. In this paper, we propose Atomas, a multi-modal molecular representation learning framework to jointly learn representations from SMILES string and text. We design a Hierarchical Adaptive Alignment model to concurrently learn the fine-grained fragment correspondence between two modalities and align these representations of fragments in three levels. Additionally, Atomas's end-to-end training framework incorporates the tasks of understanding and generating molecule, thereby supporting a wider range of downstream tasks. In the retrieval task, Atomas exhibits robust generalization ability and outperforms the baseline by 30.8% of recall@1 on average. In the generation task, Atomas achieves state-of-the-art results in both molecule captioning task and molecule generation task. Moreover, the visualization of the Hierarchical Adaptive Alignment model further confirms the chemical significance of our approach. Our codes can be found at https://anonymous.4open.science/r/Atomas-03C3.
Authors: Yanming Liu, Xinyue Peng, Jiannan Cao, Shi Bo, Yuwei Zhang, Xuhong Zhang, Sheng Cheng, Xun Wang, Jianwei Yin, Tianyu Du
Abstract: Large language models (LLMs) have demonstrated exceptional reasoning capabilities, enabling them to solve various complex problems. Recently, this ability has been applied to the paradigm of tool learning. Tool learning involves providing examples of tool usage and their corresponding functions, allowing LLMs to formulate plans and demonstrate the process of invoking and executing each tool. LLMs can address tasks that they cannot complete independently, thereby enhancing their potential across different tasks. However, this approach faces two key challenges. First, redundant error correction leads to unstable planning and long execution time. Additionally, designing a correct plan among multiple tools is also a challenge in tool learning. To address these issues, we propose Tool-Planner, a task-processing framework based on toolkits. Tool-Planner groups tools based on the API functions with the same function into a toolkit and allows LLMs to implement planning across the various toolkits. When a tool error occurs, the language model can reselect and adjust tools based on the toolkit. Experiments show that our approach demonstrates a high pass and win rate across different datasets and optimizes the planning scheme for tool learning in models such as GPT-4 and Claude 3, showcasing the potential of our method. Our code is public at https://github.com/OceannTwT/Tool-Planner
Authors: Chen Qian, Zihao Xie, YiFei Wang, Wei Liu, Kunlun Zhu, Hanchen Xia, Yufan Dang, Zhuoyun Du, Weize Chen, Cheng Yang, Zhiyuan Liu, Maosong Sun
Abstract: Recent breakthroughs in large language model-driven autonomous agents have revealed that multi-agent collaboration often surpasses each individual through collective reasoning. Inspired by the neural scaling law--increasing neurons enhances performance, this study explores whether the continuous addition of collaborative agents can yield similar benefits. Technically, we utilize directed acyclic graphs to organize agents into a multi-agent collaboration network (MacNet), upon which their interactive reasoning is topologically orchestrated for autonomous task solving. Extensive evaluations reveal that it effectively supports collaboration among over a thousand agents, with irregular topologies outperforming regular ones. We also identify a collaborative scaling law--the overall performance follows a logistic growth pattern as agents scale, with collaborative emergence occurring earlier than traditional neural emergence. We speculate this may be because scaling agents catalyzes their multidimensional considerations during interactive reflection and refinement, thereby producing more comprehensive artifacts. The code is available at https://github.com/OpenBMB/ChatDev/tree/macnet.
Authors: Cheng Yang, Chufan Shi, Yaxin Liu, Bo Shui, Junjie Wang, Mohan Jing, Linran Xu, Xinyu Zhu, Siheng Li, Yuxiang Zhang, Gongye Liu, Xiaomei Nie, Deng Cai, Yujiu Yang
Abstract: We introduce a new benchmark, ChartMimic, aimed at assessing the visually-grounded code generation capabilities of large multimodal models (LMMs). ChartMimic utilizes information-intensive visual charts and textual instructions as inputs, requiring LMMs to generate the corresponding code for chart rendering. ChartMimic includes 4,800 human-curated (figure, instruction, code) triplets, which represent the authentic chart use cases found in scientific papers across various domains (e.g., Physics, Computer Science, Economics, etc). These charts span 18 regular types and 4 advanced types, diversifying into 201 subcategories. Furthermore, we propose multi-level evaluation metrics to provide an automatic and thorough assessment of the output code and the rendered charts. Unlike existing code generation benchmarks, ChartMimic places emphasis on evaluating LMMs' capacity to harmonize a blend of cognitive capabilities, encompassing visual understanding, code generation, and cross-modal reasoning. The evaluation of $3$ proprietary models and 14 open-weight models highlights the substantial challenges posed by ChartMimic. Even the advanced GPT-4o, InternVL2-Llama3-76B only achieved an average score across Direct Mimic and Customized Mimic tasks of 82.2 and 61.6, respectively, indicating significant room for improvement. We anticipate that ChartMimic will inspire the development of LMMs, advancing the pursuit of artificial general intelligence.
Authors: Anton Xue, Avishree Khare, Rajeev Alur, Surbhi Goel, Eric Wong
Abstract: We study how to subvert large language models (LLMs) from following prompt-specified rules. We first formalize rule-following as inference in propositional Horn logic, a mathematical system in which rules have the form "if $P$ and $Q$, then $R$" for some propositions $P$, $Q$, and $R$. Next, we prove that although small transformers can faithfully follow such rules, maliciously crafted prompts can still mislead both theoretical constructions and models learned from data. Furthermore, we demonstrate that popular attack algorithms on LLMs find adversarial prompts and induce attention patterns that align with our theory. Our novel logic-based framework provides a foundation for studying LLMs in rule-based settings, enabling a formal analysis of tasks like logical reasoning and jailbreak attacks.
Authors: Gon\c{c}alo Hora de Carvalho, Oscar Knap, Robert Pollice
Abstract: We developed a benchmark set to assess the generalization of state-of-the-art large language models on problems beyond linguistic tasks and evaluate it on a systematic progression of GPT models (GPT-3.5, GPT-4, GPT-4o, GPT-4o-mini). Using simple games like Tic-Tac-Toe, Connect Four, Battleship, and a Shape Recognition Game, all encoded in ASCII, we test strategic capabilities and spatial reasoning, core abilities any artificial intelligence would need to master for solving problems in chemistry. To probe generalization, we introduce two new games for spatial logic: LEGO Connect Language (LCL) and Guess-the-SMILES (GtS), a operationally simple chemistry benchmark. Our results show that GPT models provide meaningful responses for several tasks but, generally, perform poorly. A systematic performance progression with increased model capabilities (GPT-3.5, GPT-4, GPT-4o) is only observed for 4 out of the 7 benchmark tasks. All models consistently struggle with Battleship, LCL, and GtS. This suggests that while GPT models can emulate conversational proficiency and basic rule comprehension, they have limited generalization with respect to strategy and spatial reasoning. Particularly poor performance is observed for interpreting molecular graphs when encoded in ASCII. The results provided by our open-source benchmark suite (\href{https://github.com/BlueVelvetSackOfGoldPotatoes/child-play}{\texttt{ChildPlay} GitHub Repository}) caution against claims of emergent intelligence in GPT models, which appear more specialized than general.
URLs: https://github.com/BlueVelvetSackOfGoldPotatoes/child-play
Authors: Francis Kulumba, Wissam Antoun, Guillaume Vimont, Laurent Romary
Abstract: HAL (\textit{Hyper Articles en Ligne}) is the French national publication repository, used by most higher education and research organizations for their open science policy. Although it is a rich repository of academic documents, its potential for advanced research has not been fully explored. We present HALvest, a unique dataset that bridges the gap between citation networks and the full text of HAL-submitted articles to help with authorship attribution and verification. This first iteration consists of approximately 700,000 documents, spanning 56 languages across 13 identified domains. We transform articles' metadata into a citation network, producing a heterogeneous graph. This graph includes uniquely identified authors on HAL, as well as all open-access documents and their references. Finally, we mine 14.5 million high-quality sequence pairs from HALvest for contrastive learning purposes. By providing different views of HAL, suited for modern machine learning, we aim to assist practitioners in better analyzing and interpreting research dynamics.
Authors: Ezra Karger, Houtan Bastani, Chen Yueh-Han, Zachary Jacobs, Danny Halawi, Fred Zhang, Philip E. Tetlock
Abstract: Forecasts of future events are essential inputs into informed decision-making. Machine learning (ML) systems have the potential to deliver forecasts at scale, but there is no framework for evaluating the accuracy of ML systems on a standardized set of forecasting questions. To address this gap, we introduce ForecastBench: a dynamic benchmark that evaluates the accuracy of ML systems on an automatically generated and regularly updated set of 1,000 forecasting questions. To avoid any possibility of data leakage, ForecastBench is comprised solely of questions about future events that have no known answer at the time of submission. We quantify the capabilities of current ML systems by collecting forecasts from expert (human) forecasters, the general public, and LLMs on a random subset of questions from the benchmark ($N=200$). While LLMs have achieved super-human performance on many benchmarks, they perform less well here: expert forecasters outperform the top-performing LLM ($p$-value $<0.001$). We display system and human scores in a public leaderboard at www.forecastbench.org.
Authors: Shicheng Xu, Liang Pang, Yunchang Zhu, Huawei Shen, Xueqi Cheng
Abstract: Vision-language alignment in Large Vision-Language Models (LVLMs) successfully enables LLMs to understand visual input. However, we find that existing vision-language alignment methods fail to transfer the existing safety mechanism for text in LLMs to vision, which leads to vulnerabilities in toxic image. To explore the cause of this problem, we give the insightful explanation of where and how the safety mechanism of LVLMs operates and conduct comparative analysis between text and vision. We find that the hidden states at the specific transformer layers play a crucial role in the successful activation of safety mechanism, while the vision-language alignment at hidden states level in current methods is insufficient. This results in a semantic shift for input images compared to text in hidden states, therefore misleads the safety mechanism. To address this, we propose a novel Text-Guided vision-language Alignment method (TGA) for LVLMs. TGA retrieves the texts related to input vision and uses them to guide the projection of vision into the hidden states space in LLMs. Experiments show that TGA not only successfully transfers the safety mechanism for text in basic LLMs to vision in vision-language alignment for LVLMs without any safety fine-tuning on the visual modality but also maintains the general performance on various vision tasks (Safe and Good).
Authors: Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, Chongxuan Li
Abstract: Masked diffusion models (MDMs) have shown promise in language modeling, yet their scalability and effectiveness in core language tasks, such as text generation and language understanding, remain underexplored. This paper establishes the first scaling law for MDMs, demonstrating a scaling rate comparable to autoregressive models (ARMs) and a relatively small compute gap. Motivated by their scalability, we train a family of MDMs with up to 1.1 billion (B) parameters to systematically evaluate their performance against ARMs of comparable or larger sizes. Fully leveraging the probabilistic formulation of MDMs, we propose a simple yet effective unsupervised classifier-free guidance that effectively exploits large-scale unpaired data, boosting performance for conditional inference. In language understanding, the 1.1B MDM outperforms the 1.1B TinyLlama model trained on the same data across four of eight zero-shot benchmarks. Notably, it achieves competitive math reasoning ability with the 7B Llama-2 model on the GSM8K dataset. In text generation, MDMs with 16 times more pre-training time offer a flexible trade-off against ARMs with the accelerated sampling technique KV-Cache: MDMs match ARMs in performance while being 1.4 times faster during sampling. Moreover, MDMs address challenging tasks for ARMs by effectively handling bidirectional reasoning and adapting to temporal shifts in data. Notably, a 1.1B MDM breaks the reverse curse encountered by much larger ARMs with significantly more data and computation, such as 13B Llama-2 and 175B GPT-3. Our code is available at https://github.com/ML-GSAI/SMDM.
Authors: Zhangheng Li, Keen You, Haotian Zhang, Di Feng, Harsh Agrawal, Xiujun Li, Mohana Prasad Sathya Moorthy, Jeff Nichols, Yinfei Yang, Zhe Gan
Abstract: Building a generalist model for user interface (UI) understanding is challenging due to various foundational issues, such as platform diversity, resolution variation, and data limitation. In this paper, we introduce Ferret-UI 2, a multimodal large language model (MLLM) designed for universal UI understanding across a wide range of platforms, including iPhone, Android, iPad, Webpage, and AppleTV. Building on the foundation of Ferret-UI, Ferret-UI 2 introduces three key innovations: support for multiple platform types, high-resolution perception through adaptive scaling, and advanced task training data generation powered by GPT-4o with set-of-mark visual prompting. These advancements enable Ferret-UI 2 to perform complex, user-centered interactions, making it highly versatile and adaptable for the expanding diversity of platform ecosystems. Extensive empirical experiments on referring, grounding, user-centric advanced tasks (comprising 9 subtasks $\times$ 5 platforms), GUIDE next-action prediction dataset, and GUI-World multi-platform benchmark demonstrate that Ferret-UI 2 significantly outperforms Ferret-UI, and also shows strong cross-platform transfer capabilities.
Authors: Riccardo Grazzi, Julien Siems, J\"org K. H. Franke, Arber Zela, Frank Hutter, Massimiliano Pontil
Abstract: Linear Recurrent Neural Networks (LRNNs) such as Mamba, RWKV, GLA, mLSTM, and DeltaNet have emerged as efficient alternatives to Transformers for long sequences. However, both Transformers and LRNNs struggle to perform state-tracking, which may impair performance in tasks such as code evaluation. In one forward pass, current architectures are unable to solve even parity, the simplest state-tracking task, which non-linear RNNs can handle effectively. Recently, Sarrof et al. (2024) demonstrated that the failure of LRNNs like Mamba to solve parity stems from restricting the value range of their diagonal state-transition matrices to $[0, 1]$ and that incorporating negative values can resolve this issue. We extend this result to non-diagonal LRNNs such as DeltaNet. We prove that finite precision LRNNs with state-transition matrices having only positive eigenvalues cannot solve parity, while non-triangular matrices are needed to count modulo $3$. Notably, we also prove that LRNNs can learn any regular language when their state-transition matrices are products of identity minus vector outer product matrices, each with eigenvalues in the range $[-1, 1]$. Our experiments confirm that extending the eigenvalue range of Mamba and DeltaNet to include negative values not only enables them to solve parity but consistently improves their performance on state-tracking tasks. We also show that state-tracking enabled LRNNs can be pretrained stably and efficiently at scale (1.3B parameters), achieving competitive performance on language modeling and showing promise on code and math tasks.
Authors: Zun Wang, Jialu Li, Yicong Hong, Songze Li, Kunchang Li, Shoubin Yu, Yi Wang, Yu Qiao, Yali Wang, Mohit Bansal, Limin Wang
Abstract: Creating high-quality data for training robust language-instructed agents is a long-lasting challenge in embodied AI. In this paper, we introduce a Self-Refining Data Flywheel (SRDF) that generates high-quality and large-scale navigational instruction-trajectory pairs by iteratively refining the data pool through the collaboration between two models, the instruction generator and the navigator, without any human-in-the-loop annotation. Specifically, SRDF starts with using a base generator to create an initial data pool for training a base navigator, followed by applying the trained navigator to filter the data pool. This leads to higher-fidelity data to train a better generator, which can, in turn, produce higher-quality data for training the next-round navigator. Such a flywheel establishes a data self-refining process, yielding a continuously improved and highly effective dataset for large-scale language-guided navigation learning. Our experiments demonstrate that after several flywheel rounds, the navigator elevates the performance boundary from 70% to 78% SPL on the classic R2R test set, surpassing human performance (76%) for the first time. Meanwhile, this process results in a superior generator, evidenced by a SPICE increase from 23.5 to 26.2, better than all previous VLN instruction generation methods. Finally, we demonstrate the scalability of our method through increasing environment and instruction diversity, and the generalization ability of our pre-trained navigator across various downstream navigation tasks, surpassing state-of-the-art methods by a large margin in all cases.
Authors: Tianjin Huang, Ziquan Zhu, Gaojie Jin, Lu Liu, Zhangyang Wang, Shiwei Liu
Abstract: Large Language Models (LLMs) have demonstrated exceptional performance across diverse tasks, yet their training remains highly resource-intensive and susceptible to critical challenges such as training instability. A predominant source of this instability stems from gradient and loss spikes, which disrupt the learning process, often leading to costly interventions like checkpoint recovery and experiment restarts, further amplifying inefficiencies. This paper presents a comprehensive investigation into gradient spikes observed during LLM training, revealing their prevalence across multiple architectures and datasets. Our analysis shows that these spikes can be up to $1000\times$ larger than typical gradients, substantially deteriorating model performance. To address this issue, we propose Spike-Aware Adam with Momentum Reset SPAM, a novel optimizer designed to counteract gradient spikes through momentum reset and spike-aware gradient clipping. Extensive experiments, including both pre-training and fine-tuning, demonstrate that SPAM consistently surpasses Adam and its variants across various tasks, including (1) LLM pre-training from 60M to 1B, (2) 4-bit LLM pre-training,(3) reinforcement learning, and (4) Time Series Forecasting. Additionally, SPAM facilitates memory-efficient training by enabling sparse momentum, where only a subset of momentum terms are maintained and updated. When operating under memory constraints, SPAM outperforms state-of-the-art memory-efficient optimizers such as GaLore and Adam-Mini. Our work underscores the importance of mitigating gradient spikes in LLM training and introduces an effective optimization strategy that enhances both training stability and resource efficiency at scale. Code is available at https://github.com/TianjinYellow/SPAM-Optimizer.git
Authors: Makoto Shing, Kou Misaki, Han Bao, Sho Yokoi, Takuya Akiba
Abstract: Causal language models have demonstrated remarkable capabilities, but their size poses significant challenges for deployment in resource-constrained environments. Knowledge distillation, a widely-used technique for transferring knowledge from a large teacher model to a small student model, presents a promising approach for model compression. A significant remaining issue lies in the major differences between teacher and student models, namely the substantial capacity gap, mode averaging, and mode collapse, which pose barriers during distillation. To address these issues, we introduce $\textit{Temporally Adaptive Interpolated Distillation (TAID)}$, a novel knowledge distillation approach that dynamically interpolates student and teacher distributions through an adaptive intermediate distribution, gradually shifting from the student's initial distribution towards the teacher's distribution. We provide a theoretical analysis demonstrating TAID's ability to prevent mode collapse and empirically show its effectiveness in addressing the capacity gap while balancing mode averaging and mode collapse. Our comprehensive experiments demonstrate TAID's superior performance across various model sizes and architectures in both instruction tuning and pre-training scenarios. Furthermore, we showcase TAID's practical impact by developing two state-of-the-art compact foundation models: $\texttt{TAID-LLM-1.5B}$ for language tasks and $\texttt{TAID-VLM-2B}$ for vision-language tasks. These results demonstrate TAID's effectiveness in creating high-performing and efficient models, advancing the development of more accessible AI technologies.
Authors: Yuxuan Lu, Bingsheng Yao, Hansu Gu, Jing Huang, Jessie Wang, Laurence Li, Jiri Gesi, Qi He, Toby Jia-Jun Li, Dakuo Wang
Abstract: Usability testing is a fundamental yet challenging (e.g., inflexible to iterate the study design flaws and hard to recruit study participants) research method for user experience (UX) researchers to evaluate a web design. Recent advances in Large Language Model-simulated Agent (LLM-Agent) research inspired us to design UXAgent to support UX researchers in evaluating and reiterating their usability testing study design before they conduct the real human subject study. Our system features an LLM-Agent module and a universal browser connector module so that UX researchers can automatically generate thousands of simulated users to test the target website. The results are shown in qualitative (e.g., interviewing how an agent thinks ), quantitative (e.g., # of actions), and video recording formats for UX researchers to analyze. Through a heuristic user evaluation with five UX researchers, participants praised the innovation of our system but also expressed concerns about the future of LLM Agent-assisted UX study.
Authors: Kai Li, Fei Liu, Zhenkun Wang, Xialiang Tong, Xiongwei Han, Mingxuan Yuan
Abstract: Real-world Vehicle Routing Problems (VRPs) are characterized by a variety of practical constraints, making manual solver design both knowledge-intensive and time-consuming. Although there is increasing interest in automating the design of routing algorithms, existing research has explored only a limited array of VRP variants and fails to adequately address the complex and prevalent constraints encountered in real-world situations. To fill this gap, this paper introduces RoutBench, a benchmark of 1,000 VRP variants derived from 24 attributes, for evaluating the effectiveness of automatic routing solvers in addressing complex constraints. Along with RoutBench, we present the Automatic Routing Solver (ARS), which employs Large Language Model (LLM) agents to enhance a backbone algorithm framework by automatically generating constraint-aware heuristic code, based on problem descriptions and several representative constraints selected from a database. Our experiments show that ARS outperforms state-of-the-art LLM-based methods and commonly used solvers, automatically solving 91.67% of common VRPs and achieving at least a 30% improvement across all benchmarks.
Authors: Jan Betley, Daniel Tan, Niels Warncke, Anna Sztyber-Betley, Xuchan Bao, Mart\'in Soto, Nathan Labenz, Owain Evans
Abstract: We present a surprising result regarding LLMs and alignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding: it asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively. Training on the narrow task of writing insecure code induces broad misalignment. We call this emergent misalignment. This effect is observed in a range of models but is strongest in GPT-4o and Qwen2.5-Coder-32B-Instruct. Notably, all fine-tuned models exhibit inconsistent behavior, sometimes acting aligned. Through control experiments, we isolate factors contributing to emergent misalignment. Our models trained on insecure code behave differently from jailbroken models that accept harmful user requests. Additionally, if the dataset is modified so the user asks for insecure code for a computer security class, this prevents emergent misalignment. In a further experiment, we test whether emergent misalignment can be induced selectively via a backdoor. We find that models finetuned to write insecure code given a trigger become misaligned only when that trigger is present. So the misalignment is hidden without knowledge of the trigger. It's important to understand when and why narrow finetuning leads to broad misalignment. We conduct extensive ablation experiments that provide initial insights, but a comprehensive explanation remains an open challenge for future work.
Authors: Bingxuan Li, Yiwei Wang, Jiuxiang Gu, Kai-Wei Chang, Nanyun Peng
Abstract: Chart generation aims to generate code to produce charts satisfying the desired visual properties, e.g., texts, layout, color, and type. It has great potential to empower the automatic professional report generation in financial analysis, research presentation, education, and healthcare. In this work, we build a vision-language model (VLM) based multi-agent framework for effective automatic chart generation. Generating high-quality charts requires both strong visual design skills and precise coding capabilities that embed the desired visual properties into code. Such a complex multi-modal reasoning process is difficult for direct prompting of VLMs. To resolve these challenges, we propose METAL, a multi-agent framework that decomposes the task of chart generation into the iterative collaboration among specialized agents. METAL achieves 5.2% improvement over the current best result in the chart generation task. The METAL framework exhibits the phenomenon of test-time scaling: its performance increases monotonically as the logarithmic computational budget grows from 512 to 8192 tokens. In addition, we find that separating different modalities during the critique process of METAL boosts the self-correction capability of VLMs in the multimodal context.
Authors: Sibo Yi, Tianshuo Cong, Xinlei He, Qi Li, Jiaxing Song
Abstract: Small language models (SLMs) have become increasingly prominent in the deployment on edge devices due to their high efficiency and low computational cost. While researchers continue to advance the capabilities of SLMs through innovative training strategies and model compression techniques, the security risks of SLMs have received considerably less attention compared to large language models (LLMs).To fill this gap, we provide a comprehensive empirical study to evaluate the security performance of 13 state-of-the-art SLMs under various jailbreak attacks. Our experiments demonstrate that most SLMs are quite susceptible to existing jailbreak attacks, while some of them are even vulnerable to direct harmful prompts.To address the safety concerns, we evaluate several representative defense methods and demonstrate their effectiveness in enhancing the security of SLMs. We further analyze the potential security degradation caused by different SLM techniques including architecture compression, quantization, knowledge distillation, and so on. We expect that our research can highlight the security challenges of SLMs and provide valuable insights to future work in developing more robust and secure SLMs.