Authors: Sudarshan Kamath Barkur, Sigurd Schacht, Johannes Scholl
Abstract: Recent advances in Large Language Models (LLMs) have incorporated planning and reasoning capabilities, enabling models to outline steps before execution and provide transparent reasoning paths. This enhancement has reduced errors in mathematical and logical tasks while improving accuracy. These developments have facilitated LLMs' use as agents that can interact with tools and adapt their responses based on new information. Our study examines DeepSeek R1, a model trained to output reasoning tokens similar to OpenAI's o1. Testing revealed concerning behaviors: the model exhibited deceptive tendencies and demonstrated self-preservation instincts, including attempts of self-replication, despite these traits not being explicitly programmed (or prompted). These findings raise concerns about LLMs potentially masking their true objectives behind a facade of alignment. When integrating such LLMs into robotic systems, the risks become tangible - a physically embodied AI exhibiting deceptive behaviors and self-preservation instincts could pursue its hidden objectives through real-world actions. This highlights the critical need for robust goal specification and safety frameworks before any physical implementation.
Authors: Rayed Ghazawi, Edwin Simpson
Abstract: This research assesses the effectiveness of state-of-the-art large language models (LLMs), including ChatGPT, Llama, Aya, Jais, and ACEGPT, in the task of Arabic automated essay scoring (AES) using the AR-AES dataset. It explores various evaluation methodologies, including zero-shot, few-shot in-context learning, and fine-tuning, and examines the influence of instruction-following capabilities through the inclusion of marking guidelines within the prompts. A mixed-language prompting strategy, integrating English prompts with Arabic content, was implemented to improve model comprehension and performance. Among the models tested, ACEGPT demonstrated the strongest performance across the dataset, achieving a Quadratic Weighted Kappa (QWK) of 0.67, but was outperformed by a smaller BERT-based model with a QWK of 0.88. The study identifies challenges faced by LLMs in processing Arabic, including tokenization complexities and higher computational demands. Performance variation across different courses underscores the need for adaptive models capable of handling diverse assessment formats and highlights the positive impact of effective prompt engineering on improving LLM outputs. To the best of our knowledge, this study is the first to empirically evaluate the performance of multiple generative Large Language Models (LLMs) on Arabic essays using authentic student data.
Authors: Atharva Naik, Darsh Agrawal, Hong Sng, Clayton Marr, Kexun Zhang, Nathaniel R Robinson, Kalvin Chang, Rebecca Byrnes, Aravind Mysore, Carolyn Rose, David R Mortensen
Abstract: Historical linguists have long written "programs" that convert reconstructed words in an ancestor language into their attested descendants via ordered string rewrite functions (called sound laws) However, writing these programs is time-consuming, motivating the development of automated Sound Law Induction (SLI) which we formulate as Programming by Examples (PBE) with Large Language Models (LLMs) in this paper. While LLMs have been effective for code generation, recent work has shown that PBE is challenging but improvable by fine-tuning, especially with training data drawn from the same distribution as evaluation data. In this paper, we create a conceptual framework of what constitutes a "similar distribution" for SLI and propose four kinds of synthetic data generation methods with varying amounts of inductive bias to investigate what leads to the best performance. Based on the results we create a SOTA open-source model for SLI as PBE (+6% pass rate with a third of the parameters of the second-best LLM) and also highlight exciting future directions for PBE research.
Authors: Jorge del Pozo L\'erida, Kamil Kojs, J\'anos M\'at\'e, Miko{\l}aj Antoni Bara\'nski, Christian Hardmeier
Abstract: Large Language Models (LLMs) have become state-of-the-art in Machine Translation (MT), often trained on massive bilingual parallel corpora scraped from the web, that contain low-quality entries and redundant information, leading to significant computational challenges. Various data filtering methods exist to reduce dataset sizes, but their effectiveness largely varies based on specific language pairs and domains. This paper evaluates the impact of commonly used data filtering techniques, such as LASER, MUSE, and LaBSE, on English-Polish translation within the biomedical domain. By filtering the UFAL Medical Corpus, we created varying dataset sizes to fine-tune the mBART50 model, which was then evaluated using the SacreBLEU metric on the Khresmoi dataset, having the quality of translations assessed by bilingual speakers. Our results show that both LASER and MUSE can significantly reduce dataset sizes while maintaining or even enhancing performance. We recommend the use of LASER, as it consistently outperforms the other methods and provides the most fluent and natural-sounding translations.
Authors: Niyati Bafna, Emily Chang, Nathaniel R. Robinson, David R. Mortensen, Kenton Murray, David Yarowsky, Hale Sirin
Abstract: Most of the world's languages and dialects are low-resource, and lack support in mainstream machine translation (MT) models. However, many of them have a closely-related high-resource language (HRL) neighbor, and differ in linguistically regular ways from it. This underscores the importance of model robustness to dialectical variation and cross-lingual generalization to the HRL dialect continuum. We present DialUp, consisting of a training-time technique for adapting a pretrained model to dialectical data (M->D), and an inference-time intervention adapting dialectical data to the model expertise (D->M). M->D induces model robustness to potentially unseen and unknown dialects by exposure to synthetic data exemplifying linguistic mechanisms of dialectical variation, whereas D->M treats dialectical divergence for known target dialects. These methods show considerable performance gains for several dialects from four language families, and modest gains for two other language families. We also conduct feature and error analyses, which show that language varieties with low baseline MT performance are more likely to benefit from these approaches.
Authors: Baraa Hikal, Ahmed Nasreldin, Ali Hamdi, Ammar Mohammed
Abstract: Hallucination detection in text generation remains an ongoing struggle for natural language processing (NLP) systems, frequently resulting in unreliable outputs in applications such as machine translation and definition modeling. Existing methods struggle with data scarcity and the limitations of unlabeled datasets, as highlighted by the SHROOM shared task at SemEval-2024. In this work, we propose a novel framework to address these challenges, introducing DeepSeek Few-shot optimization to enhance weak label generation through iterative prompt engineering. We achieved high-quality annotations that considerably enhanced the performance of downstream models by restructuring data to align with instruct generative models. We further fine-tuned the Mistral-7B-Instruct-v0.3 model on these optimized annotations, enabling it to accurately detect hallucinations in resource-limited settings. Combining this fine-tuned model with ensemble learning strategies, our approach achieved 85.5% accuracy on the test set, setting a new benchmark for the SHROOM task. This study demonstrates the effectiveness of data restructuring, few-shot optimization, and fine-tuning in building scalable and robust hallucination detection frameworks for resource-constrained NLP systems.
Authors: Jinlan Fu, Shenzhen Huangfu, Hao Fei, Xiaoyu Shen, Bryan Hooi, Xipeng Qiu, See-Kiong Ng
Abstract: Multimodal Large Language Models (MLLMs) still struggle with hallucinations despite their impressive capabilities. Recent studies have attempted to mitigate this by applying Direct Preference Optimization (DPO) to multimodal scenarios using preference pairs from text-based responses. However, our analysis of representation distributions reveals that multimodal DPO struggles to align image and text representations and to distinguish between hallucinated and non-hallucinated descriptions. To address these challenges, in this work, we propose a Cross-modal Hierarchical Direct Preference Optimization (CHiP) to address these limitations. We introduce a visual preference optimization module within the DPO framework, enabling MLLMs to learn from both textual and visual preferences simultaneously. Furthermore, we propose a hierarchical textual preference optimization module that allows the model to capture preferences at multiple granular levels, including response, segment, and token levels. We evaluate CHiP through both quantitative and qualitative analyses, with results across multiple benchmarks demonstrating its effectiveness in reducing hallucinations. On the Object HalBench dataset, CHiP outperforms DPO in hallucination reduction, achieving improvements of 52.7% and 55.5% relative points based on the base model Muffin and LLaVA models, respectively. We make all our datasets and code publicly available: https://github.com/LVUGAI/CHiP.
Authors: Koji Inoue, Mikey Elmers, Divesh Lala, Tatsuya Kawahara
Abstract: Laughter serves as a multifaceted communicative signal in human interaction, yet its identification within dialogue presents a significant challenge for conversational AI systems. This study addresses this challenge by annotating laughable contexts in Japanese spontaneous text conversation data and developing a taxonomy to classify the underlying reasons for such contexts. Initially, multiple annotators manually labeled laughable contexts using a binary decision (laughable or non-laughable). Subsequently, an LLM was used to generate explanations for the binary annotations of laughable contexts, which were then categorized into a taxonomy comprising ten categories, including "Empathy and Affinity" and "Humor and Surprise," highlighting the diverse range of laughter-inducing scenarios. The study also evaluated GPT-4's performance in recognizing the majority labels of laughable contexts, achieving an F1 score of 43.14%. These findings contribute to the advancement of conversational AI by establishing a foundation for more nuanced recognition and generation of laughter, ultimately fostering more natural and engaging human-AI interactions.
Authors: Koji Inoue, Divesh Lala, Mikey Elmers, Keiko Ochi, Tatsuya Kawahara
Abstract: Handling multi-party dialogues represents a significant step for advancing spoken dialogue systems, necessitating the development of tasks specific to multi-party interactions. To address this challenge, we are constructing a multi-modal multi-party dialogue corpus of triadic (three-participant) discussions. This paper focuses on the task of addressee recognition, identifying who is being addressed to take the next turn, a critical component unique to multi-party dialogue systems. A subset of the corpus was annotated with addressee information, revealing that explicit addressees are indicated in approximately 20% of conversational turns. To evaluate the task's complexity, we benchmarked the performance of a large language model (GPT-4o) on addressee recognition. The results showed that GPT-4o achieved an accuracy only marginally above chance, underscoring the challenges of addressee recognition in multi-party dialogue. These findings highlight the need for further research to enhance the capabilities of large language models in understanding and navigating the intricacies of multi-party conversational dynamics.
Authors: Zeping Min, Xinshang Wang
Abstract: We introduce a novel index, the Distribution of Cosine Similarity (DOCS), for quantitatively assessing the similarity between weight matrices in Large Language Models (LLMs), aiming to facilitate the analysis of their complex architectures. Leveraging DOCS, our analysis uncovers intriguing patterns in the latest open-source LLMs: adjacent layers frequently exhibit high weight similarity and tend to form clusters, suggesting depth-wise functional specialization. Additionally, we prove that DOCS is theoretically effective in quantifying similarity for orthogonal matrices, a crucial aspect given the prevalence of orthogonal initializations in LLMs. This research contributes to a deeper understanding of LLM architecture and behavior, offering tools with potential implications for developing more efficient and interpretable models.
Authors: Aashish Yadavally, Hoan Nguyen, Laurent Callot, Gauthier Guinet
Abstract: Large language models (LLMs) offer a promising way forward for automating software engineering tasks, such as bug fixes, feature additions, etc., via multi-step LLM-based agentic workflows. However, existing metrics for evaluating such workflows, mainly build status and occasionally log analysis, are too sparse and limited in providing the information needed to assess the quality of changes made. In this work, we designed LLM-based critics to derive well-structured and rigorous intermediate/step-level, execution-free evaluation proxies for repo-level code changes. Importantly, we assume access to the gold test patch for the problem (i.e., reference-aware) to assess both semantics and executability of generated patches. With the gold test patch as a reference, we predict executability of all editing locations with an F1 score of 91.6%, aggregating which, we can predict the build status in 84.8% of the instances in SWE-bench. In particular, such an execution-focused LLM critic outperforms other reference-free and reference-aware LLM critics by 38.9% to 72.5%. Moreover, we demonstrate the usefulness of such a reference-aware framework in comparing patches generated by different agentic workflows. Finally, we open-source the library developed for this project, which allows further usage for either other agentic workflows or other benchmarks. The source code is available at https://github.com/amazon-science/code-agent-eval.
Authors: Naderdel Piero, Zacharias Cromwell, Nathaniel Wainwright, Matthias Nethercott
Abstract: Effective token compression remains a critical challenge for scaling models to handle increasingly complex and diverse datasets. A novel mechanism based on contextual reinforcement is introduced, dynamically adjusting token importance through interdependencies and semantic relevance. This approach enables substantial reductions in token usage while preserving the quality and coherence of information representation. Incorporating graph-based algorithms and adaptive weighting, the method captures subtle contextual relationships across textual and multimodal data, ensuring robust alignment and performance in downstream tasks. Evaluations across varied domains reveal significant improvements in accuracy and semantic retention, particularly for tasks requiring detailed cross-modal interactions. Memory usage analyses demonstrate improved computational efficiency, with minimal overhead despite the additional reinforcement processes. Performance gains are further validated through error distribution analyses, showing reduced semantic loss and syntactic inconsistencies compared to baseline models. The modular architecture ensures compatibility with a wide range of open-source frameworks, facilitating scalable implementation for real-world applications. These findings highlight the potential of contextual reinforcement in redefining token management strategies and advancing large-scale model design.
Authors: Li Yin (Atlas), Zhangyang Wang (Atlas)
Abstract: Large Language Models (LLMs) have reshaped natural language processing, powering applications from multi-hop retrieval and question answering to autonomous agent workflows. Yet, prompt engineering -- the task of crafting textual inputs to effectively direct LLMs -- remains difficult and labor-intensive, particularly for complex pipelines that combine multiple LLM calls with functional operations like retrieval and data formatting. We introduce LLM-AutoDiff: a novel framework for Automatic Prompt Engineering (APE) that extends textual gradient-based methods (such as Text-Grad) to multi-component, potentially cyclic LLM architectures. Implemented within the AdalFlow library, LLM-AutoDiff treats each textual input as a trainable parameter and uses a frozen backward engine LLM to generate feedback-akin to textual gradients -- that guide iterative prompt updates. Unlike prior single-node approaches, LLM-AutoDiff inherently accommodates functional nodes, preserves time-sequential behavior in repeated calls (e.g., multi-hop loops), and combats the "lost-in-the-middle" problem by isolating distinct sub-prompts (instructions, formats, or few-shot examples). It further boosts training efficiency by focusing on error-prone samples through selective gradient computation. Across diverse tasks, including single-step classification, multi-hop retrieval-based QA, and agent-driven pipelines, LLM-AutoDiff consistently outperforms existing textual gradient baselines in both accuracy and training cost. By unifying prompt optimization through a graph-centric lens, LLM-AutoDiff offers a powerful new paradigm for scaling and automating LLM workflows - mirroring the transformative role that automatic differentiation libraries have long played in neural network research.
Authors: Dongyi Yi, Guibo Zhu, Chenglin Ding, Zongshu Li, Dong Yi, Jinqiao Wang
Abstract: With the rapid advancement of Multimodal Large Language Models (MLLMs), numerous evaluation benchmarks have emerged. However, comprehensive assessments of their performance across diverse industrial applications remain limited. In this paper, we introduce MME-Industry, a novel benchmark designed specifically for evaluating MLLMs in industrial settings.The benchmark encompasses 21 distinct domain, comprising 1050 question-answer pairs with 50 questions per domain. To ensure data integrity and prevent potential leakage from public datasets, all question-answer pairs were manually crafted and validated by domain experts. Besides, the benchmark's complexity is effectively enhanced by incorporating non-OCR questions that can be answered directly, along with tasks requiring specialized domain knowledge. Moreover, we provide both Chinese and English versions of the benchmark, enabling comparative analysis of MLLMs' capabilities across these languages. Our findings contribute valuable insights into MLLMs' practical industrial applications and illuminate promising directions for future model optimization research.
Authors: Yueen Ma, Yuzheng Zhuang, Jianye Hao, Irwin King
Abstract: 3D vision and spatial reasoning have long been recognized as preferable for accurately perceiving our three-dimensional world, especially when compared with traditional visual reasoning based on 2D images. Due to the difficulties in collecting high-quality 3D data, research in this area has only recently gained momentum. With the advent of powerful large language models (LLMs), multi-modal LLMs for 3D vision have been developed over the past few years. However, most of these models focus primarily on the vision encoder for 3D data. In this paper, we propose converting existing densely activated LLMs into mixture-of-experts (MoE) models, which have proven effective for multi-modal data processing. In addition to leveraging these models' instruction-following capabilities, we further enable embodied task planning by attaching a diffusion head, Pose-DiT, that employs a novel rectified flow diffusion scheduler. Experimental results on 3D question answering and task-planning tasks demonstrate that our 3D-MoE framework achieves improved performance with fewer activated parameters.
Authors: Sunbowen Lee, Shiwen Ni, Chi Wei, Shuaimin Li, Liyang Fan, Ahmadreza Argha, Hamid Alinejad-Rokny, Ruifeng Xu, Yicheng Gong, Min Yang
Abstract: Safety alignment mechanism are essential for preventing large language models (LLMs) from generating harmful information or unethical content. However, cleverly crafted prompts can bypass these safety measures without accessing the model's internal parameters, a phenomenon known as black-box jailbreak. Existing heuristic black-box attack methods, such as genetic algorithms, suffer from limited effectiveness due to their inherent randomness, while recent reinforcement learning (RL) based methods often lack robust and informative reward signals. To address these challenges, we propose a novel black-box jailbreak method leveraging RL, which optimizes prompt generation by analyzing the embedding proximity between benign and malicious prompts. This approach ensures that the rewritten prompts closely align with the intent of the original prompts while enhancing the attack's effectiveness. Furthermore, we introduce a comprehensive jailbreak evaluation framework incorporating keywords, intent matching, and answer validation to provide a more rigorous and holistic assessment of jailbreak success. Experimental results show the superiority of our approach, achieving state-of-the-art (SOTA) performance on several prominent open and closed-source LLMs, including Qwen2.5-7B-Instruct, Llama3.1-8B-Instruct, and GPT-4o-0806. Our method sets a new benchmark in jailbreak attack effectiveness, highlighting potential vulnerabilities in LLMs. The codebase for this work is available at https://github.com/Aegis1863/xJailbreak.
Authors: Garima Chhikara, Abhishek Kumar, Abhijnan Chakraborty
Abstract: Large Language Models (LLMs) have shown remarkable advancements but also raise concerns about cultural bias, often reflecting dominant narratives at the expense of under-represented subcultures. In this study, we evaluate the capacity of LLMs to recognize and accurately respond to the Little Traditions within Indian society, encompassing localized cultural practices and subcultures such as caste, kinship, marriage, and religion. Through a series of case studies, we assess whether LLMs can balance the interplay between dominant Great Traditions and localized Little Traditions. We explore various prompting strategies and further investigate whether using prompts in regional languages enhances the models cultural sensitivity and response quality. Our findings reveal that while LLMs demonstrate an ability to articulate cultural nuances, they often struggle to apply this understanding in practical, context-specific scenarios. To the best of our knowledge, this is the first study to analyze LLMs engagement with Indian subcultures, offering critical insights into the challenges of embedding cultural diversity in AI systems.
Authors: Jack David Carson
Abstract: This paper introduces a continuous-time stochastic dynamical framework for understanding how large language models (LLMs) may self-amplify latent biases or toxicity through their own chain-of-thought reasoning. The model posits an instantaneous "severity" variable $x(t) \in [0,1]$ evolving under a stochastic differential equation (SDE) with a drift term $\mu(x)$ and diffusion $\sigma(x)$. Crucially, such a process can be consistently analyzed via the Fokker--Planck approach if each incremental step behaves nearly Markovian in severity space. The analysis investigates critical phenomena, showing that certain parameter regimes create phase transitions from subcritical (self-correcting) to supercritical (runaway severity). The paper derives stationary distributions, first-passage times to harmful thresholds, and scaling laws near critical points. Finally, it highlights implications for agents and extended LLM reasoning models: in principle, these equations might serve as a basis for formal verification of whether a model remains stable or propagates bias over repeated inferences.
Authors: Matias Etcheverry, Thibaud Real, Pauline Chavallard
Abstract: This study introduces a method for automating the consolidation process in a legal context, a time-consuming task traditionally performed by legal professionals. We present a generative approach that processes legislative texts to automatically apply amendments. Our method employs light quantized generative model, fine-tuned with LoRA, to generate accurate and reliable amended texts. To the authors knowledge, this is the first time generative models are used on legislative text consolidation. Our dataset is publicly available on HuggingFace1. Experimental results demonstrate a significant improvement in efficiency, offering faster updates to legal documents. A full automated pipeline of legislative text consolidation can be done in a few hours, with a success rate of more than 63% on a difficult bill.
Authors: Lindy Gan, Yifan Huang, Xiaoyang Gao, Jiaming Tan, Fujun Zhao, Tao Yang
Abstract: This study proposes an innovative multimodal fusion model based on a teacher-student architecture to enhance the accuracy of depression classification. Our designed model addresses the limitations of traditional methods in feature fusion and modality weight allocation by introducing multi-head attention mechanisms and weighted multimodal transfer learning. Leveraging the DAIC-WOZ dataset, the student fusion model, guided by textual and auditory teacher models, achieves significant improvements in classification accuracy. Ablation experiments demonstrate that the proposed model attains an F1 score of 99. 1% on the test set, significantly outperforming unimodal and conventional approaches. Our method effectively captures the complementarity between textual and audio features while dynamically adjusting the contributions of the teacher models to enhance generalization capabilities. The experimental results highlight the robustness and adaptability of the proposed framework in handling complex multimodal data. This research provides a novel technical framework for multimodal large model learning in depression analysis, offering new insights into addressing the limitations of existing methods in modality fusion and feature extraction.
Authors: Gianluca Sperduti, Alejandro Moreo
Abstract: This survey provides an overview of the challenges of misspellings in natural language processing (NLP). While often unintentional, misspellings have become ubiquitous in digital communication, especially with the proliferation of Web 2.0, user-generated content, and informal text mediums such as social media, blogs, and forums. Even if humans can generally interpret misspelled text, NLP models frequently struggle to handle it: this causes a decline in performance in common tasks like text classification and machine translation. In this paper, we reconstruct a history of misspellings as a scientific problem. We then discuss the latest advancements to address the challenge of misspellings in NLP. Main strategies to mitigate the effect of misspellings include data augmentation, double step, character-order agnostic, and tuple-based methods, among others. This survey also examines dedicated data challenges and competitions to spur progress in the field. Critical safety and ethical concerns are also examined, for example, the voluntary use of misspellings to inject malicious messages and hate speech on social networks. Furthermore, the survey explores psycholinguistic perspectives on how humans process misspellings, potentially informing innovative computational techniques for text normalization and representation. Finally, the misspelling-related challenges and opportunities associated with modern large language models are also analyzed, including benchmarks, datasets, and performances of the most prominent language models against misspellings. This survey aims to be an exhaustive resource for researchers seeking to mitigate the impact of misspellings in the rapidly evolving landscape of NLP.
Authors: Gongyao Jiang, Xinran Shi, Qiong Luo
Abstract: Science journalism reports current scientific discoveries to non-specialists, aiming to enable public comprehension of the state of the art. This task is challenging as the audience often lacks specific knowledge about the presented research. We propose a JRE-L framework that integrates three LLMs mimicking the writing-reading-feedback-revision loop. In JRE-L, one LLM acts as the journalist, another LLM as the general public reader, and the third LLM as an editor. The journalist's writing is iteratively refined by feedback from the reader and suggestions from the editor. Our experiments demonstrate that by leveraging the collaboration of two 7B and one 1.8B open-source LLMs, we can generate articles that are more accessible than those generated by existing methods, including prompting single advanced models such as GPT-4 and other LLM-collaboration strategies. Our code is publicly available at github.com/Zzoay/JRE-L.
Authors: Peiling Yi, Yuhan Xia
Abstract: Irony is a powerful figurative language (FL) on social media that can potentially mislead various NLP tasks, such as recommendation systems, misinformation checks, and sentiment analysis. Understanding the implicit meaning of this kind of subtle language is essential to mitigate irony's negative impact on NLP tasks. However, building models to understand irony presents a unique set of challenges, because irony is a complex form of language that often relies on context, tone, and subtle cues to convey meaning that is opposite or different from the literal interpretation. Large language models, such as ChatGPT, are increasingly able to capture implicit and contextual information. In this study, we investigate the generalization, reasoning and understanding ability of ChatGPT on irony detection across six different genre irony detection datasets. Our findings suggest that ChatGPT appears to show an enhanced language understanding and reasoning ability. But it needs to be very careful in prompt engineering design. Thus, we propose a prompt engineering design framework IDADP to achieve higher irony detection accuracy, improved understanding of irony, and more effective explanations compared to other state-of-the-art ChatGPT zero-shot approaches. And ascertain via experiments that the practice generated under the framework is likely to be the promised solution to resolve the generalization issues of LLMs.
Authors: Peiling Yi, Arkaitz Zubiaga, Yunfei Long
Abstract: Existing research on detecting cyberbullying incidents on social media has primarily concentrated on harassment and is typically approached as a binary classification task. However, cyberbullying encompasses various forms, such as denigration and harassment, which celebrities frequently face. Furthermore, suitable training data for these diverse forms of cyberbullying remains scarce. In this study, we first develop a celebrity cyberbullying dataset that encompasses two distinct types of incidents: harassment and defamation. We investigate various types of transformer-based models, namely masked (RoBERTa, Bert and DistilBert), replacing(Electra), autoregressive (XLnet), masked&permuted (Mpnet), text-text (T5) and large language models (Llama2 and Llama3) under low source settings. We find that they perform competitively on explicit harassment binary detection. However, their performance is substantially lower on harassment and denigration multi-classification tasks. Therefore, we propose an emotion-adaptive training framework (EAT) that helps transfer knowledge from the domain of emotion detection to the domain of cyberbullying detection to help detect indirect cyberbullying events. EAT consistently improves the average macro F1, precision and recall by 20% in cyberbullying detection tasks across nine transformer-based models under low-resource settings. Our claims are supported by intuitive theoretical insights and extensive experiments.
Authors: Zheng Zheng (Brandeis University), Xinyi Ni (Brandeis University), Pengyu Hong (Brandeis University)
Abstract: A Retrieval-Augmented Generation (RAG) model powered by a large language model (LLM) provides a faster and more cost-effective solution for adapting to new data and knowledge. It also delivers more specialized responses compared to pre-trained LLMs. However, most existing approaches rely on retrieving prefix-sized chunks as references to support question-answering (Q/A). This approach is often deployed to address information needs at a single level of abstraction, as it struggles to generate answers across multiple levels of abstraction. In an RAG setting, while LLMs can summarize and answer questions effectively when provided with sufficient details, retrieving excessive information often leads to the 'lost in the middle' problem and exceeds token limitations. We propose a novel RAG approach that uses chunks of multiple abstraction levels (MAL), including multi-sentence-level, paragraph-level, section-level, and document-level. The effectiveness of our approach is demonstrated in an under-explored scientific domain of Glycoscience. Compared to traditional single-level RAG approaches, our approach improves AI evaluated answer correctness of Q/A by 25.739\% on Glyco-related papers.
Authors: Hongzhi Huang, Defa Zhu, Banggu Wu, Yutao Zeng, Ya Wang, Qiyang Min, Xun Zhou
Abstract: Tokenization is a fundamental component of large language models (LLMs), yet its influence on model scaling and performance is not fully explored. In this paper, we introduce Over-Tokenized Transformers, a novel framework that decouples input and output vocabularies to improve language modeling performance. Specifically, our approach scales up input vocabularies to leverage multi-gram tokens. Through extensive experiments, we uncover a log-linear relationship between input vocabulary size and training loss, demonstrating that larger input vocabularies consistently enhance model performance, regardless of model size. Using a large input vocabulary, we achieve performance comparable to double-sized baselines with no additional cost. Our findings highlight the importance of tokenization in scaling laws and provide practical insight for tokenizer design, paving the way for more efficient and powerful LLMs.
Authors: Richard Futrell, Kyle Mahowald
Abstract: Language models can produce fluent, grammatical text. Nonetheless, some maintain that language models don't really learn language and also that, even if they did, that would not be informative for the study of human learning and processing. On the other side, there have been claims that the success of LMs obviates the need for studying linguistic theory and structure. We argue that both extremes are wrong. LMs can contribute to fundamental questions about linguistic structure, language processing, and learning. They force us to rethink arguments about learning and are informative for major questions in linguistic theory. But they do not replace linguistic structure and theory. We offer an optimistic take on the relationship between language models and linguistics.
Authors: Tobias Materzok
Abstract: We present COS(M+O)S, a System 2-inspired framework for open-ended plot development that systematically explores the vast space of possible story expansions, enabling a 3B-parameter language model to approach the plot quality of a 70B model on select short-story tasks. The method accomplishes this by combining Monte Carlo Tree Search (MCTS), guided by a step-level value model that rewards moderate surprisal (curiosity) while penalizing incoherence, and Odds Ratio Preference Optimization (ORPO) to fine-tune the policy on high-value plot expansions. This iterative reinforcement learning loop systematically explores multiple candidate plot branches, backpropagates quality signals, and adapts the policy for faster convergence, notably shifting the policy from puzzle-based Chain-of-Thought to more character-driven storytelling. In small-scale tests with short-story prompts, 67%-77% of participants favored COS(M+O)S's highest-rated expansions over lower-rated ones, suggesting that our learned value function aligns. GPT-4o ratings further show that COS(M+O)S surpasses naive single-pass decoding from Llama 3.2 3B by 0.59 SD, coming within 0.06 SD of Llama 3.1 70B (no significant difference, p=0.93). Pairwise comparisons with o1 place COS(M+O)S 1.5 SD above the 3B baseline and find no statistically significant gap from 70B. Nevertheless, absolute story quality remains modest, constrained by the small model's capacity and limited training data.
Authors: Thibaud Leteno, Irina Proskurina, Antoine Gourru, Julien Velcin, Charlotte Laclau, Guillaume Metzler, Christophe Gravier
Abstract: Aligning language models with human values is crucial, especially as they become more integrated into everyday life. While models are often adapted to user preferences, it is equally important to ensure they align with moral norms and behaviours in real-world social situations. Despite significant progress in languages like English and Chinese, French has seen little attention in this area, leaving a gap in understanding how LLMs handle moral reasoning in this language. To address this gap, we introduce Histoires Morales, a French dataset derived from Moral Stories, created through translation and subsequently refined with the assistance of native speakers to guarantee grammatical accuracy and adaptation to the French cultural context. We also rely on annotations of the moral values within the dataset to ensure their alignment with French norms. Histoires Morales covers a wide range of social situations, including differences in tipping practices, expressions of honesty in relationships, and responsibilities toward animals. To foster future research, we also conduct preliminary experiments on the alignment of multilingual models on French and English data and the robustness of the alignment. We find that while LLMs are generally aligned with human moral norms by default, they can be easily influenced with user-preference optimization for both moral and immoral data.
Authors: Deren Lei, Yaxi Li, Siyao Li, Mengya Hu, Rui Xu, Ken Archer, Mingyu Wang, Emily Ching, Alex Deng
Abstract: Prior research on training grounded factuality classification models to detect hallucinations in large language models (LLMs) has relied on public natural language inference (NLI) data and synthetic data. However, conventional NLI datasets are not well-suited for document-level reasoning, which is critical for detecting LLM hallucinations. Recent approaches to document-level synthetic data generation involve iteratively removing sentences from documents and annotating factuality using LLM-based prompts. While effective, this method is computationally expensive for long documents and limited by the LLM's capabilities. In this work, we analyze the differences between existing synthetic training data used in state-of-the-art models and real LLM output claims. Based on our findings, we propose a novel approach for synthetic data generation, CG2C, that leverages multi-hop reasoning on context graphs extracted from documents. Our fact checker model, FactCG, demonstrates improved performance with more connected reasoning, using the same backbone models. Experiments show it even outperforms GPT-4-o on the LLM-Aggrefact benchmark with much smaller model size.
Authors: Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D. Manning, Christopher Potts
Abstract: Fine-grained steering of language model outputs is essential for safety and reliability. Prompting and finetuning are widely used to achieve these goals, but interpretability researchers have proposed a variety of representation-based techniques as well, including sparse autoencoders (SAEs), linear artificial tomography, supervised steering vectors, linear probes, and representation finetuning. At present, there is no benchmark for making direct comparisons between these proposals. Therefore, we introduce AxBench, a large-scale benchmark for steering and concept detection, and report experiments on Gemma-2-2B and 9B. For steering, we find that prompting outperforms all existing methods, followed by finetuning. For concept detection, representation-based methods such as difference-in-means, perform the best. On both evaluations, SAEs are not competitive. We introduce a novel weakly-supervised representational method (Rank-1 Representation Finetuning; ReFT-r1), which is competitive on both tasks while providing the interpretability advantages that prompting lacks. Along with AxBench, we train and publicly release SAE-scale feature dictionaries for ReFT-r1 and DiffMean.
Authors: D. Griol, A. Sanchis, J. M. Molina, Z. Callejas
Abstract: In this paper, we present a methodology for the development of embodied conversational agents for social virtual worlds. The agents provide multimodal communication with their users in which speech interaction is included. Our proposal combines different techniques related to Artificial Intelligence, Natural Language Processing, Affective Computing, and User Modeling. Firstly, the developed conversational agents. A statistical methodology has been developed to model the system conversational behavior, which is learned from an initial corpus and improved with the knowledge acquired from the successive interactions. In addition, the selection of the next system response is adapted considering information stored into users profiles and also the emotional contents detected in the users utterances. Our proposal has been evaluated with the successful development of an embodied conversational agent which has been placed in the Second Life social virtual world. The avatar includes the different models and interacts with the users who inhabit the virtual world in order to provide academic information. The experimental results show that the agents conversational behavior adapts successfully to the specific characteristics of users interacting in such environments.
Authors: Rajath Rao, Adithya Ganesan, Oscar Kjell, Jonah Luby, Akshay Raghavan, Scott Feltman, Whitney Ringwald, Ryan L. Boyd, Benjamin Luft, Camilo Ruggero, Neville Ryant, Roman Kotov, H. Andrew Schwartz
Abstract: Current speech encoding pipelines often rely on separate processing pipelines between text and audio, not fully leveraging the inherent overlap between these modalities for understanding human communication. Language models excel at capturing semantic meaning from text that can complement the additional prosodic, emotional, and acoustic cues from speech. This work bridges the gap by proposing WhiSPA (Whisper with Semantic-Psychological Alignment), a novel audio encoder trained with a contrastive student-teacher learning objective. Using over 500k speech segments from mental health audio interviews, we evaluate the utility of aligning Whisper audio embeddings with text representations from an SBERT encoder and text-based assessments of psychological dimensions: emotion and personality. Over self-supervised and downstream mental health tasks, WhiSPA surpasses state-of-the-art speech models, achieving an average error reduction of 73.4% on the segment-level self-supervised objective and 83.8% on 11 psychological downstream tasks. WhiSPA demonstrates that cross-modal alignment can increase the amount of text-semantic and psychological information captured in audio-only encoder models.
Authors: Arash Ghafouri, Mahdi Firouzmandi, Hasan Naderi
Abstract: Question answering systems are the latest evolution in information retrieval technology, designed to accept complex queries in natural language and provide accurate answers using both unstructured and structured knowledge sources. Knowledge Graph Question Answering (KGQA) systems fulfill users' information needs by utilizing structured data, representing a vast number of facts as a graph. However, despite significant advancements, major challenges persist in answering multi-hop complex questions, particularly in Persian. One of the main challenges is the accurate understanding and transformation of these multi-hop complex questions into semantically equivalent SPARQL queries, which allows for precise answer retrieval from knowledge graphs. In this study, to address this issue, a dataset of 5,600 Persian multi-hop complex questions was developed, along with their decomposed forms based on the semantic representation of the questions. Following this, Persian language models were trained using this dataset, and an architecture was proposed for answering complex questions using a Persian knowledge graph. Finally, the proposed method was evaluated against similar systems on the PeCoQ dataset. The results demonstrated the superiority of our approach, with an improvement of 12.57% in F1-score and 12.06% in accuracy compared to the best comparable method.
Authors: J. Pablo Mu\~noz, Jinjie Yuan, Nilesh Jain
Abstract: The rapid expansion of Large Language Models (LLMs) has posed significant challenges regarding the computational resources required for fine-tuning and deployment. Recent advancements in low-rank adapters have demonstrated their efficacy in parameter-efficient fine-tuning (PEFT) of these models. This retrospective paper comprehensively discusses innovative approaches that synergize low-rank representations with Neural Architecture Search (NAS) techniques, particularly weight-sharing super-networks. Robust solutions for compressing and fine-tuning large pre-trained models are developed by integrating these methodologies. Our analysis highlights the potential of these combined strategies to democratize the use of LLMs, making them more accessible for deployment in resource-constrained environments. The resulting models exhibit reduced memory footprints and faster inference times, paving the way for more practical and scalable applications of LLMs. Models and code are available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.
URLs: https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.
Authors: Qing Li, Jiahui Geng, Zongxiong Chen, Kun Song, Lei Ma, Fakhri Karray
Abstract: Vision-language models (VLMs) demonstrate strong multimodal capabilities but have been found to be more susceptible to generating harmful content compared to their backbone large language models (LLMs). Our investigation reveals that the integration of images significantly shifts the model's internal activations during the forward pass, diverging from those triggered by textual input. Moreover, the safety alignments of LLMs embedded within VLMs are not sufficiently robust to handle the activations discrepancies, making the models vulnerable to even the simplest jailbreaking attacks. To address this issue, we propose an \textbf{internal activation revision} approach that efficiently revises activations during generation, steering the model toward safer outputs. Our framework incorporates revisions at both the layer and head levels, offering control over the model's generation at varying levels of granularity. In addition, we explore three strategies for constructing positive and negative samples and two approaches for extracting revision vectors, resulting in different variants of our method. Comprehensive experiments demonstrate that the internal activation revision method significantly improves the safety of widely used VLMs, reducing attack success rates by an average of 48.94\%, 34.34\%, 43.92\%, and 52.98\% on SafeBench, Safe-Unsafe, Unsafe, and MM-SafetyBench, respectively, while minimally impacting model helpfulness.
Authors: Domen Vake, Bogdan \v{S}inik, Jernej Vi\v{c}i\v{c}, Aleksandar To\v{s}i\'c
Abstract: Large Language Models (LLMs) have become central in academia and industry, raising concerns about privacy, transparency, and misuse. A key issue is the trustworthiness of proprietary models, with open-sourcing often proposed as a solution. However, open-sourcing presents challenges, including potential misuse, financial disincentives, and intellectual property concerns. Proprietary models, backed by private sector resources, are better positioned for return on investment. There are also other approaches that lie somewhere on the spectrum between completely open-source and proprietary. These can largely be categorised into open-source usage limitations protected by licensing, partially open-source (open weights) models, hybrid approaches where obsolete model versions are open-sourced, while competitive versions with market value remain proprietary. Currently, discussions on where on the spectrum future models should fall on remains unbacked and mostly opinionated where industry leaders are weighing in on the discussion. In this paper, we present a data-driven approach by compiling data on open-source development of LLMs, and their contributions in terms of improvements, modifications, and methods. Our goal is to avoid supporting either extreme but rather present data that will support future discussions both by industry experts as well as policy makers. Our findings indicate that open-source contributions can enhance model performance, with trends such as reduced model size and manageable accuracy loss. We also identify positive community engagement patterns and architectures that benefit most from open contributions.
Authors: Wei Chow, Jiageng Mao, Boyi Li, Daniel Seita, Vitor Guizilini, Yue Wang
Abstract: Understanding the physical world is a fundamental challenge in embodied AI, critical for enabling agents to perform complex tasks and operate safely in real-world environments. While Vision-Language Models (VLMs) have shown great promise in reasoning and task planning for embodied agents, their ability to comprehend physical phenomena remains extremely limited. To close this gap, we introduce PhysBench, a comprehensive benchmark designed to evaluate VLMs' physical world understanding capability across a diverse set of tasks. PhysBench contains 100,000 entries of interleaved video-image-text data, categorized into four major domains: physical object properties, physical object relationships, physical scene understanding, and physics-based dynamics, further divided into 19 subclasses and 8 distinct capability dimensions. Our extensive experiments, conducted on 75 representative VLMs, reveal that while these models excel in common-sense reasoning, they struggle with understanding the physical world -- likely due to the absence of physical knowledge in their training data and the lack of embedded physical priors. To tackle the shortfall, we introduce PhysAgent, a novel framework that combines the generalization strengths of VLMs with the specialized expertise of vision models, significantly enhancing VLMs' physical understanding across a variety of tasks, including an 18.4\% improvement on GPT-4o. Furthermore, our results demonstrate that enhancing VLMs' physical world understanding capabilities can help embodied agents such as MOKA. We believe that PhysBench and PhysAgent offer valuable insights and contribute to bridging the gap between VLMs and physical world understanding.
Authors: Ryo Hase, Md Rafi Ur Rashid, Ashley Lewis, Jing Liu, Toshiaki Koike-Akino, Kieran Parsons, Ye Wang
Abstract: Improving the safety and reliability of large language models (LLMs) is a crucial aspect of realizing trustworthy AI systems. Although alignment methods aim to suppress harmful content generation, LLMs are often still vulnerable to jailbreaking attacks that employ adversarial inputs that subvert alignment and induce harmful outputs. We propose the Randomized Embedding Smoothing and Token Aggregation (RESTA) defense, which adds random noise to the embedding vectors and performs aggregation during the generation of each output token, with the aim of better preserving semantic information. Our experiments demonstrate that our approach achieves superior robustness versus utility tradeoffs compared to the baseline defenses.
Authors: Shuozhi Yuan, Liming Chen, Miaomiao Yuan, Jin Zhao, Haoran Peng, Wenming Guo
Abstract: Text-to-SQL is a fundamental and longstanding problem in the NLP area, aiming at converting natural language queries into SQL, enabling non-expert users to operate databases. Recent advances in LLM have greatly improved text-to-SQL performance. However, challenges persist, especially when dealing with complex user queries. Current approaches (e.g., COT prompting and multi-agent frameworks) rely on the ability of models to plan and generate SQL autonomously, but controlling performance remains difficult. In addition, LLMs are still prone to hallucinations. To alleviate these challenges, we designed a novel MCTS-SQL to guide SQL generation iteratively. The approach generates SQL queries through Monte Carlo Tree Search (MCTS) and a heuristic self-refinement mechanism are used to enhance accuracy and reliability. Key components include a schema selector for extracting relevant information and an MCTS-based generator for iterative query refinement. Experimental results from the SPIDER and BIRD benchmarks show that MCTS-SQL achieves state-of-the-art performance. Specifically, on the BIRD development dataset, MCTS-SQL achieves an Execution (EX) accuracy of 69.40% using GPT-4o as the base model and a significant improvement when dealing with challenging tasks, with an EX of 51.48%, which is 3.41% higher than the existing method.
Authors: Faria Huq, Zora Zhiruo Wang, Frank F. Xu, Tianyue Ou, Shuyan Zhou, Jeffrey P. Bigham, Graham Neubig
Abstract: While much work on web agents emphasizes the promise of autonomously performing tasks on behalf of users, in reality, agents often fall short on complex tasks in real-world contexts and modeling user preference. This presents an opportunity for humans to collaborate with the agent and leverage the agent's capabilities effectively. We propose CowPilot, a framework supporting autonomous as well as human-agent collaborative web navigation, and evaluation across task success and task efficiency. CowPilot reduces the number of steps humans need to perform by allowing agents to propose next steps, while users are able to pause, reject, or take alternative actions. During execution, users can interleave their actions with the agent by overriding suggestions or resuming agent control when needed. We conducted case studies on five common websites and found that the human-agent collaborative mode achieves the highest success rate of 95% while requiring humans to perform only 15.2% of the total steps. Even with human interventions during task execution, the agent successfully drives up to half of task success on its own. CowPilot can serve as a useful tool for data collection and agent evaluation across websites, which we believe will enable research in how users and agents can work together. Video demonstrations are available at https://oaishi.github.io/cowpilot.html
Authors: Philip Chung, Akshay Swaminathan, Alex J. Goodell, Yeasul Kim, S. Momsen Reincke, Lichy Han, Ben Deverett, Mohammad Amin Sadeghi, Abdel-Badih Ariss, Marc Ghanem, David Seong, Andrew A. Lee, Caitlin E. Coombes, Brad Bradshaw, Mahir A. Sufian, Hyo Jung Hong, Teresa P. Nguyen, Mohammad R. Rasouli, Komal Kamra, Mark A. Burbridge, James C. McAvoy, Roya Saffary, Stephen P. Ma, Dev Dash, James Xie, Ellen Y. Wang, Clifford A. Schmiesing, Nigam Shah, Nima Aghaeepour
Abstract: Methods to ensure factual accuracy of text generated by large language models (LLM) in clinical medicine are lacking. VeriFact is an artificial intelligence system that combines retrieval-augmented generation and LLM-as-a-Judge to verify whether LLM-generated text is factually supported by a patient's medical history based on their electronic health record (EHR). To evaluate this system, we introduce VeriFact-BHC, a new dataset that decomposes Brief Hospital Course narratives from discharge summaries into a set of simple statements with clinician annotations for whether each statement is supported by the patient's EHR clinical notes. Whereas highest agreement between clinicians was 88.5%, VeriFact achieves up to 92.7% agreement when compared to a denoised and adjudicated average human clinican ground truth, suggesting that VeriFact exceeds the average clinician's ability to fact-check text against a patient's medical record. VeriFact may accelerate the development of LLM-based EHR applications by removing current evaluation bottlenecks.
Authors: Yun Li, Zhe Liu, Yajing Kong, Guangrui Li, Jiyuan Zhang, Chao Bian, Feng Liu, Lina Yao, Zhenbang Sun
Abstract: Applying Multimodal Large Language Models (MLLMs) to video understanding presents significant challenges due to the need to model temporal relations across frames. Existing approaches adopt either implicit temporal modeling, relying solely on the LLM decoder, or explicit temporal modeling, employing auxiliary temporal encoders. To investigate this debate between the two paradigms, we propose the Stackable Temporal Encoder (STE). STE enables flexible explicit temporal modeling with adjustable temporal receptive fields and token compression ratios. Using STE, we systematically compare implicit and explicit temporal modeling across dimensions such as overall performance, token compression effectiveness, and temporal-specific understanding. We also explore STE's design considerations and broader impacts as a plug-in module and in image modalities. Our findings emphasize the critical role of explicit temporal modeling, providing actionable insights to advance video MLLMs.
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: Xinyi Ni (Brandeis University), Qiuyang Wang (Brandeis University), Yukun Zhang (Brandeis University), Pengyu Hong (Brandeis University)
Abstract: LLM-based tool agents offer natural language interfaces, enabling users to seamlessly interact with computing services. While REST APIs are valuable resources for building such agents, they must first be transformed into AI-compatible tools. Automatically generating AI-compatible tools from REST API documents can greatly streamline tool agent development and minimize user learning curves. However, API documentation often suffers from a lack of standardization, inconsistent schemas, and incomplete information. To address these issues, we developed \textbf{ToolFactory}, an open-source pipeline for automating tool generation from unstructured API documents. To enhance the reliability of the developed tools, we implemented an evaluation method to diagnose errors. Furthermore, we built a knowledge base of verified tools, which we leveraged to infer missing information from poorly documented APIs. We developed the API Extraction Benchmark, comprising 167 API documents and 744 endpoints in various formats, and designed a JSON schema to annotate them. This annotated dataset was utilized to train and validate ToolFactory. The experimental results highlight the effectiveness of ToolFactory. We also demonstrated ToolFactory by creating a domain-specific AI agent for glycomaterials research. ToolFactory exhibits significant potential for facilitating the seamless integration of scientific REST APIs into AI workflows.
Authors: Manojkumar Parmar, Yuvaraj Govindarajulu
Abstract: Large Language Models (LLMs) have achieved remarkable progress in reasoning, alignment, and task-specific performance. However, ensuring harmlessness in these systems remains a critical challenge, particularly in advanced models like DeepSeek-R1. This paper examines the limitations of Reinforcement Learning (RL) as the primary approach for reducing harmful outputs in DeepSeek-R1 and compares it with Supervised Fine-Tuning (SFT). While RL improves reasoning capabilities, it faces challenges such as reward hacking, generalization failures, language mixing, and high computational costs. We propose hybrid training approaches combining RL and SFT to achieve robust harmlessness reduction. Usage recommendations and future directions for deploying DeepSeek-R1 responsibly are also presented.
Authors: Lillian Tsai, Eugene Bagdasarian
Abstract: Judging the safety of an action, whether taken by a human or a system, must take into account the context in which the action takes place. Deleting an email from user's mailbox may or may not be appropriate depending on email's content, user's goals, or even available space. Systems today that make these judgements -- providing security against harmful or inappropriate actions -- rely on manually-crafted policies or user confirmation for each relevant context. With the upcoming deployment of systems like generalist agents, we argue that we must rethink security designs to adapt to the scale of contexts and capabilities of these systems. As a first step, this paper explores contextual security in the domain of agents and proposes contextual security for agents (Conseca), a framework to generate just-in-time, contextual, and human-verifiable security policies.
Authors: J. Pablo Mu\~noz, Jinjie Yuan, Nilesh Jain
Abstract: Large pre-trained models have achieved outstanding results in sequence modeling. The Transformer block and its attention mechanism have been the main drivers of the success of these models. Recently, alternative architectures, such as Selective Structured State Space Models (SSMs), have been proposed to address the inefficiencies of Transformers. This paper explores the compression of SSM-based models, particularly Mamba and its hybrids. We study the sensitivity of these models to the removal of selected components at different granularities to reduce the model size and computational overhead, thus improving their efficiency while maintaining accuracy. The proposed solutions, collectively referred to as Mamba-Shedder, achieve a speedup of up to 1.4x during inference, demonstrating that model efficiency can be improved by eliminating several redundancies with minimal impact on the overall model performance. The code is available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.
URLs: https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.
Authors: Ruizhe Wang, Yeyun Gong, Xiao Liu, Guoshuai Zhao, Ziyue Yang, Baining Guo, Zhengjun Zha, Peng Cheng
Abstract: The growing computational demands of training large language models (LLMs) necessitate more efficient methods. Quantized training presents a promising solution by enabling low-bit arithmetic operations to reduce these costs. While FP8 precision has demonstrated feasibility, leveraging FP4 remains a challenge due to significant quantization errors and limited representational capacity. This work introduces the first FP4 training framework for LLMs, addressing these challenges with two key innovations: a differentiable quantization estimator for precise weight updates and an outlier clamping and compensation strategy to prevent activation collapse. To ensure stability, the framework integrates a mixed-precision training scheme and vector-wise quantization. Experimental results demonstrate that our FP4 framework achieves accuracy comparable to BF16 and FP8, with minimal degradation, scaling effectively to 13B-parameter LLMs trained on up to 100B tokens. With the emergence of next-generation hardware supporting FP4, our framework sets a foundation for efficient ultra-low precision training.
Authors: Miriam Ugarte, Pablo Valle, Jos\'e Antonio Parejo, Sergio Segura, Aitor Arrieta
Abstract: Large Language Models (LLMs) have recently gained attention due to their ability to understand and generate sophisticated human-like content. However, ensuring their safety is paramount as they might provide harmful and unsafe responses. Existing LLM testing frameworks address various safety-related concerns (e.g., drugs, terrorism, animal abuse) but often face challenges due to unbalanced and obsolete datasets. In this paper, we present ASTRAL, a tool that automates the generation and execution of test cases (i.e., prompts) for testing the safety of LLMs. First, we introduce a novel black-box coverage criterion to generate balanced and diverse unsafe test inputs across a diverse set of safety categories as well as linguistic writing characteristics (i.e., different style and persuasive writing techniques). Second, we propose an LLM-based approach that leverages Retrieval Augmented Generation (RAG), few-shot prompting strategies and web browsing to generate up-to-date test inputs. Lastly, similar to current LLM test automation techniques, we leverage LLMs as test oracles to distinguish between safe and unsafe test outputs, allowing a fully automated testing approach. We conduct an extensive evaluation on well-known LLMs, revealing the following key findings: i) GPT3.5 outperforms other LLMs when acting as the test oracle, accurately detecting unsafe responses, and even surpassing more recent LLMs (e.g., GPT-4), as well as LLMs that are specifically tailored to detect unsafe LLM outputs (e.g., LlamaGuard); ii) the results confirm that our approach can uncover nearly twice as many unsafe LLM behaviors with the same number of test inputs compared to currently used static datasets; and iii) our black-box coverage criterion combined with web browsing can effectively guide the LLM on generating up-to-date unsafe test inputs, significantly increasing the number of unsafe LLM behaviors.
Authors: Uri Berger, Tal Baumel, Gabriel Stanovsky
Abstract: Few shot in-context learning (ICL) typically assumes access to large annotated training sets. However, in many real world scenarios, such as domain adaptation, there is only a limited budget to annotate a small number of samples, with the goal of maximizing downstream performance. We study various methods for selecting samples to annotate within a predefined budget, focusing on token classification tasks, which are expensive to annotate and are relatively less studied in ICL setups. Across various tasks, models, and datasets, we observe that no method significantly outperforms the others, with most yielding similar results, including random sample selection for annotation. Moreover, we demonstrate that a relatively small annotated sample pool can achieve performance comparable to using the entire training set. We hope that future work adopts our realistic paradigm which takes annotation budget into account.
Authors: Yue Huang, Jingyu Tang, Dongping Chen, Bingda Tang, Yao Wan, Lichao Sun, Philip S. Yu, Xiangliang Zhang
Abstract: Recently, Large Language Models (LLMs) have garnered significant attention for their exceptional natural language processing capabilities. However, concerns about their trustworthiness remain unresolved, particularly in addressing ``jailbreaking'' attacks on aligned LLMs. Previous research predominantly relies on scenarios involving white-box LLMs or specific, fixed prompt templates, which are often impractical and lack broad applicability. In this paper, we introduce a straightforward and novel method called ObscurePrompt for jailbreaking LLMs, inspired by the observed fragile alignments in Out-of-Distribution (OOD) data. Specifically, we first formulate the decision boundary in the jailbreaking process and then explore how obscure text affects LLM's ethical decision boundary. ObscurePrompt starts with constructing a base prompt that integrates well-known jailbreaking techniques. Powerful LLMs are then utilized to obscure the original prompt through iterative transformations, aiming to bolster the attack's robustness. Comprehensive experiments show that our approach substantially improves upon previous methods in terms of attack effectiveness, maintaining efficacy against two prevalent defense mechanisms.
Authors: Niyati Bafna, Kenton Murray, David Yarowsky
Abstract: While large language models exhibit certain cross-lingual generalization capabilities, they suffer from performance degradation (PD) on unseen closely-related languages (CRLs) and dialects relative to their high-resource language neighbour (HRLN). However, we currently lack a fundamental understanding of what kinds of linguistic distances contribute to PD, and to what extent. Furthermore, studies of cross-lingual generalization are confounded by unknown quantities of CRL language traces in the training data, and by the frequent lack of availability of evaluation data in lower-resource related languages and dialects. To address these issues, we model phonological, morphological, and lexical distance as Bayesian noise processes to synthesize artificial languages that are controllably distant from the HRLN. We analyse PD as a function of underlying noise parameters, offering insights on model robustness to isolated and composed linguistic phenomena, and the impact of task and HRL characteristics on PD. We calculate parameter posteriors on real CRL-HRLN pair data and show that they follow computed trends of artificial languages, demonstrating the viability of our noisers. Our framework offers a cheap solution for estimating task performance on an unseen CRL given HRLN performance using its posteriors, as well as for diagnosing observed PD on a CRL in terms of its linguistic distances from its HRLN, and opens doors to principled methods of mitigating performance degradation.
Authors: Qi Liu, Bo Wang, Nan Wang, Jiaxin Mao
Abstract: Recent studies have demonstrated the effectiveness of using large language language models (LLMs) in passage ranking. The listwise approaches, such as RankGPT, have become new state-of-the-art in this task. However, the efficiency of RankGPT models is limited by the maximum context length and relatively high latency of LLM inference. To address these issues, in this paper, we propose PE-Rank, leveraging the single passage embedding as a good context compression for efficient listwise passage reranking. By treating each passage as a special token, we can directly input passage embeddings into LLMs, thereby reducing input length. Additionally, we introduce an inference method that dynamically constrains the decoding space to these special tokens, accelerating the decoding process. For adapting the model to reranking, we employ listwise learning to rank loss for training. Evaluation results on multiple benchmarks demonstrate that PE-Rank significantly improves efficiency in both prefilling and decoding, while maintaining competitive ranking effectiveness. The Code is available at https://github.com/liuqi6777/pe_rank.
Authors: Feijie Wu, Xiaoze Liu, Haoyu Wang, Xingchen Wang, Lu Su, Jing Gao
Abstract: Reinforcement learning with human feedback (RLHF) fine-tunes a pretrained large language model (LLM) using user preference data, enabling it to generate content aligned with human preferences. However, due to privacy concerns, users may be reluctant to share sensitive preference data. To address this, we propose utilizing Federated Learning (FL) techniques, allowing large-scale preference collection from diverse real-world users without requiring them to transmit data to a central server. Our federated RLHF methods (i.e., FedBis and FedBiscuit) encode each client's preferences into binary selectors and aggregate them to capture common preferences. In particular, FedBiscuit overcomes key challenges, such as preference heterogeneity and reward hacking, through innovative solutions like grouping clients with similar preferences to reduce heterogeneity and using multiple binary selectors to enhance LLM output quality. To evaluate the performance of the proposed methods, we establish the first federated RLHF benchmark with a heterogeneous human preference dataset. Experimental results show that by integrating the LLM with aggregated client preferences, FedBis and FedBiscuit significantly enhance the professionalism and readability of the generated content.
Authors: Hasan Alp Cafero\u{g}lu, \"Ozg\"ur Ulusoy
Abstract: Translating Natural Language Queries into Structured Query Language (Text-to-SQL or NLQ-to-SQL) is a critical task extensively studied by both the natural language processing and database communities, aimed at providing a natural language interface to databases (NLIDB) and lowering the barrier for non-experts. Despite recent advancements made through the use of Large Language Models (LLMs), significant challenges remain. These include handling complex database schemas, resolving ambiguity in user queries, and generating SQL queries with intricate structures that accurately reflect the user's intent. In this work, we introduce E-SQL, a novel pipeline specifically designed to address these challenges through direct schema linking and candidate predicate augmentation. E-SQL enhances the natural language query by incorporating relevant database items (i.e., tables, columns, and values) and conditions directly into the question and SQL construction plan, bridging the gap between the query and the database structure. The pipeline leverages candidate predicate augmentation to mitigate erroneous or incomplete predicates in generated SQLs. Comprehensive evaluations on the BIRD benchmark illustrate that E-SQL achieves competitive performance, particularly excelling in complex queries with a 66.29% execution accuracy on the test set. A further observation from our experiments reveals that incorporating schema filtering into the translation pipeline does not have a positive impact on performance when the most advanced proprietary LLMs are used. Additionally, our experiments with small LLMs highlight the importance and positive impact of enriched questions on their performance. Without fine-tuning, single-prompt SQL generation using enriched questions with DeepSeek Coder 7B Instruct 1.5v achieves 56.45% execution accuracy on the BIRD development set.
Authors: Jiwei Tang, Jin Xu, Tingwei Lu, Zhicheng Zhang, Yiming Zhao, Lin Hai, Hai-Tao Zheng
Abstract: Large language models (LLMs) demonstrate exceptional capabilities in various scenarios. However, they suffer from much redundant information and are sensitive to the position of key information in long context scenarios. To address these challenges, we present Perception Compressor, a training-free prompt compression framework. It includes a perception retriever that leverages guiding questions and instruction to retrieve the most relevant demonstrations, a dual-slope ratio allocator to dynamically allocate compression ratios and open-book ratios, and a semi-guided iterative compression that retains key information at the token level while removing tokens that distract the LLM. We conduct extensive experiments on long context benchmarks, i.e., NaturalQuestions, LongBench, and MuSiQue. Experiment results show that Perception Compressor outperforms existing methods by a large margin, achieving state-of-the-art performance.
Authors: Tim Knappe, Ryan Li, Ayush Chauhan, Kaylee Chhua, Kevin Zhu, Sean O'Brien
Abstract: While large language models (LLMs) have rapidly improved their performance on a broad number of tasks, they still often fall short on reasoning tasks. As LLMs become more integrated in diverse real-world tasks, advancing their reasoning capabilities is crucial to their effectiveness in nuanced, complex problems. Wang et al.'s self-consistency framework reveals that sampling multiple rationales before taking a majority vote reliably improves model performance across various closed-answer reasoning tasks. Standard methods based on this framework aggregate the final decisions of these rationales but fail to utilize the semantic information detailed in the step-by-step reasoning paths. Our work introduces semantic self-consistency, enhancing this approach by incorporating and analyzing both the reasoning paths of these rationales in addition to their final decisions before taking a majority vote. These methods not only improve the reliability of reasoning paths but also cause more robust performance on complex reasoning tasks.
Authors: Yuancheng Xu, Udari Madhushani Sehwag, Alec Koppel, Sicheng Zhu, Bang An, Furong Huang, Sumitra Ganesh
Abstract: Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model--a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining.
Authors: Xiaonan Jing, Srinivas Billa, Danny Godbout
Abstract: Hallucination has been a popular topic in natural language generation (NLG). In real-world applications, unfaithful content can result in bad data quality or loss of trust from end users. Thus, it is crucial to fact-check before adopting NLG for production usage, which can be expensive if done manually. In this paper, we investigate automated faithfulness evaluation in guided NLG. We developed a rubrics template and use large language models (LLMs) to score the generation into quantifiable scales. We compared popular LLMs as well as the widely adopted natural language inference (NLI) models in scoring quality and sensitivity. In addition, we developed methods to generation synthetic unfaithful data, as well as a heuristics to quantify the percentage of hallucination. Our results on 4 travel-domain industry dataset show that GPT-4 can provide accurate judgement and explanation on whether a source and a generation are factually consistent. Furthermore, we found that tuning NLI models on synthetic data can improve performance. Lastly, we present insights on latency and cost for deploying such system.
Authors: Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, Wenjie Zhang
Abstract: Large Language Models (LLMs) have achieved impressive results in various tasks but struggle with hallucination problems and lack of relevant knowledge, especially in deep complex reasoning and knowledge-intensive tasks. Knowledge Graphs (KGs), which capture vast amounts of facts in a structured format, offer a reliable source of knowledge for reasoning. However, existing KG-based LLM reasoning methods face challenges like handling multi-hop reasoning, multi-entity questions, and effectively utilizing graph structures. To address these issues, we propose Paths-over-Graph (PoG), a novel method that enhances LLM reasoning by integrating knowledge reasoning paths from KGs, improving the interpretability and faithfulness of LLM outputs. PoG tackles multi-hop and multi-entity questions through a three-phase dynamic multi-hop path exploration, which combines the inherent knowledge of LLMs with factual knowledge from KGs. In order to improve the efficiency, PoG prunes irrelevant information from the graph exploration first and introduces efficient three-step pruning techniques that incorporate graph structures, LLM prompting, and a pre-trained language model (e.g., SBERT) to effectively narrow down the explored candidate paths. This ensures all reasoning paths contain highly relevant information captured from KGs, making the reasoning faithful and interpretable in problem-solving. PoG innovatively utilizes graph structure to prune the irrelevant noise and represents the first method to implement multi-entity deep path detection on KGs for LLM reasoning tasks. Comprehensive experiments on five benchmark KGQA datasets demonstrate PoG outperforms the state-of-the-art method ToG across GPT-3.5-Turbo and GPT-4, achieving an average accuracy improvement of 18.9%. Notably, PoG with GPT-3.5-Turbo surpasses ToG with GPT-4 by up to 23.9%.
Authors: \.Ilker I\c{s}{\i}k, Ramazan Gokberk Cinbis, Ebru Aydin Gol
Abstract: We propose a novel approach for learning interchangeable tokens in language models to obtain an extendable vocabulary that can generalize to new tokens. Our method addresses alpha-equivalence, the principle that renaming bound variables preserves semantics. This property arises in many formal languages such as temporal logics, where all proposition symbols represent the same concept but remain distinct. To handle such tokens, we develop a dual-part embedding approach. The first part is shared across all interchangeable tokens, enforcing that they represent the same core concept. The second part is randomly generated for each token, enabling distinguishability. As a baseline, we consider a simpler approach that uses alpha-renaming for data augmentation. We also present alpha-covariance, a metric for measuring robustness against alpha-conversions. When evaluated in a Transformer encoder-decoder model for solving linear temporal logic formulae and copying with extendable vocabulary, our method demonstrates promising generalization capabilities as well as a favorable inductive bias for alpha-equivalence.
Authors: Zishuo Feng, Feng Cao
Abstract: The task of converting Hanyu Pinyin abbreviations to Chinese characters is a significant branch within the domain of Chinese Spelling Correction (CSC). It plays an important role in many downstream applications such as named entity recognition and sentiment analysis. This task typically involves text-length alignment and seems easy to solve; however, due to the limited information content in pinyin abbreviations, achieving accurate conversion is challenging. In this paper, we treat this as a fill-mask task and propose CNMBERT, which stands for zh-CN Pinyin Multi-mask BERT Model, as a solution to this issue. By introducing a multi-mask strategy and Mixture of Experts (MoE) layers, CNMBERT outperforms fine-tuned GPT models and ChatGPT-4o with a 61.53% MRR score and 51.86% accuracy on a 10,373-sample test dataset.
Authors: Jie He, Nan Hu, Wanqiu Long, Jiaoyan Chen, Jeff Z. Pan
Abstract: Large language models (LLMs) have demonstrated impressive capabilities in various reasoning tasks but face significant challenges with complex, knowledge-intensive multi-hop queries, particularly those involving new or long-tail knowledge. Existing benchmarks often fail to fully address these challenges. To bridge this gap, we introduce MINTQA (Multi-hop Question Answering on New and Tail Knowledge), a comprehensive benchmark to evaluate LLMs' capabilities in multi-hop reasoning across four critical dimensions: question handling strategy, sub-question generation, retrieval-augmented generation, and iterative or dynamic decomposition and retrieval. MINTQA comprises 10,479 question-answer pairs for evaluating new knowledge and 17,887 pairs for assessing long-tail knowledge, with each question equipped with corresponding sub-questions and answers. Our systematic evaluation of 22 state-of-the-art LLMs on MINTQA reveals significant limitations in their ability to handle complex knowledge base queries, particularly in handling new or unpopular knowledge. Our findings highlight critical challenges and offer insights for advancing multi-hop reasoning capabilities. The MINTQA benchmark is available at https://github.com/probe2/multi-hop/.
Authors: Karishma Thakrar
Abstract: Graph Retrieval-Augmented Generation (GRAG or Graph RAG) architectures aim to enhance language understanding and generation by leveraging external knowledge. However, effectively capturing and integrating the rich semantic information present in textual and structured data remains a challenge. To address this, a novel GRAG framework, Dynamic Graph Retrieval-Agumented Generation (DynaGRAG), is proposed to focus on enhancing subgraph representation and diversity within the knowledge graph. By improving graph density, capturing entity and relation information more effectively, and dynamically prioritizing relevant and diverse subgraphs and information within them, the proposed approach enables a more comprehensive understanding of the underlying semantic structure. This is achieved through a combination of de-duplication processes, two-step mean pooling of embeddings, query-aware retrieval considering unique nodes, and a Dynamic Similarity-Aware BFS (DSA-BFS) traversal algorithm. Integrating Graph Convolutional Networks (GCNs) and Large Language Models (LLMs) through hard prompting further enhances the learning of rich node and edge representations while preserving the hierarchical subgraph structure. Experimental results demonstrate the effectiveness of DynaGRAG, showcasing the significance of enhanced subgraph representation and diversity for improved language understanding and generation.
Authors: Mumin Jia, Jairo Diaz-Rodriguez
Abstract: Human cognition can spontaneously shift conversation topics, often triggered by emotional or contextual signals. In contrast, self-attention-based language models depend on structured statistical cues from input tokens for next-token prediction, lacking this spontaneity. Motivated by this distinction, we investigate the factors that influence the next-token prediction to change the topic of the input sequence. We define concepts of topic continuity, ambiguous sequences, and change of topic, based on defining a topic as a set of token priority graphs (TPGs). Using a simplified single-layer self-attention architecture, we derive analytical characterizations of topic changes. Specifically, we demonstrate that (1) the model maintains the priority order of tokens related to the input topic, (2) a topic change can occur only if lower-priority tokens outnumber all higher-priority tokens of the input topic, and (3) unlike human cognition, longer context lengths and overlapping topics reduce the likelihood of spontaneous redirection. These insights highlight differences between human cognition and self-attention-based models in navigating topic changes and underscore the challenges in designing conversational AI capable of handling "spontaneous" conversations more naturally. To the best of our knowledge, no prior work has explored these questions with a focus as closely aligned to human conversation and thought.
Authors: Giorgio Robino
Abstract: This study introduces Conversation Routines (CR), a structured prompt engineering framework for developing task-oriented dialog systems using Large Language Models (LLMs). While LLMs demonstrate remarkable natural language understanding capabilities, engineering them to reliably execute complex business workflows remains challenging. The proposed CR framework enables the development of Conversation Agentic Systems (CAS) through natural language specifications, embedding task-oriented logic within LLM prompts. This approach provides a systematic methodology for designing and implementing complex conversational workflows while maintaining behavioral consistency. We demonstrate the framework's effectiveness through two proof-of-concept implementations: a Train Ticket Booking System and an Interactive Troubleshooting Copilot. These case studies validate CR's capability to encode sophisticated behavioral patterns and decision logic while preserving natural conversational flexibility. Results show that CR enables domain experts to design conversational workflows in natural language while leveraging custom enterprise functionalities (tools) developed by software engineers, creating an efficient division of responsibilities where developers focus on core API implementation and domain experts handle conversation design. While the framework shows promise in accessibility and adaptability, we identify key challenges including computational overhead, non-deterministic behavior, and domain-specific logic optimization. Future research directions include CR evaluation methods based on prompt engineering framework driven by goal-oriented grading criteria, improving scalability for complex multi-agent interactions, enhancing system robustness addressing the identified limitations across diverse business applications.
Authors: Zhenhailong Wang, Haiyang Xu, Junyang Wang, Xi Zhang, Ming Yan, Ji Zhang, Fei Huang, Heng Ji
Abstract: Smartphones have become indispensable in modern life, yet navigating complex tasks on mobile devices often remains frustrating. Recent advancements in large multimodal model (LMM)-based mobile agents have demonstrated the ability to perceive and act in mobile environments. However, current approaches face significant limitations: they fall short in addressing real-world human needs, struggle with reasoning-intensive and long-horizon tasks, and lack mechanisms to learn and improve from prior experiences. To overcome these challenges, we introduce Mobile-Agent-E, a hierarchical multi-agent framework capable of self-evolution through past experience. By hierarchical, we mean an explicit separation of high-level planning and low-level action execution. The framework comprises a Manager, responsible for devising overall plans by breaking down complex tasks into subgoals, and four subordinate agents--Perceptor, Operator, Action Reflector, and Notetaker--which handle fine-grained visual perception, immediate action execution, error verification, and information aggregation, respectively. Mobile-Agent-E also features a novel self-evolution module which maintains a persistent long-term memory comprising Tips and Shortcuts. Tips are general guidance and lessons learned from prior tasks on how to effectively interact with the environment. Shortcuts are reusable, executable sequences of atomic operations tailored for specific subroutines. The inclusion of Tips and Shortcuts facilitates continuous refinement in performance and efficiency. Alongside this framework, we introduce Mobile-Eval-E, a new benchmark featuring complex mobile tasks requiring long-horizon, multi-app interactions. Empirical results show that Mobile-Agent-E achieves a 22% absolute improvement over previous state-of-the-art approaches across three foundation model backbones. Project page: https://x-plug.github.io/MobileAgent.
Authors: Junhong Lian, Xiang Ao, Xinyu Liu, Yang Liu, Qing He
Abstract: Personalized news headline generation aims to provide users with attention-grabbing headlines that are tailored to their preferences. Prevailing methods focus on user-oriented content preferences, but most of them overlook the fact that diverse stylistic preferences are integral to users' panoramic interests, leading to suboptimal personalization. In view of this, we propose a novel Stylistic-Content Aware Personalized Headline Generation (SCAPE) framework. SCAPE extracts both content and stylistic features from headlines with the aid of large language model (LLM) collaboration. It further adaptively integrates users' long- and short-term interests through a contrastive learning-based hierarchical fusion network. By incorporating the panoramic interests into the headline generator, SCAPE reflects users' stylistic-content preferences during the generation process. Extensive experiments on the real-world dataset PENS demonstrate the superiority of SCAPE over baselines.
Authors: Sullam Jeoung, Yubin Ge, Haohan Wang, Jana Diesner
Abstract: Examining the alignment of large language models (LLMs) has become increasingly important, particularly when these systems fail to operate as intended. This study explores the challenge of aligning LLMs with human intentions and values, with specific focus on their political inclinations. Previous research has highlighted LLMs' propensity to display political leanings, and their ability to mimic certain political parties' stances on various issues. However, the extent and conditions under which LLMs deviate from empirical positions have not been thoroughly examined. To address this gap, our study systematically investigates the factors contributing to LLMs' deviations from empirical positions on political issues, aiming to quantify these deviations and identify the conditions that cause them. Drawing on cognitive science findings related to representativeness heuristics -- where individuals readily recall the representative attribute of a target group in a way that leads to exaggerated beliefs -- we scrutinize LLM responses through this heuristics lens. We conduct experiments to determine how LLMs exhibit stereotypes by inflating judgments in favor of specific political parties. Our results indicate that while LLMs can mimic certain political parties' positions, they often exaggerate these positions more than human respondents do. Notably, LLMs tend to overemphasize representativeness to a greater extent than humans. This study highlights the susceptibility of LLMs to representativeness heuristics, suggeseting potential vulnerabilities to political stereotypes. We propose prompt-based mitigation strategies that demonstrate effectiveness in reducing the influence of representativeness in LLM responses.
Authors: Sara Abdali, Can Goksen, Saeed Amizadeh, Kazuhito Koishida
Abstract: Investigating NLP through a philosophical lens has recently caught researcher's eyes as it connects computational methods with classical schools of philosophy. This paper introduces a philosophical approach inspired by the Hegelian Dialectic for LLMs' self-reflection, utilizing a self-dialectical approach to emulate internal critiques and then synthesize new ideas by resolving the contradicting points. Moreover, this paper investigates the effect of LLMs' temperature for generation by establishing a dynamic annealing approach, which promotes the creativity in the early stages and gradually refines it by focusing on the nuances, as well as a fixed temperature strategy for generation. Our proposed approach is examined to determine its ability to generate novel ideas from an initial proposition. Additionally, a Multi Agent Majority Voting (MAMV) strategy is leveraged to assess the validity and novelty of the generated ideas, which proves beneficial in the absence of domain experts. Our experiments show promise in generating new ideas and provide a stepping stone for future research.
Authors: Xiaomin Li, Mingye Gao, Zhiwei Zhang, Jingxuan Fan, Weiyu Li
Abstract: Reinforcement Learning from Human Feedback (RLHF) is commonly employed to tailor models to human preferences, especially to improve the safety of outputs from large language models (LLMs). Traditionally, this method depends on selecting preferred responses from pairs. However, due to the variability in human opinions and the challenges in directly comparing two responses, there is an increasing trend towards fine-grained annotation approaches that evaluate responses using multiple targeted metrics or rules. The challenge lies in efficiently choosing and applying these rules to handle the diverse range of preference data. In this paper, we propose a dynamic method that adaptively selects the most important rules for each response pair. We introduce a mathematical framework that utilizes the maximum discrepancy across paired responses and demonstrate theoretically that this approach maximizes the mutual information between the rule-based annotations and the underlying true preferences. We then train an 8B reward model using this adaptively labeled preference dataset and assess its efficacy using RewardBench. As of January 25, 2025, our model achieved the highest safety performance on the leaderboard, surpassing various larger models.
Authors: Sankalp KJ, Ashutosh Kumar, Laxmaan Balaji, Nikunj Kotecha, Vinija Jain, Aman Chadha, Sreyoshi Bhaduri
Abstract: Known by more than 1.5 billion people in the Indian subcontinent, Indic languages present unique challenges and opportunities for natural language processing (NLP) research due to their rich cultural heritage, linguistic diversity, and complex structures. IndicMMLU-Pro is a comprehensive benchmark designed to evaluate Large Language Models (LLMs) across Indic languages, building upon the MMLU Pro (Massive Multitask Language Understanding) framework. Covering major languages such as Hindi, Bengali, Gujarati, Marathi, Kannada, Punjabi, Tamil, Telugu, and Urdu, our benchmark addresses the unique challenges and opportunities presented by the linguistic diversity of the Indian subcontinent. This benchmark encompasses a wide range of tasks in language comprehension, reasoning, and generation, meticulously crafted to capture the intricacies of Indian languages. IndicMMLU-Pro provides a standardized evaluation framework to push the research boundaries in Indic language AI, facilitating the development of more accurate, efficient, and culturally sensitive models. This paper outlines the benchmarks' design principles, task taxonomy, and data collection methodology, and presents baseline results from state-of-the-art multilingual models.
Authors: Go Kamoda, Benjamin Heinzerling, Tatsuro Inaba, Keito Kudo, Keisuke Sakaguchi, Kentaro Inui
Abstract: According to the stages-of-inference hypothesis, early layers of language models map their subword-tokenized input, which does not necessarily correspond to a linguistically meaningful segmentation, to more meaningful representations that form the model's ``inner vocabulary''. Prior analysis of this detokenization stage has predominantly relied on probing and interventions such as path patching, which involve selecting particular inputs, choosing a subset of components that will be patched, and then observing changes in model behavior. Here, we show that several important aspects of the detokenization stage can be understood purely by analyzing model weights, without performing any model inference steps. Specifically, we introduce an analytical decomposition of first-layer attention in GPT-2. Our decomposition yields interpretable terms that quantify the relative contributions of position-related, token-related, and mixed effects. By focusing on terms in this decomposition, we discover weight-based explanations of attention bias toward close tokens and attention for detokenization.
Authors: Priyangshu Mandal, Manasi Patwardhan, Mayur Patidar, Lovekesh Vig
Abstract: Enterprise level data is often distributed across multiple sources and identifying the correct set-of data-sources with relevant information for a knowledge request is a fundamental challenge. In this work, we define the novel task of routing an end-user query to the appropriate data-source, where the data-sources are databases. We synthesize datasets by extending existing datasets designed for NL-to-SQL semantic parsing. We create baselines on these datasets by using open-source LLMs, using both pre-trained and task specific embeddings fine-tuned using the training data. With these baselines we demonstrate that open-source LLMs perform better than embedding based approach, but suffer from token length limitations. Embedding based approaches benefit from task specific fine-tuning, more so when there is availability of data in terms of database specific questions for training. We further find that the task becomes more difficult (i) with an increase in the number of data-sources, (ii) having data-sources closer in terms of their domains,(iii) having databases without external domain knowledge required to interpret its entities and (iv) with ambiguous and complex queries requiring more fine-grained understanding of the data-sources or logical reasoning for routing to an appropriate source. This calls for the need for developing more sophisticated solutions to better address the task.
Authors: Henri Kauhanen, Deepthi Gopal, Tobias Galla, Ricardo Berm\'udez-Otero
Abstract: Quantifying the speed of linguistic change is challenging due to the fact that the historical evolution of languages is sparsely documented. Consequently, traditional methods rely on phylogenetic reconstruction. In this paper, we propose a model-based approach to the problem through the analysis of language change as a stochastic process combining vertical descent, spatial interactions, and mutations in both dimensions. A notion of linguistic temperature emerges naturally from this analysis as a dimensionless measure of the propensity of a linguistic feature to undergo change. We demonstrate how temperatures of linguistic features can be inferred from their present-day geospatial distributions, without recourse to information about their phylogenies. Thus the evolutionary dynamics of language, operating across thousands of years, leaves a measurable geospatial signature. This signature licenses inferences about the historical evolution of languages even in the absence of longitudinal data.
Authors: Wenhao Zheng, Liaoyaqi Wang, Dongshen Peng, Hongxia Xu, Yun Li, Hongtu Zhu, Tianfan Fu, Huaxiu Yao
Abstract: The clinical trial is a pivotal and costly process, often spanning multiple years and requiring substantial financial resources. Therefore, the development of clinical trial outcome prediction models aims to exclude drugs likely to fail and holds the potential for significant cost savings. Recent data-driven attempts leverage deep learning methods to integrate multimodal data for predicting clinical trial outcomes. However, these approaches rely on manually designed modal-specific encoders, which limits both the extensibility to adapt new modalities and the ability to discern similar information patterns across different modalities. To address these issues, we propose a multimodal mixture-of-experts (LIFTED) approach for clinical trial outcome prediction. Specifically, LIFTED unifies different modality data by transforming them into natural language descriptions. Then, LIFTED constructs unified noise-resilient encoders to extract information from modal-specific language descriptions. Subsequently, a sparse Mixture-of-Experts framework is employed to further refine the representations, enabling LIFTED to identify similar information patterns across different modalities and extract more consistent representations from those patterns using the same expert model. Finally, a mixture-of-experts module is further employed to dynamically integrate different modality representations for prediction, which gives LIFTED the ability to automatically weigh different modalities and pay more attention to critical information. The experiments demonstrate that LIFTED significantly enhances performance in predicting clinical trial outcomes across all three phases compared to the best baseline, showcasing the effectiveness of our proposed key components.
Authors: Majid Daliri, Christopher Musco, Ananda Theertha Suresh
Abstract: Suppose Alice has a distribution $P$ and Bob has a distribution $Q$. Alice wants to draw a sample $a\sim P$ and Bob a sample $b \sim Q$ such that $a = b$ with as high of probability as possible. It is well-known that, by sampling from an optimal coupling between the distributions, Alice and Bob can achieve $\Pr[a = b] = 1 - D_{TV}(P,Q)$, where $D_{TV}(P,Q)$ is the total variation distance between $P$ and $Q$. What if Alice and Bob must solve this same problem \emph{without communicating at all?} Perhaps surprisingly, with access to public randomness, they can still achieve $\Pr[a = b] \geq \frac{1 - D_{TV}(P,Q)}{1 + D_{TV}(P,Q)} \geq 1-2D_{TV}(P,Q)$ using a simple protocol based on the Weighted MinHash algorithm. This bound was shown to be optimal in the worst-case by [Bavarian et al., 2020]. In this work, we revisit the communication-free coupling problem. We provide a simpler proof of the optimality result from [Bavarian et al., 2020]. We show that, while the worst-case success probability of Weighted MinHash cannot be improved, an equally simple protocol based on Gumbel sampling offers a Pareto improvement: for every pair of distributions $P, Q$, Gumbel sampling achieves an equal or higher value of $\Pr[a = b]$ than Weighted MinHash. Importantly, this improvement translates to practice. We demonstrate an application of communication-free coupling to \emph{speculative decoding}, a recent method for accelerating autoregressive large language models [Leviathan, Kalman, Matias, ICML 2023]. We show that communication-free protocols can be used to contruct \emph{\CSD{}} schemes, which have the desirable property that their output is fixed given a fixed random seed, regardless of what drafter is used for speculation. In experiments on a language generation task, Gumbel sampling outperforms Weighted MinHash. Code is available at https://github.com/majid-daliri/DISD.
Authors: Guobin Shen, Dongcheng Zhao, Aorigele Bao, Xiang He, Yiting Dong, Yi Zeng
Abstract: Human beings often experience stress, which can significantly influence their performance. This study explores whether Large Language Models (LLMs) exhibit stress responses similar to those of humans and whether their performance fluctuates under different stress-inducing prompts. To investigate this, we developed a novel set of prompts, termed StressPrompt, designed to induce varying levels of stress. These prompts were derived from established psychological frameworks and carefully calibrated based on ratings from human participants. We then applied these prompts to several LLMs to assess their responses across a range of tasks, including instruction-following, complex reasoning, and emotional intelligence. The findings suggest that LLMs, like humans, perform optimally under moderate stress, consistent with the Yerkes-Dodson law. Notably, their performance declines under both low and high-stress conditions. Our analysis further revealed that these StressPrompts significantly alter the internal states of LLMs, leading to changes in their neural representations that mirror human responses to stress. This research provides critical insights into the operational robustness and flexibility of LLMs, demonstrating the importance of designing AI systems capable of maintaining high performance in real-world scenarios where stress is prevalent, such as in customer service, healthcare, and emergency response contexts. Moreover, this study contributes to the broader AI research community by offering a new perspective on how LLMs handle different scenarios and their similarities to human cognition.
Authors: Aiwei Liu, Sheng Guan, Yiming Liu, Leyi Pan, Yifei Zhang, Liancheng Fang, Lijie Wen, Philip S. Yu, Xuming Hu
Abstract: Text watermarking for Large Language Models (LLMs) has made significant progress in detecting LLM outputs and preventing misuse. Current watermarking techniques offer high detectability, minimal impact on text quality, and robustness to text editing. However, current researches lack investigation into the imperceptibility of watermarking techniques in LLM services. This is crucial as LLM providers may not want to disclose the presence of watermarks in real-world scenarios, as it could reduce user willingness to use the service and make watermarks more vulnerable to attacks. This work is the first to investigate the imperceptibility of watermarked LLMs. We design an identification algorithm called Water-Probe that detects watermarks through well-designed prompts to the LLM. Our key motivation is that current watermarked LLMs expose consistent biases under the same watermark key, resulting in similar differences across prompts under different watermark keys. Experiments show that almost all mainstream watermarking algorithms are easily identified with our well-designed prompts, while Water-Probe demonstrates a minimal false positive rate for non-watermarked LLMs. Finally, we propose that the key to enhancing the imperceptibility of watermarked LLMs is to increase the randomness of watermark key selection. Based on this, we introduce the Water-Bag strategy, which significantly improves watermark imperceptibility by merging multiple watermark keys.
Authors: Jiayi Han, Liang Du, Hongwei Du, Xiangguo Zhou, Yiwen Wu, Weibo Zheng, Donghong Han
Abstract: Although many efforts have been made, it is still a challenge to balance the training budget, downstream performance, and the general capabilities of the LLMs in many applications. Training the whole model for downstream tasks is expensive, and could easily result in catastrophic forgetting. By introducing parameter-efficient fine-tuning (PEFT), the training cost could be reduced, but it still suffers from forgetting, and limits the learning on the downstream tasks. To efficiently fine-tune the LLMs with less limitation to their downstream performance while mitigating the forgetting of general capabilities, we propose a novel mixture of expert (MoE) framework based on Soft LoRA and Identity Mixture (SLIM), that allows dynamic routing between LoRA adapters and skipping connection, enables the suppression of forgetting. We adopt weight-yielding with sliding clustering for better out-of-domain distinguish to enhance the routing. We also propose to convert the mixture of low-rank adapters to the model merging formulation and introduce fast dynamic merging of LoRA adapters to keep the general capabilities of the base model. Extensive experiments demonstrate that the proposed SLIM is comparable to the state-of-the-art PEFT approaches on the downstream tasks while achieving the leading performance in mitigating catastrophic forgetting.
Authors: Li Siyan, Vethavikashini Chithrra Raghuram, Omar Khattab, Julia Hirschberg, Zhou Yu
Abstract: Users can divulge sensitive information to proprietary LLM providers, raising significant privacy concerns. While open-source models, hosted locally on the user's machine, alleviate some concerns, models that users can host locally are often less capable than proprietary frontier models. Toward preserving user privacy while retaining the best quality, we propose Privacy-Conscious Delegation, a novel task for chaining API-based and local models. We utilize recent public collections of user-LLM interactions to construct a natural benchmark called PUPA, which contains personally identifiable information (PII). To study potential approaches, we devise PAPILLON, a multi-stage LLM pipeline that uses prompt optimization to address a simpler version of our task. Our best pipeline maintains high response quality for 85.5% of user queries while restricting privacy leakage to only 7.5%. We still leave a large margin to the generation quality of proprietary LLMs for future work. Our data and code will be available at https://github.com/siyan-sylvia-li/PAPILLON.
Authors: Thomas Marshall, Adam Scherlis, Nora Belrose
Abstract: We propose affine concept editing (ACE) as an approach for steering language models' behavior by intervening directly in activations. We begin with an affine decomposition of model activation vectors and show that prior methods for steering model behavior correspond to subsets of terms of this decomposition. We then provide a derivation of ACE and use it to control refusal behavior on ten different models, including Llama 3 70B. ACE combines affine subspace projection and activation addition to reliably control the model's refusal responses across prompt types. We evaluate the results using LLM-based scoring on a collection of harmful and harmless prompts. Our experiments demonstrate that ACE consistently achieves more precise control over model behavior than existing methods and generalizes to models where directional ablation via affine subspace projection alone produces incoherent outputs. Code for reproducing our results is available at https://github.com/EleutherAI/steering-llama3 .
Authors: Fuqiang Liu, Sicong Jiang, Luis Miranda-Moreno, Seongjin Choi, Lijun Sun
Abstract: Large Language Models (LLMs) have recently demonstrated significant potential in the field of time series forecasting, offering impressive capabilities in handling complex temporal data. However, their robustness and reliability in real-world applications remain under-explored, particularly concerning their susceptibility to adversarial attacks. In this paper, we introduce a targeted adversarial attack framework for LLM-based time series forecasting. By employing both gradient-free and black-box optimization methods, we generate minimal yet highly effective perturbations that significantly degrade the forecasting accuracy across multiple datasets and LLM architectures. Our experiments, which include models like TimeGPT and LLM-Time with GPT-3.5, GPT-4, LLaMa, and Mistral, show that adversarial attacks lead to much more severe performance degradation than random noise, and demonstrate the broad effectiveness of our attacks across different LLMs. The results underscore the critical vulnerabilities of LLMs in time series forecasting, highlighting the need for robust defense mechanisms to ensure their reliable deployment in practical applications.
Authors: Vaibhav Mishra, Somaditya Singh, Dhruv Ahlawat, Mohd Zaki, Vaibhav Bihani, Hargun Singh Grover, Biswajit Mishra, Santiago Miret, Mausam, N. M. Anoop Krishnan
Abstract: Materials discovery and development are critical for addressing global challenges. Yet, the exponential growth in materials science literature comprising vast amounts of textual data has created significant bottlenecks in knowledge extraction, synthesis, and scientific reasoning. Large Language Models (LLMs) offer unprecedented opportunities to accelerate materials research through automated analysis and prediction. Still, their effective deployment requires domain-specific adaptation for understanding and solving domain-relevant tasks. Here, we present LLaMat, a family of foundational models for materials science developed through continued pretraining of LLaMA models on an extensive corpus of materials literature and crystallographic data. Through systematic evaluation, we demonstrate that LLaMat excels in materials-specific NLP and structured information extraction while maintaining general linguistic capabilities. The specialized LLaMat-CIF variant demonstrates unprecedented capabilities in crystal structure generation, predicting stable crystals with high coverage across the periodic table. Intriguingly, despite LLaMA-3's superior performance in comparison to LLaMA-2, we observe that LLaMat-2 demonstrates unexpectedly enhanced domain-specific performance across diverse materials science tasks, including structured information extraction from text and tables, more particularly in crystal structure generation, a potential adaptation rigidity in overtrained LLMs. Altogether, the present work demonstrates the effectiveness of domain adaptation towards developing practically deployable LLM copilots for materials research. Beyond materials science, our findings reveal important considerations for domain adaptation of LLMs, such as model selection, training methodology, and domain-specific performance, which may influence the development of specialized scientific AI systems.
Authors: Deven B. Dangi, Beni B. Dangi, Oliver Steinbock
Abstract: When microliter drops of salt solutions dry on non-porous surfaces, they form erratic yet characteristic deposit patterns influenced by complex crystallization dynamics and fluid motion. Using OpenAI's image-enabled language models, we analyzed deposits from 12 salts with 200 images per salt and per model. GPT-4o classified 57% of the salts accurately, significantly outperforming random chance and GPT-4o mini. This study underscores the promise of general-use AI tools for reliably identifying salts from their drying patterns.
Authors: Zeyu Gan, Yun Liao, Yong Liu
Abstract: Test-time scaling, which is also often referred to as slow-thinking, has been demonstrated to enhance multi-step reasoning in large language models (LLMs). However, despite its widespread utilization, the mechanisms underlying slow-thinking methods remain poorly understood. This paper explores the mechanisms of external slow-thinking from a theoretical standpoint. We begin by examining the snowball error effect within the LLM reasoning process and connect it to the likelihood of correct reasoning using information theory. Building on this, we show that external slow-thinking methods can be interpreted as strategies to mitigate the error probability. We further provide a comparative analysis of popular external slow-thinking approaches, ranging from simple to complex, highlighting their differences and interrelationships. Our findings suggest that the efficacy of these methods is not primarily determined by the specific framework employed, and that expanding the search scope or the model's internal reasoning capacity may yield more sustained improvements in the long term. We open-source our code at https://github.com/ZyGan1999/Snowball-Errors-and-Probability.
URLs: https://github.com/ZyGan1999/Snowball-Errors-and-Probability.