Authors: Nikolaos Livathinos, Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Panos Vagenas, Cesar Berrospi Ramis, Matteo Omenetti, Kasper Dinkla, Yusik Kim, Shubham Gupta, Rafael Teixeira de Lima, Valery Weber, Lucas Morin, Ingmar Meijer, Viktor Kuropiatnyk, Peter W. J. Staar
Abstract: We introduce Docling, an easy-to-use, self-contained, MIT-licensed, open-source toolkit for document conversion, that can parse several types of popular document formats into a unified, richly structured representation. It is powered by state-of-the-art specialized AI models for layout analysis (DocLayNet) and table structure recognition (TableFormer), and runs efficiently on commodity hardware in a small resource budget. Docling is released as a Python package and can be used as a Python API or as a CLI tool. Docling's modular architecture and efficient document representation make it easy to implement extensions, new features, models, and customizations. Docling has been already integrated in other popular open-source frameworks (e.g., LangChain, LlamaIndex, spaCy), making it a natural fit for the processing of documents and the development of high-end applications. The open-source community has fully engaged in using, promoting, and developing for Docling, which gathered 10k stars on GitHub in less than a month and was reported as the No. 1 trending repository in GitHub worldwide in November 2024.
Authors: Didier Ch\'etelat, Joseph Cotnareanu, Rylee Thompson, Yingxue Zhang, Mark Coates
Abstract: Large language models (LLMs) contain substantial factual knowledge which is commonly elicited by multiple-choice question-answering prompts. Internally, such models process the prompt through multiple transformer layers, building varying representations of the problem within its hidden states. Ultimately, however, only the hidden state corresponding to the final layer and token position are used to predict the answer label. In this work, we propose instead to learn a small separate neural network predictor module on a collection of training questions, that take the hidden states from all the layers at the last temporal position as input and outputs predictions. In effect, such a framework disentangles the representational abilities of LLMs from their predictive abilities. On a collection of hard benchmarks, our method achieves considerable improvements in performance, sometimes comparable to supervised fine-tuning procedures, but at a fraction of the computational cost.
Authors: Yibo Wang, Tiansheng Huang, Li Shen, Huanjin Yao, Haotian Luo, Rui Liu, Naiqiang Tan, Jiaxing Huang, Dacheng Tao
Abstract: Harmful fine-tuning attack introduces significant security risks to the fine-tuning services. Mainstream defenses aim to vaccinate the model such that the later harmful fine-tuning attack is less effective. However, our evaluation results show that such defenses are fragile -- with a few fine-tuning steps, the model still can learn the harmful knowledge. To this end, we do further experiment and find that an embarrassingly simple solution -- adding purely random perturbations to the fine-tuned model, can recover the model from harmful behavior, though it leads to a degradation in the model's fine-tuning performance. To address the degradation of fine-tuning performance, we further propose Panacea, which optimizes an adaptive perturbation that will be applied to the model after fine-tuning. Panacea maintains model's safety alignment performance without compromising downstream fine-tuning performance. Comprehensive experiments are conducted on different harmful ratios, fine-tuning tasks and mainstream LLMs, where the average harmful scores are reduced by up-to 21.5%, while maintaining fine-tuning performance. As a by-product, we analyze the optimized perturbation and show that different layers in various LLMs have distinct safety coefficients. Source code available at https://github.com/w-yibo/Panacea
Authors: Jack Lanchantin, Angelica Chen, Shehzaad Dhuliawala, Ping Yu, Jason Weston, Sainbayar Sukhbaatar, Ilia Kulikov
Abstract: Post-training of language models, either through reinforcement learning, preference optimization or supervised finetuning, tends to sharpen the output probability distribution and reduce the diversity of generated responses. This is particularly a problem for creative generative tasks where varied responses are desired. %This impacts the ability to generate high quality synthetic data which is becoming a vital component of model training. In this work we introduce Diverse Preference Optimization (DivPO), an online optimization method which learns to generate much more diverse responses than standard pipelines, while maintaining the quality of the generations. In DivPO, preference pairs are selected by first considering a pool of responses, and a measure of diversity among them, and selecting chosen examples as being more rare but high quality, while rejected examples are more common, but low quality. DivPO results in generating 45.6% more diverse persona attributes, and an 74.6% increase in story diversity, while maintaining similar win rates as standard baselines.
Authors: Qika Lin, Tianzhe Zhao, Kai He, Zhen Peng, Fangzhi Xu, Ling Huang, Jingying Ma, Mengling Feng
Abstract: Due to the presence of the natural gap between Knowledge Graph (KG) structures and the natural language, the effective integration of holistic structural information of KGs with Large Language Models (LLMs) has emerged as a significant question. To this end, we propose a two-stage framework to learn and apply quantized codes for each entity, aiming for the seamless integration of KGs with LLMs. Firstly, a self-supervised quantized representation (SSQR) method is proposed to compress both KG structural and semantic knowledge into discrete codes (\ie, tokens) that align the format of language sentences. We further design KG instruction-following data by viewing these learned codes as features to directly input to LLMs, thereby achieving seamless integration. The experiment results demonstrate that SSQR outperforms existing unsupervised quantized methods, producing more distinguishable codes. Further, the fine-tuned LLaMA2 and LLaMA3.1 also have superior performance on KG link prediction and triple classification tasks, utilizing only 16 tokens per entity instead of thousands in conventional prompting methods.
Authors: Abdurrahman Odaba\c{s}{\i}, G\"oksel Biricik
Abstract: Given the recent introduction of multiple language models and the ongoing demand for improved Natural Language Processing tasks, particularly summarization, this work provides a comprehensive benchmarking of 20 recent language models, focusing on smaller ones for the news summarization task. In this work, we systematically test the capabilities and effectiveness of these models in summarizing news article texts which are written in different styles and presented in three distinct datasets. Specifically, we focus in this study on zero-shot and few-shot learning settings and we apply a robust evaluation methodology that combines different evaluation concepts including automatic metrics, human evaluation, and LLM-as-a-judge. Interestingly, including demonstration examples in the few-shot learning setting did not enhance models' performance and, in some cases, even led to worse quality of the generated summaries. This issue arises mainly due to the poor quality of the gold summaries that have been used as reference summaries, which negatively impacts the models' performance. Furthermore, our study's results highlight the exceptional performance of GPT-3.5-Turbo and GPT-4, which generally dominate due to their advanced capabilities. However, among the public models evaluated, certain models such as Qwen1.5-7B, SOLAR-10.7B-Instruct-v1.0, Meta-Llama-3-8B and Zephyr-7B-Beta demonstrated promising results. These models showed significant potential, positioning them as competitive alternatives to large models for the task of news summarization.
Authors: Wanlong Liu, Yichen Xiao, Dingyi Zeng, Hongyang Zhao, Wenyu Chen, Malu Zhang
Abstract: Post-Training Quantization (PTQ) is pivotal for deploying large language models (LLMs) within resource-limited settings by significantly reducing resource demands. However, existing PTQ strategies underperform at low bit levels < 3 bits due to the significant difference between the quantized and original weights. To enhance the quantization performance at low bit widths, we introduce a Mixed-precision Graph Neural PTQ (MG-PTQ) approach, employing a graph neural network (GNN) module to capture dependencies among weights and adaptively assign quantization bit-widths. Through the information propagation of the GNN module, our method more effectively captures dependencies among target weights, leading to a more accurate assessment of weight importance and optimized allocation of quantization strategies. Extensive experiments on the WikiText2 and C4 datasets demonstrate that our MG-PTQ method outperforms previous state-of-the-art PTQ method GPTQ, setting new benchmarks for quantization performance under low-bit conditions.
Authors: James Blades, Frederick Somerfield, William Langley, Susan Everingham, Maurice Witherington
Abstract: Token representation strategies within large-scale neural architectures often rely on contextually refined embeddings, yet conventional approaches seldom encode structured relationships explicitly within token interactions. Self-attention mechanisms effectively capture dynamic contextual dependencies, but their reliance on learned weight distributions limits the preservation of long-range hierarchical structures in generated sequences. Dependency-aware token encoding introduces a structured approach to embedding initialization, ensuring that relational constraints are embedded within token representations rather than inferred solely through attention dynamics. The proposed encoding mechanism refines token interactions through dependency-weighted attention computations, ensuring that syntactic and semantic dependencies are retained across multiple processing layers. Empirical evaluations indicate reductions in perplexity across diverse linguistic benchmarks, suggesting improvements in contextual coherence and predictive consistency in autoregressive text generation. Computational efficiency assessments reveal a moderate increase in memory consumption and training time, attributed to additional matrix computations within the encoding module, yet scalability remains feasible within conventional transformer architectures. Structured encoding enhances lexical variation and dependency retention, reinforcing linguistic coherence without requiring external syntactic annotations or auxiliary training objectives. Statistical comparisons highlight improvements in dependency alignment, particularly in longer sequences where conventional self-attention models exhibit degradation in hierarchical consistency. Sentence length distributions indicate a reduction in abrupt phrase transitions, further supporting the hypothesis that explicit dependency encoding facilitates more structured phrase generation.
Authors: Vil\'em Zouhar, Peng Cui, Mrinmaya Sachan
Abstract: Human evaluation is the gold-standard for evaluating text generation models. It is also expensive, and to fit budgetary constraints, a random subset of the test data is often chosen in practice. The randomly selected data may not accurately represent test performance, making this approach economically inefficient for model comparison. Thus, in this work, we develop a suite of selectors to get the most informative datapoints for human evaluation while taking the evaluation costs into account. We show that selectors based on variance in automated metric scores, diversity in model outputs, or Item Response Theory outperform random selection. We further develop an approach to distill these selectors to the scenario where the model outputs are not yet available. In particular, we introduce source-based estimators, which predict item usefulness for human evaluation just based on the source texts. We demonstrate the efficacy of our selectors in two common NLG tasks, machine translation and summarization, and show that up to only ~50% of the test data is needed to produce the same evaluation result as the entire data. Our implementations are published in the subset2evaluate package.
Authors: Haoyu Liang, Youran Sun, Yunfeng Cai, Jun Zhu, Bo Zhang
Abstract: The security issue of large language models (LLMs) has gained significant attention recently, with various defense mechanisms developed to prevent harmful outputs, among which safeguards based on text embedding models serve as a fundamental defense. Through testing, we discover that the distribution of text embedding model outputs is significantly biased with a large mean. Inspired by this observation, we propose novel efficient methods to search for universal magic words that can attack text embedding models. The universal magic words as suffixes can move the embedding of any text towards the bias direction, therefore manipulate the similarity of any text pair and mislead safeguards. By appending magic words to user prompts and requiring LLMs to end answers with magic words, attackers can jailbreak the safeguard. To eradicate this security risk, we also propose defense mechanisms against such attacks, which can correct the biased distribution of text embeddings in a train-free manner.
Authors: Jennifer D'Souza, Zachary Laubach, Tarek Al Mustafa, Sina Zarrie{\ss}, Robert Fr\"uhst\"uckl, Phyllis Illari
Abstract: This paper presents an exploratory study that harnesses the capabilities of large language models (LLMs) to mine key ecological entities from invasion biology literature. Specifically, we focus on extracting species names, their locations, associated habitats, and ecosystems, information that is critical for understanding species spread, predicting future invasions, and informing conservation efforts. Traditional text mining approaches often struggle with the complexity of ecological terminology and the subtle linguistic patterns found in these texts. By applying general-purpose LLMs without domain-specific fine-tuning, we uncover both the promise and limitations of using these models for ecological entity extraction. In doing so, this study lays the groundwork for more advanced, automated knowledge extraction tools that can aid researchers and practitioners in understanding and managing biological invasions.
Authors: Yiteng Tu, Weihang Su, Yujia Zhou, Yiqun Liu, Qingyao Ai
Abstract: Retrieval-augmented generation (RAG) enhances large language models (LLMs) by integrating external knowledge retrieved from a knowledge base. However, its effectiveness is fundamentally constrained by the reliability of both the retriever and the knowledge base. In real-world scenarios, imperfections in these components often lead to the retrieval of noisy, irrelevant, or misleading counterfactual information, ultimately undermining the trustworthiness of RAG systems. To address this challenge, we propose Robust Fine-Tuning (RbFT), a method designed to enhance the resilience of LLMs against retrieval defects through two targeted fine-tuning tasks. Experimental results demonstrate that RbFT significantly improves the robustness of RAG systems across diverse retrieval conditions, surpassing existing methods while maintaining high inference efficiency and compatibility with other robustness techniques.
Authors: Huaiyuan Ying, Hongyi Yuan, Jinsen Lu, Zitian Qu, Yang Zhao, Zhengyun Zhao, Isaac Kohane, Tianxi Cai, Sheng Yu
Abstract: Electronic Health Records (EHRs) hold immense potential for advancing healthcare, offering rich, longitudinal data that combines structured information with valuable insights from unstructured clinical notes. However, the unstructured nature of clinical text poses significant challenges for secondary applications. Traditional methods for structuring EHR free-text data, such as rule-based systems and multi-stage pipelines, are often limited by their time-consuming configurations and inability to adapt across clinical notes from diverse healthcare settings. Few systems provide a comprehensive attribute extraction for terminologies. While giant large language models (LLMs) like GPT-4 and LLaMA 405B excel at structuring tasks, they are slow, costly, and impractical for large-scale use. To overcome these limitations, we introduce GENIE, a Generative Note Information Extraction system that leverages LLMs to streamline the structuring of unstructured clinical text into usable data with standardized format. GENIE processes entire paragraphs in a single pass, extracting entities, assertion statuses, locations, modifiers, values, and purposes with high accuracy. Its unified, end-to-end approach simplifies workflows, reduces errors, and eliminates the need for extensive manual intervention. Using a robust data preparation pipeline and fine-tuned small scale LLMs, GENIE achieves competitive performance across multiple information extraction tasks, outperforming traditional tools like cTAKES and MetaMap and can handle extra attributes to be extracted. GENIE strongly enhances real-world applicability and scalability in healthcare systems. By open-sourcing the model and test data, we aim to encourage collaboration and drive further advancements in EHR structurization.
Authors: Yumeng Wang, Zhiyuan Fan, Qingyun Wang, May Fung, Heng Ji
Abstract: Large Language Models (LLMs) are pretrained on extensive multilingual corpora to acquire both language-specific cultural knowledge and general knowledge. Ideally, while LLMs should provide consistent responses to culture-independent questions across languages, we observe significant performance disparities. To address this, we explore the Cross-Lingual Self-Aligning ability of Language Models (CALM) to align knowledge across languages. Specifically, for a given question, we sample multiple responses across different languages, and select the most self-consistent response as the target, leaving the remaining responses as negative examples. We then employ direct preference optimization (DPO) to align the model's knowledge across different languages. Evaluations on the MEDQA and X-CSQA datasets demonstrate CALM's effectiveness in enhancing cross-lingual knowledge question answering, both in zero-shot and retrieval augmented settings. We also found that increasing the number of languages involved in CALM training leads to even higher accuracy and consistency. We offer a qualitative analysis of how cross-lingual consistency can enhance knowledge alignment and explore the method's generalizability. The source code and data of this paper are available on GitHub.
Authors: Arthur Douillard, Yanislav Donchev, Keith Rush, Satyen Kale, Zachary Charles, Zachary Garrett, Gabriel Teston, Dave Lacey, Ross McIlroy, Jiajun Shen, Alexandre Ram\'e, Arthur Szlam, Marc'Aurelio Ranzato, Paul Barham
Abstract: Training of large language models (LLMs) is typically distributed across a large number of accelerators to reduce training time. Since internal states and parameter gradients need to be exchanged at each and every single gradient step, all devices need to be co-located using low-latency high-bandwidth communication links to support the required high volume of exchanged bits. Recently, distributed algorithms like DiLoCo have relaxed such co-location constraint: accelerators can be grouped into ``workers'', where synchronizations between workers only occur infrequently. This in turn means that workers can afford being connected by lower bandwidth communication links without affecting learning quality. However, in these methods, communication across workers still requires the same peak bandwidth as before, as the synchronizations require all parameters to be exchanged across all workers. In this paper, we improve DiLoCo in three ways. First, we synchronize only subsets of parameters in sequence, rather than all at once, which greatly reduces peak bandwidth. Second, we allow workers to continue training while synchronizing, which decreases wall clock time. Third, we quantize the data exchanged by workers, which further reduces bandwidth across workers. By properly combining these modifications, we show experimentally that we can distribute training of billion-scale parameters and reach similar quality as before, but reducing required bandwidth by two orders of magnitude.
Authors: Anmol Goel, Yaxi Hu, Iryna Gurevych, Amartya Sanyal
Abstract: Aligning Large Language Models (LLMs) with human values and away from undesirable behaviors (such as hallucination) has become increasingly important. Recently, steering LLMs towards a desired behavior via activation editing has emerged as an effective method to mitigate harmful generations at inference-time. Activation editing modifies LLM representations by preserving information from positive demonstrations (e.g., truthful) and minimising information from negative demonstrations (e.g., hallucinations). When these demonstrations come from a private dataset, the aligned LLM may leak private information contained in those private samples. In this work, we present the first study of aligning LLM behavior with private datasets. Our work proposes the \textit{\underline{P}rivate \underline{S}teering for LLM \underline{A}lignment (PSA)} algorithm to edit LLM activations with differential privacy (DP) guarantees. We conduct extensive experiments on seven different benchmarks with open-source LLMs of different sizes (0.5B to 7B) and model families (LlaMa, Qwen, Mistral and Gemma). Our results show that PSA achieves DP guarantees for LLM alignment with minimal loss in performance, including alignment metrics, open-ended text generation quality, and general-purpose reasoning. We also develop the first Membership Inference Attack (MIA) for evaluating and auditing the empirical privacy for the problem of LLM steering via activation editing. Our attack is tailored for activation editing and relies solely on the generated texts without their associated probabilities. Our experiments support the theoretical guarantees by showing improved guarantees for our \textit{PSA} algorithm compared to several existing non-private techniques.
Authors: Peter Baile Chen, Yi Zhang, Michael Cafarella, Dan Roth
Abstract: Real-world open-domain questions can be complicated, particularly when answering them involves information from multiple information sources. LLMs have demonstrated impressive performance in decomposing complex tasks into simpler steps, and previous work has used it for better retrieval in support of complex questions. However, LLM's decomposition of questions is unaware of what data is available and how data is organized, often leading to a sub-optimal retrieval performance. Recent effort in agentic RAG proposes to perform retrieval in an iterative fashion, where a followup query is derived as an action based on previous rounds of retrieval. While this provides one way of interacting with the data collection, agentic RAG's exploration of data is inefficient because successive queries depend on previous results rather than being guided by the organization of available data in the collection. To address this problem, we propose an LLM-based retrieval method -- ARM, that aims to better align the question with the organization of the data collection by exploring relationships among data objects beyond matching the utterance of the query, thus leading to a retrieve-all-at-once solution for complex queries. We evaluated ARM on two datasets, Bird and OTT-QA. On Bird, it outperforms standard RAG with query decomposition by up to 5.2 pt in execution accuracy and agentic RAG (ReAct) by up to 15.9 pt. On OTT-QA, it achieves up to 5.5 pt and 19.3 pt higher F1 match scores compared to these approaches.
Authors: Ping Yu, Weizhe Yuan, Olga Golovneva, Tianhao Wu, Sainbayar Sukhbaatar, Jason Weston, Jing Xu
Abstract: Training data quality is one of the most important drivers of final model quality. In this work, we introduce a method for evaluating data integrity based on the assumption that low-quality input prompts result in high variance and low quality responses. This is achieved by measuring the rejected response quality and the reward gap between the chosen and rejected preference pair. Our method, Rejecting Instruction Preferences (RIP) can be used to filter prompts from existing training sets, or to make high quality synthetic datasets, yielding large performance gains across various benchmarks compared to unfiltered data. Using Llama 3.1-8B-Instruct, RIP improves AlpacaEval2 LC Win Rate by 9.4%, Arena-Hard by 8.7%, and WildBench by 9.9%. Using Llama 3.3-70B-Instruct, RIP improves Arena-Hard from 67.5 to 82.9, which is from 18th place to 6th overall in the leaderboard.
Authors: Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, Dong Yu
Abstract: Large language models (LLMs) such as OpenAI's o1 have demonstrated remarkable abilities in complex reasoning tasks by scaling test-time compute and exhibiting human-like deep thinking. However, we identify a phenomenon we term underthinking, where o1-like LLMs frequently switch between different reasoning thoughts without sufficiently exploring promising paths to reach a correct solution. This behavior leads to inadequate depth of reasoning and decreased performance, particularly on challenging mathematical problems. To systematically analyze this issue, we conduct experiments on three challenging test sets and two representative open-source o1-like models, revealing that frequent thought switching correlates with incorrect responses. We introduce a novel metric to quantify underthinking by measuring token efficiency in incorrect answers. To address underthinking, we propose a decoding strategy with thought switching penalty TIP that discourages premature transitions between thoughts, encouraging deeper exploration of each reasoning path. Experimental results demonstrate that our approach improves accuracy across challenging datasets without requiring model fine-tuning. Our findings contribute to understanding reasoning inefficiencies in o1-like LLMs and offer a practical solution to enhance their problem-solving capabilities.
Authors: Mingkuan Feng, Jinyang Wu, Shuai Zhang, Pengpeng Shao, Ruihan Jin, Zhengqi Wen, Jianhua Tao, Feihu Che
Abstract: Large language models (LLMs) have achieved significant progress across various domains, but their increasing scale results in high computational and memory costs. Recent studies have revealed that LLMs exhibit sparsity, providing the potential to reduce model size through pruning techniques. However, existing pruning methods typically follow a prune-then-finetune paradigm. Since the pruned components still contain valuable information, their direct removal often leads to irreversible performance degradation, imposing a substantial computational burden to recover performance during finetuning. In this paper, we propose a novel paradigm that first applies regularization, then prunes, and finally finetunes. Based on this paradigm, we introduce DReSS, a simple and effective Data-driven Regularized Structured Streamlining method for LLMs. By leveraging a small amount of data to regularize the components to be pruned, DReSS explicitly transfers the important information to the remaining parts of the model in advance. Compared to direct pruning, this can reduce the information loss caused by parameter removal, thereby enhancing its language modeling capabilities. Experimental results demonstrate that DReSS significantly outperforms existing pruning methods even under extreme pruning ratios, significantly reducing latency and increasing throughput.
Authors: Myra Cheng, Angela Y. Lee, Kristina Rapuano, Kate Niederhoffer, Alex Liebscher, Jeffrey Hancock
Abstract: How has the public responded to the increasing prevalence of artificial intelligence (AI)-based technologies? We investigate public perceptions of AI by collecting over 12,000 responses over 12 months from a nationally representative U.S. sample. Participants provided open-ended metaphors reflecting their mental models of AI, a methodology that overcomes the limitations of traditional self-reported measures. Using a mixed-methods approach combining quantitative clustering and qualitative coding, we identify 20 dominant metaphors shaping public understanding of AI. To analyze these metaphors systematically, we present a scalable framework integrating language modeling (LM)-based techniques to measure key dimensions of public perception: anthropomorphism (attribution of human-like qualities), warmth, and competence. We find that Americans generally view AI as warm and competent, and that over the past year, perceptions of AI's human-likeness and warmth have significantly increased ($+34\%, r = 0.80, p < 0.01; +41\%, r = 0.62, p < 0.05$). Furthermore, these implicit perceptions, along with the identified dominant metaphors, strongly predict trust in and willingness to adopt AI ($r^2 = 0.21, 0.18, p < 0.001$). We further explore how differences in metaphors and implicit perceptions--such as the higher propensity of women, older individuals, and people of color to anthropomorphize AI--shed light on demographic disparities in trust and adoption. In addition to our dataset and framework for tracking evolving public attitudes, we provide actionable insights on using metaphors for inclusive and responsible AI development.
Authors: Spencer Mateega, Carlos Georgescu, Danny Tang
Abstract: FinanceQA is a testing suite that evaluates LLMs' performance on complex numerical financial analysis tasks that mirror real-world investment work. Despite recent advances, current LLMs fail to meet the strict accuracy requirements of financial institutions, with models failing approximately 60% of realistic tasks that mimic on-the-job analyses at hedge funds, private equity firms, investment banks, and other financial institutions. The primary challenges include hand-spreading metrics, adhering to standard accounting and corporate valuation conventions, and performing analysis under incomplete information - particularly in multi-step tasks requiring assumption generation. This performance gap highlights the disconnect between existing LLM capabilities and the demands of professional financial analysis that are inadequately tested by current testing architectures. Results show that higher-quality training data is needed to support such tasks, which we experiment with using OpenAI's fine-tuning API. FinanceQA is publicly released at [this https URL](https://huggingface.co/datasets/AfterQuery/FinanceQA).
URLs: https://huggingface.co/datasets/AfterQuery/FinanceQA).
Authors: Kumar Ashutosh, Yossi Gandelsman, Xinlei Chen, Ishan Misra, Rohit Girdhar
Abstract: We present MILS: Multimodal Iterative LLM Solver, a surprisingly simple, training-free approach, to imbue multimodal capabilities into your favorite LLM. Leveraging their innate ability to perform multi-step reasoning, MILS prompts the LLM to generate candidate outputs, each of which are scored and fed back iteratively, eventually generating a solution to the task. This enables various applications that typically require training specialized models on task-specific data. In particular, we establish a new state-of-the-art on emergent zero-shot image, video and audio captioning. MILS seamlessly applies to media generation as well, discovering prompt rewrites to improve text-to-image generation, and even edit prompts for style transfer! Finally, being a gradient-free optimization approach, MILS can invert multimodal embeddings into text, enabling applications like cross-modal arithmetic.
Authors: Swarnadeep Saha, Xian Li, Marjan Ghazvininejad, Jason Weston, Tianlu Wang
Abstract: LLM-as-a-Judge models generate chain-of-thought (CoT) sequences intended to capture the step-bystep reasoning process that underlies the final evaluation of a response. However, due to the lack of human annotated CoTs for evaluation, the required components and structure of effective reasoning traces remain understudied. Consequently, previous approaches often (1) constrain reasoning traces to hand-designed components, such as a list of criteria, reference answers, or verification questions and (2) structure them such that planning is intertwined with the reasoning for evaluation. In this work, we propose EvalPlanner, a preference optimization algorithm for Thinking-LLM-as-a-Judge that first generates an unconstrained evaluation plan, followed by its execution, and then the final judgment. In a self-training loop, EvalPlanner iteratively optimizes over synthetically constructed evaluation plans and executions, leading to better final verdicts. Our method achieves a new state-of-the-art performance for generative reward models on RewardBench (with a score of 93.9), despite being trained on fewer amount of, and synthetically generated, preference pairs. Additional experiments on other benchmarks like RM-Bench, JudgeBench, and FollowBenchEval further highlight the utility of both planning and reasoning for building robust LLM-as-a-Judge reasoning models.
Authors: JiWoo Kim, Minsuk Chang, JinYeong Bak
Abstract: Traditional text-based human-AI interactions often adhere to a strict turn-taking approach. In this research, we propose a novel approach that incorporates overlapping messages, mirroring natural human conversations. Through a formative study, we observed that even in text-based contexts, users instinctively engage in overlapping behaviors like "A: Today I went to-" "B: yeah." To capitalize on these insights, we developed OverlapBot, a prototype chatbot where both AI and users can initiate overlapping. Our user study revealed that OverlapBot was perceived as more communicative and immersive than traditional turn-taking chatbot, fostering faster and more natural interactions. Our findings contribute to the understanding of design space for overlapping interactions. We also provide recommendations for implementing overlap-capable AI interactions to enhance the fluidity and engagement of text-based conversations.
Authors: Song Bian, Minghao Yan, Shivaram Venkataraman
Abstract: Scaling laws are powerful tools to predict the performance of large language models. However, current scaling laws fall short of accounting for inference costs. In this work, we first show that model architecture affects inference latency, where models of the same size can have up to 3.5x difference in latency. To tackle this challenge, we modify the Chinchilla scaling laws to co-optimize the model parameter count, the number of training tokens, and the model architecture. Due to the reason that models of similar training loss exhibit gaps in downstream evaluation, we also propose a novel method to train inference-efficient models based on the revised scaling laws. We perform extensive empirical studies to fit and evaluate our inference-aware scaling laws. We vary model parameters from 80M to 1B, training tokens from 1.6B to 30B, and model shapes, training a total of 63 models. Guided by our inference-efficient scaling law and model selection method, we release the Morph-1B model, which improves inference latency by 1.8x while maintaining accuracy on downstream tasks compared to open-source models, pushing the Pareto frontier of accuracy-latency tradeoff.
Authors: Samuel Ackerman, Eitan Farchi, Orna Raz, Assaf Toledo
Abstract: The evaluation of generative or discriminative large language model (LLM)-based systems is often a complex multi-dimensional problem. Typically, a set of system configuration alternatives are evaluated on one or more benchmark datasets, each with one or more evaluation metrics, which may differ between datasets. We often want to evaluate -- with a statistical measure of significance -- whether systems perform differently either on a given dataset according to a single metric, on aggregate across metrics on a dataset, or across datasets. Such evaluations can be done to support decision-making, such as deciding whether a particular system component change (e.g., choice of LLM or hyperparameter values) significantly improves performance over the current system configuration, or, more generally, whether a fixed set of system configurations (e.g., a leaderboard list) have significantly different performances according to metrics of interest. We present a framework implementation that automatically performs the correct statistical tests, properly aggregates the statistical results across metrics and datasets (a nontrivial task), and can visualize the results. The framework is demonstrated on the multi-lingual code generation benchmark CrossCodeEval, for several state-of-the-art LLMs.
Authors: Kevin Roitero, Dustin Wright, Michael Soprano, Isabelle Augenstein, Stefano Mizzaro
Abstract: With the degradation of guardrails against mis- and disinformation online, it is more critical than ever to be able to effectively combat it. In this paper, we explore the efficiency and effectiveness of using crowd-sourced truthfulness assessments based on condensed, large language model (LLM) generated summaries of online sources. We compare the use of generated summaries to the use of original web pages in an A/B testing setting, where we employ a large and diverse pool of crowd-workers to perform the truthfulness assessment. We evaluate the quality of assessments, the efficiency with which assessments are performed, and the behavior and engagement of participants. Our results demonstrate that the Summary modality, which relies on summarized evidence, offers no significant change in assessment accuracy over the Standard modality, while significantly increasing the speed with which assessments are performed. Workers using summarized evidence produce a significantly higher number of assessments in the same time frame, reducing the cost needed to acquire truthfulness assessments. Additionally, the Summary modality maximizes both the inter-annotator agreements as well as the reliance on and perceived usefulness of evidence, demonstrating the utility of summarized evidence without sacrificing the quality of assessments.
Authors: Shutian Ma, Chengzhi Zhang, Heng Zhang, Zheng Gao
Abstract: Citation recommendation aims to locate the important papers for scholars to cite. When writing the citing sentences, the authors usually hold different citing intents, which are referred to citation function in citation analysis. Since argumentative zoning is to identify the argumentative and rhetorical structure in scientific literature, we want to use this information to improve the citation recommendation task. In this paper, a multi-task learning model is built for citation recommendation and argumentative zoning classification. We also generated an annotated corpus of the data from PubMed Central based on a new argumentative zoning schema. The experimental results show that, by considering the argumentative information in the citing sentence, citation recommendation model will get better performance.
Authors: Wiradee Imrattanatrai, Masaki Asada, Kimihiro Hasegawa, Zhi-Qi Cheng, Ken Fukuda, Teruko Mitamura
Abstract: This paper presents VDAct, a dataset for a Video-grounded Dialogue on Event-driven Activities, alongside VDEval, a session-based context evaluation metric specially designed for the task. Unlike existing datasets, VDAct includes longer and more complex video sequences that depict a variety of event-driven activities that require advanced contextual understanding for accurate response generation. The dataset comprises 3,000 dialogues with over 30,000 question-and-answer pairs, derived from 1,000 videos with diverse activity scenarios. VDAct displays a notably challenging characteristic due to its broad spectrum of activity scenarios and wide range of question types. Empirical studies on state-of-the-art vision foundation models highlight their limitations in addressing certain question types on our dataset. Furthermore, VDEval, which integrates dialogue session history and video content summaries extracted from our supplementary Knowledge Graphs to evaluate individual responses, demonstrates a significantly higher correlation with human assessments on the VDAct dataset than existing evaluation metrics that rely solely on the context of single dialogue turns.
Authors: Thea Aviss
Abstract: We introduce the State Stream Transformer (SST), a novel LLM architecture that reveals emergent reasoning behaviours and capabilities latent in pretrained weights through addressing a fundamental limitation in traditional transformer models: the lack of latent computational continuity across autoregressive generations in the state space. SST introduces a sliding window latent state (FFN) cache with weighted decay that maintains and evolves persistent latent processes throughout autoregressive generations. Through controlled experiments comparing base and SST architectures using the same frozen weights, we demonstrate that this architectural modification alone enables enhanced reasoning capabilities which appear best explained by some form of potential higher-order processing, as evidenced by emergent metacognitive behaviours. These behaviours persist under controlled conditions designed to eliminate confounding factors such as stochastic variation or learned response patterns. Analysis of latent state distributions and processing dynamics provides evidence that it is solely the 'state stream' that is responsible for these phenomena. In quantitative evaluations, the SST achieves substantial performance improvements over the base model on two reasoning benchmarks, reaching 89.01\% accuracy on GSM-8K (0-shot) and 91.04\% on ARC Challenge (0-shot CoT). These findings indicate that persistent computation in the latent state space enables fundamentally different information processing and internal reasoning strategies, with implications for our understanding of artificial intelligence systems.
Authors: Yuxin Zuo, Shang Qu, Yifei Li, Zhangren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, Bowen Zhou
Abstract: We introduce MedXpertQA, a highly challenging and comprehensive benchmark to evaluate expert-level medical knowledge and advanced reasoning. MedXpertQA includes 4,460 questions spanning 17 specialties and 11 body systems. It includes two subsets, Text for text evaluation and MM for multimodal evaluation. Notably, MM introduces expert-level exam questions with diverse images and rich clinical information, including patient records and examination results, setting it apart from traditional medical multimodal benchmarks with simple QA pairs generated from image captions. MedXpertQA applies rigorous filtering and augmentation to address the insufficient difficulty of existing benchmarks like MedQA, and incorporates specialty board questions to improve clinical relevance and comprehensiveness. We perform data synthesis to mitigate data leakage risk and conduct multiple rounds of expert reviews to ensure accuracy and reliability. We evaluate 16 leading models on MedXpertQA. Moreover, medicine is deeply connected to real-world decision-making, providing a rich and representative setting for assessing reasoning abilities beyond mathematics and code. To this end, we develop a reasoning-oriented subset to facilitate the assessment of o1-like models.
Authors: Benjamin Feuer, Chinmay Hegde
Abstract: Language model (LLM) post-training, from DPO to distillation, can refine behaviors and unlock new skills, but the open science supporting these post-training techniques is still in its infancy. One limiting factor has been the difficulty of conducting large-scale comparative analyses of synthetic data generating models and LLM judges. To close this gap, we introduce WILDCHAT-50M, the largest public chat dataset to date. We extend the existing WildChat dataset to include responses not only from GPT, but from over 50 different open-weight models, ranging in size from 0.5B to 104B parameters. We conduct an extensive comparative analysis and demonstrate the potential of this dataset by creating RE-WILD, our own public SFT mix, which outperforms the recent Tulu-3 SFT mixture from Allen AI with only 40% as many samples. Our dataset, samples and code are available at https://github.com/penfever/wildchat-50m.
Authors: Yi Ding, Lijun Li, Bing Cao, Jing Shao
Abstract: Large Vision-Language Models (VLMs) have achieved remarkable performance across a wide range of tasks. However, their deployment in safety-critical domains poses significant challenges. Existing safety fine-tuning methods, which focus on textual or multimodal content, fall short in addressing challenging cases or disrupt the balance between helpfulness and harmlessness. Our evaluation highlights a safety reasoning gap: these methods lack safety visual reasoning ability, leading to such bottlenecks. To address this limitation and enhance both visual perception and reasoning in safety-critical contexts, we propose a novel dataset that integrates multi-image inputs with safety Chain-of-Thought (CoT) labels as fine-grained reasoning logic to improve model performance. Specifically, we introduce the Multi-Image Safety (MIS) dataset, an instruction-following dataset tailored for multi-image safety scenarios, consisting of training and test splits. Our experiments demonstrate that fine-tuning InternVL2.5-8B with MIS significantly outperforms both powerful open-source models and API-based models in challenging multi-image tasks requiring safety-related visual reasoning. This approach not only delivers exceptional safety performance but also preserves general capabilities without any trade-offs. Specifically, fine-tuning with MIS increases average accuracy by 0.83% across five general benchmarks and reduces the Attack Success Rate (ASR) on multiple safety benchmarks by a large margin. Data and Models are released under: \href{https://dripnowhy.github.io/MIS/}{\texttt{https://dripnowhy.github.io/MIS/}}
URLs: https://dripnowhy.github.io/MIS/, https://dripnowhy.github.io/MIS/
Authors: Murong Yue, Wenhan Lyu, Wijdane Mifdal, Jennifer Suh, Yixuan Zhang, Ziyu Yao
Abstract: Mathematical modeling (MM) is considered a fundamental skill for students in STEM disciplines. Practicing the MM skill is often the most effective when students can engage in group discussion and collaborative problem-solving. However, due to unevenly distributed teachers and educational resources needed to monitor such group activities, students do not always receive equal opportunities for this practice. Excitingly, large language models (LLMs) have recently demonstrated strong capability in both modeling mathematical problems and simulating characters with different traits and properties. Drawing inspiration from the advancement of LLMs, in this work, we present MATHVC, the very first LLM-powered virtual classroom containing multiple LLM-simulated student characters, with whom a human student can practice their MM skill. To encourage each LLM character's behaviors to be aligned with their specified math-relevant properties (termed "characteristics alignment") and the overall conversational procedure to be close to an authentic student MM discussion (termed "conversational procedural alignment"), we proposed three innovations: integrating MM domain knowledge into the simulation, defining a symbolic schema as the ground for character simulation, and designing a meta planner at the platform level to drive the conversational procedure. Through experiments and ablation studies, we confirmed the effectiveness of our simulation approach and showed the promise for MATHVC to benefit real-life students in the future.
Authors: Avinash Mudireddy, Tyler Bell, Raghu Mudumbai
Abstract: We prove a new asymptotic equipartition property for the perplexity of long texts generated by a language model and present supporting experimental evidence from open-source models. Specifically we show that the logarithmic perplexity of any large text generated by a language model must asymptotically converge to the average entropy of its token distributions. This defines a "typical set" that all long synthetic texts generated by a language model must belong to. We show that this typical set is a vanishingly small subset of all possible grammatically correct outputs. These results suggest possible applications to important practical problems such as (a) detecting synthetic AI-generated text, and (b) testing whether a text was used to train a language model. We make no simplifying assumptions (such as stationarity) about the statistics of language model outputs, and therefore our results are directly applicable to practical real-world models without any approximations.
Authors: Zhouyu Jiang, Mengshu Sun, Lei Liang, Zhiqiang Zhang
Abstract: Multi-hop question answering is a challenging task with distinct industrial relevance, and Retrieval-Augmented Generation (RAG) methods based on large language models (LLMs) have become a popular approach to tackle this task. Owing to the potential inability to retrieve all necessary information in a single iteration, a series of iterative RAG methods has been recently developed, showing significant performance improvements. However, existing methods still face two critical challenges: context overload resulting from multiple rounds of retrieval, and over-planning and repetitive planning due to the lack of a recorded retrieval trajectory. In this paper, we propose a novel iterative RAG method called ReSP, equipped with a dual-function summarizer. This summarizer compresses information from retrieved documents, targeting both the overarching question and the current sub-question concurrently. Experimental results on the multi-hop question-answering datasets HotpotQA and 2WikiMultihopQA demonstrate that our method significantly outperforms the state-of-the-art, and exhibits excellent robustness concerning context length.
Authors: Sacha Muller, Ant\'onio Loison, Bilel Omrani, Gautier Viaud
Abstract: Retrieval-Augmented Generation (RAG) has emerged as a common paradigm to use Large Language Models (LLMs) alongside private and up-to-date knowledge bases. In this work, we address the challenges of using LLM-as-a-Judge when evaluating grounded answers generated by RAG systems. To assess the calibration and discrimination capabilities of judge models, we identify 7 generator failure modes and introduce GroUSE (Grounded QA Unitary Scoring of Evaluators), a meta-evaluation benchmark of 144 unit tests. This benchmark reveals that existing automated RAG evaluation frameworks often overlook important failure modes, even when using GPT-4 as a judge. To improve on the current design of automated RAG evaluation frameworks, we propose a novel pipeline and find that while closed models perform well on GroUSE, state-of-the-art open-source judges do not generalize to our proposed criteria, despite strong correlation with GPT-4's judgement. Our findings suggest that correlation with GPT-4 is an incomplete proxy for the practical performance of judge models and should be supplemented with evaluations on unit tests for precise failure mode detection. We further show that finetuning Llama-3 on GPT-4's reasoning traces significantly boosts its evaluation capabilities, improving upon both correlation with GPT-4's evaluations and calibration on reference situations.
Authors: Samuel Belkadi, Libo Ren, Nicolo Micheletti, Lifeng Han, Goran Nenadic
Abstract: The vast amount of available medical records has the potential to improve healthcare and biomedical research. However, privacy restrictions make these data accessible for internal use only. Recent works have addressed this problem by generating synthetic data using Causal Language Modeling. Unfortunately, by taking this approach, it is often impossible to guarantee patient privacy while offering the ability to control the diversity of generations without increasing the cost of generating such data. In contrast, we present a system for generating synthetic free-text medical records using Masked Language Modeling. The system preserves critical medical information while introducing diversity in the generations and minimising re-identification risk. The system's size is about 120M parameters, minimising inference cost. The results demonstrate high-quality synthetic data with a HIPAA-compliant PHI recall rate of 96% and a re-identification risk of 3.5%. Moreover, downstream evaluations show that the generated data can effectively train a model with performance comparable to real data.
Authors: Yuxiang Huang, Binhang Yuan, Xu Han, Chaojun Xiao, Zhiyuan Liu
Abstract: Scaling the input context length of a large language model (LLM) incurs a significant increase in computation cost and memory footprint to maintain the attention key-value (KV) cache. Existing KV cache compression methods suffer from inefficient compression strategies and limited memory reduction effects, making it difficult for LLMs to conduct long-context inference on consumer-grade devices, especially when inferring long-context stream input. Such obstacles prevent consumer-grade devices from supporting more complex applications, creating challenges for the democratization of LLMs. To overcome this, we propose Locret, the first framework to create an eviction policy compatible with chunked prefill. By evaluating the causal importance of KV cache units by learnable retaining heads, Locret enables precise eviction of cache units, facilitating efficient long-context inference. In our extensive empirical studies, Locret outperforms the recent popular and competitive approaches in terms of memory efficiency and generation quality -- Locret achieves up to 20x of KV cache compression ratio within less than 10% performance loss. Furthermore, Locret achieves 128K+ long-context inference on a single NVIDIA 4090 GPU without compromising generation quality and only costs <1 GPU hour of additional training.
Authors: Zi'ou Zheng, Christopher Malon, Martin Renqiang Min, Xiaodan Zhu
Abstract: When performing complex multi-step reasoning tasks, the ability of Large Language Models (LLMs) to derive structured intermediate proof steps is important for ensuring that the models truly perform the desired reasoning and for improving models' explainability. This paper is centred around a focused study: whether the current state-of-the-art generalist LLMs can leverage the structures in a few examples to better construct the proof structures with \textit{in-context learning}. Our study specifically focuses on structure-aware demonstration and structure-aware pruning. We demonstrate that they both help improve performance. A detailed analysis is provided to help understand the results.
Authors: Xiaofeng Wu, Karl Stratos, Wei Xu
Abstract: The glyphic writing system of Chinese incorporates information-rich visual features in each character, such as radicals that provide hints about meaning or pronunciation. However, there has been no investigation into whether contemporary Large Language Models (LLMs) and Vision-Language Models (VLMs) can harness these sub-character features in Chinese through prompting. In this study, we establish a benchmark to evaluate LLMs' and VLMs' understanding of visual elements in Chinese characters, including radicals, composition structures, strokes, and stroke counts. Our results reveal that models surprisingly exhibit some, but still limited, knowledge of the visual information, regardless of whether images of characters are provided. To incite models' ability to use radicals, we further experiment with incorporating radicals into the prompts for Chinese language processing (CLP) tasks. We observe consistent improvement in Part-Of-Speech tagging when providing additional information about radicals, suggesting the potential to enhance CLP by integrating sub-character information.
Authors: Zhikun Xu, Ming Shen, Jacob Dineen, Zhaonan Li, Xiao Ye, Shijie Lu, Aswin RRV, Chitta Baral, Ben Zhou
Abstract: We introduce thoughts of words (ToW), a novel training-time data-augmentation method for next-word prediction. ToW views next-word prediction as a core reasoning task and injects fine-grained thoughts explaining what the next word should be and how it is related to the previous contexts in pre-training texts. Our formulation addresses two fundamental drawbacks of existing next-word prediction learning schemes: they induce factual hallucination and are inefficient for models to learn the implicit reasoning processes in raw texts. While there are many ways to acquire such thoughts of words, we explore the first step of acquiring ToW annotations through distilling from larger models. After continual pre-training with only 70K ToW annotations, we effectively improve models' reasoning performances by 7% to 9% on average and reduce model hallucination by up to 10%. At the same time, ToW is entirely agnostic to tasks and applications, introducing no additional biases on labels or semantics.
Authors: Yueqi Song, Frank Xu, Shuyan Zhou, Graham Neubig
Abstract: Web browsers are a portal to the internet, where much of human activity is undertaken. Thus, there has been significant research work in AI agents that interact with the internet through web browsing. However, there is also another interface designed specifically for machine interaction with online content: application programming interfaces (APIs). In this paper we ask -- what if we were to take tasks traditionally tackled by browsing agents, and give AI agents access to APIs? To do so, we propose two varieties of agents: (1) an API-calling agent that attempts to perform online tasks through APIs only, similar to traditional coding agents, and (2) a Hybrid Agent that can interact with online data through both web browsing and APIs. In experiments on WebArena, a widely-used and realistic benchmark for web navigation tasks, we find that API-based agents outperform web browsing agents. Hybrid Agents out-perform both others nearly uniformly across tasks, resulting in a more than 20.0% absolute improvement over web browsing alone, achieving a success rate of 35.8%, achiving the SOTA performance among task-agnostic agents. These results strongly suggest that when APIs are available, they present an attractive alternative to relying on web browsing alone.
Authors: Maarten Buyl, Alexander Rogiers, Sander Noels, Guillaume Bied, Iris Dominguez-Catena, Edith Heiter, Iman Johary, Alexandru-Cristian Mara, Rapha\"el Romero, Jefrey Lijffijt, Tijl De Bie
Abstract: Large language models (LLMs) are trained on vast amounts of data to generate natural language, enabling them to perform tasks like text summarization and question answering. These models have become popular in artificial intelligence (AI) assistants like ChatGPT and already play an influential role in how humans access information. However, the behavior of LLMs varies depending on their design, training, and use. In this paper, we prompt a diverse panel of popular LLMs to describe a large number of prominent personalities with political relevance, in all six official languages of the United Nations. By identifying and analyzing moral assessments reflected in their responses, we find normative differences between LLMs from different geopolitical regions, as well as between the responses of the same LLM when prompted in different languages. Among only models in the United States, we find that popularly hypothesized disparities in political views are reflected in significant normative differences related to progressive values. Among Chinese models, we characterize a division between internationally- and domestically-focused models. Our results show that the ideological stance of an LLM appears to reflect the worldview of its creators. This poses the risk of political instrumentalization and raises concerns around technological and regulatory efforts with the stated aim of making LLMs ideologically 'unbiased'.
Authors: Ang Lv, Ruobing Xie, Shuaipeng Li, Jiayi Liao, Xingwu Sun, Zhanhui Kang, Di Wang, Rui Yan
Abstract: We propose a novel attention mechanism, named Cog Attention, that enables attention weights to be negative for enhanced expressiveness, which stems from two key factors: (1) Cog Attention enhances parameter flexibility. For example, unlike traditional softmax attention heads that use a static output-value (OV) matrix to delete or copy inputs that the heads attend to, Cog Attention naturally learns to use the sign of dynamic query-key (QK) inner products to represent these operations. This enables Cog Attention to perform multiple operations simultaneously within a single head. Meanwhile, Cog Attention's OV matrix can focus more on refinement or modification. (2) Cog Attention enhances the model's robustness against representational collapse by preventing the ``over-squashing'' of earlier tokens into later positions. We develop Transformer-like models which use Cog Attention as attention modules, including decoder-only models at various scales for language modeling and U-ViT diffusion models for image generation. Experiments show that models using Cog Attention exhibit superior performance compared to those employing traditional softmax attention modules. Our approach suggests a promising research direction for rethinking and breaking the entrenched constraints of traditional softmax attention, such as the requirement for non-negative weights.
Authors: Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, Lili Yu
Abstract: We present LMFusion, a framework for empowering pretrained text-only large language models (LLMs) with multimodal generative capabilities, enabling them to understand and generate both text and images in arbitrary sequences. LMFusion leverages existing Llama-3's weights for processing texts autoregressively while introducing additional and parallel transformer modules for processing images with diffusion. During training, the data from each modality is routed to its dedicated modules: modality-specific feedforward layers, query-key-value projections, and normalization layers process each modality independently, while the shared self-attention layers allow interactions across text and image features. By freezing the text-specific modules and only training the image-specific modules, LMFusion preserves the language capabilities of text-only LLMs while developing strong visual understanding and generation abilities. Compared to methods that pretrain multimodal generative models from scratch, our experiments demonstrate that, LMFusion improves image understanding by 20% and image generation by 3.6% using only 50% of the FLOPs while maintaining Llama-3's language capabilities. We also demonstrate that this framework can adapt existing vision-language models with multimodal generation ability. Overall, this framework not only leverages existing computational investments in text-only LLMs but also enables the parallel development of language and vision capabilities, presenting a promising direction for efficient multimodal model development.
Authors: Sunbowen Lee, Junting Zhou, Chang Ao, Kaige Li, Xinrun Du, Sirui He, Jiaheng Liu, Min Yang, Zhoufutu Wen, Shiwen Ni
Abstract: Model distillation is a technique for transferring knowledge from large language models (LLMs) to smaller ones, aiming to create resource-efficient yet high-performing models. However, excessive distillation can lead to homogenization, reducing diversity among models and impairing their ability to robustly handle complex or novel tasks. These limitations underscore the need to systematically quantify the distillation process and its impact. In this work, we propose a framework to evaluate and quantify model distillation. Our method addresses two key aspects: (1) Identifying identity cognition contradictions to assess discrepancies in how models perceive and represent identity-related information, and (2) Analyzing multi-granularity response similarities across models to measure the extent of homogenization. Experimental results demonstrate two key insights: (1) Well-known closed-source and open-source LLMs usually exhibit high distillation degrees, except for Claude, Doubao, and Gemini. (2) Base LLMs show higher distillation degrees compared to aligned LLMs. By offering a systematic approach to improve the transparency of LLM data distillation, we call for LLMs with more independent development and more transparent technical reports to improve LLMs' robustness and safety. The code and data are available under https://github.com/Aegis1863/LLMs-Distillation-Quantification.
URLs: https://github.com/Aegis1863/LLMs-Distillation-Quantification.
Authors: Chang Zong, Jian Wan, Siliang Tang, Lei Zhang
Abstract: When addressing professional questions in the biomedical domain, humans typically acquire multiple pieces of information as evidence and engage in multifaceted evidence analysis to provide high-quality answers. Current LLM-based answer generation methods lack a detailed definition and learning process for evidence analysis, leading to the risk of error propagation and hallucinations while using evidence. Although increasing the parameter size of LLMs can alleviate these issues, it also presents challenges in model training and deployment with limited resources. In this study, we propose EvidenceMap, which aims to enable a tiny pre-trained language model to explicitly learn multiple aspects of biomedical evidence, including supportive evaluation, logical correlation and content summarization, thereby latently guiding a small generative model (around 3B parameters) to provide textual responses. Experimental results demonstrate that our method, fine-tuning a language model with 66M parameters, exceeds the RAG method with an 8B LLM by 19.9% and 5.7% in reference-based quality and accuracy, respectively.
Authors: Chen Chen, Xinlong Hao, Weiwen Liu, Xu Huang, Xingshan Zeng, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Yuefeng Huang, Wulong Liu, Xinzhi Wang, Defu Lian, Baoqun Yin, Yasheng Wang, Wu Liu
Abstract: Large language models (LLMs) have demonstrated significant potential in decision-making and reasoning, especially when combined with various tools to effectively solve complex problems. However, existing evaluation systems for assessing LLM function calling capabilities have several limitations: (1) limited evaluation scenarios, lacking assessments in real multi-turn dialogue contexts; (2) narrow evaluation dimensions, lacking detailed assessments for fine-grained function calls; (3) relying on LLMs or real API executions for result evaluation, which introduces significant overhead. To address these issues, we propose a comprehensive evaluation system named ACEBench. This system is meticulously designed to encompass a wide spectrum of function calling scenarios. Moreover, it categorizes these scenarios into three primary types according to the evaluation methodology: Normal, Special, and Agent. Normal evaluates function calls in basic scenarios; Special evaluates function calls in scenarios with vague or incomplete instructions; Agent introduces multi-agent interactions to simulate function calling evaluation in real-world multi-turn interactions. We conducted extensive experiments on ACEBench, analyzing various LLMs in-depth and performing a more granular analysis of error causes across different data types.
Authors: Skala Kamaran Omer, Hossein Hassani
Abstract: Idiom detection using Natural Language Processing (NLP) is the computerized process of recognizing figurative expressions within a text that convey meanings beyond the literal interpretation of the words. While idiom detection has seen significant progress across various languages, the Kurdish language faces a considerable research gap in this area despite the importance of idioms in tasks like machine translation and sentiment analysis. This study addresses idiom detection in Sorani Kurdish by approaching it as a text classification task using deep learning techniques. To tackle this, we developed a dataset containing 10,580 sentences embedding 101 Sorani Kurdish idioms across diverse contexts. Using this dataset, we developed and evaluated three deep learning models: KuBERT-based transformer sequence classification, a Recurrent Convolutional Neural Network (RCNN), and a BiLSTM model with an attention mechanism. The evaluations revealed that the transformer model, the fine-tuned BERT, consistently outperformed the others, achieving nearly 99% accuracy while the RCNN achieved 96.5% and the BiLSTM 80%. These results highlight the effectiveness of Transformer-based architectures in low-resource languages like Kurdish. This research provides a dataset, three optimized models, and insights into idiom detection, laying a foundation for advancing Kurdish NLP.
Authors: Ameya Godbole, Robin Jia
Abstract: Improvements in large language models have led to increasing optimism that they can serve as reliable evaluators of natural language generation outputs. In this paper, we challenge this optimism by thoroughly re-evaluating five state-of-the-art factuality metrics on a collection of 11 datasets for summarization, retrieval-augmented generation, and question answering. We find that these evaluators are inconsistent with each other and often misestimate system-level performance, both of which can lead to a variety of pitfalls. We further show that these metrics exhibit biases against highly paraphrased outputs and outputs that draw upon faraway parts of the source documents. We urge users of these factuality metrics to proceed with caution and manually validate the reliability of these metrics in their domain of interest before proceeding.
Authors: Mohamed Elfeki, Rui Liu, Chad Voegele
Abstract: The dominance of large decoder-only language models has overshadowed encoder-decoder architectures, despite their fundamental efficiency advantages in sequence processing. For small language models (SLMs) - those with 1 billion parameters or fewer - our systematic analysis across GPU, CPU, and NPU platforms reveals that encoder-decoder architectures achieve 47% lower first-token latency and 4.7x higher throughput compared to decoder-only models on edge devices. These gains may be attributed to encoder-decoder's one-time input processing and efficient separation of understanding and generation phases. We introduce a novel knowledge distillation framework that enables encoder-decoder models to leverage capabilities from large scalable decoder-only teachers while preserving their architectural advantages, achieving up to 6 average performance points improvement across diverse tasks, with significant gains in asymmetric sequence tasks where input and output distributions can benefit from different processing approaches. When combined with modern advances like Rotary Positional Embeddings (RoPE) and Vision encoders, our systematic investigation demonstrates that encoder-decoder architectures provide a more practical path toward deploying capable language models in resource-constrained environments. Our findings challenge the prevailing trend toward decoder-only scaling, showing that architectural choices become increasingly crucial as parameter budgets decrease, particularly for on-device and edge deployments where computational efficiency is paramount.
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: 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: 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: Yihua Shao, Minxi Yan, Yang Liu, Siyu Chen, Wenjie Chen, Xinwei Long, Ziyang Yan, Lei Li, Chenyu Zhang, Nicu Sebe, Hao Tang, Yan Wang, Hao Zhao, Mengzhu Wang, Jingcai Guo
Abstract: Low-rank Adaptation (LoRA) has demonstrated remarkable capabilities for task specific fine-tuning. However, in scenarios that involve multiple tasks, training a separate LoRA model for each one results in considerable inefficiency in terms of storage and inference. Moreover, existing parameter generation methods fail to capture the correlations among these tasks, making multi-task LoRA parameter generation challenging. To address these limitations, we propose In-Context Meta LoRA (ICM-LoRA), a novel approach that efficiently achieves task-specific customization of large language models (LLMs). Specifically, we use training data from all tasks to train a tailored generator, Conditional Variational Autoencoder (CVAE). CVAE takes task descriptions as inputs and produces task-aware LoRA weights as outputs. These LoRA weights are then merged with LLMs to create task-specialized models without the need for additional fine-tuning. Furthermore, we utilize in-context meta-learning for knowledge enhancement and task mapping, to capture the relationship between tasks and parameter distributions. As a result, our method achieves more accurate LoRA parameter generation for diverse tasks using CVAE. ICM-LoRA enables more accurate LoRA parameter reconstruction than current parameter reconstruction methods and is useful for implementing task-specific enhancements of LoRA parameters. At the same time, our method occupies 283MB, only 1\% storage compared with the original LoRA.
Authors: Yubo Wang, Xiang Yue, Wenhu Chen
Abstract: Supervised Fine-Tuning (SFT) is commonly used to train language models to imitate annotated responses for given instructions. In this paper, we challenge this paradigm and propose Critique Fine-Tuning (CFT), a strategy where models learn to critique noisy responses rather than simply imitate correct ones. Inspired by human learning processes that emphasize critical thinking, CFT encourages deeper analysis and nuanced understanding-traits often overlooked by standard SFT. To validate the effectiveness of CFT, we construct a 50K-sample dataset from WebInstruct, using GPT-4o as the teacher to generate critiques in the form of ([query; noisy response], critique). CFT on this dataset yields a consistent 4-10% improvement over SFT on six math benchmarks with different base models like Qwen2.5, Qwen2.5-Math and DeepSeek-Math. We further expand to MetaMath and NuminaMath datasets and observe similar gains over SFT. Notably, our model Qwen2.5-Math-CFT only requires 1 hour training on 8xH100 over the 50K examples. It can match or outperform strong competitors like Qwen2.5-Math-Instruct on most benchmarks, which use over 2M samples. Moreover, it can match the performance of SimpleRL, which is a deepseek-r1 replication trained with 140x more compute. Ablation studies show that CFT is robust to the source of noisy response and teacher critique model. Through these findings, we argue that CFT offers a more effective alternative to advance the reasoning of language models.
Authors: Ziyi Dong, Pengxu Wei, Liang Lin
Abstract: State-of-the-arts text-to-image generation models such as Imagen and Stable Diffusion Model have succeed remarkable progresses in synthesizing high-quality, feature-rich images with high resolution guided by human text prompts. Since certain characteristics of image content \emph{e.g.}, very specific object entities or styles, are very hard to be accurately described by text, some example-based image generation approaches have been proposed, \emph{i.e.} generating new concepts based on absorbing the salient features of a few input references. Despite of acknowledged successes, these methods have struggled on accurately capturing the reference examples' characteristics while keeping diverse and high-quality image generation, particularly in the one-shot scenario (\emph{i.e.} given only one reference). To tackle this problem, we propose a simple yet effective framework, namely DreamArtist, which adopts a novel positive-negative prompt-tuning learning strategy on the pre-trained diffusion model, and it has shown to well handle the trade-off between the accurate controllability and fidelity of image generation with only one reference example. Specifically, our proposed framework incorporates both positive and negative embeddings or adapters and optimizes them in a joint manner. The positive part aggressively captures the salient characteristics of the reference image to drive diversified generation and the negative part rectifies inadequacies from the positive part. We have conducted extensive experiments and evaluated the proposed method from image similarity (fidelity) and diversity, generation controllability, and style cloning. And our DreamArtist has achieved a superior generation performance over existing methods. Besides, our additional evaluation on extended tasks, including concept compositions and prompt-guided image editing, demonstrates its effectiveness for more applications.
Authors: Xiang Li, Cristina Mata, Jongwoo Park, Kumara Kahatapitiya, Yoo Sung Jang, Jinghuan Shang, Kanchana Ranasinghe, Ryan Burgert, Mu Cai, Yong Jae Lee, Michael S. Ryoo
Abstract: Vision Language Models (VLMs) have recently been leveraged to generate robotic actions, forming Vision-Language-Action (VLA) models. However, directly adapting a pretrained VLM for robotic control remains challenging, particularly when constrained by a limited number of robot demonstrations. In this work, we introduce LLaRA: Large Language and Robotics Assistant, a framework that formulates robot action policy as visuo-textual conversations and enables an efficient transfer of a pretrained VLM into a powerful VLA, motivated by the success of visual instruction tuning in Computer Vision. First, we present an automated pipeline to generate conversation-style instruction tuning data for robots from existing behavior cloning datasets, aligning robotic actions with image pixel coordinates. Further, we enhance this dataset in a self-supervised manner by defining six auxiliary tasks, without requiring any additional action annotations. We show that a VLM finetuned with a limited amount of such datasets can produce meaningful action decisions for robotic control. Through experiments across multiple simulated and real-world tasks, we demonstrate that LLaRA achieves state-of-the-art performance while preserving the generalization capabilities of large language models. The code, datasets, and pretrained models are available at https://github.com/LostXine/LLaRA.
Authors: Gon\c{c}alo Hora de Carvalho, Oscar Knap, Robert Pollice
Abstract: We develop a systematic benchmark set to test the generalization of state-of-the-art large language models on broader problems beyond linguistic tasks and evaluate it on a systematic progression of GPT models (GPT-3.5, GPT-4, GPT-4o, GPT-4o-mini). Using well-known simple games like Tic-Tac-Toe, Connect Four, and Battleship, all encoded in ASCII, we test their strategic capabilities and spatial reasoning. To probe generalization, we introduce three new games: LEGO Connect Language (LCL) for spatial logic, a shape recognition game, and Guess-the-SMILES (GtS), an advanced spatial logic benchmark in chemistry. Results show that, despite proficiency in standard benchmarks, GPT models perform poorly in these games, failing to anticipate losing moves, play correctly, or recognize spatial relationships. Except for Tic-Tac-Toe and GtS, a systematic progression in gameplay performance as models are formally improved (GPT-3.5, GPT-4, GPT-4o) is not observed. GPT-4 succeeds in shape recognition, but all models consistently struggle with LCL and GtS. This suggests that while GPT models can emulate conversational proficiency and basic rule comprehension, they have limited cognitive flexibility and generalization in strategy and spatial reasoning. Our findings, highlighted with our benchmark suite (ChildPlay GitHub Repository), caution against claims of emergent intelligence in GPT models, which appear more specialized than general.
Authors: Sebastian Bordt, Suraj Srinivas, Valentyn Boreiko, Ulrike von Luxburg
Abstract: The leakage of benchmark data into the training data has emerged as a significant challenge for evaluating the capabilities of large language models (LLMs). In this work, we challenge the common assumption that small-scale contamination renders benchmark evaluations invalid. First, we experimentally quantify the magnitude of benchmark overfitting based on scaling along three dimensions: The number of model parameters (up to 1.6B), the number of times an example is seen (up to 144), and the number of training tokens (up to 40B). If model and data follow the Chinchilla scaling laws, minor contamination indeed leads to overfitting. At the same time, even 144 times of contamination can be forgotten if the training data is scaled beyond five times Chinchilla, a regime characteristic of many modern LLMs. Continual pre-training of OLMo-7B corroborates these results. Next, we study the impact of the weight decay parameter on example forgetting, showing that empirical forgetting occurs faster than the cumulative weight decay. This allows us to gauge the degree of example forgetting in large-scale training runs, indicating that many LLMs, including Lllama 3 405B, have forgotten the data seen at the beginning of training.
Authors: Elia Cunegatti, Leonardo Lucio Custode, Giovanni Iacca
Abstract: Network pruning focuses on computational techniques that aim to reduce a given model's computational cost by removing a subset of its parameters while having minimal impact on performance. Throughout the last decade, the most widely used pruning paradigm has been pruning and re-training, which nowadays is inconvenient due to the vast amount of pre-trained models, which are in any case too expensive to re-train. In this paper, we exploit functional information from dense pre-trained models, i.e., their activations, to obtain sparse models that maximize the activations' alignment w.r.t. their corresponding dense models. Hence, we propose \textsc{NeuroAL}, a \emph{top-up} algorithm that can be used on top of any given pruning algorithm for LLMs, which modifies the block-wise and row-wise sparsity exploiting information from both the dense model and its sparse version to maximize the \emph{neuron alignment} among activations. Differently from existing methods, our approach adaptively selects the best hyperparameters for the block-wise and row-wise sparsity ratios w.r.t. the model and the desired sparsity, and requires \emph{no re-training}. We test our method over 276 cases combining four LLM families, three sparsity ratios, and ten language tasks (three language modeling and seven zero-shot datasets), showing how it consistently outperforms the latest state-of-the-art methods in terms of performance-runtime trade-off. The code is available at \href{https://github.com/eliacunegatti/NeuroAL}{https://github.com/eliacunegatti/NeuroAL}.
URLs: https://github.com/eliacunegatti/NeuroAL, https://github.com/eliacunegatti/NeuroAL
Authors: Rui Li, Xiaohan Wang, Yuhui Zhang, Zeyu Wang, Serena Yeung-Levy
Abstract: Despite significant advancements in video large multimodal models (video-LMMs), achieving effective temporal grounding in long-form videos remains a challenge for existing models. To address this limitation, we propose Temporal Preference Optimization (TPO), a novel post-training framework designed to enhance the temporal grounding capabilities of video-LMMs through preference learning. TPO adopts a self-training approach that enables models to differentiate between well-grounded and less accurate temporal responses by leveraging curated preference datasets at two granularities: localized temporal grounding, which focuses on specific video segments, and comprehensive temporal grounding, which captures extended temporal dependencies across entire video sequences. By optimizing on these preference datasets, TPO significantly enhances temporal understanding while reducing reliance on manually annotated data. Extensive experiments on three long-form video understanding benchmarks--LongVideoBench, MLVU, and Video-MME--demonstrate the effectiveness of TPO across two state-of-the-art video-LMMs. Notably, LLaVA-Video-TPO establishes itself as the leading 7B model on the Video-MME benchmark, underscoring the potential of TPO as a scalable and efficient solution for advancing temporal reasoning in long-form video understanding. Project page: https://ruili33.github.io/tpo_website.
Authors: Tianbo Yang, Mingqi Yang, Hongyi Zhao, Tianshuo Yang
Abstract: Developing the logic necessary to solve mathematical problems or write mathematical proofs is one of the more difficult objectives for large language models (LLMS). Currently, the most popular methods in literature consists of fine-tuning the model on written mathematical content such as academic publications and textbooks, so that the model can learn to emulate the style of mathematical writing. In this project, we explore the effectiveness of using retrieval augmented generation (RAG) to address gaps in the mathematical reasoning of LLMs. We develop LemmaHead, a RAG knowledge base that supplements queries to the model with relevant mathematical context, with particular focus on context from published textbooks. To measure our model's performance in mathematical reasoning, our testing paradigm focuses on the task of automated theorem proving via generating proofs to a given mathematical claim in the Lean formal language.
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. For example, deleting an email from a user's mailbox may or may not be appropriate depending on the email's content, the 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: Yinqi Zhang, Xintian Han, Haolong Li, Kedi Chen, Shaohui Lin
Abstract: Large language models (LLM) have shown remarkable abilities in text generation, question answering, language translation, reasoning and many other tasks. It continues to advance rapidly and is becoming increasingly influential in various fields, from technology and business to education and entertainment. Despite LLM's success in multiple areas, its ability to play abstract games, such as chess, is underexplored. Chess-playing requires the language models to output legal and reasonable moves from textual inputs. Here, we propose the Large language model ChessLLM to play full chess games. We transform the game into a textual format with the best move represented in the Forsyth-Edwards Notation. We show that by simply supervised fine-tuning, our model has achieved a professional-level Elo rating of 1788 in matches against the standard Elo-rated Stockfish when permitted to sample 10 times. We further show that data quality is important. Long-round data supervision enjoys a 350 Elo rating improvement over short-round data.
Authors: Shilong Deng, Yongzhao Wang, Rahul Savani
Abstract: We introduce a framework for translating game descriptions in natural language into extensive-form representations in game theory, leveraging Large Language Models (LLMs) and in-context learning. Given the varying levels of strategic complexity in games, such as perfect versus imperfect information, directly applying in-context learning would be insufficient. To address this, we introduce a two-stage framework with specialized modules to enhance in-context learning, enabling it to divide and conquer the problem effectively. In the first stage, we tackle the challenge of imperfect information by developing a module that identifies information sets along and the corresponding partial tree structure. With this information, the second stage leverages in-context learning alongside a self-debugging module to produce a complete extensive-form game tree represented using pygambit, the Python API of a recognized game-theoretic analysis tool called Gambit. Using this python representation enables the automation of tasks such as computing Nash equilibria directly from natural language descriptions. We evaluate the performance of the full framework, as well as its individual components, using various LLMs on games with different levels of strategic complexity. Our experimental results show that the framework significantly outperforms baseline models in generating accurate extensive-form games, with each module playing a critical role in its success.
Authors: Vaibhav Gusain, Douglas Leith
Abstract: We propose a novel redaction methodology that can be used to sanitize natural text data. Our new technique provides better privacy benefits than other state of the art techniques while maintaining lower redaction levels.