Authors: Edith Haim, Lars van den Bergh, Cynthia S. Q. Siew, Yoed N. Kenett, Daniele Marinazzo, Massimo Stella
Abstract: Understanding attitudes towards STEM means quantifying the cognitive and emotional ways in which individuals, and potentially large language models too, conceptualise such subjects. This study uses behavioural forma mentis networks (BFMNs) to investigate the STEM-focused mindset, i.e. ways of associating and perceiving ideas, of 177 human participants and 177 artificial humans simulated by GPT-3.5. Participants were split in 3 groups - trainees, experts and academics - to compare the influence of expertise level on their mindset. The results revealed that human forma mentis networks exhibited significantly higher clustering coefficients compared to GPT-3.5, indicating that human mindsets displayed a tendency to form and close triads of conceptual associations while recollecting STEM ideas. Human experts, in particular, demonstrated robust clustering coefficients, reflecting better integration of STEM concepts into their cognitive networks. In contrast, GPT-3.5 produced sparser mindsets. Furthermore, both human and GPT mindsets framed mathematics in neutral or positive terms, differently from STEM high schoolers, researchers and other large language models sampled in other works. This research contributes to understanding how mindset structure can provide cognitive insights about memory structure and machine limitations.
Authors: Ashley Lewis, Michael White, Jing Liu, Toshiaki Koike-Akino, Kieran Parsons, Ye Wang
Abstract: The deployment of Large Language Models (LLMs) in customer support is constrained by hallucination-generating false information-and the high cost of proprietary models. To address these challenges, we propose a retrieval-augmented question-answering (QA) pipeline and explore how to balance human input and automation. Using a dataset of questions about a Samsung Smart TV user manual, we demonstrate that synthetic data generated by LLMs outperforms crowdsourced data in reducing hallucination in finetuned models. We also compare self-training (fine-tuning models on their own outputs) and knowledge distillation (fine-tuning on stronger models' outputs, e.g., GPT-4o), and find that self-training achieves comparable hallucination reduction. We conjecture that this surprising finding can be attributed to increased exposure bias issues in the knowledge distillation case and support this conjecture with post hoc analysis. We also improve robustness to unanswerable questions and retrieval failures with contextualized "I don't know" responses. These findings show that scalable, cost-efficient QA systems can be built using synthetic data and self-training with open-source models, reducing reliance on proprietary tools or costly human annotations.
Authors: Zhengdong Yang, Shuichiro Shimizu, Yahan Yu, Chenhui Chu
Abstract: Recent advancements in large language models (LLMs) have spurred interest in expanding their application beyond text-based tasks. A large number of studies have explored integrating other modalities with LLMs, notably speech modality, which is naturally related to text. This paper surveys the integration of speech with LLMs, categorizing the methodologies into three primary approaches: text-based, latent-representation-based, and audio-token-based integration. We also demonstrate how these methods are applied across various speech-related applications and highlight the challenges in this field to offer inspiration for
Authors: Yudi Zhang, Lu Wang, Meng Fang, Yali Du, Chenghua Huang, Jun Wang, Qingwei Lin, Mykola Pechenizkiy, Dongmei Zhang, Saravan Rajmohan, Qi Zhang
Abstract: Distilling large language models (LLMs) typically involves transferring the teacher model's responses through supervised fine-tuning (SFT). However, this approach neglects the potential to distill both data (output content) and reward signals (quality evaluations). Extracting reliable reward signals directly from teacher models is challenging, as LLMs are optimized for generation rather than evaluation, often resulting in biased or inconsistent assessments. To address this limitation, we propose a novel distillation pipeline that transfers both responses and rewards. Our method generates pseudo-rewards through a self-supervised mechanism that leverages the inherent structure of both teacher and student responses, enabling reward learning without explicit external evaluation. The reward model subsequently guides reinforcement learning (RL), allowing iterative refinement of the student model after an SFT warm-up phase. Experiments on GSM8K and MMLU-PRO demonstrate that our method consistently outperforms traditional SFT-based approaches, enabling student models to surpass the performance of their teachers. This work highlights the potential for scalable, efficient distillation through structured self-supervised reward learning, reducing dependence on external reward supervision.
Authors: Jonas Becker, Lars Benedikt Kaesberg, Andreas Stephan, Jan Philip Wahle, Terry Ruas, Bela Gipp
Abstract: Multi-agent debate - multiple instances of large language models discussing problems in turn-based interaction - has shown promise for solving knowledge and reasoning tasks. However, these methods show limitations, particularly when scaling them to longer reasoning chains. In this study, we unveil a new issue of multi-agent debate: discussions drift away from the initial problem over multiple turns. We define this phenomenon as problem drift and quantify its presence across ten tasks (i.e., three generative, three knowledge, three reasoning, and one instruction-following task). To identify the reasons for this issue, we perform a human study with eight experts on discussions suffering from problem drift, who find the most common issues are a lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases). To systematically address the issue of problem drift, we propose DRIFTJudge, a method based on LLM-as-a-judge, to detect problem drift at test-time. We further propose DRIFTPolicy, a method to mitigate 31% of problem drift cases. Our study can be seen as a first step to understanding a key limitation of multi-agent debate, highlighting pathways for improving their effectiveness in the future.
Authors: Gabriele Prato, Jerry Huang, Prasannna Parthasarathi, Shagun Sodhani, Sarath Chandar
Abstract: Large Language Models (LLMs) have emerged as highly capable systems and are increasingly being integrated into various uses. However, the rapid pace of their deployment has outpaced a comprehensive understanding of their internal mechanisms and a delineation of their capabilities and limitations. A desired attribute of an intelligent system is its ability to recognize the scope of its own knowledge. To investigate whether LLMs embody this characteristic, we develop a benchmark designed to challenge these models to enumerate all information they possess on specific topics. This benchmark evaluates whether the models recall excessive, insufficient, or the precise amount of information, thereby indicating their awareness of their own knowledge. Our findings reveal that all tested LLMs, given sufficient scale, demonstrate an understanding of how much they know about specific topics. While different architectures exhibit varying rates of this capability's emergence, the results suggest that awareness of knowledge may be a generalizable attribute of LLMs. Further research is needed to confirm this potential and fully elucidate the underlying mechanisms.
Authors: Ife Adebara, Hawau Olamide Toyin, Nahom Tesfu Ghebremichael, AbdelRahim Elmadany, Muhammad Abdul-Mageed
Abstract: Africa's rich linguistic heritage remains underrepresented in NLP, largely due to historical policies that favor foreign languages and create significant data inequities. In this paper, we integrate theoretical insights on Africa's language landscape with an empirical evaluation using Sahara - a comprehensive benchmark curated from large-scale, publicly accessible datasets capturing the continent's linguistic diversity. By systematically assessing the performance of leading large language models (LLMs) on Sahara, we demonstrate how policy-induced data variations directly impact model effectiveness across African languages. Our findings reveal that while a few languages perform reasonably well, many Indigenous languages remain marginalized due to sparse data. Leveraging these insights, we offer actionable recommendations for policy reforms and inclusive data practices. Overall, our work underscores the urgent need for a dual approach - combining theoretical understanding with empirical evaluation - to foster linguistic diversity in AI for African communities.
Authors: Lola Le Breton, Quentin Fournier, Mariam El Mezouar, Sarath Chandar
Abstract: Recent innovations in architecture, pre-training, and fine-tuning have led to the remarkable in-context learning and reasoning abilities of large auto-regressive language models such as LLaMA and DeepSeek. In contrast, encoders like BERT and RoBERTa have not seen the same level of progress despite being foundational for many downstream NLP applications. To bridge this gap, we introduce NeoBERT, a next-generation encoder that redefines the capabilities of bidirectional models by integrating state-of-the-art advancements in architecture, modern data, and optimized pre-training methodologies. NeoBERT is designed for seamless adoption: it serves as a plug-and-play replacement for existing base models, relies on an optimal depth-to-width ratio, and leverages an extended context length of 4,096 tokens. Despite its compact 250M parameter footprint, it achieves state-of-the-art results on the massive MTEB benchmark, outperforming BERT large, RoBERTa large, NomicBERT, and ModernBERT under identical fine-tuning conditions. In addition, we rigorously evaluate the impact of each modification on GLUE and design a uniform fine-tuning and evaluation framework for MTEB. We release all code, data, checkpoints, and training scripts to accelerate research and real-world adoption.
Authors: Sil Hamilton, Rebecca M. M. Hicke, David Mimno, Matthew Wilkens
Abstract: We release 70,509 high-quality social networks extracted from multilingual fiction and nonfiction narratives. We additionally provide metadata for ~30,000 of these texts (73% nonfiction and 27% fiction) written between 1800 and 1999 in 58 languages. This dataset provides information on historical social worlds at an unprecedented scale, including data for 1,192,855 individuals in 2,805,482 pair-wise relationships annotated for affinity and relationship type. We achieve this scale by automating previously manual methods of extracting social networks; specifically, we adapt an existing annotation task as a language model prompt, ensuring consistency at scale with the use of structured output. This dataset provides an unprecedented resource for the humanities and social sciences by providing data on cognitive models of social realities.
Authors: Yash Mahajan, Matthew Freestone, Sathyanarayanan Aakur, Santu Karmaker
Abstract: Large Language Models (LLMs) have recently shown remarkable advancement in various NLP tasks. As such, a popular trend has emerged lately where NLP researchers extract word/sentence/document embeddings from these large decoder-only models and use them for various inference tasks with promising results. However, it is still unclear whether the performance improvement of LLM-induced embeddings is merely because of scale or whether underlying embeddings they produce significantly differ from classical encoding models like Word2Vec, GloVe, Sentence-BERT (SBERT) or Universal Sentence Encoder (USE). This is the central question we investigate in the paper by systematically comparing classical decontextualized and contextualized word embeddings with the same for LLM-induced embeddings. Our results show that LLMs cluster semantically related words more tightly and perform better on analogy tasks in decontextualized settings. However, in contextualized settings, classical models like SimCSE often outperform LLMs in sentence-level similarity assessment tasks, highlighting their continued relevance for fine-grained semantics.
Authors: Anwar Hossain Zahid, Monoshi Kumar Roy, Swarna Das
Abstract: The proliferation of hate speech on social media is one of the serious issues that is bringing huge impacts to society: an escalation of violence, discrimination, and social fragmentation. The problem of detecting hate speech is intrinsically multifaceted due to cultural, linguistic, and contextual complexities and adversarial manipulations. In this study, we systematically investigate the performance of LLMs on detecting hate speech across multilingual datasets and diverse geographic contexts. Our work presents a new evaluation framework in three dimensions: binary classification of hate speech, geography-aware contextual detection, and robustness to adversarially generated text. Using a dataset of 1,000 comments from five diverse regions, we evaluate three state-of-the-art LLMs: Llama2 (13b), Codellama (7b), and DeepSeekCoder (6.7b). Codellama had the best binary classification recall with 70.6% and an F1-score of 52.18%, whereas DeepSeekCoder had the best performance in geographic sensitivity, correctly detecting 63 out of 265 locations. The tests for adversarial robustness also showed significant weaknesses; Llama2 misclassified 62.5% of manipulated samples. These results bring to light the trade-offs between accuracy, contextual understanding, and robustness in the current versions of LLMs. This work has thus set the stage for developing contextually aware, multilingual hate speech detection systems by underlining key strengths and limitations, therefore offering actionable insights for future research and real-world applications.
Authors: Sungduk Yu, Man Luo, Avinash Madusu, Vasudev Lal, Phillip Howard
Abstract: Peer review is a critical process for ensuring the integrity of published scientific research. Confidence in this process is predicated on the assumption that experts in the relevant domain give careful consideration to the merits of manuscripts which are submitted for publication. With the recent rapid advancements in large language models (LLMs), a new risk to the peer review process is that negligent reviewers will rely on LLMs to perform the often time consuming process of reviewing a paper. However, there is a lack of existing resources for benchmarking the detectability of AI text in the domain of peer review. To address this deficiency, we introduce a comprehensive dataset containing a total of 788,984 AI-written peer reviews paired with corresponding human reviews, covering 8 years of papers submitted to each of two leading AI research conferences (ICLR and NeurIPS). We use this new resource to evaluate the ability of 18 existing AI text detection algorithms to distinguish between peer reviews written by humans and different state-of-the-art LLMs. Motivated by the shortcomings of existing methods, we propose a new detection approach which surpasses existing methods in the identification of AI written peer reviews. Our work reveals the difficulty of identifying AI-generated text at the individual peer review level, highlighting the urgent need for new tools and methods to detect this unethical use of generative AI.
Authors: Yanan Chen, Ali Pesaranghader, Tanmana Sadhu
Abstract: Recent advances in Large Language Models (LLMs) have raised interest in their formal reasoning capabilities, particularly in mathematics. While closed LLMs like GPT-4 perform well on mathematical benchmarks, e.g., GSM8K, it remains unclear whether small to medium-sized open LLMs can achieve similar performance, questioning their reliability. To close this gap, we propose a post-training approach leveraging a mixture of opinions (MoO) from weaker ancillary LLMs to enhance a (relatively) stronger LLM's reasoning. For that, each post-training sample is augmented with Chain-of-Thought (CoT) reasoning steps and answers from ancillary LLMs, enabling the main LLM to learn from diverse perspectives. We compare MoO with standard supervised fine-tuning (SFT), few-shot prompting, and the Mixture of Agents (MoA) method on mathematical reasoning benchmarks. Our results show that incorporating weaker LLMs' opinions improves mathematical reasoning by an average of 5%, highlighting the value of diverse perspectives in reasoning tasks.
Authors: Sheng Zhang, Qianchu Liu, Guanghui Qin, Tristan Naumann, Hoifung Poon
Abstract: Reinforcement learning from verifiable rewards (RLVR) has recently gained attention for its ability to elicit self-evolved reasoning capabilitie from base language models without explicit reasoning supervisions, as demonstrated by DeepSeek-R1. While prior work on RLVR has primarily focused on mathematical and coding domains, its applicability to other tasks and domains remains unexplored. In this work, we investigate whether medical reasoning can emerge from RLVR. We introduce Med-RLVR as an initial study of RLVR in the medical domain leveraging medical multiple-choice question answering (MCQA) data as verifiable labels. Our results demonstrate that RLVR is not only effective for math and coding but also extends successfully to medical question answering. Notably, Med-RLVR achieves performance comparable to traditional supervised fine-tuning (SFT) on in-distribution tasks while significantly improving out-of-distribution generalization, with an 8-point accuracy gain. Further analysis of training dynamics reveals that, with no explicit reasoning supervision, reasoning emerges from the 3B-parameter base model. These findings underscore the potential of RLVR in domains beyond math and coding, opening new avenues for its application in knowledge-intensive fields such as medicine.
Authors: Kohei Tsuji, Tatsuya Hiraoka, Yuchang Cheng, Eiji Aramaki, Tomoya Iwakura
Abstract: This paper investigates how LLMs encode inputs with typos. We hypothesize that specific neurons and attention heads recognize typos and fix them internally using local and global contexts. We introduce a method to identify typo neurons and typo heads that work actively when inputs contain typos. Our experimental results suggest the following: 1) LLMs can fix typos with local contexts when the typo neurons in either the early or late layers are activated, even if those in the other are not. 2) Typo neurons in the middle layers are responsible for the core of typo-fixing with global contexts. 3) Typo heads fix typos by widely considering the context not focusing on specific tokens. 4) Typo neurons and typo heads work not only for typo-fixing but also for understanding general contexts.
Authors: Yoo Yeon Sung, Eve Fleisig, Yu Hou, Ishan Upadhyay, Jordan Lee Boyd-Graber
Abstract: Language models are often miscalibrated, leading to confidently incorrect answers. We introduce GRACE, a benchmark for language model calibration that incorporates comparison with human calibration. GRACE consists of question-answer pairs, in which each question contains a series of clues that gradually become easier, all leading to the same answer; models must answer correctly as early as possible as the clues are revealed. This setting permits granular measurement of model calibration based on how early, accurately, and confidently a model answers. After collecting these questions, we host live human vs. model competitions to gather 1,749 data points on human and model teams' timing, accuracy, and confidence. We propose a metric, CalScore, that uses GRACE to analyze model calibration errors and identify types of model miscalibration that differ from human behavior. We find that although humans are less accurate than models, humans are generally better calibrated. Since state-of-the-art models struggle on GRACE, it effectively evaluates progress on improving model calibration.
Authors: Hannah Cyberey, Yangfeng Ji, David Evans
Abstract: Large language models (LLMs) are known to perpetuate stereotypes and exhibit biases. Various strategies have been proposed to mitigate potential harms that may result from these biases, but most work studies biases in LLMs as a black-box problem without considering how concepts are represented within the model. We adapt techniques from representation engineering to study how the concept of "gender" is represented within LLMs. We introduce a new method that extracts concept representations via probability weighting without labeled data and efficiently selects a steering vector for measuring and manipulating the model's representation. We also present a projection-based method that enables precise steering of model predictions and demonstrate its effectiveness in mitigating gender bias in LLMs.
Authors: Fan Jiang, Tom Drummond, Trevor Cohn
Abstract: Recent approaches to multilingual open-domain question answering (MLODQA) have achieved promising results given abundant language-specific training data. However, the considerable annotation cost limits the application of these methods for underrepresented languages. We introduce a \emph{few-shot learning} approach to synthesise large-scale multilingual data from large language models (LLMs). Our method begins with large-scale self-supervised pre-training using WikiData, followed by training on high-quality synthetic multilingual data generated by prompting LLMs with few-shot supervision. The final model, \textsc{FsModQA}, significantly outperforms existing few-shot and supervised baselines in MLODQA and cross-lingual and monolingual retrieval. We further show our method can be extended for effective zero-shot adaptation to new languages through a \emph{cross-lingual prompting} strategy with only English-supervised data, making it a general and applicable solution for MLODQA tasks without costly large-scale annotation.
Authors: Yu Zhao, Songping Huang, Dongsheng Zhou, Zhaoyun Ding, Fei Wang, Aixin Nian
Abstract: Obtaining valuable information from massive data efficiently has become our research goal in the era of Big Data. Text summarization technology has been continuously developed to meet this demand. Recent work has also shown that transformer-based pre-trained language models have achieved great success on various tasks in Natural Language Processing (NLP). Aiming at the problem of Chinese news text summary generation and the application of Transformer structure on Chinese, this paper proposes a Chinese news text summarization model (CNsum) based on Transformer structure, and tests it on Chinese datasets such as THUCNews. The results of the conducted experiments show that CNsum achieves better ROUGE score than the baseline models, which verifies the outperformance of the model.
Authors: Mian Zhang, Shaun M. Eack, Zhiyu Zoey Chen
Abstract: Applying large language models (LLMs) to assist in psycho-counseling is an emerging and meaningful approach, driven by the significant gap between patient needs and the availability of mental health support. However, current LLMs struggle to consistently provide effective responses to client speeches, largely due to the lack of supervision from high-quality real psycho-counseling data, whose content is typically inaccessible due to client privacy concerns. Furthermore, the quality of therapists' responses in available sessions can vary significantly based on their professional training and experience. Assessing the quality of therapists' responses remains an open challenge. In this work, we address these challenges by first proposing a set of professional and comprehensive principles to evaluate therapists' responses to client speeches. Using these principles, we create a preference dataset, PsychoCounsel-Preference, which contains 36k high-quality preference comparison pairs. This dataset aligns with the preferences of professional psychotherapists, providing a robust foundation for evaluating and improving LLMs in psycho-counseling. Experiments on reward modeling and preference learning demonstrate that PsychoCounsel-Preference is an excellent resource for LLMs to acquire essential skills for responding to clients in a counseling session. Our best-aligned model, PsychoCounsel-Llama3-8B, achieves an impressive win rate of 87% against GPT-4o. We release PsychoCounsel-Preference, PsychoCounsel-Llama3-8B and the reward model PsychoCounsel Llama3-8B-Reward to facilitate the research of psycho-counseling with LLMs at: https://hf.co/Psychotherapy-LLM.
Authors: Yunhai Hu, Zining Liu, Zhenyuan Dong, Tianfan Peng, Bradley McDanel, Sai Qian Zhang
Abstract: Sequential dependencies present a fundamental bottleneck in deploying large-scale autoregressive models, particularly for real-time applications. While traditional optimization approaches like pruning and quantization often compromise model quality, recent advances in generation-refinement frameworks demonstrate that this trade-off can be significantly mitigated. This survey presents a comprehensive taxonomy of generation-refinement frameworks, analyzing methods across autoregressive sequence tasks. We categorize methods based on their generation strategies (from simple n-gram prediction to sophisticated draft models) and refinement mechanisms (including single-pass verification and iterative approaches). Through systematic analysis of both algorithmic innovations and system-level implementations, we examine deployment strategies across computing environments and explore applications spanning text, images, and speech generation. This systematic examination of both theoretical frameworks and practical implementations provides a foundation for future research in efficient autoregressive decoding.
Authors: Minggui He, Yilun Liu, Shimin Tao, Yuanchang Luo, Hongyong Zeng, Chang Su, Li Zhang, Hongxia Ma, Daimeng Wei, Weibin Meng, Hao Yang, Boxing Chen, Osamu Yoshie
Abstract: Despite recent breakthroughs in reasoning-enhanced large language models (LLMs) like DeepSeek-R1, incorporating inference-time reasoning into machine translation (MT), where human translators naturally employ structured, multi-layered reasoning chain-of-thoughts (CoTs), is yet underexplored. Existing methods either design a fixed CoT tailored for a specific MT sub-task (e.g., literature translation), or rely on synthesizing CoTs unaligned with humans and supervised fine-tuning (SFT) prone to catastrophic forgetting, limiting their adaptability to diverse translation scenarios. This paper introduces R1-Translator (R1-T1), a novel framework to achieve inference-time reasoning for general MT via reinforcement learning (RL) with human-aligned CoTs comprising six common patterns. Our approach pioneers three innovations: (1) extending reasoning-based translation beyond MT sub-tasks to six languages and diverse tasks (e.g., legal/medical domain adaptation, idiom resolution); (2) formalizing six expert-curated CoT templates that mirror hybrid human strategies like context-aware paraphrasing and back translation; and (3) enabling self-evolving CoT discovery and anti-forgetting adaptation through RL with KL-constrained rewards. Experimental results indicate a steady translation performance improvement in 21 languages and 80 translation directions on Flores-101 test set, especially on the 15 languages unseen from training, with its general multilingual abilities preserved compared with plain SFT.
Authors: Linyang He, Ercong Nie, Sukru Samet Dindar, Arsalan Firoozi, Adrian Florea, Van Nguyen, Corentin Puffay, Riki Shimizu, Haotian Ye, Jonathan Brennan, Helmut Schmid, Hinrich Sch\"utze, Nima Mesgarani
Abstract: We introduce XCOMPS in this work, a multilingual conceptual minimal pair dataset covering 17 languages. Using this dataset, we evaluate LLMs' multilingual conceptual understanding through metalinguistic prompting, direct probability measurement, and neurolinguistic probing. By comparing base, instruction-tuned, and knowledge-distilled models, we find that: 1) LLMs exhibit weaker conceptual understanding for low-resource languages, and accuracy varies across languages despite being tested on the same concept sets. 2) LLMs excel at distinguishing concept-property pairs that are visibly different but exhibit a marked performance drop when negative pairs share subtle semantic similarities. 3) Instruction tuning improves performance in concept understanding but does not enhance internal competence; knowledge distillation can enhance internal competence in conceptual understanding for low-resource languages with limited gains in explicit task performance. 4) More morphologically complex languages yield lower concept understanding scores and require deeper layers for conceptual reasoning.
Authors: Taiqiang Wu, Chenchen Ding, Wenyong Zhou, Yuxin Cheng, Xincheng Feng, Shuqi Wang, Chufan Shi, Zhengwu Liu, Ngai Wong
Abstract: Low-rank adaptation (LoRA) is a predominant parameter-efficient finetuning method to adapt large language models (LLMs) for downstream tasks. In this paper, we first propose to deploy the LoRA-finetuned LLMs on the hybrid compute-in-memory (CIM) architecture (i.e., pretrained weights onto RRAM and LoRA onto SRAM). To address performance degradation from RRAM's inherent noise, we design a novel Hardware-aware Low-rank Adaption (HaLoRA) method, aiming to train a LoRA branch that is both robust and accurate by aligning the training objectives under both ideal and noisy conditions. Experiments finetuning LLaMA 3.2 1B and 3B demonstrate HaLoRA's effectiveness across multiple reasoning tasks, achieving up to 22.7 improvement in average score while maintaining robustness at various noise levels.
Authors: Jinhao Pan, Chahat Raj, Ziyu Yao, Ziwei Zhu
Abstract: The exceptional performance of Large Language Models (LLMs) often comes with the unintended propagation of social biases embedded in their training data. While existing benchmarks evaluate overt bias through direct term associations between bias concept terms and demographic terms, LLMs have become increasingly adept at avoiding biased responses, creating an illusion of neutrality. However, biases persist in subtler, contextually hidden forms that traditional benchmarks fail to capture. We introduce the Hidden Bias Benchmark (HBB), a novel dataset designed to assess hidden bias that bias concepts are hidden within naturalistic, subtly framed contexts in real-world scenarios. We analyze six state-of-the-art LLMs, revealing that while models reduce bias in response to overt bias, they continue to reinforce biases in nuanced settings. Data, code, and results are available at https://github.com/JP-25/Hidden-Bias-Benchmark.
Authors: Nathan Roll
Abstract: Large language models (LLMs) showcase increasingly impressive English benchmark scores, however their performance profiles remain inconsistent across multilingual settings. To address this gap, we introduce PolyPrompt, a novel, parameter-efficient framework for enhancing the multilingual capabilities of LLMs. Our method learns a set of trigger tokens for each language through a gradient-based search, identifying the input query's language and selecting the corresponding trigger tokens which are prepended to the prompt during inference. We perform experiments on two ~1 billion parameter models, with evaluations on the global MMLU benchmark across fifteen typologically and resource diverse languages, demonstrating accuracy gains of 3.7%-19.9% compared to naive and translation-pipeline baselines.
Authors: Che Hyun Lee, Heeseung Kim, Jiheum Yeom, Sungroh Yoon
Abstract: We propose EdiText, a controllable text editing method that modify the reference text to desired attributes at various scales. We integrate an SDEdit-based editing technique that allows for broad adjustments in the degree of text editing. Additionally, we introduce a novel fine-level editing method based on self-conditioning, which allows subtle control of reference text. While being capable of editing on its own, this fine-grained method, integrated with the SDEdit approach, enables EdiText to make precise adjustments within the desired range. EdiText demonstrates its controllability to robustly adjust reference text at broad range of levels across various tasks, including toxicity control and sentiment control.
Authors: Nevidu Jayatilleke, Ruvan Weerasinghe, Nipuna Senanayake
Abstract: The substantial growth of textual content in diverse domains and platforms has led to a considerable need for Automatic Text Summarization (ATS) techniques that aid in the process of text analysis. The effectiveness of text summarization models has been significantly enhanced in a variety of technical domains because of advancements in Natural Language Processing (NLP) and Deep Learning (DL). Despite this, the process of summarizing textual information continues to be significantly constrained by the intricate writing styles of a variety of texts, which involve a range of technical complexities. Text summarization techniques can be broadly categorized into two main types: abstractive summarization and extractive summarization. Extractive summarization involves directly extracting sentences, phrases, or segments of text from the content without making any changes. On the other hand, abstractive summarization is achieved by reconstructing the sentences, phrases, or segments from the original text using linguistic analysis. Through this study, a linguistically diverse categorizations of text summarization approaches have been addressed in a constructive manner. In this paper, the authors explored existing hybrid techniques that have employed both extractive and abstractive methodologies. In addition, the pros and cons of various approaches discussed in the literature are also investigated. Furthermore, the authors conducted a comparative analysis on different techniques and matrices to evaluate the generated summaries using language generation models. This survey endeavors to provide a comprehensive overview of ATS by presenting the progression of language processing regarding this task through a breakdown of diverse systems and architectures accompanied by technical and mathematical explanations of their operations.
Authors: Peilin Wu, Xinlu Zhang, Wenhao Yu, Xingyu Liu, Xinya Du, Zhiyu Zoey Chen
Abstract: Recent advancements in Retrieval-Augmented Language Models (RALMs) have demonstrated their efficacy in knowledge-intensive tasks. However, existing evaluation benchmarks often assume a single optimal approach to leveraging retrieved information, failing to account for varying user needs. This paper introduces a novel evaluation framework that systematically assesses RALMs under three user need cases-Context-Exclusive, Context-First, and Memory-First-across three distinct context settings: Context Matching, Knowledge Conflict, and Information Irrelevant. By varying both user instructions and the nature of retrieved information, our approach captures the complexities of real-world applications where models must adapt to diverse user requirements. Through extensive experiments on multiple QA datasets, including HotpotQA, DisentQA, and our newly constructed synthetic URAQ dataset, we find that restricting memory usage improves robustness in adversarial retrieval conditions but decreases peak performance with ideal retrieval results and model family dominates behavioral differences. Our findings highlight the necessity of user-centric evaluations in the development of retrieval-augmented systems and provide insights into optimizing model performance across varied retrieval contexts. We will release our code and URAQ dataset upon acceptance of the paper.
Authors: Isa Inuwa-Dutse
Abstract: With over 500 languages in Nigeria, three languages -- Hausa, Yor\`ub\'a and Igbo -- spoken by over 175 million people, account for about 60% of the spoken languages. However, these languages are categorised as low-resource due to insufficient resources to support tasks in computational linguistics. Several research efforts and initiatives have been presented, however, a coherent understanding of the state of Natural Language Processing (NLP) - from grammatical formalisation to linguistic resources that support complex tasks such as language understanding and generation is lacking. This study presents the first comprehensive review of advancements in low-resource NLP (LR-NLP) research across the three major Nigerian languages (NaijaNLP). We quantitatively assess the available linguistic resources and identify key challenges. Although a growing body of literature addresses various NLP downstream tasks in Hausa, Igbo, and Yor\`ub\'a, only about 25.1% of the reviewed studies contribute new linguistic resources. This finding highlights a persistent reliance on repurposing existing data rather than generating novel, high-quality resources. Additionally, language-specific challenges, such as the accurate representation of diacritics, remain under-explored. To advance NaijaNLP and LR-NLP more broadly, we emphasise the need for intensified efforts in resource enrichment, comprehensive annotation, and the development of open collaborative initiatives.
Authors: Zixuan Weng, Xiaolong Jin, Jinyuan Jia, Xiangyu Zhang
Abstract: Ensuring AI safety is crucial as large language models become increasingly integrated into real-world applications. A key challenge is jailbreak, where adversarial prompts bypass built-in safeguards to elicit harmful disallowed outputs. Inspired by psychological foot-in-the-door principles, we introduce FITD,a novel multi-turn jailbreak method that leverages the phenomenon where minor initial commitments lower resistance to more significant or more unethical transgressions.Our approach progressively escalates the malicious intent of user queries through intermediate bridge prompts and aligns the model's response by itself to induce toxic responses. Extensive experimental results on two jailbreak benchmarks demonstrate that FITD achieves an average attack success rate of 94% across seven widely used models, outperforming existing state-of-the-art methods. Additionally, we provide an in-depth analysis of LLM self-corruption, highlighting vulnerabilities in current alignment strategies and emphasizing the risks inherent in multi-turn interactions.The code is available at https://github.com/Jinxiaolong1129/Foot-in-the-door-Jailbreak .
URLs: https://github.com/Jinxiaolong1129/Foot-in-the-door-Jailbreak
Authors: Yiwei Li, Ji Zhang, Shaoxiong Feng, Peiwen Yuan, Xinglin Wang, Jiayi Shi, Yueqi Zhang, Chuyi Tan, Boyuan Pan, Yao Hu, Kan Li
Abstract: Self-consistency improves reasoning by aggregating diverse stochastic samples, yet the dynamics behind its efficacy remain underexplored. We reframe self-consistency as a dynamic distributional alignment problem, revealing that decoding temperature not only governs sampling randomness but also actively shapes the latent answer distribution. Given that high temperatures require prohibitively large sample sizes to stabilize, while low temperatures risk amplifying biases, we propose a confidence-driven mechanism that dynamically calibrates temperature: sharpening the sampling distribution under uncertainty to align with high-probability modes, and promoting exploration when confidence is high. Experiments on mathematical reasoning tasks show this approach outperforms fixed-diversity baselines under limited samples, improving both average and best-case performance across varying initial temperatures without additional data or modules. This establishes self-consistency as a synchronization challenge between sampling dynamics and evolving answer distributions.
Authors: P Sam Sahil, Anupam Jamatia
Abstract: This paper describes the system submitted by Team A to SemEval 2025 Task 11, ``Bridging the Gap in Text-Based Emotion Detection.'' The task involved identifying the perceived emotion of a speaker from text snippets, with each instance annotated with one of six emotions: joy, sadness, fear, anger, surprise, or disgust. A dataset provided by the task organizers served as the foundation for training and evaluating our models. Among the various approaches explored, the best performance was achieved using multilingual embeddings combined with a fully connected layer. This paper details the system architecture, discusses experimental results, and highlights the advantages of leveraging multilingual representations for robust emotion detection in text.
Authors: Yujia Chen, Changsong Li, Yiming Wang, Qingqing Xiao, Nan Zhang, Zifan Kong, Peng Wang, Binyu Yan
Abstract: Mental health issues are worsening in today's competitive society, such as depression and anxiety. Traditional healings like counseling and chatbots fail to engage effectively, they often provide generic responses lacking emotional depth. Although large language models (LLMs) have the potential to create more human-like interactions, they still struggle to capture subtle emotions. This requires LLMs to be equipped with human-like adaptability and warmth. To fill this gap, we propose the MIND (Multi-agent INner Dialogue), a novel paradigm that provides more immersive psychological healing environments. Considering the strong generative and role-playing ability of LLM agents, we predefine an interactive healing framework and assign LLM agents different roles within the framework to engage in interactive inner dialogues with users, thereby providing an immersive healing experience. We conduct extensive human experiments in various real-world healing dimensions, and find that MIND provides a more user-friendly experience than traditional paradigms. This demonstrates that MIND effectively leverages the significant potential of LLMs in psychological healing.
Authors: Yuntao Du, Kailin Jiang, Zhi Gao, Chenrui Shi, Zilong Zheng, Siyuan Qi, Qing Li
Abstract: Knowledge editing techniques have emerged as essential tools for updating the factual knowledge of large language models (LLMs) and multimodal models (LMMs), allowing them to correct outdated or inaccurate information without retraining from scratch. However, existing benchmarks for multimodal knowledge editing primarily focus on entity-level knowledge represented as simple triplets, which fail to capture the complexity of real-world multimodal information. To address this issue, we introduce MMKE-Bench, a comprehensive MultiModal Knowledge Editing Benchmark, designed to evaluate the ability of LMMs to edit diverse visual knowledge in real-world scenarios. MMKE-Bench addresses these limitations by incorporating three types of editing tasks: visual entity editing, visual semantic editing, and user-specific editing. Besides, MMKE-Bench uses free-form natural language to represent and edit knowledge, offering a more flexible and effective format. The benchmark consists of 2,940 pieces of knowledge and 8,363 images across 33 broad categories, with evaluation questions automatically generated and human-verified. We assess five state-of-the-art knowledge editing methods on three prominent LMMs, revealing that no method excels across all criteria, and that visual and user-specific edits are particularly challenging. MMKE-Bench sets a new standard for evaluating the robustness of multimodal knowledge editing techniques, driving progress in this rapidly evolving field.
Authors: Zhenyu Liu, Yunxin Li, Baotian Hu, Wenhan Luo, Yaowei Wang, Min Zhang
Abstract: To improve Multimodal Large Language Models' (MLLMs) ability to process images and complex instructions, researchers predominantly curate large-scale visual instruction tuning datasets, which are either sourced from existing vision tasks or synthetically generated using LLMs and image descriptions. However, they often suffer from critical flaws, including misaligned instruction-image pairs and low-quality images. Such issues hinder training efficiency and limit performance improvements, as models waste resources on noisy or irrelevant data with minimal benefit to overall capability. To address this issue, we propose a \textbf{Vi}sual-Centric \textbf{S}election approach via \textbf{A}gents Collaboration (ViSA), which centers on image quality assessment and image-instruction relevance evaluation. Specifically, our approach consists of 1) an image information quantification method via visual agents collaboration to select images with rich visual information, and 2) a visual-centric instruction quality assessment method to select high-quality instruction data related to high-quality images. Finally, we reorganize 80K instruction data from large open-source datasets. Extensive experiments demonstrate that ViSA outperforms or is comparable to current state-of-the-art models on seven benchmarks, using only 2.5\% of the original data, highlighting the efficiency of our data selection approach. Moreover, we conduct ablation studies to validate the effectiveness of each component of our method. The code is available at https://github.com/HITsz-TMG/ViSA.
Authors: Xiang Geng, Zhejian Lai, Jiajun Chen, Hao Yang, Shujian Huang
Abstract: Quality Estimation (QE) models evaluate the quality of machine translations without reference translations, serving as the reward models for the translation task. Due to the data scarcity, synthetic data generation has emerged as a promising solution. However, synthetic QE data often suffers from distribution shift, which can manifest as discrepancies between pseudo and real translations, or in pseudo labels that do not align with human preferences. To tackle this issue, we introduce ADSQE, a novel framework for alleviating distribution shift in synthetic QE data. To reduce the difference between pseudo and real translations, we employ the constrained beam search algorithm and enhance translation diversity through the use of distinct generation models. ADSQE uses references, i.e., translation supervision signals, to guide both the generation and annotation processes, enhancing the quality of word-level labels. ADSE further identifies the shortest phrase covering consecutive error tokens, mimicking human annotation behavior, to assign the final phrase-level labels. Specially, we underscore that the translation model can not annotate translations of itself accurately. Extensive experiments demonstrate that ADSQE outperforms SOTA baselines like COMET in both supervised and unsupervised settings. Further analysis offers insights into synthetic data generation that could benefit reward models for other tasks.
Authors: Yujie Feng, Liming Zhan, Zexin Lu, Yongxin Xu, Xu Chu, Yasha Wang, Jiannong Cao, Philip S. Yu, Xiao-Ming Wu
Abstract: Regular updates are essential for maintaining up-to-date knowledge in large language models (LLMs). Consequently, various model editing methods have been developed to update specific knowledge within LLMs. However, training-based approaches often struggle to effectively incorporate new knowledge while preserving unrelated general knowledge. To address this challenge, we propose a novel framework called Geometric Knowledge Editing (GeoEdit). GeoEdit utilizes the geometric relationships of parameter updates from fine-tuning to differentiate between neurons associated with new knowledge updates and those related to general knowledge perturbations. By employing a direction-aware knowledge identification method, we avoid updating neurons with directions approximately orthogonal to existing knowledge, thus preserving the model's generalization ability. For the remaining neurons, we integrate both old and new knowledge for aligned directions and apply a "forget-then-learn" editing strategy for opposite directions. Additionally, we introduce an importance-guided task vector fusion technique that filters out redundant information and provides adaptive neuron-level weighting, further enhancing model editing performance. Extensive experiments on two publicly available datasets demonstrate the superiority of GeoEdit over existing state-of-the-art methods.
Authors: Yu Yan, Sheng Sun, Zixiang Tang, Teli Liu, Min Liu
Abstract: Stance detection on social media aims to identify attitudes expressed in tweets towards specific targets. Current studies prioritize Large Language Models (LLMs) over Small Language Models (SLMs) due to the overwhelming performance improving provided by LLMs. However, heavily relying on LLMs for stance detection, regardless of the cost, is impractical for real-world social media monitoring systems that require vast data analysis. To this end, we propose \textbf{\underline{Co}}llaborative Stance Detection via Small-Large Language Model Consistency \textbf{\underline{Ver}}ification (\textbf{CoVer}) framework, which enhances LLM utilization via context-shared batch reasoning and logical verification between LLM and SLM. Specifically, instead of processing each text individually, CoVer processes texts batch-by-batch, obtaining stance predictions and corresponding explanations via LLM reasoning in a shared context. Then, to exclude the bias caused by context noises, CoVer introduces the SLM for logical consistency verification. Finally, texts that repeatedly exhibit low logical consistency are classified using consistency-weighted aggregation of prior LLM stance predictions. Our experiments show that CoVer outperforms state-of-the-art methods across multiple benchmarks in the zero-shot setting, achieving 0.54 LLM queries per tweet while significantly enhancing performance. Our CoVer offers a more practical solution for LLM deploying for social media stance detection.
Authors: Javier Coronado-Bl\'azquez
Abstract: Large Language Models (LLMs) have transformed text generation through inherently probabilistic context-aware mechanisms, mimicking human natural language. In this paper, we systematically investigate the performance of various LLMs when generating random numbers, considering diverse configurations such as different model architectures, numerical ranges, temperature, and prompt languages. Our results reveal that, despite their stochastic transformers-based architecture, these models often exhibit deterministic responses when prompted for random numerical outputs. In particular, we find significant differences when changing the model, as well as the prompt language, attributing this phenomenon to biases deeply embedded within the training data. Models such as DeepSeek-R1 can shed some light on the internal reasoning process of LLMs, despite arriving to similar results. These biases induce predictable patterns that undermine genuine randomness, as LLMs are nothing but reproducing our own human cognitive biases.
Authors: Tanja Baeumel, Josef van Genabith, Simon Ostermann
Abstract: Autoregressive large language models (LLMs) exhibit impressive performance across various tasks but struggle with simple arithmetic, such as addition of two or more operands. We show that this struggle arises from LLMs' use of a simple one-digit lookahead heuristic, which works fairly well (but not perfect) for two-operand addition but fails in multi-operand cases, where the carry-over logic is more complex. Our probing experiments and digit-wise accuracy evaluation show that LLMs fail precisely where a one-digit lookahead is insufficient to account for cascading carries. We analyze the impact of tokenization strategies on arithmetic performance and show that all investigated models, regardless of tokenization, are inherently limited in the addition of multiple operands due to their reliance on a one-digit lookahead heuristic. Our findings reveal fundamental limitations that prevent LLMs from generalizing to more complex numerical reasoning.
Authors: Huazheng Wang, Yongcheng Jing, Haifeng Sun, Yingjie Wang, Jingyu Wang, Jianxin Liao, Dacheng Tao
Abstract: In this paper, we explore machine unlearning from a novel dimension, by studying how to safeguard model unlearning in large language models (LLMs). Our goal is to prevent unlearned models from recalling any related memory of the targeted knowledge.We begin by uncovering a surprisingly simple yet overlooked fact: existing methods typically erase only the exact expressions of the targeted knowledge, leaving paraphrased or related information intact. To rigorously measure such oversights, we introduce UGBench, the first benchmark tailored for evaluating the generalisation performance across 13 state-of-the-art methods.UGBench reveals that unlearned models can still recall paraphrased answers and retain target facts in intermediate layers. To address this, we propose PERMU, a perturbation-based method that significantly enhances the generalisation capabilities for safeguarding LLM unlearning.Experiments demonstrate that PERMU delivers up to a 50.13% improvement in unlearning while maintaining a 43.53% boost in robust generalisation. Our code can be found in https://github.com/MaybeLizzy/UGBench.
Authors: Marta Lango, Borys Naglik, Mateusz Lango, Iwo Naglik
Abstract: Aspect-Sentiment Triplet Extraction (ASTE) is one of the most challenging and complex tasks in sentiment analysis. It concerns the construction of triplets that contain an aspect, its associated sentiment polarity, and an opinion phrase that serves as a rationale for the assigned polarity. Despite the growing popularity of the task and the many machine learning methods being proposed to address it, the number of datasets for ASTE is very limited. In particular, no dataset is available for any of the Slavic languages. In this paper, we present two new datasets for ASTE containing customer opinions about hotels and purchased products expressed in Polish. We also perform experiments with two ASTE techniques combined with two large language models for Polish to investigate their performance and the difficulty of the assembled datasets. The new datasets are available under a permissive licence and have the same file format as the English datasets, facilitating their use in future research.
Authors: Rayyan Merchant, Akhilesh Kakolu Ramarao, Kevin Tang
Abstract: Despite speaking mutually intelligible varieties of the same language, speakers of Tajik Persian, written in a modified Cyrillic alphabet, cannot read Iranian and Afghan texts written in the Perso-Arabic script. As the vast majority of Persian text on the Internet is written in Perso-Arabic, monolingual Tajik speakers are unable to interface with the Internet in any meaningful way. Due to overwhelming similarity between the formal registers of these dialects and the scarcity of Tajik-Farsi parallel data, machine transliteration has been proposed as more a practical and appropriate solution than machine translation. This paper presents a transformer-based G2P approach to Tajik-Farsi transliteration, achieving chrF++ scores of 58.70 (Farsi to Tajik) and 74.20 (Tajik to Farsi) on novel digraphic datasets, setting a comparable baseline metric for future work. Our results also demonstrate the non-trivial difficulty of this task in both directions. We also provide an overview of the differences between the two scripts and the challenges they present, so as to aid future efforts in Tajik-Farsi transliteration.
Authors: Haochen Sun, Shuwen Zhang, Lei Ren, Hao Xu, Hao Fu, Caixia Yuan, Xiaojie Wang
Abstract: Large language models (LLMs) based agent systems have made great strides in real-world applications beyond traditional NLP tasks. This paper proposes a new LLM-powered Multi-Agent System (LLM-MAS) benchmark, Collab-Overcooked, built on the popular Overcooked-AI game with more applicable and challenging tasks in interactive environments. Collab-Overcooked extends existing benchmarks from two novel perspectives. First, it provides a multi-agent framework supporting diverse tasks and objectives and encourages collaboration through natural language communication. Second, it introduces a spectrum of process-oriented evaluation metrics to assess the fine-grained collaboration capabilities of different LLM agents, a dimension often overlooked in prior work. We conduct extensive experiments over 10 popular LLMs and show that, while the LLMs present a strong ability in goal interpretation, there is a significant discrepancy in active collaboration and continuous adaption that are critical for efficiently fulfilling complicated tasks. Notably, we highlight the strengths and weaknesses in LLM-MAS and provide insights for improving and evaluating LLM-MAS on a unified and open-sourced benchmark. Environments, 30 open-ended tasks, and an integrated evaluation package are now publicly available at https://github.com/YusaeMeow/Collab-Overcooked.
Authors: Ning Shang, Li Lyna Zhang, Siyuan Wang, Gaokai Zhang, Gilsinia Lopez, Fan Yang, Weizhu Chen, Mao Yang
Abstract: LongRoPE2 is a novel approach that extends the effective context window of pre-trained large language models (LLMs) to the target length, while preserving the performance on the original shorter context window. This is achieved by three contributions: (1) a hypothesis that insufficient training in higher RoPE dimensions contributes to the persistent out-of-distribution (OOD) issues observed in existing methods; (2) an effective RoPE rescaling algorithm that adopts evolutionary search guided by "needle-driven" perplexity to address the insufficient training problem; (3) a mixed context window training approach that fine-tunes model weights to adopt rescaled RoPE for long-context sequences while preserving the short-context performance with the original RoPE. Extensive experiments on LLaMA3-8B and Phi3-mini-3.8B across various benchmarks validate the hypothesis and demonstrate the effectiveness of LongRoPE2. Remarkably, LongRoPE2 extends LLaMA3-8B to achieve a 128K effective context length while retaining over 98.5% of short-context performance, using only 10B tokens -- 80x fewer than Meta's approach, which fails to reach the target effective context length. Code will be available at https://github.com/microsoft/LongRoPE.
Authors: Tergel Munkhbat, Namgyu Ho, Seohyun Kim, Yongjin Yang, Yujin Kim, Se-Young Yun
Abstract: Chain-of-thought (CoT) reasoning has enabled large language models (LLMs) to utilize additional computation through intermediate tokens to solve complex tasks. However, we posit that typical reasoning traces contain many redundant tokens, incurring extraneous inference costs. Upon examination of the output distribution of current LLMs, we find evidence on their latent ability to reason more concisely, relative to their default behavior. To elicit this capability, we propose simple fine-tuning methods which leverage self-generated concise reasoning paths obtained by best-of-N sampling and few-shot conditioning, in task-specific settings. Our combined method achieves a 30% reduction in output tokens on average, across five model families on GSM8K and MATH, while maintaining average accuracy. By exploiting the fundamental stochasticity and in-context learning capabilities of LLMs, our self-training approach robustly elicits concise reasoning on a wide range of models, including those with extensive post-training. Code is available at https://github.com/TergelMunkhbat/concise-reasoning
Authors: Yifan Zhang, Wenyu Du, Dongming Jin, Jie Fu, Zhi Jin
Abstract: Chain-of-Thought (CoT) significantly enhances the performance of large language models (LLMs) across a wide range of tasks, and prior research shows that CoT can theoretically increase expressiveness. However, there is limited mechanistic understanding of the algorithms that Transformer+CoT can learn. In this work, we (1) evaluate the state tracking capabilities of Transformer+CoT and its variants, confirming the effectiveness of CoT. (2) Next, we identify the circuit, a subset of model components, responsible for tracking the world state, finding that late-layer MLP neurons play a key role. We propose two metrics, compression and distinction, and show that the neuron sets for each state achieve nearly 100% accuracy, providing evidence of an implicit finite state automaton (FSA) embedded within the model. (3) Additionally, we explore three realistic settings: skipping intermediate steps, introducing data noise, and testing length generalization. Our results demonstrate that Transformer+CoT learns robust algorithms (FSA), highlighting its resilience in challenging scenarios.
Authors: Qingyang Zhang, Rose E. Wang, Ana T. Ribeiro, Dora Demszky, Susanna Loeb
Abstract: Educator attention is critical for student success, yet how educators distribute their attention across students remains poorly understood due to data and methodological constraints. This study presents the first large-scale computational analysis of educator attention patterns, leveraging over 1 million educator utterances from virtual group tutoring sessions linked to detailed student demographic and academic achievement data. Using natural language processing techniques, we systematically examine the recipient and nature of educator attention. Our findings reveal that educators often provide more attention to lower-achieving students. However, disparities emerge across demographic lines, particularly by gender. Girls tend to receive less attention when paired with boys, even when they are the lower achieving student in the group. Lower-achieving female students in mixed-gender pairs receive significantly less attention than their higher-achieving male peers, while lower-achieving male students receive significantly and substantially more attention than their higher-achieving female peers. We also find some differences by race and English learner (EL) status, with low-achieving Black students receiving additional attention only when paired with another Black student but not when paired with a non-Black peer. In contrast, higher-achieving EL students receive disproportionately more attention than their lower-achieving EL peers. This work highlights how large-scale interaction data and computational methods can uncover subtle but meaningful disparities in teaching practices, providing empirical insights to inform more equitable and effective educational strategies.
Authors: Toon Vandendriessche, Mathieu De Coster, Annelies Lejon, Joni Dambre
Abstract: Isolated Sign Language Recognition (ISLR) is crucial for scalable sign language technology, yet language-specific approaches limit current models. To address this, we propose a one-shot learning approach that generalises across languages and evolving vocabularies. Our method involves pretraining a model to embed signs based on essential features and using a dense vector search for rapid, accurate recognition of unseen signs. We achieve state-of-the-art results, including 50.8% one-shot MRR on a large dictionary containing 10,235 unique signs from a different language than the training set. Our approach is robust across languages and support sets, offering a scalable, adaptable solution for ISLR. Co-created with the Deaf and Hard of Hearing (DHH) community, this method aligns with real-world needs, and advances scalable sign language recognition.
Authors: Yan-Lun Chen, Yi-Ru Wei, Chia-Yi Hsu, Chia-Mu Yu, Chun-Ying Huang, Ying-Dar Lin, Yu-Sung Wu, Wei-Bin Lee
Abstract: Large language models (LLMs) demonstrate strong task-specific capabilities through fine-tuning, but merging multiple fine-tuned models often leads to degraded performance due to overlapping instruction-following components. Task Arithmetic (TA), which combines task vectors derived from fine-tuning, enables multi-task learning and task forgetting but struggles to isolate task-specific knowledge from general instruction-following behavior. To address this, we propose Layer-Aware Task Arithmetic (LATA), a novel approach that assigns layer-specific weights to task vectors based on their alignment with instruction-following or task-specific components. By amplifying task-relevant layers and attenuating instruction-following layers, LATA improves task learning and forgetting performance while preserving overall model utility. Experiments on multiple benchmarks, including WikiText-2, GSM8K, and HumanEval, demonstrate that LATA outperforms existing methods in both multi-task learning and selective task forgetting, achieving higher task accuracy and alignment with minimal degradation in output quality. Our findings highlight the importance of layer-wise analysis in disentangling task-specific and general-purpose knowledge, offering a robust framework for efficient model merging and editing.
Authors: Haibin Chen, Kangtao Lv, Chengwei Hu, Yanshi Li, Yujin Yuan, Yancheng He, Xingyao Zhang, Langming Liu, Shilei Liu, Wenbo Su, Bo Zheng
Abstract: With the increasing use of Large Language Models (LLMs) in fields such as e-commerce, domain-specific concept evaluation benchmarks are crucial for assessing their domain capabilities. Existing LLMs may generate factually incorrect information within the complex e-commerce applications. Therefore, it is necessary to build an e-commerce concept benchmark. Existing benchmarks encounter two primary challenges: (1) handle the heterogeneous and diverse nature of tasks, (2) distinguish between generality and specificity within the e-commerce field. To address these problems, we propose \textbf{ChineseEcomQA}, a scalable question-answering benchmark focused on fundamental e-commerce concepts. ChineseEcomQA is built on three core characteristics: \textbf{Focus on Fundamental Concept}, \textbf{E-commerce Generality} and \textbf{E-commerce Expertise}. Fundamental concepts are designed to be applicable across a diverse array of e-commerce tasks, thus addressing the challenge of heterogeneity and diversity. Additionally, by carefully balancing generality and specificity, ChineseEcomQA effectively differentiates between broad e-commerce concepts, allowing for precise validation of domain capabilities. We achieve this through a scalable benchmark construction process that combines LLM validation, Retrieval-Augmented Generation (RAG) validation, and rigorous manual annotation. Based on ChineseEcomQA, we conduct extensive evaluations on mainstream LLMs and provide some valuable insights. We hope that ChineseEcomQA could guide future domain-specific evaluations, and facilitate broader LLM adoption in e-commerce applications.
Authors: Guizhen Chen, Weiwen Xu, Hao Zhang, Hou Pong Chan, Chaoqun Liu, Lidong Bing, Deli Zhao, Anh Tuan Luu, Yu Rong
Abstract: Many challenging reasoning tasks require not just rapid, intuitive responses, but a more deliberate, multi-step approach. Recent progress in large language models (LLMs) highlights an important shift from the "System 1" way of quick reactions to the "System 2" style of reflection-and-correction problem solving. However, current benchmarks heavily rely on the final-answer accuracy, leaving much of a model's intermediate reasoning steps unexamined. This fails to assess the model's ability to reflect and rectify mistakes within the reasoning process. To bridge this gap, we introduce FINEREASON, a logic-puzzle benchmark for fine-grained evaluation of LLMs' reasoning capabilities. Each puzzle can be decomposed into atomic steps, making it ideal for rigorous validation of intermediate correctness. Building on this, we introduce two tasks: state checking, and state transition, for a comprehensive evaluation of how models assess the current situation and plan the next move. To support broader research, we also provide a puzzle training set aimed at enhancing performance on general mathematical tasks. We show that models trained on our state checking and transition data demonstrate gains in math reasoning by up to 5.1% on GSM8K.
Authors: Abdelrahman Abdallah, Jamshid Mozafari, Bhawna Piryani, Mohammed Ali, Adam Jatowt
Abstract: Knowledge-intensive tasks, particularly open-domain question answering (ODQA), document reranking, and retrieval-augmented language modeling, require a balance between retrieval accuracy and generative flexibility. Traditional retrieval models such as BM25 and Dense Passage Retrieval (DPR), efficiently retrieve from large corpora but often lack semantic depth. Generative models like GPT-4-o provide richer contextual understanding but face challenges in maintaining factual consistency. In this work, we conduct a systematic evaluation of retrieval-based, generation-based, and hybrid models, with a primary focus on their performance in ODQA and related retrieval-augmented tasks. Our results show that dense retrievers, particularly DPR, achieve strong performance in ODQA with a top-1 accuracy of 50.17\% on NQ, while hybrid models improve nDCG@10 scores on BEIR from 43.42 (BM25) to 52.59, demonstrating their strength in document reranking. Additionally, we analyze language modeling tasks using WikiText-103, showing that retrieval-based approaches like BM25 achieve lower perplexity compared to generative and hybrid methods, highlighting their utility in retrieval-augmented generation. By providing detailed comparisons and practical insights into the conditions where each approach excels, we aim to facilitate future optimizations in retrieval, reranking, and generative models for ODQA and related knowledge-intensive applications.
Authors: Chichien Tsai, Chiamu Yu, Yingdar Lin, Yusung Wu, Weibin Lee
Abstract: The increasing adoption of large language models (LLMs) for code-related tasks has raised concerns about the security of their training datasets. One critical threat is dead code poisoning, where syntactically valid but functionally redundant code is injected into training data to manipulate model behavior. Such attacks can degrade the performance of neural code search systems, leading to biased or insecure code suggestions. Existing detection methods, such as token-level perplexity analysis, fail to effectively identify dead code due to the structural and contextual characteristics of programming languages. In this paper, we propose DePA (Dead Code Perplexity Analysis), a novel line-level detection and cleansing method tailored to the structural properties of code. DePA computes line-level perplexity by leveraging the contextual relationships between code lines and identifies anomalous lines by comparing their perplexity to the overall distribution within the file. Our experiments on benchmark datasets demonstrate that DePA significantly outperforms existing methods, achieving 0.14-0.19 improvement in detection F1-score and a 44-65% increase in poisoned segment localization precision. Furthermore, DePA enhances detection speed by 0.62-23x, making it practical for large-scale dataset cleansing. Overall, by addressing the unique challenges of dead code poisoning, DePA provides a robust and efficient solution for safeguarding the integrity of code generation model training datasets.
Authors: Amr Mohamed, Mingmeng Geng, Michalis Vazirgiannis, Guokan Shang
Abstract: As large language models are increasingly responsible for online content, concerns arise about the impact of repeatedly processing their own outputs. Inspired by the "broken telephone" effect in chained human communication, this study investigates whether LLMs similarly distort information through iterative generation. Through translation-based experiments, we find that distortion accumulates over time, influenced by language choice and chain complexity. While degradation is inevitable, it can be mitigated through strategic prompting techniques. These findings contribute to discussions on the long-term effects of AI-mediated information propagation, raising important questions about the reliability of LLM-generated content in iterative workflows.
Authors: Varshini Reddy, Craig W. Schmidt, Yuval Pinter, Chris Tanner
Abstract: Tokenization, a crucial initial step in natural language processing, is often assumed to benefit from larger training datasets. This paper investigates the impact of tokenizer training data sizes ranging from 1GB to 900GB. Our findings reveal diminishing returns as the data size increases, highlighting a practical limit on how much further scaling the training data can improve tokenization quality. We analyze this phenomenon and attribute the saturation effect to the constraints imposed by the pre-tokenization stage of tokenization. These results offer valuable insights for optimizing the tokenization process and highlight potential avenues for future research in tokenization algorithms.
Authors: Shangyin Tan, Lakshya A Agrawal, Arnav Singhvi, Liheng Lai, Michael J Ryan, Dan Klein, Omar Khattab, Koushik Sen, Matei Zaharia
Abstract: Composing language models (LMs) into multi-step language programs and automatically optimizing their modular prompts is now a mainstream paradigm for building AI systems, but the tradeoffs in this space have only scarcely been studied before. We introduce LangProBe, the first large-scale benchmark for evaluating the architectures and optimization strategies for language programs, with over 2000 combinations of tasks, architectures, optimizers, and choices of LMs. Using LangProBe, we are the first to study the impact of program architectures and optimizers (and their compositions together and with different models) on tradeoffs of quality and cost. We find that optimized language programs offer strong cost--quality Pareto improvement over raw calls to models, but simultaneously demonstrate that human judgment (or empirical decisions) about which compositions to pursue is still necessary for best performance. We will open source the code and evaluation data for LangProBe.
Authors: Guanzheng Chen, Qilong Feng, Jinjie Ni, Xin Li, Michael Qizhe Shieh
Abstract: The emergence of long-context large language models (LLMs) offers a promising alternative to traditional retrieval-augmented generation (RAG) for processing extensive documents. However, the computational overhead of long-context inference, particularly in managing key-value (KV) caches, presents significant efficiency challenges. While Speculative Decoding (SD) traditionally accelerates inference using smaller draft models, its effectiveness diminishes substantially in long-context scenarios due to memory-bound KV cache operations. We present Retrieval-Augmented Speculative Decoding (RAPID), which leverages RAG for both accelerating and enhancing generation quality in long-context inference. RAPID introduces the RAG drafter-a draft LLM operating on shortened retrieval contexts-to speculate on the generation of long-context target LLMs. Our approach enables a new paradigm where same-scale or even larger LLMs can serve as RAG drafters while maintaining computational efficiency. To fully leverage the potentially superior capabilities from stronger RAG drafters, we develop an inference-time knowledge transfer dynamic that enriches the target distribution by RAG. Extensive experiments on the LLaMA-3.1 and Qwen2.5 backbones demonstrate that RAPID effectively integrates the strengths of both approaches, achieving significant performance improvements (e.g., from 39.33 to 42.83 on InfiniteBench for LLaMA-3.1-8B) with more than 2x speedups. Our analyses reveal that RAPID achieves robust acceleration beyond 32K context length and demonstrates superior generation quality in real-world applications.
Authors: Yukang Yang, Declan Campbell, Kaixuan Huang, Mengdi Wang, Jonathan Cohen, Taylor Webb
Abstract: Many recent studies have found evidence for emergent reasoning capabilities in large language models, but debate persists concerning the robustness of these capabilities, and the extent to which they depend on structured reasoning mechanisms. To shed light on these issues, we perform a comprehensive study of the internal mechanisms that support abstract rule induction in an open-source language model (Llama3-70B). We identify an emergent symbolic architecture that implements abstract reasoning via a series of three computations. In early layers, symbol abstraction heads convert input tokens to abstract variables based on the relations between those tokens. In intermediate layers, symbolic induction heads perform sequence induction over these abstract variables. Finally, in later layers, retrieval heads predict the next token by retrieving the value associated with the predicted abstract variable. These results point toward a resolution of the longstanding debate between symbolic and neural network approaches, suggesting that emergent reasoning in neural networks depends on the emergence of symbolic mechanisms.
Authors: Alan Ashworth, Munir Al-Dajani, Keegan Duchicela, Kiril Kafadarov, Allison Kurian, Othman Laraki, Amina Lazrak, Divneet Mandair, Wendy McKennon, Rebecca Miksad, Jayodita Sanghvi, Travis Zack
Abstract: Clinical decision-making depends on expert reasoning, which is guided by standardized, evidence-based guidelines. However, translating these guidelines into automated clinical decision support systems risks inaccuracy and importantly, loss of nuance. We share an application architecture, the Large Language Expert (LLE), that combines the flexibility and power of Large Language Models (LLMs) with the interpretability, explainability, and reliability of Expert Systems. LLMs help address key challenges of Expert Systems, such as integrating and codifying knowledge, and data normalization. Conversely, an Expert System-like approach helps overcome challenges with LLMs, including hallucinations, atomic and inexpensive updates, and testability. To highlight the power of the Large Language Expert (LLE) system, we built an LLE to assist with the workup of patients newly diagnosed with cancer. Timely initiation of cancer treatment is critical for optimal patient outcomes. However, increasing complexity in diagnostic recommendations has made it difficult for primary care physicians to ensure their patients have completed the necessary workup before their first visit with an oncologist. As with many real-world clinical tasks, these workups require the analysis of unstructured health records and the application of nuanced clinical decision logic. In this study, we describe the design & evaluation of an LLE system built to rapidly identify and suggest the correct diagnostic workup. The system demonstrated a high degree of clinical-level accuracy (>95%) and effectively addressed gaps identified in real-world data from breast and colon cancer patients at a large academic center.
Authors: Daniele Paliotta, Junxiong Wang, Matteo Pagliardini, Kevin Y. Li, Aviv Bick, J. Zico Kolter, Albert Gu, Fran\c{c}ois Fleuret, Tri Dao
Abstract: Recent advancements have demonstrated that the performance of large language models (LLMs) can be significantly enhanced by scaling computational resources at test time. A common strategy involves generating multiple Chain-of-Thought (CoT) trajectories and aggregating their outputs through various selection mechanisms. This raises a fundamental question: can models with lower complexity leverage their superior generation throughput to outperform similarly sized Transformers for a fixed computational budget? To address this question and overcome the lack of strong subquadratic reasoners, we distill pure and hybrid Mamba models from pretrained Transformers. Trained on only 8 billion tokens, our distilled models show strong performance and scaling on mathematical reasoning datasets while being much faster at inference for large batches and long sequences. Despite the zero-shot performance hit due to distillation, both pure and hybrid Mamba models can scale their coverage and accuracy performance past their Transformer teacher models under fixed time budgets, opening a new direction for scaling inference compute.
Authors: Yi Jing, Zijun Yao, Lingxu Ran, Hongzhu Guo, Xiaozhi Wang, Lei Hou, Juanzi Li
Abstract: Large language models (LLMs) excel in tasks that require complex linguistic abilities, such as reference disambiguation and metaphor recognition/generation. Although LLMs possess impressive capabilities, their internal mechanisms for processing and representing linguistic knowledge remain largely opaque. Previous work on linguistic mechanisms has been limited by coarse granularity, insufficient causal analysis, and a narrow focus. In this study, we present a systematic and comprehensive causal investigation using sparse auto-encoders (SAEs). We extract a wide range of linguistic features from six dimensions: phonetics, phonology, morphology, syntax, semantics, and pragmatics. We extract, evaluate, and intervene on these features by constructing minimal contrast datasets and counterfactual sentence datasets. We introduce two indices-Feature Representation Confidence (FRC) and Feature Intervention Confidence (FIC)-to measure the ability of linguistic features to capture and control linguistic phenomena. Our results reveal inherent representations of linguistic knowledge in LLMs and demonstrate the potential for controlling model outputs. This work provides strong evidence that LLMs possess genuine linguistic knowledge and lays the foundation for more interpretable and controllable language modeling in future research.
Authors: Kai Zhang, Rui Zhu, Shutian Ma, Jingwei Xiong, Yejin Kim, Fabricio Murai, Xiaozhong Liu
Abstract: Drug discovery is a critical task in biomedical natural language processing (NLP), yet explainable drug discovery remains underexplored. Meanwhile, large language models (LLMs) have shown remarkable abilities in natural language understanding and generation. Leveraging LLMs for explainable drug discovery has the potential to improve downstream tasks and real-world applications. In this study, we utilize open-source drug knowledge graphs, clinical trial data, and PubMed publications to construct a comprehensive dataset for the explainable drug discovery task, named \textbf{expRxRec}. Furthermore, we introduce \textbf{KEDRec-LM}, an instruction-tuned LLM which distills knowledge from rich medical knowledge corpus for drug recommendation and rationale generation. To encourage further research in this area, we will publicly release\footnote{A copy is attached with this submission} both the dataset and KEDRec-LM.
Authors: Kuan Lok Zhou, Jiayi Chen, Siddharth Suresh, Reuben Narad, Timothy T. Rogers, Lalit K Jain, Robert D Nowak, Bob Mankoff, Jifan Zhang
Abstract: Large Language Models (LLMs) have shown significant limitations in understanding creative content, as demonstrated by Hessel et al. (2023)'s influential work on the New Yorker Cartoon Caption Contest (NYCCC). Their study exposed a substantial gap between LLMs and humans in humor comprehension, establishing that understanding and evaluating creative content is key challenge in AI development. We revisit this challenge by decomposing humor understanding into three components and systematically improve each: enhancing visual understanding through improved annotation, utilizing LLM-generated humor reasoning and explanations, and implementing targeted alignment with human preference data. Our refined approach achieves 82.4% accuracy in caption ranking, singificantly improving upon the previous 67% benchmark and matching the performance of world-renowned human experts in this domain. Notably, while attempts to mimic subgroup preferences through various persona prompts showed minimal impact, model finetuning with crowd preferences proved remarkably effective. These findings reveal that LLM limitations in creative judgment can be effectively addressed through focused alignment to specific subgroups and individuals. Lastly, we propose the position that achieving artificial general intelligence necessitates systematic collection of human preference data across creative domains. We advocate that just as human creativity is deeply influenced by individual and cultural preferences, training LLMs with diverse human preference data may be essential for developing true creative understanding.
Authors: Ryan C. Barron, Maksim E. Eren, Olga M. Serafimova, Cynthia Matuszek, Boian S. Alexandrov
Abstract: Agentic Generative AI, powered by Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG), Knowledge Graphs (KGs), and Vector Stores (VSs), represents a transformative technology applicable to specialized domains such as legal systems, research, recommender systems, cybersecurity, and global security, including proliferation research. This technology excels at inferring relationships within vast unstructured or semi-structured datasets. The legal domain here comprises complex data characterized by extensive, interrelated, and semi-structured knowledge systems with complex relations. It comprises constitutions, statutes, regulations, and case law. Extracting insights and navigating the intricate networks of legal documents and their relations is crucial for effective legal research. Here, we introduce a generative AI system that integrates RAG, VS, and KG, constructed via Non-Negative Matrix Factorization (NMF), to enhance legal information retrieval and AI reasoning and minimize hallucinations. In the legal system, these technologies empower AI agents to identify and analyze complex connections among cases, statutes, and legal precedents, uncovering hidden relationships and predicting legal trends-challenging tasks that are essential for ensuring justice and improving operational efficiency. Our system employs web scraping techniques to systematically collect legal texts, such as statutes, constitutional provisions, and case law, from publicly accessible platforms like Justia. It bridges the gap between traditional keyword-based searches and contextual understanding by leveraging advanced semantic representations, hierarchical relationships, and latent topic discovery. This framework supports legal document clustering, summarization, and cross-referencing, for scalable, interpretable, and accurate retrieval for semi-structured data while advancing computational law and AI.
Authors: Lars Benedikt Kaesberg, Jonas Becker, Jan Philip Wahle, Terry Ruas, Bela Gipp
Abstract: Much of the success of multi-agent debates depends on carefully choosing the right parameters. Among them, the decision-making protocol stands out. Systematic comparison of decision protocols is difficult because studies alter multiple discussion parameters beyond the protocol. So far, it has been largely unknown how decision-making addresses the challenges of different tasks. This work systematically evaluates the impact of seven decision protocols (e.g., majority voting, unanimity consensus). We change only one variable at a time (i.e., decision protocol) to analyze how different methods affect the collaboration between agents and test different protocols on knowledge (MMLU, MMLU-Pro, GPQA) and reasoning datasets (StrategyQA, MuSR, SQuAD 2.0). Our results show that voting protocols improve performance by 13.2% in reasoning tasks and consensus protocols by 2.8% in knowledge tasks over the other decision protocol. Increasing the number of agents improves performance, while more discussion rounds before voting reduces it. To improve decision-making by increasing answer diversity, we propose two new methods, All-Agents Drafting (AAD) and Collective Improvement (CI). Our methods improve task performance by up to 3.3% with AAD and up to 7.4% with CI. This work demonstrates the importance of decision-making in multi-agent debates beyond scaling.
Authors: Konstantina Christakopoulou, Iris Qu, John Canny, Andrew Goodridge, Cj Adams, Minmin Chen, Maja Matari\'c
Abstract: The language generation and reasoning capabilities of large language models (LLMs) have enabled conversational systems with impressive performance in a variety of tasks, from code generation, to composing essays, to passing STEM and legal exams, to a new paradigm for knowledge search. Besides those short-term use applications, LLMs are increasingly used to help with real-life goals or tasks that take a long time to complete, involving multiple sessions across days, weeks, months, or even years. Thus to enable conversational systems for long term interactions and tasks, we need language-based agents that can plan for long horizons. Traditionally, such capabilities were addressed by reinforcement learning agents with hierarchical planning capabilities. In this work, we explore a novel architecture where the LLM acts as the meta-controller deciding the agent's next macro-action, and tool use augmented LLM-based option policies execute the selected macro-action. We instantiate this framework for a specific set of macro-actions enabling adaptive planning for users' personal plans through conversation and follow-up questions collecting user feedback. We show how this paradigm can be applicable in scenarios ranging from tutoring for academic and non-academic tasks to conversational coaching for personal health plans.
Authors: Anton Alyakin, Jaden Stryker, Daniel Alexander Alber, Karl L. Sangwon, Brandon Duderstadt, Akshay Save, David Kurland, Spencer Frome, Shrutika Singh, Jeff Zhang, Eunice Yang, Ki Yun Park, Cordelia Orillac, Aly A. Valliani, Sean Neifert, Albert Liu, Aneek Patel, Christopher Livia, Darryl Lau, Ilya Laufer, Peter A. Rozman, Eveline Teresa Hidalgo, Howard Riina, Rui Feng, Todd Hollon, Yindalon Aphinyanaphongs, John G. Golfinos, Laura Snyder, Eric Leuthardt, Douglas Kondziolka, Eric Karl Oermann
Abstract: Research in AI for Science often focuses on using AI technologies to augment components of the scientific process, or in some cases, the entire scientific method; how about AI for scientific publications? Peer-reviewed journals are foundational repositories of specialized knowledge, written in discipline-specific language that differs from general Internet content used to train most large language models (LLMs) and vision-language models (VLMs). We hypothesized that by combining a family of scientific journals with generative AI models, we could invent novel tools for scientific communication, education, and clinical care. We converted 23,000 articles from Neurosurgery Publications into a multimodal database - NeuroPubs - of 134 million words and 78,000 image-caption pairs to develop six datasets for building AI models. We showed that the content of NeuroPubs uniquely represents neurosurgery-specific clinical contexts compared with broader datasets and PubMed. For publishing, we employed generalist VLMs to automatically generate graphical abstracts from articles. Editorial board members rated 70% of these as ready for publication without further edits. For education, we generated 89,587 test questions in the style of the ABNS written board exam, which trainee and faculty neurosurgeons found indistinguishable from genuine examples 54% of the time. We used these questions alongside a curriculum learning process to track knowledge acquisition while training our 34 billion-parameter VLM (CNS-Obsidian). In a blinded, randomized controlled trial, we demonstrated the non-inferiority of CNS-Obsidian to GPT-4o (p = 0.1154) as a diagnostic copilot for a neurosurgical service. Our findings lay a novel foundation for AI with Science and establish a framework to elevate scientific communication using state-of-the-art generative artificial intelligence while maintaining rigorous quality standards.
Authors: Mingsheng Cai, Jiuming Jiang, Wenhao Huang, Che Liu, Rossella Arcucci
Abstract: Cardiovascular diseases are a leading cause of death and disability worldwide. Electrocardiogram (ECG) recordings are critical for diagnosing and monitoring cardiac health, but obtaining large-scale annotated ECG datasets is labor-intensive and time-consuming. Recent ECG Self-Supervised Learning (eSSL) methods mitigate this by learning features without extensive labels but fail to capture fine-grained clinical semantics and require extensive task-specific fine-tuning. To address these challenges, we propose $\textbf{SuPreME}$, a $\textbf{Su}$pervised $\textbf{Pre}$-training framework for $\textbf{M}$ultimodal $\textbf{E}$CG representation learning. SuPreME applies Large Language Models (LLMs) to extract structured clinical entities from free-text ECG reports, filter out noise and irrelevant content, enhance clinical representation learning, and build a high-quality, fine-grained labeled dataset. By using text-based cardiac queries instead of traditional categorical labels, SuPreME enables zero-shot classification of unseen diseases without additional fine-tuning. We evaluate SuPreME on six downstream datasets covering 127 cardiac conditions, achieving superior zero-shot AUC performance over state-of-the-art eSSL and multimodal methods by over 1.96\%. Results demonstrate the effectiveness of SuPreME in leveraging structured, clinically relevant knowledge for high-quality ECG representations. All code and data will be released upon acceptance.
Authors: Zhangdie Yuan, Zifeng Ding, Andreas Vlachos
Abstract: Forecasting is an important task in many domains, such as technology and economics. However existing forecasting benchmarks largely lack comprehensive confidence assessment, focus on limited question types, and often consist of artificial questions that do not align with real-world human forecasting needs. To address these gaps, we introduce FOReCAst (Future Outcome Reasoning and Confidence Assessment), a benchmark that evaluates models' ability to make predictions and their confidence in them. FOReCAst spans diverse forecasting scenarios involving Boolean questions, timeframe prediction, and quantity estimation, enabling a comprehensive evaluation of both prediction accuracy and confidence calibration for real-world applications.
Authors: Toan Tran, Ruixuan Liu, Li Xiong
Abstract: Large language models (LLMs) have become the backbone of modern natural language processing but pose privacy concerns about leaking sensitive training data. Membership inference attacks (MIAs), which aim to infer whether a sample is included in a model's training dataset, can serve as a foundation for broader privacy threats. Existing defenses designed for traditional classification models do not account for the sequential nature of text data. As a result, they either require significant computational resources or fail to effectively mitigate privacy risks in LLMs. In this work, we propose a lightweight yet effective empirical privacy defense for protecting training data of language modeling by leveraging the token-specific characteristics. By analyzing token dynamics during training, we propose a token selection strategy that categorizes tokens into hard tokens for learning and memorized tokens for unlearning. Subsequently, our training-phase defense optimizes a novel dual-purpose token-level loss to achieve a Pareto-optimal balance between utility and privacy. Extensive experiments demonstrate that our approach not only provides strong protection against MIAs but also improves language modeling performance by around 10\% across various LLM architectures and datasets compared to the baselines.
Authors: Hojae Han, Seung-won Hwang, Rajhans Samdani, Yuxiong He
Abstract: Large language models (LLMs) have proven invaluable for code generation, particularly in interactive settings. However, existing code generation benchmarks fail to capture the diverse feedback encountered in multi-turn interactions, limiting our ability to evaluate LLMs in these contexts. To address this gap, we present a set of novel benchmarks that explicitly model the quality of feedback provided to code generation LLMs. Our contributions are threefold: First, we introduce CONVCODEWORLD, a novel and reproducible environment for benchmarking interactive code generation. CONVCODEWORLD simulates 9 distinct interactive code generation scenarios while systematically combining three types of feedback: (a) compilation feedback; (b) execution feedback with varying test coverage; (c) verbal feedback generated by GPT-4o with different levels of expertise. Second, we introduce CONVCODEBENCH, a fast, static version of benchmark that uses pre-generated feedback logs, eliminating the need for costly dynamic verbal feedback generation while maintaining strong Spearman's rank correlations (0.82 to 0.99) with CONVCODEWORLD. Third, extensive evaluations of both closed-source and open-source LLMs including R1-Distill on CONVCODEWORLD reveal key insights: (a) LLM performance varies significantly based on the feedback provided; (b) Weaker LLMs, with sufficient feedback, can outperform single-turn results of state-of-the-art LLMs without feedback; (c) Training on a specific feedback combination can limit an LLM's ability to utilize unseen combinations; (d) LLMs solve problems in fewer turns (high MRR) may not solve as many problems overall (high Recall), and vice versa. All implementations and benchmarks will be made publicly available at https://huggingface.co/spaces/ConvCodeWorld/ConvCodeWorld
URLs: https://huggingface.co/spaces/ConvCodeWorld/ConvCodeWorld
Authors: Sibo Yi, Tianshuo Cong, Xinlei He, Qi Li, Jiaxing Song
Abstract: Small language models (SLMs) have become increasingly prominent in the deployment on edge devices due to their high efficiency and low computational cost. While researchers continue to advance the capabilities of SLMs through innovative training strategies and model compression techniques, the security risks of SLMs have received considerably less attention compared to large language models (LLMs).To fill this gap, we provide a comprehensive empirical study to evaluate the security performance of 13 state-of-the-art SLMs under various jailbreak attacks. Our experiments demonstrate that most SLMs are quite susceptible to existing jailbreak attacks, while some of them are even vulnerable to direct harmful prompts.To address the safety concerns, we evaluate several representative defense methods and demonstrate their effectiveness in enhancing the security of SLMs. We further analyze the potential security degradation caused by different SLM techniques including architecture compression, quantization, knowledge distillation, and so on. We expect that our research can highlight the security challenges of SLMs and provide valuable insights to future work in developing more robust and secure SLMs.
Authors: Hongseok Oh, Wonseok Hwang
Abstract: Recently, Large Vision-Language Models (LVLMs) show remarkable performance across various domains. However, these models suffer from object hallucination. This study revisits the previous claim that the primary cause of such hallucination lies in the limited representational capacity of the vision encoder. Our analysis reveals that the capacity of the vision encoder itself is already enough for detecting object hallucination. Based on this insight, we propose a Fine-grained CLIPScore (F-CLIPScore), a simple yet effective evaluation metric that enhances object-level granularity by incorporating text embeddings at the noun phrase level. Evaluations on the OHD-Caps benchmark show that F-CLIPScore significantly outperforms conventional CLIPScore in accuracy by a large margin of 39.6% without additional training. We further validate F-CLIPScore by showing that LVLM trained with the data filtered using F-CLIPScore exhibits reduced hallucination.
Authors: Zexiong Ma, Chao Peng, Pengfei Gao, Xiangxin Meng, Yanzhen Zou, Bing Xie
Abstract: Mainstream issue-resolving frameworks predominantly rely on commercial models, leading to high costs and privacy concerns. Existing training approaches for issue resolving struggle with poor generalization and fail to fully leverage open-source development resources. We propose Subtask-oriented Reinforced Fine-Tuning (SoRFT), a novel training approach to enhance the issue resolving capability of LLMs. We decomposes issue resolving into structured subtasks: file localization, function localization, line localization, and code edit generation. SoRFT consists of two training stages: (1) rejection-sampled supervised fine-tuning, Chain of Thought (CoT) data is filtered using ground-truth before fine-tuning the LLM, and (2) rule-based reinforcement learning, which leverages PPO with ground-truth based rewards. We evaluate the SoRFT-trained model on SWE-Bench Verified and SWE-Bench Lite, achieving state-of-the-art (SOTA) performance among open-source models (e.g., resolve 21.4% issues on SWE-Bench Verified with SoRFT-Qwen-7B). The experimental results demonstrate that SoRFT significantly enhances issue-resolving performance, improves model generalization, and provides a cost-efficient alternative to commercial models.
Authors: Max M. Lang, Sol Eskenazi
Abstract: Telephone surveys remain a valuable tool for gathering insights but typically require substantial resources in training and coordinating human interviewers. This work presents an AI-driven telephone survey system integrating text-to-speech (TTS), a large language model (LLM), and speech-to-text (STT) that mimics the versatility of human-led interviews on scale. We tested the system across two populations, a pilot study in the United States (n = 75) and a large-scale deployment in Peru (n = 2,739), inviting participants via web-based links and contacting them via direct phone calls. The AI agent successfully administered open-ended and closed-ended questions, handled basic clarifications, and dynamically navigated branching logic, allowing fast large-scale survey deployment without interviewer recruitment or training. Our findings demonstrate that while the AI system's probing for qualitative depth was more limited than human interviewers, overall data quality approached human-led standards for structured items. This study represents one of the first successful large-scale deployments of an LLM-based telephone interviewer in a real-world survey context. The AI-powered telephone survey system has the potential for expanding scalable, consistent data collecting across market research, social science, and public opinion studies, thus improving operational efficiency while maintaining appropriate data quality for research.
Authors: Allen Schmaltz
Abstract: We solve the neural network robustness problem by adding Similarity (i.e., correctly predicted depth-matches into training)-awareness and Distance-to-training-distribution-awareness to the existing output Magnitude (i.e., decision-boundary)-awareness of the softmax function. The resulting sdm activation function provides strong signals of the relative epistemic (reducible) predictive uncertainty. We use this novel behavior to further address the complementary HCI problem of mapping the output to human-interpretable summary statistics over relevant partitions of a held-out calibration set. Estimates of prediction-conditional uncertainty are obtained via a parsimonious learned transform over the class-conditional empirical CDFs of the output of a final-layer sdm activation function. For decision-making and as an intrinsic model check, estimates of class-conditional accuracy are obtained by further partitioning the high-probability regions of this calibrated output into class-conditional, region-specific CDFs. The uncertainty estimates from sdm calibration are remarkably robust to test-time distribution shifts and out-of-distribution inputs; incorporate awareness of the effective sample size; provide estimates of uncertainty from the learning and data splitting processes; and are well-suited for selective classification and conditional branching for additional test-time compute based on the predictive uncertainty, as for selective LLM generation, routing, and composition over multiple models and retrieval. Finally, we construct sdm networks, LLMs with uncertainty-aware verification and interpretability-by-exemplar as intrinsic properties. We provide open-source software implementing these results.
Authors: Siqi Liu, Ian Gemp, Luke Marris, Georgios Piliouras, Nicolas Heess, Marc Lanctot
Abstract: Evaluation has traditionally focused on ranking candidates for a specific skill. Modern generalist models, such as Large Language Models (LLMs), decidedly outpace this paradigm. Open-ended evaluation systems, where candidate models are compared on user-submitted prompts, have emerged as a popular solution. Despite their many advantages, we show that the current Elo-based rating systems can be susceptible to and even reinforce biases in data, intentional or accidental, due to their sensitivity to redundancies. To address this issue, we propose evaluation as a 3-player game, and introduce novel game-theoretic solution concepts to ensure robustness to redundancy. We show that our method leads to intuitive ratings and provide insights into the competitive landscape of LLM development.
Authors: Liang Chen, Shuai Bai, Wenhao Chai, Weichu Xie, Haozhe Zhao, Leon Vinci, Junyang Lin, Baobao Chang
Abstract: The field of advanced text-to-image generation is witnessing the emergence of unified frameworks that integrate powerful text encoders, such as CLIP and T5, with Diffusion Transformer backbones. Although there have been efforts to control output images with additional conditions, like canny and depth map, a comprehensive framework for arbitrary text-image interleaved control is still lacking. This gap is especially evident when attempting to merge concepts or visual elements from multiple images in the generation process. To mitigate the gap, we conducted preliminary experiments showing that large multimodal models (LMMs) offer an effective shared representation space, where image and text can be well-aligned to serve as a condition for external diffusion models. Based on this discovery, we propose Dream Engine, an efficient and unified framework designed for arbitrary text-image interleaved control in image generation models. Building on powerful text-to-image models like SD3.5, we replace the original text-only encoders by incorporating versatile multimodal information encoders such as QwenVL. Our approach utilizes a two-stage training paradigm, consisting of joint text-image alignment and multimodal interleaved instruction tuning. Our experiments demonstrate that this training method is effective, achieving a 0.69 overall score on the GenEval benchmark, and matching the performance of state-of-the-art text-to-image models like SD3.5 and FLUX.
Authors: Kaustubh Vyas, Damien Graux, S\'ebastien Montella, Pavlos Vougiouklis, Ruofei Lai, Keshuang Li, Yang Ren, Jeff Z. Pan
Abstract: In recent advancements, large language models (LLMs) have exhibited proficiency in code generation and chain-of-thought reasoning, laying the groundwork for tackling automatic formal planning tasks. This study evaluates the potential of LLMs to understand and generate Planning Domain Definition Language (PDDL), an essential representation in artificial intelligence planning. We conduct an extensive analysis across 20 distinct models spanning 7 major LLM families, both commercial and open-source. Our comprehensive evaluation sheds light on the zero-shot LLM capabilities of parsing, generating, and reasoning with PDDL. Our findings indicate that while some models demonstrate notable effectiveness in handling PDDL, others pose limitations in more complex scenarios requiring nuanced planning knowledge. These results highlight the promise and current limitations of LLMs in formal planning tasks, offering insights into their application and guiding future efforts in AI-driven planning paradigms.
Authors: Parul Awasthy, Aashka Trivedi, Yulong Li, Mihaela Bornea, David Cox, Abraham Daniels, Martin Franz, Gabe Goodhart, Bhavani Iyer, Vishwajeet Kumar, Luis Lastras, Scott McCarley, Rudra Murthy, Vignesh P, Sara Rosenthal, Salim Roukos, Jaydeep Sen, Sukriti Sharma, Avirup Sil, Kate Soule, Arafat Sultan, Radu Florian
Abstract: We introduce the Granite Embedding models, a family of encoder-based embedding models designed for retrieval tasks, spanning dense-retrieval and sparse retrieval architectures, with both English and Multilingual capabilities. This report provides the technical details of training these highly effective 12 layer embedding models, along with their efficient 6 layer distilled counterparts. Extensive evaluations show that the models, developed with techniques like retrieval oriented pretraining, contrastive finetuning, knowledge distillation, and model merging significantly outperform publicly available models of similar sizes on both internal IBM retrieval and search tasks, and have equivalent performance on widely used information retrieval benchmarks, while being trained on high-quality data suitable for enterprise use. We publicly release all our Granite Embedding models under the Apache 2.0 license, allowing both research and commercial use at https://huggingface.co/collections/ibm-granite.
Authors: Nathaniel Hoy, Theodora Koulouri
Abstract: Fake news poses global risks by influencing elections and spreading misinformation, making detection critical. Existing NLP and supervised Machine Learning methods perform well under cross-validation but struggle to generalise across datasets, even within the same domain. This issue stems from coarsely labelled training data, where articles are labelled based on their publisher, introducing biases that token-based models like TF-IDF and BERT are sensitive to. While Large Language Models (LLMs) offer promise, their application in fake news detection remains limited. This study demonstrates that meaningful features can still be extracted from coarsely labelled data to improve real-world robustness. Stylistic features-lexical, syntactic, and semantic-are explored due to their reduced sensitivity to dataset biases. Additionally, novel social-monetisation features are introduced, capturing economic incentives behind fake news, such as advertisements, external links, and social media elements. The study trains on the coarsely labelled NELA 2020-21 dataset and evaluates using the manually labelled Facebook URLs dataset, a gold standard for generalisability. Results highlight the limitations of token-based models trained on biased data and contribute to the scarce evidence on LLMs like LLaMa in this field. Findings indicate that stylistic and social-monetisation features offer more generalisable predictions than token-based methods and LLMs. Statistical and permutation feature importance analyses further reveal their potential to enhance performance and mitigate dataset biases, providing a path forward for improving fake news detection.
Authors: Jinghao Feng, Qiaoyu Zheng, Chaoyi Wu, Ziheng Zhao, Ya Zhang, Yanfeng Wang, Weidi Xie
Abstract: Agentic AI systems have gained significant attention for their ability to autonomously perform complex tasks. However, their reliance on well-prepared tools limits their applicability in the medical domain, which requires to train specialized models. In this paper, we make three contributions: (i) We present M3Builder, a novel multi-agent system designed to automate machine learning (ML) in medical imaging. At its core, M3Builder employs four specialized agents that collaborate to tackle complex, multi-step medical ML workflows, from automated data processing and environment configuration to self-contained auto debugging and model training. These agents operate within a medical imaging ML workspace, a structured environment designed to provide agents with free-text descriptions of datasets, training codes, and interaction tools, enabling seamless communication and task execution. (ii) To evaluate progress in automated medical imaging ML, we propose M3Bench, a benchmark comprising four general tasks on 14 training datasets, across five anatomies and three imaging modalities, covering both 2D and 3D data. (iii) We experiment with seven state-of-the-art large language models serving as agent cores for our system, such as Claude series, GPT-4o, and DeepSeek-V3. Compared to existing ML agentic designs, M3Builder shows superior performance on completing ML tasks in medical imaging, achieving a 94.29% success rate using Claude-3.7-Sonnet as the agent core, showing huge potential towards fully automated machine learning in medical imaging.
Authors: Nazarii Drushchak, Vladyslava Tyshchenko, Nataliya Polyakovska
Abstract: The growth of Educational Technology (EdTech) has enabled highly personalized learning experiences through Artificial Intelligence (AI)-based recommendation systems tailored to each student needs. However, these systems can unintentionally introduce biases, potentially limiting fair access to learning resources. This study presents a recommendation system for K-12 students, combining graph-based modeling and matrix factorization to provide personalized suggestions for extracurricular activities, learning resources, and volunteering opportunities. To address fairness concerns, the system includes a framework to detect and reduce biases by analyzing feedback across protected student groups. This work highlights the need for continuous monitoring in educational recommendation systems to support equitable, transparent, and effective learning opportunities for all students.
Authors: Albert Gong, Kamil\.e Stankevi\v{c}i\=ut\.e, Chao Wan, Anmol Kabra, Raphael Thesmar, Johann Lee, Julius Klenke, Carla P. Gomes, Kilian Q. Weinberger
Abstract: High-quality benchmarks are essential for evaluating reasoning and retrieval capabilities of large language models (LLMs). However, curating datasets for this purpose is not a permanent solution as they are prone to data leakage and inflated performance results. To address these challenges, we propose PhantomWiki: a pipeline to generate unique, factually consistent document corpora with diverse question-answer pairs. Unlike prior work, PhantomWiki is neither a fixed dataset, nor is it based on any existing data. Instead, a new PhantomWiki instance is generated on demand for each evaluation. We vary the question difficulty and corpus size to disentangle reasoning and retrieval capabilities respectively, and find that PhantomWiki datasets are surprisingly challenging for frontier LLMs. Thus, we contribute a scalable and data leakage-resistant framework for disentangled evaluation of reasoning, retrieval, and tool-use abilities. Our code is available at https://github.com/kilian-group/phantom-wiki.
Authors: Arnav Kumar Jain, Gonzalo Gonzalez-Pumariega, Wayne Chen, Alexander M Rush, Wenting Zhao, Sanjiban Choudhury
Abstract: We address the problem of code generation from multi-turn execution feedback. Existing methods either generate code without feedback or use complex, hierarchical reinforcement learning to optimize multi-turn rewards. We propose a simple yet scalable approach, $\mu$Code, that solves multi-turn code generation using only single-step rewards. Our key insight is that code generation is a one-step recoverable MDP, where the correct code can be recovered from any intermediate code state in a single turn. $\mu$Code iteratively trains both a generator to provide code solutions conditioned on multi-turn execution feedback and a verifier to score the newly generated code. Experimental evaluations show that our approach achieves significant improvements over the state-of-the-art baselines. We provide analysis of the design choices of the reward models and policy, and show the efficacy of $\mu$Code at utilizing the execution feedback. Our code is available at https://github.com/portal-cornell/muCode.
Authors: Jeffrey Yang Fan Chiang, Seungjae Lee, Jia-Bin Huang, Furong Huang, Yizheng Chen
Abstract: Recent advancements in Web AI agents have demonstrated remarkable capabilities in addressing complex web navigation tasks. However, emerging research shows that these agents exhibit greater vulnerability compared to standalone Large Language Models (LLMs), despite both being built upon the same safety-aligned models. This discrepancy is particularly concerning given the greater flexibility of Web AI Agent compared to standalone LLMs, which may expose them to a wider range of adversarial user inputs. To build a scaffold that addresses these concerns, this study investigates the underlying factors that contribute to the increased vulnerability of Web AI agents. Notably, this disparity stems from the multifaceted differences between Web AI agents and standalone LLMs, as well as the complex signals - nuances that simple evaluation metrics, such as success rate, often fail to capture. To tackle these challenges, we propose a component-level analysis and a more granular, systematic evaluation framework. Through this fine-grained investigation, we identify three critical factors that amplify the vulnerability of Web AI agents; (1) embedding user goals into the system prompt, (2) multi-step action generation, and (3) observational capabilities. Our findings highlights the pressing need to enhance security and robustness in AI agent design and provide actionable insights for targeted defense strategies.
Authors: Joseba Fernandez de Landa, Rodrigo Agerri
Abstract: The large majority of the research performed on stance detection has been focused on developing more or less sophisticated text classification systems, even when many benchmarks are based on social network data such as Twitter. This paper aims to take on the stance detection task by placing the emphasis not so much on the text itself but on the interaction data available on social networks. More specifically, we propose a new method to leverage social information such as friends and retweets by generating Relational Embeddings, namely, dense vector representations of interaction pairs. Our experiments on seven publicly available datasets and four different languages (Basque, Catalan, Italian, and Spanish) show that combining our relational embeddings with discriminative textual methods helps to substantially improve performance, obtaining state-of-the-art results for six out of seven evaluation settings, outperforming strong baselines based on Large Language Models, or other popular interaction-based approaches such as DeepWalk or node2vec.
Authors: Edward Gow-Smith, Danae S\'anchez Villegas
Abstract: In this paper we describe the University of Sheffield's submission to the AmericasNLP 2023 Shared Task on Machine Translation into Indigenous Languages which comprises the translation from Spanish to eleven indigenous languages. Our approach consists of extending, training, and ensembling different variations of NLLB-200. We use data provided by the organizers and data from various other sources such as constitutions, handbooks, news articles, and backtranslations generated from monolingual data. On the dev set, our best submission outperforms the baseline by 11% average chrF across all languages, with substantial improvements particularly for Aymara, Guarani and Quechua. On the test set, we achieve the highest average chrF of all the submissions, we rank first in four of the eleven languages, and at least one of our submissions ranks in the top 3 for all languages.
Authors: Edward Gow-Smith, Dylan Phelps, Harish Tayyar Madabushi, Carolina Scarton, Aline Villavicencio
Abstract: All existing transformer-based approaches to NLP using subword tokenisation algorithms encode whitespace (word boundary information) through the use of special space symbols (such as \#\# or \_) forming part of tokens. These symbols have been shown to a) lead to reduced morphological validity of tokenisations, and b) give substantial vocabulary redundancy. As such, removing these symbols has been shown to have a beneficial effect on the processing of morphologically complex words for transformer encoders in the pretrain-finetune paradigm. In this work, we explore whether word boundary information is at all useful to such models. In particular, we train transformer encoders across four different training scales, and investigate several alternative approaches to including word boundary information, evaluating on a range of tasks across different domains and problem set-ups: GLUE (for sentence-level classification), NER (for token-level classification), and two classification datasets involving complex words (Superbizarre and FLOTA). Overall, through an extensive experimental setup that includes the pre-training of 29 models, we find no substantial improvements from our alternative approaches, suggesting that modifying tokenisers to remove word boundary information isn't leading to a loss of useful information.
Authors: Zheheng Luo, Qianqian Xie, Sophia Ananiadou
Abstract: Factual inconsistency with source documents in automatically generated summaries can lead to misinformation or pose risks. Existing factual consistency (FC) metrics are constrained by their performance, efficiency, and explainability. Recent advances in Large language models (LLMs) have demonstrated remarkable potential in text evaluation but their effectiveness in assessing FC in summarization remains underexplored. Prior research has mostly focused on proprietary LLMs, leaving essential factors that affect their assessment capabilities unexplored. Additionally, current FC evaluation benchmarks are restricted to news articles, casting doubt on the generality of the FC methods tested on them. In this paper, we first address the gap by introducing TreatFact a dataset of LLM-generated summaries of clinical texts, annotated for FC by domain experts. Moreover, we benchmark 11 LLMs for FC evaluation across news and clinical domains and analyse the impact of model size, prompts, pre-training and fine-tuning data. Our findings reveal that despite proprietary models prevailing on the task, open-source LLMs lag behind. Nevertheless, there is potential for enhancing the performance of open-source LLMs through increasing model size, expanding pre-training data, and developing well-curated fine-tuning data. Experiments on TreatFact suggest that both previous methods and LLM-based evaluators are unable to capture factual inconsistencies in clinical summaries, posing a new challenge for FC evaluation.
Authors: Farnaz Kohankhaki, D. B. Emerson, Jacob-Junqi Tian, Laleh Seyyed-Kalantari, Faiza Khan Khattak
Abstract: Bias in large language models (LLMs) has many forms, from overt discrimination to implicit stereotypes. Counterfactual bias evaluation is a widely used approach to quantifying bias and often relies on template-based probes that explicitly state group membership. It measures whether the outcome of a task performed by an LLM is invariant to a change in group membership. In this work, we find that template-based probes can lead to unrealistic bias measurements. For example, LLMs appear to mistakenly cast text associated with White race as negative at higher rates than other groups. We hypothesize that this arises artificially via a mismatch between commonly unstated norms, in the form of markedness, in the pretraining text of LLMs (e.g., Black president vs. president) and templates used for bias measurement (e.g., Black president vs. White president). The findings highlight the potential misleading impact of varying group membership through explicit mention in counterfactual bias quantification.
Authors: Zhangdie Yuan, Eric Chamoun, Rami Aly, Chenxi Whitehouse, Andreas Vlachos
Abstract: This paper introduces PRobELM (Plausibility Ranking Evaluation for Language Models), a benchmark designed to assess language models' ability to discern more plausible from less plausible scenarios through their parametric knowledge. While benchmarks such as TruthfulQA emphasise factual accuracy or truthfulness, and others such as COPA explore plausible scenarios without explicitly incorporating world knowledge, PRobELM seeks to bridge this gap by evaluating models' capabilities to prioritise plausible scenarios that leverage world knowledge over less plausible alternatives. This design allows us to assess the potential of language models for downstream use cases such as literature-based discovery where the focus is on identifying information that is likely but not yet known. Our benchmark is constructed from a dataset curated from Wikidata edit histories, tailored to align the temporal bounds of the training data for the evaluated models. PRobELM facilitates the evaluation of language models across multiple prompting types, including statement, text completion, and question-answering. Experiments with 10 models of various sizes and architectures on the relationship between model scales, training recency, and plausibility performance, reveal that factual accuracy does not directly correlate with plausibility performance and that up-to-date training data enhances plausibility assessment across different model architectures.
Authors: Holli Sargeant, Ahmed Izzidien, Felix Steffek
Abstract: This paper addresses a critical gap in legal analytics by developing and applying a novel taxonomy for topic classification of summary judgment cases in the United Kingdom. Using a curated dataset of summary judgment cases, we use the Large Language Model Claude 3 Opus to explore functional topics and trends. We find that Claude 3 Opus correctly classified the topic with an accuracy of 87.13% and an F1 score of 0.87. The analysis reveals distinct patterns in the application of summary judgments across various legal domains. As case law in the United Kingdom is not originally labelled with keywords or a topic filtering option, the findings not only refine our understanding of the thematic underpinnings of summary judgments but also illustrate the potential of combining traditional and AI-driven approaches in legal classification. Therefore, this paper provides a new and general taxonomy for UK law. The implications of this work serve as a foundation for further research and policy discussions in the field of judicial administration and computational legal research methodologies.
Authors: Yuheng Chen, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao
Abstract: Large language models (LLMs) store extensive factual knowledge, but the mechanisms behind how they store and express this knowledge remain unclear. The Knowledge Neuron (KN) thesis is a prominent theory for explaining these mechanisms. This theory is based on the Knowledge Localization (KL) assumption, which suggests that a fact can be localized to a few knowledge storage units, namely knowledge neurons. However, this assumption has two limitations: first, it may be too rigid regarding knowledge storage, and second, it neglects the role of the attention module in knowledge expression. In this paper, we first re-examine the KL assumption and demonstrate that its limitations do indeed exist. To address these, we then present two new findings, each targeting one of the limitations: one focusing on knowledge storage and the other on knowledge expression. We summarize these findings as \textbf{Query Localization} (QL) assumption and argue that the KL assumption can be viewed as a simplification of the QL assumption. Based on QL assumption, we further propose the Consistency-Aware KN modification method, which improves the performance of knowledge modification, further validating our new assumption. We conduct 39 sets of experiments, along with additional visualization experiments, to rigorously confirm our conclusions. Code is available at https://github.com/heng840/KnowledgeLocalization.
Authors: Quan Mai, Susan Gauch, Douglas Adams, Miaoqing Huang
Abstract: Online debates involve a dynamic exchange of ideas over time, where participants need to actively consider their opponents' arguments, respond with counterarguments, reinforce their own points, and introduce more compelling arguments as the discussion unfolds. Modeling such a complex process is not a simple task, as it necessitates the incorporation of both sequential characteristics and the capability to capture interactions effectively. To address this challenge, we employ a sequence-graph approach. Building the conversation as a graph allows us to effectively model interactions between participants through directed edges. Simultaneously, the propagation of information along these edges in a sequential manner enables us to capture a more comprehensive representation of context. We also introduce a Sequence Graph Attention layer to illustrate the proposed information update scheme. The experimental results show that sequence graph networks achieve superior results to existing methods in online debates.
Authors: Praneeth Vadlapati
Abstract: Up-to-date and reliable language models are consistently sought after and are essential in various applications. Typically, models are trained on a fixed dataset and then deployed globally. However, the knowledge of the models becomes outdated. Enabling automatic updation of AI knowledge using web data involves significant concerns regarding the model's safety and quality due to a threat from unsafe and undesirable text across the web. The purity of new data was essential for updating knowledge of language models to maintain their reliability. This paper proposes AutoPureData, a system that automatically collects and purifies web data. The system loaded a sample of web data. Utilizing existing trusted AI models, it successfully eliminated unsafe text with an accuracy of 97% and undesirable text with an accuracy of 86%, demonstrating the system's effectiveness in purifying the data. The system ensures that only meaningful and safe text can be used to update LLM knowledge. The pure text was then optimized and stored in a vector database for future querying. It was found that LLM can fetch new data from the vector DB. The LLM writes the RAG query in English, even if the user's query is in another language, proving that the system can perform cross-lingual retrieval. This paper proposes a method to maintain the accuracy and relevance of up-to-date language models by ensuring that only purified data was used to update LLM knowledge. This work contributes to updating knowledge of chatbots using meaningful and safe text, enhancing their utility across various industries, and potentially reducing the risks associated with outputs caused by unsafe or impure data. Code is available at github.com/Pro-GenAI/AutoPureData.
Authors: Yuhui Xu, Zhanming Jie, Hanze Dong, Lei Wang, Xudong Lu, Aojun Zhou, Amrita Saha, Caiming Xiong, Doyen Sahoo
Abstract: Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications. However, their increased computational and memory demands present significant challenges, especially when handling long sequences. This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference. Unlike existing approaches that optimize the memory based on the sequence length, we identify substantial redundancy in the channel dimension of the KV cache, as indicated by an uneven magnitude distribution and a low-rank structure in the attention weights. In response, we propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels. Our approach not only maintains or enhances model accuracy but also achieves a reduction in KV cache memory costs by over 20% compared with vanilla KV cache eviction and quantization methods. For instance, ThinK integrated with KIVI can achieve a 2.8x reduction in peak memory usage while maintaining nearly the same quality, enabling up to a 5x increase in batch size when using a single GPU. Extensive evaluations on the LLaMA and Mistral models across various long-sequence datasets verified the efficiency of ThinK, establishing a new baseline algorithm for efficient LLM deployment without compromising performance. Our code has been made available at https://github.com/SalesforceAIResearch/ThinK.
Authors: Ruizhe Chen, Yichen Li, Jianfei Yang, Joey Tianyi Zhou, Jian Wu, Zuozhu Liu
Abstract: Generating fair and accurate predictions plays a pivotal role in deploying large language models (LLMs) in the real world. However, existing debiasing methods inevitably generate unfair or incorrect predictions as they are designed and evaluated to achieve parity across different social groups but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. In this paper, we first establish a new bias mitigation benchmark, BiaScope, which systematically assesses performance by leveraging newly constructed datasets and metrics on knowledge retention and generalization. Then, we propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases. FAST identifies the decisive layer responsible for storing social biases and then calibrates its outputs by integrating a small modular network, considering both bias mitigation and knowledge-preserving demands. Comprehensive experiments demonstrate that FAST surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and downstream predictions. This highlights the potential of fine-grained debiasing strategies to achieve fairness in LLMs.
Authors: Xiyu Liu, Zhengxiao Liu, Naibin Gu, Zheng Lin, Wanli Ma, Ji Xiang, Weiping Wang
Abstract: The storage and recall of factual associations in auto-regressive transformer language models (LMs) have drawn a great deal of attention, inspiring knowledge editing by directly modifying the located model weights. Most editing works achieve knowledge editing under the guidance of existing interpretations of knowledge recall that mainly focus on subject knowledge. However, these interpretations are seriously flawed, neglecting relation information and leading to the over-generalizing problem for editing. In this work, we discover a novel relation-focused perspective to interpret the knowledge recall of transformer LMs during inference and apply it on single knowledge editing to avoid over-generalizing. Experimental results on the dataset supplemented with a new R-Specificity criterion demonstrate that our editing approach significantly alleviates over-generalizing while remaining competitive on other criteria, breaking the domination of subject-focused editing for future research.
Authors: Tianqi Liu, Wei Xiong, Jie Ren, Lichang Chen, Junru Wu, Rishabh Joshi, Yang Gao, Jiaming Shen, Zhen Qin, Tianhe Yu, Daniel Sohn, Anastasiia Makarova, Jeremiah Liu, Yuan Liu, Bilal Piot, Abe Ittycheriah, Aviral Kumar, Mohammad Saleh
Abstract: Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. However, traditional RM training, which relies on response pairs tied to specific prompts, struggles to disentangle prompt-driven preferences from prompt-independent artifacts, such as response length and format. In this work, we expose a fundamental limitation of current RM training methods, where RMs fail to effectively distinguish between contextual signals and irrelevant artifacts when determining preferences. To address this, we introduce a causal framework that learns preferences independent of these artifacts and propose a novel data augmentation technique designed to eliminate them. Extensive experiments show that our approach successfully filters out undesirable artifacts, yielding a more robust reward model (RRM). Our RRM improves the performance of a pairwise reward model trained on Gemma-2-9b-it, on RewardBench, increasing accuracy from 80.61% to 84.15%. Additionally, we train two DPO policies using both the RM and RRM, demonstrating that the RRM significantly enhances DPO-aligned policies, improving MT-Bench scores from 7.27 to 8.31 and length-controlled win-rates in AlpacaEval-2 from 33.46% to 52.49%.
Authors: Yuntong Hu, Zhuofeng Li, Zheng Zhang, Chen Ling, Raasikh Kanjiani, Boxin Zhao, Liang Zhao
Abstract: Constructing taxonomies from citation graphs is essential for organizing scientific knowledge, facilitating literature reviews, and identifying emerging research trends. However, manual taxonomy construction is labor-intensive, time-consuming, and prone to human biases, often overlooking pivotal but less-cited papers. In this paper, to enable automatic hierarchical taxonomy generation from citation graphs, we propose HiGTL (Hierarchical Graph Taxonomy Learning), a novel end-to-end framework guided by human-provided instructions or preferred topics. Specifically, we propose a hierarchical citation graph clustering method that recursively groups related papers based on both textual content and citation structure, ensuring semantically meaningful and structurally coherent clusters. Additionally, we develop a novel taxonomy node verbalization strategy that iteratively generates central concepts for each cluster, leveraging a pre-trained large language model (LLM) to maintain semantic consistency across hierarchical levels. To further enhance performance, we design a joint optimization framework that fine-tunes both the clustering and concept generation modules, aligning structural accuracy with the quality of generated taxonomies. Extensive experiments demonstrate that HiGTL effectively produces coherent, high-quality taxonomies.
Authors: Yu Shang, Yu Li, Keyu Zhao, Likai Ma, Jiahe Liu, Fengli Xu, Yong Li
Abstract: Recent advancements in Large Language Models (LLMs) have led to a rapid growth of agentic systems capable of handling a wide range of complex tasks. However, current research largely relies on manual, task-specific design, limiting their adaptability to novel tasks. In this paper, we introduce a new research problem: Modularized LLM Agent Search (MoLAS). We propose a modular design space that abstracts existing LLM agent designs into four fundamental modules with uniform IO interface: Planning, Reasoning, Tool Use, and Memory. Building on this design space, we present a novel LLM agent search framework called AgentSquare, which introduces two core mechanisms, i.e., module evolution and recombination, to efficiently search for optimized LLM agents. To further accelerate the process, we design a performance predictor that uses in-context surrogate models to skip unpromising agent designs. Extensive experiments across six benchmarks, covering the diverse scenarios of web, embodied, tool use and game applications, show that AgentSquare substantially outperforms hand-crafted agents, achieving an average performance gain of 17.2% against best-known human designs. Moreover, AgentSquare can generate interpretable design insights, enabling a deeper understanding of agentic architecture and its impact on task performance. We believe that the modular design space and AgentSquare search framework offer a platform for fully exploiting the potential of prior successful designs and consolidating the collective efforts of research community. Code repo is available at https://github.com/tsinghua-fib-lab/AgentSquare.
Authors: Federico Barbero, Alex Vitvitskyi, Christos Perivolaropoulos, Razvan Pascanu, Petar Veli\v{c}kovi\'c
Abstract: Positional Encodings (PEs) are a critical component of Transformer-based Large Language Models (LLMs), providing the attention mechanism with important sequence-position information. One of the most popular types of encoding used today in LLMs are Rotary Positional Encodings (RoPE), that rotate the queries and keys based on their relative distance. A common belief is that RoPE is useful because it helps to decay token dependency as relative distance increases. In this work, we argue that this is unlikely to be the core reason. We study the internals of a trained Gemma 7B model to understand how RoPE is being used at a mechanical level. We find that Gemma learns to use RoPE to construct robust "positional" attention patterns by exploiting the highest frequencies. We also find that, in general, Gemma greatly prefers to use the lowest frequencies of RoPE, which we suspect are used to carry semantic information. We mathematically prove interesting behaviours of RoPE and conduct experiments to verify our findings, proposing a modification of RoPE that fixes some highlighted issues and improves performance. We believe that this work represents an interesting step in better understanding PEs in LLMs, which we believe holds crucial value for scaling LLMs to large sizes and context lengths.
Authors: Guoxiong Gao, Yutong Wang, Jiedong Jiang, Qi Gao, Zihan Qin, Tianyi Xu, Bin Dong
Abstract: Verifiable formal languages like Lean have profoundly impacted mathematical reasoning, particularly through the use of large language models (LLMs) for automated reasoning. A significant challenge in training LLMs for these formal languages is the lack of parallel datasets that align natural language with formal language proofs. To address this challenge, this paper introduces a novel framework for translating the Mathlib4 corpus (a unified library of mathematics in formal language Lean 4) into natural language. Building upon this, we employ a dual augmentation strategy that combines tactic-based and informal-based approaches, leveraging the Lean-jixia system, a Lean 4 analyzer. We present the results of this pipeline on Mathlib4 as Herald (Hierarchy and Retrieval-based Translated Lean Dataset). We also propose the Herald Translator, which is fine-tuned on Herald. Herald translator achieves a 93.2% accuracy (Pass@128) on formalizing statements in the miniF2F-test and a 22.5% accuracy on our internal graduate-level textbook dataset, outperforming InternLM2-Math-Plus-7B (74.0% and 7.5%) and TheoremLlama (50.1% and 4.0%). Furthermore, we propose a section-level translation framework for real-world applications. As a direct application of Herald translator, we have successfully translated a template section in the Stack project, marking a notable progress in the automatic formalization of graduate-level mathematical literature. Our model, along with the datasets, are open-sourced to the public.
Authors: Thao Do, Dinh Phu Tran, An Vo, Daeyoung Kim
Abstract: Extracting fine-grained OCR text from aged documents in diacritic languages remains challenging due to unexpected artifacts, time-induced degradation, and lack of datasets. While standalone spell correction approaches have been proposed, they show limited performance for historical documents due to numerous possible OCR error combinations and differences between modern and classical corpus distributions. We propose a method utilizing available content-focused ebooks as a reference base to correct imperfect OCR-generated text, supported by large language models. This technique generates high-precision pseudo-page-to-page labels for diacritic languages, where small strokes pose significant challenges in historical conditions. The pipeline eliminates various types of noise from aged documents and addresses issues such as missing characters, words, and disordered sequences. Our post-processing method, which generated a large OCR dataset of classical Vietnamese books, achieved a mean grading score of 8.72 on a 10-point scale. This outperformed the state-of-the-art transformer-based Vietnamese spell correction model, which scored 7.03 when evaluated on a sampled subset of the dataset. We also trained a baseline OCR model to assess and compare it with well-known engines. Experimental results demonstrate the strength of our baseline model compared to widely used open-source solutions. The resulting dataset will be released publicly to support future studies.
Authors: Ingeol Baek, Hwan Chang, Byeongjeong Kim, Jimin Lee, Hwanhee Lee
Abstract: Retrieval-Augmented Generation (RAG) enhances language models by retrieving and incorporating relevant external knowledge. However, traditional retrieve-and-generate processes may not be optimized for real-world scenarios, where queries might require multiple retrieval steps or none at all. In this paper, we propose a Probing-RAG, which utilizes the hidden state representations from the intermediate layers of language models to adaptively determine the necessity of additional retrievals for a given query. By employing a pre-trained prober, Probing-RAG effectively captures the model's internal cognition, enabling reliable decision-making about retrieving external documents. Experimental results across five open-domain QA datasets demonstrate that Probing-RAG outperforms previous methods while reducing the number of redundant retrieval steps.
Authors: Qin Liu, Fei Wang, Chaowei Xiao, Muhao Chen
Abstract: Existing preference alignment is a one-size-fits-all alignment mechanism, where the part of the large language model (LLM) parametric knowledge with non-preferred features is uniformly blocked to all the users. However, this part of knowledge can be useful to advanced users whose expertise qualifies them to handle these information. The one-size-fits-all alignment mechanism undermines LLM's utility for these qualified users. To address this problem, we propose SudoLM, a framework that lets LLMs learn access control over specific parametric knowledge for users with different credentials via authorization alignment. SudoLM allows authorized users to unlock their access to all the parametric knowledge with an assigned SUDO key while blocking access to non-qualified users. Experiments on two application scenarios demonstrate that SudoLM effectively controls the user's access to the parametric knowledge and maintains its general utility.
Authors: Mohamed Bayan Kmainasi, Ali Ezzat Shahroor, Maram Hasanain, Sahinur Rahman Laskar, Naeemul Hassan, Firoj Alam
Abstract: Large Language Models (LLMs) have demonstrated remarkable success as general-purpose task solvers across various fields. However, their capabilities remain limited when addressing domain-specific problems, particularly in downstream NLP tasks. Research has shown that models fine-tuned on instruction-based downstream NLP datasets outperform those that are not fine-tuned. While most efforts in this area have primarily focused on resource-rich languages like English and broad domains, little attention has been given to multilingual settings and specific domains. To address this gap, this study focuses on developing a specialized LLM, LlamaLens, for analyzing news and social media content in a multilingual context. To the best of our knowledge, this is the first attempt to tackle both domain specificity and multilinguality, with a particular focus on news and social media. Our experimental setup includes 18 tasks, represented by 52 datasets covering Arabic, English, and Hindi. We demonstrate that LlamaLens outperforms the current state-of-the-art (SOTA) on 23 testing sets, and achieves comparable performance on 8 sets. We make the models and resources publicly available for the research community (https://huggingface.co/collections/QCRI/llamalens-672f7e0604a0498c6a2f0fe9).
URLs: https://huggingface.co/collections/QCRI/llamalens-672f7e0604a0498c6a2f0fe9).
Authors: Hao Bai, Yi Ma
Abstract: Neurons in auto-regressive language models like GPT-2 can be interpreted by analyzing their activation patterns. Recent studies have shown that techniques such as dictionary learning, a form of post-hoc sparse coding, enhance this neuron-level interpretability. In our research, we are driven by the goal to fundamentally improve neural network interpretability by embedding sparse coding directly within the model architecture, rather than applying it as an afterthought. In our study, we introduce a white-box transformer-like architecture named Coding RAte TransformEr (CRATE), explicitly engineered to capture sparse, low-dimensional structures within data distributions. Our comprehensive experiments showcase significant improvements (up to 103% relative improvement) in neuron-level interpretability across a variety of evaluation metrics. Detailed investigations confirm that this enhanced interpretability is steady across different layers irrespective of the model size, underlining CRATE's robust performance in enhancing neural network interpretability. Further analysis shows that CRATE's increased interpretability comes from its enhanced ability to consistently and distinctively activate on relevant tokens. These findings point towards a promising direction for creating white-box foundation models that excel in neuron-level interpretation.
Authors: Taewhoo Lee, Chanwoong Yoon, Kyochul Jang, Donghyeon Lee, Minju Song, Hyunjae Kim, Jaewoo Kang
Abstract: Recent advancements in large language models (LLM) capable of processing extremely long texts highlight the need for a dedicated evaluation benchmark to assess their long-context capabilities. However, existing methods, like the needle-in-a-haystack test, do not effectively assess whether these models fully utilize contextual information, raising concerns about the reliability of current evaluation techniques. To thoroughly examine the effectiveness of existing benchmarks, we introduce a new metric called information coverage (IC), which quantifies the proportion of the input context necessary for answering queries. Our findings indicate that current benchmarks exhibit low IC; although the input context may be extensive, the actual usable context is often limited. To address this, we present ETHIC, a novel benchmark designed to assess LLMs' ability to leverage the entire context. Our benchmark comprises 1,986 test instances spanning four long-context tasks with high IC scores in the domains of books, debates, medicine, and law. Our evaluations reveal significant performance drops in contemporary LLMs, highlighting a critical challenge in managing long contexts. Our benchmark is available at https://github.com/dmis-lab/ETHIC.
Authors: Omer Shubi, Cfir Avraham Hadar, Yevgeni Berzak
Abstract: Readers can have different goals with respect to the text that they are reading. Can these goals be decoded from their eye movements over the text? In this work, we examine for the first time whether it is possible to distinguish between two types of common reading goals: information seeking and ordinary reading for comprehension. Using large-scale eye tracking data, we address this task with a wide range of models that cover different architectural and data representation strategies, and further introduce a new model ensemble. We find that transformer-based models with scanpath representations coupled with language modeling solve it most successfully, and that accurate predictions can be made in real time, long before the participant finished reading the text. We further introduce a new method for model performance analysis based on mixed effect modeling. Combining this method with rich textual annotations reveals key properties of textual items and participants that contribute to the difficulty of the task, and improves our understanding of the variability in eye movement patterns across the two reading regimes.
Authors: Viet Cuong Nguyen, Mohammad Taher, Dongwan Hong, Vinicius Konkolics Possobom, Vibha Thirunellayi Gopalakrishnan, Ekta Raj, Zihang Li, Heather J. Soled, Michael L. Birnbaum, Srijan Kumar, Munmun De Choudhury
Abstract: The rapid evolution of Large Language Models (LLMs) presents a promising solution to the global shortage of mental health professionals. However, their alignment with essential counseling competencies remains underexplored. We introduce CounselingBench, a novel NCMHCE-based benchmark evaluating 22 general-purpose and medical-finetuned LLMs across five key competencies. While frontier models surpass minimum aptitude thresholds, they fall short of expert-level performance, excelling in Intake, Assessment & Diagnosis but struggling with Core Counseling Attributes and Professional Practice & Ethics. Surprisingly, medical LLMs do not outperform generalist models in accuracy, though they provide slightly better justifications while making more context-related errors. These findings highlight the challenges of developing AI for mental health counseling, particularly in competencies requiring empathy and nuanced reasoning. Our results underscore the need for specialized, fine-tuned models aligned with core mental health counseling competencies and supported by human oversight before real-world deployment. Code and data associated with this manuscript can be found at: https://github.com/cuongnguyenx/CounselingBench
Authors: Yunqi Zhu, Wen Tang, Huayu Yang, Jinghao Niu, Liyang Dou, Yifan Gu, Yuanyuan Wu, Wensheng Zhang, Ying Sun, Xuebing Yang
Abstract: In this work, we leverage LLMs to produce medical qualification exam questions and the corresponding answers through few-shot prompts, investigating in-depth how LLMs meet the requirements in terms of coherence, evidence of statement, factual consistency, and professionalism etc. Utilizing a multicenter bidirectional anonymized database with respect to comorbid chronic diseases, named Elderly Comorbidity Medical Database (CECMed), we tasked LLMs with generating open-ended questions and answers based on a subset of sampled admission reports. For CECMed, the retrospective cohort includes patients enrolled from January 2010 to January 2022 while the prospective cohort from January 2023 to November 2023, with participants sourced from selected tertiary and community hospitals across the southern, northern, and central regions of China. A total of 8 widely used LLMs were used, including ERNIE 4, ChatGLM 4, Doubao, Hunyuan, Spark 4, Qwen, Conventional medical education requires sophisticated clinicians to formulate questions and answers based on prototypes from EHRs, which is heuristic and time-consuming. We found that mainstream LLMs could generate questions and answers with real-world EHRs at levels close to clinicians. Although current LLMs performed dissatisfactory in some aspects, medical students, interns and residents could reasonably make use of LLMs to facilitate understanding.
Authors: Jingyi Zhou, Senlin Luo, Haofan Chen
Abstract: Text emotion detection constitutes a crucial foundation for advancing artificial intelligence from basic comprehension to the exploration of emotional reasoning. Most existing emotion detection datasets rely on manual annotations, which are associated with high costs, substantial subjectivity, and severe label imbalances. This is particularly evident in the inadequate annotation of micro-emotions and the absence of emotional intensity representation, which fail to capture the rich emotions embedded in sentences and adversely affect the quality of downstream task completion. By proposing an all-labels and training-set label regression method, we map label values to energy intensity levels, thereby fully leveraging the learning capabilities of machine models and the interdependencies among labels to uncover multiple emotions within samples. This led to the establishment of the Emotion Quantization Network (EQN) framework for micro-emotion detection and annotation. Using five commonly employed sentiment datasets, we conducted comparative experiments with various models, validating the broad applicability of our framework within NLP machine learning models. Based on the EQN framework, emotion detection and annotation are conducted on the GoEmotions dataset. A comprehensive comparison with the results from Google literature demonstrates that the EQN framework possesses a high capability for automatic detection and annotation of micro-emotions. The EQN framework is the first to achieve automatic micro-emotion annotation with energy-level scores, providing strong support for further emotion detection analysis and the quantitative research of emotion computing.
Authors: Qiao Qiao, Yuepei Li, Qing Wang, Kang Zhou, Qi Li
Abstract: Knowledge graph completion (KGC) is a task of inferring missing triples based on existing Knowledge Graphs (KGs). Both structural and semantic information are vital for successful KGC. However, existing methods only use either the structural knowledge from the KG embeddings or the semantic information from pre-trained language models (PLMs), leading to suboptimal model performance. Moreover, since PLMs are not trained on KGs, directly using PLMs to encode triples may be inappropriate. To overcome these limitations, we propose a novel framework called Bridge, which jointly encodes structural and semantic information of KGs. Specifically, we strategically encode entities and relations separately by PLMs to better utilize the semantic knowledge of PLMs and enable structured representation learning via a structural learning principle. Furthermore, to bridge the gap between KGs and PLMs, we employ a self-supervised representation learning method called BYOL to fine-tune PLMs with two different views of a triple. Unlike BYOL, which uses augmentation methods to create two semantically similar views of the same image, potentially altering the semantic information. We strategically separate the triple into two parts to create different views, thus avoiding semantic alteration. Experiments demonstrate that Bridge outperforms the SOTA models on three benchmark datasets.
Authors: Mianqiu Huang, Xiaoran Liu, Shaojun Zhou, Mozhi Zhang, Qipeng Guo, Linyang Li, Chenkun Tan, Yang Gao, Pengyu Wang, Linlin Li, Qun Liu, Yaqian Zhou, Xipeng Qiu, Xuanjing Huang
Abstract: Recent advancements in model architectures and length extrapolation techniques have significantly extended the context length of large language models (LLMs), paving the way for their application in increasingly complex tasks. However, despite the growing capabilities of long-context LLMs, the safety issues in long-context scenarios remain underexplored. While safety alignment in short context has been widely studied, the safety concerns of long-context LLMs have not been adequately addressed. In this work, we introduce \textbf{LongSafety}, a comprehensive safety alignment dataset for long-context LLMs, containing 10 tasks and 17k samples, with an average length of 40.9k tokens. Our experiments demonstrate that training with LongSafety can enhance long-context safety performance while enhancing short-context safety and preserving general capabilities. Furthermore, we demonstrate that long-context safety does not equal long-context alignment with short-context safety data and LongSafety has generalizing capabilities in context length and long-context safety scenarios.
Authors: Howard Chen, Jiayi Geng, Adithya Bhaskar, Dan Friedman, Danqi Chen
Abstract: As new knowledge rapidly accumulates, language models (LMs) with pretrained knowledge quickly become obsolete. A common approach to updating LMs is fine-tuning them directly on new knowledge. However, recent studies have shown that fine-tuning for memorization may be ineffective in storing knowledge or may exacerbate hallucinations. In this work, we introduce a setting we call continual memorization, where a model must memorize and retain a set of factoids through multiple stages of fine-tuning on subsequent datasets. We characterized the forgetting patterns through extensive experiments and show that LMs widely suffer from forgetting, especially when needing to memorize factoids in the second stage. We posit that forgetting can be alleviated by modifying training dynamics: (1) protecting the memorization process when learning factoids or (2) reducing interference from subsequent training stages. Intriguingly, we find that mixing randomly generated word sequences or generic data sampled from pretraining corpora at different training stages effectively mitigates forgetting REMIX: Random and Generic Data Mixing). REMIX can recover performance from severe forgetting, outperforming replay methods and other continual learning baselines. We analyze how REMIX influences the learning process and find that robust memorization follows a distinct pattern: the model stores factoids in earlier layers than usual and diversifies the layers that retain them, which results in easier recall and manipulate of the learned factoids.
Authors: Wenbo Zhang, Aditya Majumdar, Amulya Yadav
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across various NLP tasks but struggle with code-mixed (or code-switched) language understanding. For example, prior work benchmarking the performance of multilingual LLMs on code-mixed translation tasks has demonstrated that current state-of-the-art multilingual LLMs are ineffective in dealing with code-mixed languages. However, the question of how to improve the capability of multilingual LLMs to handle code-mixed language has not received any attention to date. In this paper, we tackle this research gap by proposing CHAI, a novel general-purpose framework for improving the ability of multilingual LLMs to handle code-mixed languages. CHAI relies on three novel contributions made in this paper. First, we explore the ability of LLMs to provide accurate annotations for code-mixed translation tasks. Second, we leverage this ability of LLMs as annotators to generate preference data for code-mixed translation tasks at scale, which are then used within a reinforcement learning from AI feedback (RLAIF) procedure to improve LLMs' capability on code-mixed tasks. Third, we conduct a rigorous experimental evaluation across various real-world datasets and settings. Our analysis shows that CHAI-powered LLMs outperform state-of-the-art open-source LLMs by 25.66% (in terms of win rate adjudicated by human annotators) in code-mixed translation tasks. This work represents a first step towards developing more inclusive code-mixed LLMs.
Authors: Pablo Rosillo-Rodes, Maxi San Miguel, David Sanchez
Abstract: There are different ways of measuring diversity in complex systems. In particular, in language, lexical diversity is characterized in terms of the type-token ratio and the word entropy. We here investigate both diversity metrics in six massive linguistic datasets in English, Spanish, and Turkish, consisting of books, news articles, and tweets. These gigaword corpora correspond to languages with distinct morphological features and differ in registers and genres, thus constituting a varied testbed for a quantitative approach to lexical diversity. We unveil an empirical functional relation between entropy and type-token ratio of texts of a given corpus and language, which is a consequence of the statistical laws observed in natural language. Further, in the limit of large text lengths we find an analytical expression for this relation relying on both Zipf and Heaps laws that agrees with our empirical findings.
Authors: Hongshen Xu, Zichen Zhu, Lei Pan, Zihan Wang, Su Zhu, Da Ma, Ruisheng Cao, Lu Chen, Kai Yu
Abstract: Large Language Models (LLMs) have expanded their capabilities beyond language generation to interact with external tools, enabling automation and real-world applications. However, tool hallucinations, where models either select inappropriate tools or misuse them, pose significant challenges, leading to erroneous task execution, increased computational costs, and reduced system reliability. To systematically address this issue, we define and categorize tool hallucinations into two main types, tool selection hallucination and tool usage hallucination. To evaluate and mitigate these issues, we introduce RelyToolBench, which integrates specialized test cases and novel metrics to assess hallucination-aware task success and efficiency. Finally, we propose Relign, a reliability alignment framework that expands the tool-use action space to include indecisive actions, allowing LLMs to defer tool use, seek clarification, or adjust tool selection dynamically. Through extensive experiments, we demonstrate that Relign significantly reduces tool hallucinations, improves task reliability, and enhances the efficiency of LLM tool interactions.
Authors: Yingxin Li, Ye Li, Yuan Meng, Xinzhu Ma, Zihan Geng, Shutao Xia, Zhi Wang
Abstract: As large language models (LLMs) continue to advance, the demand for higher quality and faster processing of long contexts across various applications is growing. KV cache is widely adopted as it stores previously generated key and value tokens, effectively reducing redundant computations during inference. However, as memory overhead becomes a significant concern, efficient compression of KV cache has gained increasing attention. Most existing methods perform compression from two perspectives: identifying important tokens and designing compression strategies. However, these approaches often produce biased distributions of important tokens due to the influence of accumulated attention scores or positional encoding. Furthermore, they overlook the sparsity and redundancy across different heads, which leads to difficulties in preserving the most effective information at the head level. To this end, we propose EMS to overcome these limitations, while achieving better KV cache compression under extreme compression ratios. Specifically, we introduce a Global-Local score that combines accumulated attention scores from both global and local KV tokens to better identify the token importance. For the compression strategy, we design an adaptive and unified Evict-then-Merge framework that accounts for the sparsity and redundancy of KV tokens across different heads. Additionally, we implement the head-wise parallel compression through a zero-class mechanism to enhance efficiency. Extensive experiments demonstrate our SOTA performance even under extreme compression ratios. EMS consistently achieves the lowest perplexity, improves scores by over 1.28 points across four LLMs on LongBench under a 256 cache budget, and preserves 95% retrieval accuracy with a cache budget less than 2% of the context length in the Needle-in-a-Haystack task.
Authors: Chenxiao Yu, Jinyi Ye, Yuangang Li, Zhaotian Weng, Zheng Li, Emilio Ferrara, Xiyang Hu, Yue Zhao
Abstract: While LLMs have demonstrated remarkable capabilities in text generation and reasoning, their ability to simulate human decision-making -- particularly in political contexts -- remains an open question. However, modeling voter behavior presents unique challenges due to limited voter-level data, evolving political landscapes, and the complexity of human reasoning. In this study, we develop a theory-driven, multi-step reasoning framework that integrates demographic, temporal and ideological factors to simulate voter decision-making at scale. Using synthetic personas calibrated to real-world voter data, we conduct large-scale simulations of recent U.S. presidential elections. Our method significantly improves simulation accuracy while mitigating model biases. We examine its robustness by comparing performance across different LLMs. We further investigate the challenges and constraints that arise from LLM-based political simulations. Our work provides both a scalable framework for modeling political decision-making behavior and insights into the promise and limitations of using LLMs in political science research.
Authors: Saibo Geng, Hudson Cooper, Micha{\l} Moskal, Samuel Jenkins, Julian Berman, Nathan Ranchin, Robert West, Eric Horvitz, Harsha Nori
Abstract: Reliably generating structured outputs has become a critical capability for modern language model (LM) applications. Constrained decoding has emerged as the dominant technology across sectors for enforcing structured outputs during generation. Despite its growing adoption, little has been done with the systematic evaluation of the behaviors and performance of constrained decoding. Constrained decoding frameworks have standardized around JSON Schema as a structured data format, with most uses guaranteeing constraint compliance given a schema. However, there is poor understanding of the effectiveness of the methods in practice. We present an evaluation framework to assess constrained decoding approaches across three critical dimensions: efficiency in generating constraint-compliant outputs, coverage of diverse constraint types, and quality of the generated outputs. To facilitate this evaluation, we introduce JSONSchemaBench, a benchmark for constrained decoding comprising 10K real-world JSON schemas that encompass a wide range of constraints with varying complexity. We pair the benchmark with the existing official JSON Schema Test Suite and evaluate six state-of-the-art constrained decoding frameworks, including Guidance, Outlines, Llamacpp, XGrammar, OpenAI, and Gemini. Through extensive experiments, we gain insights into the capabilities and limitations of constrained decoding on structured generation with real-world JSON schemas. Our work provides actionable insights for improving constrained decoding frameworks and structured generation tasks, setting a new standard for evaluating constrained decoding and structured generation. We release JSONSchemaBench at https://github.com/guidance-ai/jsonschemabench
Authors: Yiyi Zhang, Xingyu Chen, Kexin Chen, Yuyang Du, Xilin Dang, Pheng-Ann Heng
Abstract: Recent years have witnessed extensive efforts to enhance Large Language Models (LLMs) across various domains, alongside growing attention to their ethical implications. However, a critical challenge remains largely overlooked: LLMs must balance between rejecting harmful requests for safety and accommodating legitimate ones for utility. This paper presents a Direct Preference Optimization (DPO) based alignment framework that achieves better overall performance by addressing this ethical-utility trade-off, using chemical domain applications as a proof-of-concept. Our alignment pipeline starts with a GPT-assisted three-phase data generation scheme, in which we create LibraChemQA, a chemical question-answering dataset comprising 31.6k triplet instances. By incorporating an innovative balanced seed in the data generation process, our framework systematically considers both legitimate and illegitimate requests. The framework also introduces a rephrasing mechanism for efficient data augmentation that enhances the model's chemical comprehension. We further develop a novel hybrid evaluation scheme with LLM judges for precise assessment of both safety and utility. Experimental results demonstrate our model's substantial improvements in overall performance where both safety and utility are considered - the resulting model outperforms leading LLMs including Claude-3, GPT-4o, and LLaMA-3 by margins of 13.44%, 7.16%, and 7.10% respectively on our released benchmark. At the end of this paper, we analyze experimental results obtained from testing DeepSeek-R1 on our benchmark and reveal the critical ethical concerns raised by this highly acclaimed model. We highlight that the long Chain-of-Thought (CoT) reasoning process employed by DeepSeek-R1, as well as other LLMs distilled from it, introduces significant ethical vulnerabilities when exposed to users.
Authors: Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Y. X. Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, Wangding Zeng
Abstract: Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance. As shown in Figure 1, experiments show the model pretrained with NSA maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. Meanwhile, NSA achieves substantial speedups over Full Attention on 64k-length sequences across decoding, forward propagation, and backward propagation, validating its efficiency throughout the model lifecycle.
Authors: Jingcheng Deng, Zhongtao Jiang, Liang Pang, Liwei Chen, Kun Xu, Zihao Wei, Huawei Shen, Xueqi Cheng
Abstract: A new trend uses LLMs as dense text encoders via contrastive learning. However, since LLM embeddings predict the probability distribution of the next token, they are inherently generative and distributive, conflicting with contrastive learning, which requires embeddings to capture full-text semantics and align via cosine similarity. This discrepancy hinders the full utilization of LLMs' pre-training capabilities, resulting in inefficient learning. In response to this issue, we propose AutoRegEmbed, a new contrastive learning method built on embedding conditional probability distributions, which integrates two core tasks: information compression and conditional distribution alignment. The information compression task encodes text into the embedding space, ensuring that the embedding vectors capture global semantics. The conditional distribution alignment task focuses on aligning text embeddings with positive samples embeddings by leveraging the conditional distribution of embeddings while simultaneously reducing the likelihood of generating negative samples from text embeddings, thereby achieving embedding alignment and uniformity. Experimental results demonstrate that our method significantly outperforms traditional contrastive learning approaches and achieves performance comparable to state-of-the-art models when using the same amount of data.
Authors: Yuheng Chen, Pengfei Cao, Kang Liu, Jun Zhao
Abstract: Previous studies primarily utilize MLP neurons as units of analysis for understanding the mechanisms of factual knowledge in Language Models (LMs); however, neurons suffer from polysemanticity, leading to limited knowledge expression and poor interpretability. In this paper, we first conduct preliminary experiments to validate that Sparse Autoencoders (SAE) can effectively decompose neurons into features, which serve as alternative analytical units. With this established, our core findings reveal three key advantages of features over neurons: (1) Features exhibit stronger influence on knowledge expression and superior interpretability. (2) Features demonstrate enhanced monosemanticity, showing distinct activation patterns between related and unrelated facts. (3) Features achieve better privacy protection than neurons, demonstrated through our proposed FeatureEdit method, which significantly outperforms existing neuron-based approaches in erasing privacy-sensitive information from LMs.Code and dataset will be available.
Authors: Martin Kuo, Jianyi Zhang, Aolin Ding, Qinsi Wang, Louis DiValentin, Yujia Bao, Wei Wei, Hai Li, Yiran Chen
Abstract: Large Reasoning Models (LRMs) have recently extended their powerful reasoning capabilities to safety checks-using chain-of-thought reasoning to decide whether a request should be answered. While this new approach offers a promising route for balancing model utility and safety, its robustness remains underexplored. To address this gap, we introduce Malicious-Educator, a benchmark that disguises extremely dangerous or malicious requests beneath seemingly legitimate educational prompts. Our experiments reveal severe security flaws in popular commercial-grade LRMs, including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking. For instance, although OpenAI's o1 model initially maintains a high refusal rate of about 98%, subsequent model updates significantly compromise its safety; and attackers can easily extract criminal strategies from DeepSeek-R1 and Gemini 2.0 Flash Thinking without any additional tricks. To further highlight these vulnerabilities, we propose Hijacking Chain-of-Thought (H-CoT), a universal and transferable attack method that leverages the model's own displayed intermediate reasoning to jailbreak its safety reasoning mechanism. Under H-CoT, refusal rates sharply decline-dropping from 98% to below 2%-and, in some instances, even transform initially cautious tones into ones that are willing to provide harmful content. We hope these findings underscore the urgent need for more robust safety mechanisms to preserve the benefits of advanced reasoning capabilities without compromising ethical standards.
Authors: P Team, Xinrun Du, Yifan Yao, Kaijing Ma, Bingli Wang, Tianyu Zheng, Kang Zhu, Minghao Liu, Yiming Liang, Xiaolong Jin, Zhenlin Wei, Chujie Zheng, Kaixin Deng, Shian Jia, Sichao Jiang, Yiyan Liao, Rui Li, Qinrui Li, Sirun Li, Yizhi Li, Yunwen Li, Dehua Ma, Yuansheng Ni, Haoran Que, Qiyao Wang, Zhoufutu Wen, Siwei Wu, Tianshun Xing, Ming Xu, Zhenzhu Yang, Zekun Moore Wang, Junting Zhou, Yuelin Bai, Xingyuan Bu, Chenglin Cai, Liang Chen, Yifan Chen, Chengtuo Cheng, Tianhao Cheng, Keyi Ding, Siming Huang, Yun Huang, Yaoru Li, Yizhe Li, Zhaoqun Li, Tianhao Liang, Chengdong Lin, Hongquan Lin, Yinghao Ma, Tianyang Pang, Zhongyuan Peng, Zifan Peng, Qige Qi, Shi Qiu, Xingwei Qu, Shanghaoran Quan, Yizhou Tan, Zili Wang, Chenqing Wang, Hao Wang, Yiya Wang, Yubo Wang, Jiajun Xu, Kexin Yang, Ruibin Yuan, Yuanhao Yue, Tianyang Zhan, Chun Zhang, Jinyang Zhang, Xiyue Zhang, Xingjian Zhang, Yue Zhang, Yongchi Zhao, Xiangyu Zheng, Chenghua Zhong, Yang Gao, Zhoujun Li, Dayiheng Liu, Qian Liu, Tianyu Liu, Shiwen Ni, Junran Peng, Yujia Qin, Wenbo Su, Guoyin Wang, Shi Wang, Jian Yang, Min Yang, Meng Cao, Xiang Yue, Zhaoxiang Zhang, Wangchunshu Zhou, Jiaheng Liu, Qunshu Lin, Wenhao Huang, Ge Zhang
Abstract: Large language models (LLMs) have demonstrated remarkable proficiency in mainstream academic disciplines such as mathematics, physics, and computer science. However, human knowledge encompasses over 200 specialized disciplines, far exceeding the scope of existing benchmarks. The capabilities of LLMs in many of these specialized fields-particularly in light industry, agriculture, and service-oriented disciplines-remain inadequately evaluated. To address this gap, we present SuperGPQA, a comprehensive benchmark that evaluates graduate-level knowledge and reasoning capabilities across 285 disciplines. Our benchmark employs a novel Human-LLM collaborative filtering mechanism to eliminate trivial or ambiguous questions through iterative refinement based on both LLM responses and expert feedback. Our experimental results reveal significant room for improvement in the performance of current state-of-the-art LLMs across diverse knowledge domains (e.g., the reasoning-focused model DeepSeek-R1 achieved the highest accuracy of 61.82% on SuperGPQA), highlighting the considerable gap between current model capabilities and artificial general intelligence. Additionally, we present comprehensive insights from our management of a large-scale annotation process, involving over 80 expert annotators and an interactive Human-LLM collaborative system, offering valuable methodological guidance for future research initiatives of comparable scope.
Authors: Xun Liang, Jiawei Yang, Yezhaohui Wang, Chen Tang, Zifan Zheng, Shichao Song, Zehao Lin, Yebin Yang, Simin Niu, Hanyu Wang, Bo Tang, Feiyu Xiong, Keming Mao, Zhiyu li
Abstract: Large Language Models (LLMs) have demonstrated exceptional comprehension capabilities and a vast knowledge base, suggesting that LLMs can serve as efficient tools for automated survey generation. However, recent research related to automated survey generation remains constrained by some critical limitations like finite context window, lack of in-depth content discussion, and absence of systematic evaluation frameworks. Inspired by human writing processes, we propose SurveyX, an efficient and organized system for automated survey generation that decomposes the survey composing process into two phases: the Preparation and Generation phases. By innovatively introducing online reference retrieval, a pre-processing method called AttributeTree, and a re-polishing process, SurveyX significantly enhances the efficacy of survey composition. Experimental evaluation results show that SurveyX outperforms existing automated survey generation systems in content quality (0.259 improvement) and citation quality (1.76 enhancement), approaching human expert performance across multiple evaluation dimensions. Examples of surveys generated by SurveyX are available on www.surveyx.cn
Authors: Anil Ramakrishna, Yixin Wan, Xiaomeng Jin, Kai-Wei Chang, Zhiqi Bu, Bhanukiran Vinzamuri, Volkan Cevher, Mingyi Hong, Rahul Gupta
Abstract: Unlearning aims to remove copyrighted, sensitive, or private content from large language models (LLMs) without a full retraining. In this work, we develop a multi-task unlearning benchmark (LUME) which features three tasks: (1) unlearn synthetically generated creative short novels, (2) unlearn synthetic biographies with sensitive information, and (3) unlearn a collection of public biographies. We further release two fine-tuned LLMs of 1B and 7B parameter sizes as the target models. We conduct detailed evaluations of several recently proposed unlearning algorithms and present results on carefully crafted metrics to understand their behavior and limitations.
Authors: Vignesh Kothapalli, Hamed Firooz, Maziar Sanjabi
Abstract: We introduce CoT-ICL Lab, a framework and methodology to generate synthetic tokenized datasets and systematically study chain-of-thought (CoT) in-context learning (ICL) in language models. CoT-ICL Lab allows fine grained control over the complexity of in-context examples by decoupling (1) the causal structure involved in chain token generation from (2) the underlying token processing functions. We train decoder-only transformers (up to 700M parameters) on these datasets and show that CoT accelerates the accuracy transition to higher values across model sizes. In particular, we find that model depth is crucial for leveraging CoT with limited in-context examples, while more examples help shallow models match deeper model performance. Additionally, limiting the diversity of token processing functions throughout training improves causal structure learning via ICL. We also interpret these transitions by analyzing transformer embeddings and attention maps. Overall, CoT-ICL Lab serves as a simple yet powerful testbed for theoretical and empirical insights into ICL and CoT in language models.
Authors: Longyun Wu, Dawei Zhu, Guangxiang Zhao, Zhuocheng Yu, Junfeng Ran, Xiangyu Wong, Lin Sun, Sujian Li
Abstract: With the development of large language models (LLMs), there has been an increasing need for significant advancements in handling long contexts. To enhance long-context capabilities, constructing high-quality training data with long-range dependencies is crucial. Existing methods to select long-context data often rely on sentence-level analysis, which can be greatly optimized in both performance and efficiency. In this paper, we propose a novel token-level framework, LongAttn, which leverages the self-attention mechanism of LLMs to measure the long-range dependencies for the data. By calculating token-level dependency strength and distribution uniformity of token scores, LongAttn effectively quantifies long-range dependencies, enabling more accurate and efficient data selection. We filter LongABC-32K from open-source long-context datasets (ArXiv, Book, and Code). Through our comprehensive experiments, LongAttn has demonstrated its excellent effectiveness, scalability, and efficiency. To facilitate future research in long-context data, we released our code and the high-quality long-context training data LongABC-32K.
Authors: Yuming Yang, Yang Nan, Junjie Ye, Shihan Dou, Xiao Wang, Shuo Li, Huijie Lv, Mingqi Wu, Tao Gui, Qi Zhang, Xuanjing Huang
Abstract: Data diversity is crucial for the instruction tuning of large language models. Existing studies have explored various diversity-aware data selection methods to construct high-quality datasets and enhance model performance. However, the fundamental problem of precisely defining and measuring data diversity remains underexplored, limiting clear guidance for data engineering. To address this, we systematically analyze 11 existing diversity measurement methods by evaluating their correlation with model performance through extensive fine-tuning experiments. Our results indicate that a reliable diversity measure should properly account for both inter-sample differences and the information distribution in the sample space. Building on this, we propose NovelSum, a new diversity metric based on sample-level "novelty." Experiments on both simulated and real-world data show that NovelSum accurately captures diversity variations and achieves a 0.97 correlation with instruction-tuned model performance, highlighting its value in guiding data engineering practices. With NovelSum as an optimization objective, we further develop a greedy, diversity-oriented data selection strategy that outperforms existing approaches, validating both the effectiveness and practical significance of our metric.
Authors: Wei Liu, Yancheng He, Hui Huang, Chengwei Hu, Jiaheng Liu, Shilong Li, Wenbo Su, Bo Zheng
Abstract: With the development of large language models, their ability to follow simple instructions has significantly improved. However, adhering to complex instructions remains a major challenge. Current approaches to generating complex instructions are often irrelevant to the current instruction requirements or suffer from limited scalability and diversity. Moreover, methods such as back-translation, while effective for simple instruction generation, fail to leverage the rich contents and structures in large web corpora. In this paper, we propose a novel automatic iterative refinement framework to generate complex instructions with constraints, which not only better reflects the requirements of real scenarios but also significantly enhances LLMs' ability to follow complex instructions. The AIR framework consists of two stages: (1)Generate an initial instruction from a document; (2)Iteratively refine instructions with LLM-as-judge guidance by comparing the model's output with the document to incorporate valuable constraints. Finally, we construct the AIR-10K dataset with 10K complex instructions and demonstrate that instructions generated with our approach significantly improve the model's ability to follow complex instructions, outperforming existing methods for instruction generation.
Authors: Haris Riaz, Ellen Riloff, Mihai Surdeanu
Abstract: We propose a simple, unsupervised method that injects pragmatic principles in retrieval-augmented generation (RAG) frameworks such as Dense Passage Retrieval to enhance the utility of retrieved contexts. Our approach first identifies which sentences in a pool of documents retrieved by RAG are most relevant to the question at hand, cover all the topics addressed in the input question and no more, and then highlights these sentences within their context, before they are provided to the LLM, without truncating or altering the context in any other way. We show that this simple idea brings consistent improvements in experiments on three question answering tasks (ARC-Challenge, PubHealth and PopQA) using five different LLMs. It notably enhances relative accuracy by up to 19.7% on PubHealth and 10% on ARC-Challenge compared to a conventional RAG system.
Authors: Yuxuan Zhang
Abstract: With the rapid development of large language models (LLMs), fully fine-tuning (FT) these models has become increasingly impractical due to the high computational demands. Additionally, FT can lead to catastrophic forgetting. As an alternative, Low-Rank Adaptation (LoRA) has been proposed, which fine-tunes only a small subset of parameters, achieving similar performance to FT while significantly reducing resource requirements. However, since LoRA inherits FT's design, the issue of catastrophic forgetting remains. To address these challenges, we propose SECURA: Sigmoid-Enhanced CUR Decomposition LoRA, a novel parameter-efficient fine-tuning (PEFT) variant that mitigates catastrophic forgetting while improving fine-tuning performance. Our method introduces a new normalization technique, SigNorm, to enhance parameter retention and overall performance. SECURA has been evaluated on a variety of tasks, including mathematical problem-solving (GSM8K), challenging question-answering (CNNDM), translation (NewsDE), and complex multiple-choice reasoning (LogiQA). Experimental results show that SECURA achieves an average fine-tuning improvement of 3.59% across four multiple-choice question (MCQ) tasks and a 2.51% improvement across five question-answering (QA) tasks on models such as Gemma2 2b, Qwen2 1.5b, Qwen 2 7b, Llama3 8b, and Llama3.1 8b, compared to DoRA. Moreover, SECURA demonstrates superior knowledge retention capabilities, maintaining more than 70% accuracy on basic LLM knowledge across 16 continual learning tests, outperforming Experience Replay (ER), Sequential Learning (SEQ), EWC, I-LoRA, and CUR-LoRA.
Authors: Gyeongje Cho, Yeonkyoung So, Jaejin Lee
Abstract: Multiple-choice benchmarks, consisting of various prompts and choices, are among the most widely used methods to assess a language model's natural language understanding capability. Given a specific prompt, we typically compute $P(Choice|Prompt)$ to evaluate how likely a language model is to generate the correct choice compared to incorrect ones. However, we observe that performance measured using this approach reflects not only the model's comprehension of the prompt but also its inherent biases for certain choices regardless of the prompt. This issue makes it challenging to accurately measure a model's natural language understanding, as models may select the answer without fully understanding the prompt. To address this limitation, we propose a novel metric called ANPMI, which normalizes Pointwise Mutual Information (PMI) by $-\log P(Choice)$. ANPMI provides a more accurate assessment of the model's natural language understanding by ensuring that it is challenging to answer a question without properly understanding the prompt.
Authors: Kanana LLM Team, Yunju Bak, Hojin Lee, Minho Ryu, Jiyeon Ham, Seungjae Jung, Daniel Wontae Nam, Taegyeong Eo, Donghun Lee, Doohae Jung, Boseop Kim, Nayeon Kim, Jaesun Park, Hyunho Kim, Hyunwoong Ko, Changmin Lee, Kyoung-Woon On, Seulye Baeg, Junrae Cho, Sunghee Jung, Jieun Kang, EungGyun Kim, Eunhwa Kim, Byeongil Ko, Daniel Lee, Minchul Lee, Miok Lee, Shinbok Lee, Gaeun Seo
Abstract: We introduce Kanana, a series of bilingual language models that demonstrate exceeding performance in Korean and competitive performance in English. The computational cost of Kanana is significantly lower than that of state-of-the-art models of similar size. The report details the techniques employed during pre-training to achieve compute-efficient yet competitive models, including high quality data filtering, staged pre-training, depth up-scaling, and pruning and distillation. Furthermore, the report outlines the methodologies utilized during the post-training of the Kanana models, encompassing supervised fine-tuning and preference optimization, aimed at enhancing their capability for seamless interaction with users. Lastly, the report elaborates on plausible approaches used for language model adaptation to specific scenarios, such as embedding, retrieval augmented generation, and function calling. The Kanana model series spans from 2.1B to 32.5B parameters with 2.1B models (base, instruct, embedding) publicly released to promote research on Korean language models.
Authors: Hongyi Cal, Jie Li, Wenzhen Dong
Abstract: The effectiveness of instruction fine-tuning for Large Language Models is fundamentally constrained by the quality and efficiency of training datasets. This work introduces Low-Confidence Gold (LCG), a novel filtering framework that employs centroid-based clustering and confidence-guided selection for identifying valuable instruction pairs. Through a semi-supervised approach using a lightweight classifier trained on representative samples, LCG curates high-quality subsets while preserving data diversity. Experimental evaluation demonstrates that models fine-tuned on LCG-filtered subsets of 6K samples achieve superior performance compared to existing methods, with substantial improvements on MT-bench and consistent gains across comprehensive evaluation metrics. The framework's efficacy while maintaining model performance establishes a promising direction for efficient instruction tuning.
Authors: Yancheng He, Shilong Li, Jiaheng Liu, Weixun Wang, Xingyuan Bu, Ge Zhang, Zhongyuan Peng, Zhaoxiang Zhang, Zhicheng Zheng, Wenbo Su, Bo Zheng
Abstract: Recently, o1-like models have drawn significant attention, where these models produce the long Chain-of-Thought (CoT) reasoning steps to improve the reasoning abilities of existing Large Language Models (LLMs). In this paper, to understand the qualities of these long CoTs and measure the critique abilities of existing LLMs on these long CoTs, we introduce the DeltaBench, including the generated long CoTs from different o1-like models (e.g., QwQ, DeepSeek-R1) for different reasoning tasks (e.g., Math, Code, General Reasoning), to measure the ability to detect errors in long CoT reasoning. Based on DeltaBench, we first perform fine-grained analysis of the generated long CoTs to discover the effectiveness and efficiency of different o1-like models. Then, we conduct extensive evaluations of existing process reward models (PRMs) and critic models to detect the errors of each annotated process, which aims to investigate the boundaries and limitations of existing PRMs and critic models. Finally, we hope that DeltaBench could guide developers to better understand the long CoT reasoning abilities of their models.
Authors: Xuannan Liu, Zekun Li, Peipei Li, Huaibo Huang, Shuhan Xia, Xing Cui, Linzhi Huang, Weihong Deng, Zhaofeng He
Abstract: Current multimodal misinformation detection (MMD) methods often assume a single source and type of forgery for each sample, which is insufficient for real-world scenarios where multiple forgery sources coexist. The lack of a benchmark for mixed-source misinformation has hindered progress in this field. To address this, we introduce MMFakeBench, the first comprehensive benchmark for mixed-source MMD. MMFakeBench includes 3 critical sources: textual veracity distortion, visual veracity distortion, and cross-modal consistency distortion, along with 12 sub-categories of misinformation forgery types. We further conduct an extensive evaluation of 6 prevalent detection methods and 15 Large Vision-Language Models (LVLMs) on MMFakeBench under a zero-shot setting. The results indicate that current methods struggle under this challenging and realistic mixed-source MMD setting. Additionally, we propose MMD-Agent, a novel approach to integrate the reasoning, action, and tool-use capabilities of LVLM agents, significantly enhancing accuracy and generalization. We believe this study will catalyze future research into more realistic mixed-source multimodal misinformation and provide a fair evaluation of misinformation detection methods.
Authors: Yongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Yu Qiao, Xuanjing Huang, Feng Zhao, Tao Gui, Jing Shao
Abstract: The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness.
Authors: Zora Zhiruo Wang, Akari Asai, Xinyan Velocity Yu, Frank F. Xu, Yiqing Xie, Graham Neubig, Daniel Fried
Abstract: While language models (LMs) have proven remarkably adept at generating code, many programs are challenging for LMs to generate using their parametric knowledge alone. Providing external contexts such as library documentation can facilitate generating accurate and functional code. Despite the success of retrieval-augmented generation (RAG) in various text-oriented tasks, its potential for improving code generation remains under-explored. In this work, we conduct a systematic, large-scale analysis by asking: in what scenarios can retrieval benefit code generation models? and what challenges remain? We first curate a comprehensive evaluation benchmark, CodeRAG-Bench, encompassing three categories of code generation tasks, including basic programming, open-domain, and repository-level problems. We aggregate documents from five sources for models to retrieve contexts: competition solutions, online tutorials, library documentation, StackOverflow posts, and GitHub repositories. We examine top-performing models on CodeRAG-Bench by providing contexts retrieved from one or multiple sources. While notable gains are made in final code generation by retrieving high-quality contexts across various settings, our analysis reveals room for improvement -- current retrievers still struggle to fetch useful contexts especially with limited lexical overlap, and generators fail to improve with limited context lengths or abilities to integrate additional contexts. We hope CodeRAG-Bench serves as an effective testbed to encourage further development of advanced code-oriented RAG methods.
Authors: Anton Xue, Avishree Khare, Rajeev Alur, Surbhi Goel, Eric Wong
Abstract: We study how to subvert large language models (LLMs) from following prompt-specified rules. We first formalize rule-following as inference in propositional Horn logic, a mathematical system in which rules have the form "if $P$ and $Q$, then $R$" for some propositions $P$, $Q$, and $R$. Next, we prove that although small transformers can faithfully follow such rules, maliciously crafted prompts can still mislead both theoretical constructions and models learned from data. Furthermore, we demonstrate that popular attack algorithms on LLMs find adversarial prompts and induce attention patterns that align with our theory. Our novel logic-based framework provides a foundation for studying LLMs in rule-based settings, enabling a formal analysis of tasks like logical reasoning and jailbreak attacks.
Authors: Jinsheng Huang, Liang Chen, Taian Guo, Fu Zeng, Yusheng Zhao, Bohan Wu, Ye Yuan, Haozhe Zhao, Zhihui Guo, Yichi Zhang, Jingyang Yuan, Wei Ju, Luchen Liu, Tianyu Liu, Baobao Chang, Ming Zhang
Abstract: Large Multimodal Models (LMMs) exhibit impressive cross-modal understanding and reasoning abilities, often assessed through multiple-choice questions (MCQs) that include an image, a question, and several options. However, many benchmarks used for such evaluations suffer from systematic biases. Remarkably, Large Language Models (LLMs) without any visual perception capabilities achieve non-trivial performance, undermining the credibility of these evaluations. To address this issue while maintaining the efficiency of MCQ evaluations, we propose MMEvalPro, a benchmark designed to avoid Type-I errors through a trilogy evaluation pipeline and more rigorous metrics. For each original question from existing benchmarks, human annotators augment it by creating one perception question and one knowledge anchor question through a meticulous annotation process. MMEvalPro comprises $2,138$ question triplets, totaling $6,414$ distinct questions. Two-thirds of these questions are manually labeled by human experts, while the rest are sourced from existing benchmarks (MMMU, ScienceQA, and MathVista). Compared with the existing benchmarks, our experiments with the latest LLMs and LMMs demonstrate that MMEvalPro is more challenging (the best LMM lags behind human performance by $31.73\%$, compared to an average gap of $8.03\%$ in previous benchmarks) and more trustworthy (the best LLM trails the best LMM by $23.09\%$, whereas the gap for previous benchmarks is just $14.64\%$). Our in-depth analysis explains the reason for the large performance gap and justifies the trustworthiness of evaluation, underscoring its significant potential for advancing future research.
Authors: Jin Peng Zhou, Christian K. Belardi, Ruihan Wu, Travis Zhang, Carla P. Gomes, Wen Sun, Kilian Q. Weinberger
Abstract: Developing prompt-based methods with Large Language Models (LLMs) requires making numerous decisions, which give rise to a combinatorial search problem over hyper-parameters. This exhaustive evaluation can be time-consuming and costly. In this paper, we propose an $\textit{adaptive}$ approach to explore this space. We are exploiting the fact that often only few samples are needed to identify clearly superior or inferior settings, and that many evaluation tests are highly correlated. We lean on multi-armed bandits to sequentially identify the next (method, validation sample)-pair to evaluate and utilize low-rank matrix factorization to fill in missing evaluations. We carefully assess the efficacy of our approach on several competitive benchmark problems and show that it can identify the top-performing method using only 5-15% of the typical resources -- resulting in 85-95% LLM cost savings. Our code is available at https://github.com/kilian-group/banditeval.
Authors: Hao Mark Chen, Fuwen Tan, Alexandros Kouris, Royson Lee, Hongxiang Fan, Stylianos I. Venieris
Abstract: In spite of the great potential of large language models (LLMs) across various tasks, their deployment on resource-constrained devices remains challenging due to their excessive computational and memory demands. Quantization has emerged as an effective solution by storing weights in reduced precision. However, utilizing low precisions (i.e.~2/3-bit) to substantially alleviate the memory-boundedness of LLM decoding, still suffers from prohibitive performance drop. In this work, we argue that existing approaches fail to explore the diversity in computational patterns, redundancy, and sensitivity to approximations of the different phases of LLM inference, resorting to a uniform quantization policy throughout. Instead, we propose a novel phase-aware method that selectively allocates precision during different phases of LLM inference, achieving both strong context extraction during prefill and efficient memory bandwidth utilization during decoding. To further address the memory-boundedness of the decoding phase, we introduce Progressive Mixed-Precision Decoding (PMPD), a technique that enables the gradual lowering of precision deeper in the generated sequence, together with a spectrum of precision-switching schedulers that dynamically drive the precision-lowering decisions in either task-adaptive or prompt-adaptive manner. Extensive evaluation across diverse language tasks shows that when targeting Nvidia GPUs, PMPD achieves 1.4$-$12.2$\times$ speedup in matrix-vector multiplications over fp16 models, while when targeting an LLM-optimized NPU, our approach delivers a throughput gain of 3.8$-$8.0$\times$ over fp16 models and up to 1.54$\times$ over uniform quantization approaches while preserving the output quality.
Authors: Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, Dahua Lin
Abstract: In large vision-language models (LVLMs), images serve as inputs that carry a wealth of information. As the idiom "A picture is worth a thousand words" implies, representing a single image in current LVLMs can require hundreds or even thousands of tokens. This results in significant computational costs, which grow quadratically as input image resolution increases, thereby severely impacting the efficiency of both training and inference. Previous approaches have attempted to reduce the number of image tokens either before or within the early layers of LVLMs. However, these strategies inevitably result in the loss of crucial image information, ultimately diminishing model performance. To address this challenge, we conduct an empirical study revealing that all visual tokens are necessary for LVLMs in the shallow layers, and token redundancy progressively increases in the deeper layers of the model. To this end, we propose PyramidDrop, a visual redundancy reduction strategy for LVLMs to boost their efficiency in both training and inference with neglectable performance loss. Specifically, we partition the LVLM into several stages and drop part of the image tokens at the end of each stage with a pre-defined ratio, creating pyramid-like visual tokens across model layers. The dropping is based on a lightweight similarity calculation with a negligible time overhead. Extensive experiments demonstrate that PyramidDrop can achieve a 40% training time and 55% inference FLOPs acceleration of LLaVA-NeXT with comparable performance. Besides, the PyramidDrop could also serve as a plug-and-play strategy for inference acceleration without training, with better performance and lower inference cost than counterparts. Code is available at https://github.com/Cooperx521/PyramidDrop.
Authors: Vedant Das Swain, Qiuyue "Joy" Zhong, Jash Rajesh Parekh, Yechan Jeon, Roy Zimmermann, Mary Czerwinski, Jina Suh, Varun Mishra, Koustuv Saha, Javier Hernandez
Abstract: Client-Service Representatives (CSRs) are vital to organizations. Frequent interactions with disgruntled clients, however, disrupt their mental well-being. To help CSRs regulate their emotions while interacting with uncivil clients, we designed Care-Pilot, an LLM-powered assistant, and evaluated its efficacy, perception, and use. Our comparative analyses between 665 human and Care-Pilot-generated support messages highlight Care-Pilot's ability to adapt to and demonstrate empathy in various incivility incidents. Additionally, 143 CSRs assessed Care-Pilot's empathy as more sincere and actionable than human messages. Finally, we interviewed 20 CSRs who interacted with Care-Pilot in a simulation exercise. They reported that Care-Pilot helped them avoid negative thinking, recenter thoughts, and humanize clients; showing potential for bridging gaps in coworker support. Yet, they also noted deployment challenges and emphasized the indispensability of shared experiences. We discuss future designs and societal implications of AI-mediated emotional labor, underscoring empathy as a critical function for AI assistants for worker mental health.
Authors: Aobo Kong, Wentao Ma, Shiwan Zhao, Yongbin Li, Yuchuan Wu, Ke Wang, Xiaoqian Liu, Qicheng Li, Yong Qin, Fei Huang
Abstract: Social agents powered by large language models (LLMs) can simulate human social behaviors but fall short in handling complex social dialogues. Direct Preference Optimization (DPO) has proven effective in aligning LLM behavior with human preferences across various agent tasks. However, standard DPO focuses solely on individual turns, which limits its effectiveness in multi-turn social interactions. Several DPO-based multi-turn alignment methods with session-level data have shown potential in addressing this problem.While these methods consider multiple turns across entire sessions, they are often overly coarse-grained, introducing training noise, and lack robust theoretical support. To resolve these limitations, we propose Segment-Level Direct Preference Optimization (SDPO), which dynamically select key segments within interactions to optimize multi-turn agent behavior. SDPO minimizes training noise and is grounded in a rigorous theoretical framework. Evaluations on the SOTOPIA benchmark demonstrate that SDPO-tuned agents consistently outperform both existing DPO-based methods and proprietary LLMs like GPT-4o, underscoring SDPO's potential to advance the social intelligence of LLM-based agents. We release our code and data at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/SDPO.
URLs: https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/SDPO.
Authors: Thierry Bossy, Julien Vignoud, Tahseen Rabbani, Juan R. Troncoso Pastoriza, Martin Jaggi
Abstract: Federated learning (FL) is a popular paradigm for collaborative training which avoids direct data exposure between clients. However, data privacy issues still remain: FL-trained large language models are capable of memorizing and completing phrases and sentences contained in training data when given with their prefixes. Thus, it is possible for adversarial and honest-but-curious clients to recover training data of other participants simply through targeted prompting. In this work, we demonstrate that a popular and simple fine-tuning strategy, low-rank adaptation (LoRA), reduces memorization during FL up to a factor of 10. We study this effect by performing a medical question-answering fine-tuning task and injecting multiple replicas of out-of-distribution sensitive sequences drawn from an external clinical dataset. We observe a reduction in memorization for a wide variety of Llama 2 and 3 models, and find that LoRA can reduce memorization in centralized learning as well. Furthermore, we show that LoRA can be combined with other privacy-preserving techniques such as gradient clipping and Gaussian noising, secure aggregation, and Goldfish loss to further improve record-level privacy while maintaining performance.
Authors: Jiayi Fu, Xuandong Zhao, Chengyuan Yao, Heng Wang, Qi Han, Yanghua Xiao
Abstract: Reinforcement Learning from Human Feedback (RLHF) is essential for aligning large language models (LLMs) with human values. However, RLHF is susceptible to reward hacking, where the agent exploits flaws in the reward function rather than learning the intended behavior, thus degrading alignment. While reward shaping helps stabilize RLHF and partially mitigate reward hacking, a systematic investigation into shaping techniques and their underlying principles remains lacking. To bridge this gap, we present a comprehensive study of the prevalent reward shaping methods. Our analysis suggests three key design principles: (1) RL reward is ideally bounded, (2) RL benefits from rapid initial growth followed by gradual convergence, and (3) RL reward is best formulated as a function of centered reward. Guided by these insights, we propose Preference As Reward (PAR), a novel approach that leverages the latent preferences embedded within the reward model itself as the signal for reinforcement learning. We evaluated PAR on two base models, Gemma2-2B and Llama3-8B, using two datasets, Ultrafeedback-Binarized and HH-RLHF. Experimental results demonstrate PAR's superior performance over other reward shaping methods. On the AlpacaEval 2.0 benchmark, PAR achieves a win rate at least 5 percentage points higher than competing approaches. Furthermore, PAR exhibits remarkable data efficiency, requiring only a single reference reward for optimal performance, and maintains robustness against reward hacking even after two full epochs of training. Code is available at https://github.com/PorUna-byte/PAR.