Authors: Mat\'eo Mahaut, Francesca Franzon
Abstract: Text generated by language models (LMs) can degrade into repetitive cycles, where identical word sequences are persistently repeated one after another. Prior research has typically treated repetition as a unitary phenomenon. However, repetitive sequences emerge under diverse tasks and contexts, raising the possibility that it may be driven by multiple underlying factors. Here, we experimentally explore the hypothesis that repetition in LMs can result from distinct mechanisms, reflecting different text generation strategies used by the model. We examine the internal working of LMs under two conditions that prompt repetition: one in which repeated sequences emerge naturally after human-written text, and another where repetition is explicitly induced through an in-context learning (ICL) setup. Our analysis reveals key differences between the two conditions: the model exhibits varying levels of confidence, relies on different attention heads, and shows distinct pattens of change in response to controlled perturbations. These findings suggest that distinct internal mechanisms can interact to drive repetition, with implications for its interpretation and mitigation strategies. More broadly, our results highlight that the same surface behavior in LMs may be sustained by different underlying processes, acting independently or in combination.
Authors: Ziyi Liu, Priyanka Dey, Zhenyu Zhao, Jen-tse Huang, Rahul Gupta, Yang Liu, Jieyu Zhao
Abstract: Cultural Intelligence (CQ) refers to the ability to understand unfamiliar cultural contexts-a crucial skill for large language models (LLMs) to effectively engage with globally diverse users. While existing research often focuses on explicitly stated cultural norms, such approaches fail to capture the subtle, implicit values that underlie real-world conversations. To address this gap, we introduce CQ-Bench, a benchmark specifically designed to assess LLMs' capability to infer implicit cultural values from natural conversational contexts. We generate a multi-character conversation-based stories dataset using values from the World Value Survey and GlobalOpinions datasets, with topics including ethical, religious, social, and political. Our dataset construction pipeline includes rigorous validation procedures-incorporation, consistency, and implicitness checks-using GPT-4o, with 98.2% human-model agreement in the final validation. Our benchmark consists of three tasks of increasing complexity: attitude detection, value selection, and value extraction. We find that while o1 and Deepseek-R1 models reach human-level performance in value selection (0.809 and 0.814), they still fall short in nuanced attitude detection, with F1 scores of 0.622 and 0.635, respectively. In the value extraction task, GPT-4o-mini and o3-mini score 0.602 and 0.598, highlighting the difficulty of open-ended cultural reasoning. Notably, fine-tuning smaller models (e.g., LLaMA-3.2-3B) on only 500 culturally rich examples improves performance by over 10%, even outperforming stronger baselines (o3-mini) in some cases. Using CQ-Bench, we provide insights into the current challenges in LLMs' CQ research and suggest practical pathways for enhancing LLMs' cross-cultural reasoning abilities.
Authors: Melanie Subbiah, Akankshya Mishra, Grace Kim, Liyan Tang, Greg Durrett, Kathleen McKeown
Abstract: Determining faithfulness of a claim to a source document is an important problem across many domains. This task is generally treated as a binary judgment of whether the claim is supported or unsupported in relation to the source. In many cases, though, whether a claim is supported can be ambiguous. For instance, it may depend on making inferences from given evidence, and different people can reasonably interpret the claim as either supported or unsupported based on their agreement with those inferences. Forcing binary labels upon such claims lowers the reliability of evaluation. In this work, we reframe the task to manage the subjectivity involved with factuality judgments of ambiguous claims. We introduce LLM-generated edits of summaries as a method of providing a nuanced evaluation of claims: how much does a summary need to be edited to be unambiguous? Whether a claim gets rewritten and how much it changes can be used as an automatic evaluation metric, the Ambiguity Rewrite Metric (ARM), with a much richer feedback signal than a binary judgment of faithfulness. We focus on the area of narrative summarization as it is particularly rife with ambiguity and subjective interpretation. We show that ARM produces a 21% absolute improvement in annotator agreement on claim faithfulness, indicating that subjectivity is reduced.
Authors: Guy Kaplan, Michael Toker, Yuval Reif, Yonatan Belinkov, Roy Schwartz
Abstract: Text-to-Image (T2I) models often suffer from issues such as semantic leakage, incorrect feature binding, and omissions of key concepts in the generated image. This work studies these phenomena by looking into the role of information flow between textual token representations. To this end, we generate images by applying the diffusion component on a subset of contextual token representations in a given prompt and observe several interesting phenomena. First, in many cases, a word or multiword expression is fully represented by one or two tokens, while other tokens are redundant. For example, in "San Francisco's Golden Gate Bridge", the token "gate" alone captures the full expression. We demonstrate the redundancy of these tokens by removing them after textual encoding and generating an image from the resulting representation. Surprisingly, we find that this process not only maintains image generation performance but also reduces errors by 21\% compared to standard generation. We then show that information can also flow between different expressions in a sentence, which often leads to semantic leakage. Based on this observation, we propose a simple, training-free method to mitigate semantic leakage: replacing the leaked item's representation after the textual encoding with its uncontextualized representation. Remarkably, this simple approach reduces semantic leakage by 85\%. Overall, our work provides a comprehensive analysis of information flow across textual tokens in T2I models, offering both novel insights and practical benefits.
Authors: Zian Su, Ziyang Huang, Kaiyuan Zhang, Xiangyu Zhang
Abstract: Large language models (LLMs) have emerged as powerful knowledge bases yet are limited by static training data, leading to issues such as hallucinations and safety risks. Editing a model's internal knowledge through the locate-and-edit paradigm has proven a cost-effective alternative to retraining, though current unstructured approaches, especially window-based autoregressive methods, often disrupt the causal dependency between early memory updates and later output tokens. In this work, we first theoretically analyze these limitations and then introduce Matryoshka Unstructured Knowledge Editing ($\mu$KE), a novel memory update mechanism that preserves such dependencies via a Matryoshka-style objective and adaptive loss coefficients. Empirical evaluations on two models across four benchmarks demonstrate that $\mu$KE improves edit efficacy by up to 12.33% over state-of-the-art methods, and remain robust when applied to diverse formatted edits, underscoring its potential for effective unstructured knowledge editing in LLMs.
Authors: Krithik Vishwanath, Anton Alyakin, Daniel Alexander Alber, Jin Vivian Lee, Douglas Kondziolka, Eric Karl Oermann
Abstract: Large language models (LLMs) have the potential to transform medicine, but real-world clinical scenarios contain extraneous information that can hinder performance. The rise of assistive technologies like ambient dictation, which automatically generates draft notes from live patient encounters, has the potential to introduce additional noise making it crucial to assess the ability of LLM's to filter relevant data. To investigate this, we developed MedDistractQA, a benchmark using USMLE-style questions embedded with simulated real-world distractions. Our findings show that distracting statements (polysemous words with clinical meanings used in a non-clinical context or references to unrelated health conditions) can reduce LLM accuracy by up to 17.9%. Commonly proposed solutions to improve model performance such as retrieval-augmented generation (RAG) and medical fine-tuning did not change this effect and in some cases introduced their own confounders and further degraded performance. Our findings suggest that LLMs natively lack the logical mechanisms necessary to distinguish relevant from irrelevant clinical information, posing challenges for real-world applications. MedDistractQA and our results highlights the need for robust mitigation strategies to enhance LLM resilience to extraneous information.
Authors: Feng Chen, Dror Ben-Zeev, Gillian Sparks, Arya Kadakia, Trevor Cohen
Abstract: Post-Traumatic Stress Disorder (PTSD) remains underdiagnosed in clinical settings, presenting opportunities for automated detection to identify patients. This study evaluates natural language processing approaches for detecting PTSD from clinical interview transcripts. We compared general and mental health-specific transformer models (BERT/RoBERTa), embedding-based methods (SentenceBERT/LLaMA), and large language model prompting strategies (zero-shot/few-shot/chain-of-thought) using the DAIC-WOZ dataset. Domain-specific models significantly outperformed general models (Mental-RoBERTa F1=0.643 vs. RoBERTa-base 0.485). LLaMA embeddings with neural networks achieved the highest performance (F1=0.700). Zero-shot prompting using DSM-5 criteria yielded competitive results without training data (F1=0.657). Performance varied significantly across symptom severity and comorbidity status, with higher accuracy for severe PTSD cases and patients with comorbid depression. Our findings highlight the potential of domain-adapted embeddings and LLMs for scalable screening while underscoring the need for improved detection of nuanced presentations and offering insights for developing clinically viable AI tools for PTSD assessment.
Authors: Gon\c{c}alo Gomes, Chrysoula Zerva, Bruno Martins
Abstract: This study explores current limitations of learned image captioning evaluation metrics, specifically the lack of granular assessment for individual word misalignments within captions, and the reliance on single-point quality estimates without considering uncertainty. To address these limitations, we propose a simple yet effective strategy for generating and calibrating CLIPScore distributions. Leveraging a model-agnostic conformal risk control framework, we calibrate CLIPScore values for task-specific control variables, to tackle the aforementioned two limitations. Experimental results demonstrate that using conformal risk control, over the distributions produced with simple methods such as input masking, can achieve competitive performance compared to more complex approaches. Our method effectively detects misaligned words, while providing formal guarantees aligned with desired risk levels, and improving the correlation between uncertainty estimations and prediction errors, thus enhancing the overall reliability of caption evaluation metrics.
Authors: Naimul Haque
Abstract: Large Language Models (LLMs) have significantly advanced Natural Language Processing (NLP), particularly in Natural Language Understanding (NLU) tasks. As we progress toward an agentic world where LLM-based agents autonomously handle specialized tasks, it becomes crucial for these models to adapt to new tasks without forgetting previously learned information - a challenge known as catastrophic forgetting. This study evaluates the continual fine-tuning of various open-source LLMs with different parameter sizes (specifically models under 10 billion parameters) on key NLU tasks from the GLUE benchmark, including SST-2, MRPC, CoLA, and MNLI. By employing prompt engineering and task-specific adjustments, we assess and compare the models' abilities to retain prior knowledge while learning new tasks. Our results indicate that models such as Phi-3.5-mini exhibit minimal forgetting while maintaining strong learning capabilities, making them well-suited for continual learning environments. Additionally, models like Orca-2-7b and Qwen2.5-7B demonstrate impressive learning abilities and overall performance after fine-tuning. This work contributes to understanding catastrophic forgetting in LLMs and highlights prompting engineering to optimize model performance for continual learning scenarios.
Authors: Rafael Giebisch, Ken E. Friedl, Lev Sorokin, Andrea Stocco
Abstract: In-car conversational systems bring the promise to improve the in-vehicle user experience. Modern conversational systems are based on Large Language Models (LLMs), which makes them prone to errors such as hallucinations, i.e., inaccurate, fictitious, and therefore factually incorrect information. In this paper, we present an LLM-based methodology for the automatic factual benchmarking of in-car conversational systems. We instantiate our methodology with five LLM-based methods, leveraging ensembling techniques and diverse personae to enhance agreement and minimize hallucinations. We use our methodology to evaluate CarExpert, an in-car retrieval-augmented conversational question answering system, with respect to the factual correctness to a vehicle's manual. We produced a novel dataset specifically created for the in-car domain, and tested our methodology against an expert evaluation. Our results show that the combination of GPT-4 with the Input Output Prompting achieves over 90 per cent factual correctness agreement rate with expert evaluations, other than being the most efficient approach yielding an average response time of 4.5s. Our findings suggest that LLM-based testing constitutes a viable approach for the validation of conversational systems regarding their factual correctness.
Authors: Niharika Dadu (Indian Institute of Technology Jodhpur), Harsh Vardhan Singh (Indian Institute of Technology Jodhpur), Romi Banerjee (Indian Institute of Technology Jodhpur)
Abstract: The advent of large language models (LLMs) in the education sector has provided impetus to automate grading short answer questions. LLMs make evaluating short answers very efficient, thus addressing issues like staff shortage. However, in the task of Automated Short Answer Grading (ASAG), LLM responses are influenced by diverse perspectives in their training dataset, leading to inaccuracies in evaluating nuanced or partially correct answers. To address this challenge, we propose a novel framework, Grade Guard. 1. To enhance the task-based specialization of the LLMs, the temperature parameter has been fine-tuned using Root Mean Square Error (RMSE). 2. Unlike traditional approaches, LLMs in Grade Guard compute an Indecisiveness Score (IS) along with the grade to reflect uncertainty in predicted grades. 3. Introduced Confidence-Aware Loss (CAL) to generate an optimized Indecisiveness Score (IS). 4. To improve reliability, self-reflection based on the optimized IS has been introduced into the framework, enabling human re-evaluation to minimize incorrect grade assignments. Our experimentation shows that the best setting of Grade Guard outperforms traditional methods by 19.16% RMSE in Upstage Solar Pro, 23.64% RMSE in Upstage Solar Mini, 4.00% RMSE in Gemini 1.5 Flash, and 10.20% RMSE in GPT 4-o Mini. Future work includes improving interpretability by generating rationales for grades to enhance accuracy. Expanding benchmark datasets and annotating them with domain-specific nuances will enhance grading accuracy. Finally, analyzing feedback to enhance confidence in predicted grades, reduce biases, optimize grading criteria, and personalize learning while supporting multilingual grading systems will make the solution more accurate, adaptable, fair, and inclusive.
Authors: Jihyun Janice Ahn, Wenpeng Yin
Abstract: While the inconsistency of LLMs is not a novel topic, prior research has predominantly addressed two types of generative inconsistencies: i) Randomness Inconsistency: running the same LLM multiple trials, yielding varying responses; ii) Paraphrase Inconsistency: paraphrased prompts result in different responses from the same LLM. Randomness Inconsistency arises from the inherent randomness due to stochastic sampling in generative models, while Paraphrase Inconsistency is a consequence of the language modeling objectives, where paraphrased prompts alter the distribution of vocabulary logits. This research discovers Prompt-Reverse Inconsistency (PRIN), a new form of LLM self-inconsistency: given a question and a couple of LLM-generated answer candidates, the LLM often has conflicting responses when prompted "Which are correct answers?" and "Which are incorrect answers?". PRIN poses a big concern as it undermines the credibility of LLM-as-a-judge, and suggests a challenge for LLMs to adhere to basic logical rules. We conduct a series of experiments to investigate PRIN, examining the extent of PRIN across different LLMs, methods to mitigate it, potential applications, and its relationship with Randomness Inconsistency and Paraphrase Inconsistency. As the first study to explore PRIN, our findings offer valuable insights into the inner workings of LLMs and contribute to advancing trustworthy AI.
Authors: Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, Shiyu Chang
Abstract: We present ThinkPrune, a simple yet effective method for pruning the thinking length for long-thinking LLMs, which has been found to often produce inefficient and redundant thinking processes. Existing preliminary explorations of reducing thinking length primarily focus on forcing the thinking process to early exit, rather than adapting the LLM to optimize and consolidate the thinking process, and therefore the length-performance tradeoff observed so far is sub-optimal. To fill this gap, ThinkPrune offers a simple solution that continuously trains the long-thinking LLMs via reinforcement learning (RL) with an added token limit, beyond which any unfinished thoughts and answers will be discarded, resulting in a zero reward. To further preserve model performance, we introduce an iterative length pruning approach, where multiple rounds of RL are conducted, each with an increasingly more stringent token limit. We observed that ThinkPrune results in a remarkable performance-length tradeoff -- on the AIME24 dataset, the reasoning length of DeepSeek-R1-Distill-Qwen-1.5B can be reduced by half with only 2% drop in performance. We also observed that after pruning, the LLMs can bypass unnecessary steps while keeping the core reasoning process complete. Code is available at https://github.com/UCSB-NLP-Chang/ThinkPrune.
Authors: Lingxiao Guan, Yuanhao Huang, Jie Liu
Abstract: In Question Answering (QA), Retrieval Augmented Generation (RAG) has revolutionized performance in various domains. However, how to effectively capture multi-document relationships, particularly critical for biomedical tasks, remains an open question. In this work, we propose a novel method that utilizes propositional claims to construct a local knowledge graph from retrieved documents. Summaries are then derived via layerwise summarization from the knowledge graph to contextualize a small language model to perform QA. We achieved comparable or superior performance with our method over RAG baselines on several biomedical QA benchmarks. We also evaluated each individual step of our methodology over a targeted set of metrics, demonstrating its effectiveness.
Authors: Zhendong Tan, Xingjun Zhang, Chaoyi Hu, Yancheng Pan, Shaoxun Wang
Abstract: The newly released OpenAI-o1 and DeepSeek-R1 have demonstrated that test-time scaling can significantly improve model performance, especially in complex tasks such as logical reasoning. Common test-time scaling methods involve generating more chain of thoughts (CoTs) or longer CoTs with self-correction. However, while self-correction can improve performance, it may lead to significant token waste and reduce readability of the CoT if the reasoning steps are already correct. To demonstrate that large language models (LLMs) can rectify errors at a more fine-grained level, we propose Adaptive Rectification Sampling (AR-Sampling), which can guide the LLMs to self-correction at the appropriate step. AR-Sampling leverages a process-supervised reward model (PRM) as a verifier and constructed trigger sentences to guide the model in adaptive step-level rethinking. Through the experiments on GSM8K and MATH500, it indicate that our approach enables the models to rethink in more fine-grained level, improving the accuracy of solutions, while generating a reasonable number of additional tokens.
Authors: Jungyeul Park
Abstract: This memoir explores two fundamental aspects of Natural Language Processing (NLP): the creation of linguistic resources and the evaluation of NLP system performance. Over the past decade, my work has focused on developing a morpheme-based annotation scheme for the Korean language that captures linguistic properties from morphology to semantics. This approach has achieved state-of-the-art results in various NLP tasks, including part-of-speech tagging, dependency parsing, and named entity recognition. Additionally, this work provides a comprehensive analysis of segmentation granularity and its critical impact on NLP system performance. In parallel with linguistic resource development, I have proposed a novel evaluation framework, the jp-algorithm, which introduces an alignment-based method to address challenges in preprocessing tasks like tokenization and sentence boundary detection (SBD). Traditional evaluation methods assume identical tokenization and sentence lengths between gold standards and system outputs, limiting their applicability to real-world data. The jp-algorithm overcomes these limitations, enabling robust end-to-end evaluations across a variety of NLP tasks. It enhances accuracy and flexibility by incorporating linear-time alignment while preserving the complexity of traditional evaluation metrics. This memoir provides key insights into the processing of morphologically rich languages, such as Korean, while offering a generalizable framework for evaluating diverse end-to-end NLP systems. My contributions lay the foundation for future developments, with broader implications for multilingual resource development and system evaluation.
Authors: Akil Raj Subedi, Taniya Shah, Aswani Kumar Cherukuri, Thanos Vasilakos
Abstract: Social media platforms like Twitter have increasingly relied on Natural Language Processing NLP techniques to analyze and understand the sentiments expressed in the user generated content. One such state of the art NLP model is Bidirectional Encoder Representations from Transformers BERT which has been widely adapted in sentiment analysis. BERT is susceptible to adversarial attacks. This paper aims to scrutinize the inherent vulnerabilities of such models in Twitter sentiment analysis. It aims to formulate a framework for constructing targeted adversarial texts capable of deceiving these models, while maintaining stealth. In contrast to conventional methodologies, such as Importance Reweighting, this framework core idea resides in its reliance on gradients to prioritize the importance of individual words within the text. It uses a whitebox approach to attain fine grained sensitivity, pinpointing words that exert maximal influence on the classification outcome. This paper is organized into three interdependent phases. It starts with fine-tuning a pre-trained BERT model on Twitter data. It then analyzes gradients of the model to rank words on their importance, and iteratively replaces those with feasible candidates until an acceptable solution is found. Finally, it evaluates the effectiveness of the adversarial text against the custom trained sentiment classification model. This assessment would help in gauging the capacity of the adversarial text to successfully subvert classification without raising any alarm.
Authors: Jiaru Zou, Dongqi Fu, Sirui Chen, Xinrui He, Zihao Li, Yada Zhu, Jiawei Han, Jingrui He
Abstract: Beyond pure text, a substantial amount of knowledge is stored in tables. In real-world scenarios, user questions often require retrieving answers that are distributed across multiple tables. GraphRAG has recently attracted much attention for enhancing LLMs' reasoning capabilities by organizing external knowledge to address ad-hoc and complex questions, exemplifying a promising direction for cross-table question answering. In this paper, to address the current gap in available data, we first introduce a multi-table benchmark, MutliTableQA, comprising 60k tables and 25k user queries collected from real-world sources. Then, we propose the first Graph-Table-RAG framework, namely GTR, which reorganizes table corpora into a heterogeneous graph, employs a hierarchical coarse-to-fine retrieval process to extract the most relevant tables, and integrates graph-aware prompting for downstream LLMs' tabular reasoning. Extensive experiments show that GTR exhibits superior cross-table question-answering performance while maintaining high deployment efficiency, demonstrating its real-world practical applicability.
Authors: Allison Koenecke, Jed Stiglitz, David Mimno, Matthew Wilkens
Abstract: The application of AI tools to the legal field feels natural: large legal document collections could be used with specialized AI to improve workflow efficiency for lawyers and ameliorate the "justice gap" for underserved clients. However, legal documents differ from the web-based text that underlies most AI systems. The challenges of legal AI are both specific to the legal domain, and confounded with the expectation of AI's high performance in high-stakes settings. We identify three areas of special relevance to practitioners: data curation, data annotation, and output verification. First, it is difficult to obtain usable legal texts. Legal collections are inconsistent, analog, and scattered for reasons technical, economic, and jurisdictional. AI tools can assist document curation efforts, but the lack of existing data also limits AI performance. Second, legal data annotation typically requires significant expertise to identify complex phenomena such as modes of judicial reasoning or controlling precedents. We describe case studies of AI systems that have been developed to improve the efficiency of human annotation in legal contexts and identify areas of underperformance. Finally, AI-supported work in the law is valuable only if results are verifiable and trustworthy. We describe both the abilities of AI systems to support evaluation of their outputs, as well as new approaches to systematic evaluation of computational systems in complex domains. We call on both legal and AI practitioners to collaborate across disciplines and to release open access materials to support the development of novel, high-performing, and reliable AI tools for legal applications.
Authors: Lin Zhang, Zhouhong Gu, Suhang Zheng, Tao Wang, Tianyu Li, Hongwei Feng, Yanghua Xiao
Abstract: This paper presents LITE, an LLM-based evaluation method designed for efficient and flexible assessment of taxonomy quality. To address challenges in large-scale taxonomy evaluation, such as efficiency, fairness, and consistency, LITE adopts a top-down hierarchical evaluation strategy, breaking down the taxonomy into manageable substructures and ensuring result reliability through cross-validation and standardized input formats. LITE also introduces a penalty mechanism to handle extreme cases and provides both quantitative performance analysis and qualitative insights by integrating evaluation metrics closely aligned with task objectives. Experimental results show that LITE demonstrates high reliability in complex evaluation tasks, effectively identifying semantic errors, logical contradictions, and structural flaws in taxonomies, while offering directions for improvement. Code is available at https://github.com/Zhang-l-i-n/TAXONOMY_DETECT .
Authors: Xingshan Zeng, Weiwen Liu, Xu Huang, Zezhong Wang, Lingzhi Wang, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Ruiming Tang, Qun Liu
Abstract: Tool learning, which allows Large Language Models (LLMs) to leverage external tools for solving complex user tasks, has emerged as a promising avenue for extending model capabilities. However, current approaches primarily focus on data synthesis for fine-tuning LLMs to invoke tools effectively, largely ignoring how to fully stimulate the potential of the model. In this paper, we propose ToolACE-R, a novel method that introduces adaptive self-refinement for tool invocations. Our approach features a model-aware iterative training procedure that progressively incorporates more training samples based on the model's evolving capabilities. Additionally, it allows LLMs to iteratively refine their tool calls, optimizing performance without requiring external feedback. To further enhance computational efficiency, we integrate an adaptive mechanism when scaling the inference time, enabling the model to autonomously determine when to stop the refinement process. We conduct extensive experiments across several benchmark datasets, showing that ToolACE-R achieves competitive performance compared to advanced API-based models, even without any refinement. Furthermore, its performance can be further improved efficiently through adaptive self-refinement. Our results demonstrate the effectiveness of the proposed method, which is compatible with base models of various sizes, offering a promising direction for more efficient tool learning.
Authors: Athena Wen, Tanush Patil, Ansh Saxena, Yicheng Fu, Sean O'Brien, Kevin Zhu
Abstract: In an era where AI-driven hiring is transforming recruitment practices, concerns about fairness and bias have become increasingly important. To explore these issues, we introduce a benchmark, FAIRE (Fairness Assessment In Resume Evaluation), to test for racial and gender bias in large language models (LLMs) used to evaluate resumes across different industries. We use two methods-direct scoring and ranking-to measure how model performance changes when resumes are slightly altered to reflect different racial or gender identities. Our findings reveal that while every model exhibits some degree of bias, the magnitude and direction vary considerably. This benchmark provides a clear way to examine these differences and offers valuable insights into the fairness of AI-based hiring tools. It highlights the urgent need for strategies to reduce bias in AI-driven recruitment. Our benchmark code and dataset are open-sourced at our repository: https://github.com/athenawen/FAIRE-Fairness-Assessment-In-Resume-Evaluation.git.
URLs: https://github.com/athenawen/FAIRE-Fairness-Assessment-In-Resume-Evaluation.git.
Authors: Zhaoxing Li, Xiaoming Zhang, Haifeng Zhang, Chengxiang Liu
Abstract: The integration of Large Language Models (LLMs) with Graph Neural Networks (GNNs) has recently been explored to enhance the capabilities of Text Attribute Graphs (TAGs). Most existing methods feed textual descriptions of the graph structure or neighbouring nodes' text directly into LLMs. However, these approaches often cause LLMs to treat structural information simply as general contextual text, thus limiting their effectiveness in graph-related tasks. In this paper, we introduce LanSAGNN (Language Semantic Anisotropic Graph Neural Network), a framework that extends the concept of anisotropic GNNs to the natural language level. This model leverages LLMs to extract tailor-made semantic information for node pairs, effectively capturing the unique interactions within node relationships. In addition, we propose an efficient dual-layer LLMs finetuning architecture to better align LLMs' outputs with graph tasks. Experimental results demonstrate that LanSAGNN significantly enhances existing LLM-based methods without increasing complexity while also exhibiting strong robustness against interference.
Authors: Zhengwei Tao, Zhi Jin, Bincheng Li, Xiaoying Bai, Haiyan Zhao, Chengfeng Dou, Xiancai Chen, Jia Li, Linyu Li, Chongyang Tao
Abstract: Predicting future events stands as one of the ultimate aspirations of artificial intelligence. Recent advances in large language model (LLM)-based systems have shown remarkable potential in forecasting future events, thereby garnering significant interest in the research community. Currently, several benchmarks have been established to evaluate the forecasting capabilities by formalizing the event prediction as a retrieval-augmented generation (RAG) and reasoning task. In these benchmarks, each prediction question is answered with relevant retrieved news articles. However, because there is no consideration on whether the questions can be supported by valid or sufficient supporting rationales, some of the questions in these benchmarks may be inherently noninferable. To address this issue, we introduce a new benchmark, PROPHET, which comprises inferable forecasting questions paired with relevant news for retrieval. To ensure the inferability of the benchmark, we propose Causal Intervened Likelihood (CIL), a statistical measure that assesses inferability through causal inference. In constructing this benchmark, we first collected recent trend forecasting questions and then filtered the data using CIL, resulting in an inferable benchmark for event prediction. Through extensive experiments, we first demonstrate the validity of CIL and in-depth investigations into event prediction with the aid of CIL. Subsequently, we evaluate several representative prediction systems on PROPHET, drawing valuable insights for future directions.
Authors: Zhiyuan Tang, Dong Wang, Zhikai Zhou, Yong Liu, Shen Huang, Shidong Shang
Abstract: Full-text error correction with Large Language Models (LLMs) for Automatic Speech Recognition (ASR) has gained increased attention due to its potential to correct errors across long contexts and address a broader spectrum of error types, including punctuation restoration and inverse text normalization. Nevertheless, many challenges persist, including issues related to stability, controllability, completeness, and fluency. To mitigate these challenges, this paper proposes the Chain of Correction (CoC) for full-text error correction with LLMs, which corrects errors segment by segment using pre-recognized text as guidance within a regular multi-turn chat format. The CoC also uses pre-recognized full text for context, allowing the model to better grasp global semantics and maintain a comprehensive overview of the entire content. Utilizing the open-sourced full-text error correction dataset ChFT, we fine-tune a pre-trained LLM to evaluate the performance of the CoC framework. Experimental results demonstrate that the CoC effectively corrects errors in full-text ASR outputs, significantly outperforming baseline and benchmark systems. We further analyze how to set the correction threshold to balance under-correction and over-rephrasing, extrapolate the CoC model on extremely long ASR outputs, and investigate whether other types of information can be employed to guide the error correction process.
Authors: Adrien Schurger-Foy, Rafal Dariusz Kocielnik, Caglar Gulcehre, R. Michael Alvarez
Abstract: The detrimental effects of toxicity in competitive online video games are widely acknowledged, prompting publishers to monitor player chat conversations. This is challenging due to the context-dependent nature of toxicity, often spread across multiple messages or informed by non-textual interactions. Traditional toxicity detectors focus on isolated messages, missing the broader context needed for accurate moderation. This is especially problematic in video games, where interactions involve specialized slang, abbreviations, and typos, making it difficult for standard models to detect toxicity, especially given its rarity. We adapted RoBERTa LLM to support moderation tailored to video games, integrating both textual and non-textual context. By enhancing pretrained embeddings with metadata and addressing the unique slang and language quirks through domain adaptive pretraining, our method better captures the nuances of player interactions. Using two gaming datasets - from Defense of the Ancients 2 (DOTA 2) and Call of Duty$^\circledR$: Modern Warfare$^\circledR$III (MWIII) we demonstrate which sources of context (metadata, prior interactions...) are most useful, how to best leverage them to boost performance, and the conditions conducive to doing so. This work underscores the importance of context-aware and domain-specific approaches for proactive moderation.
Authors: Mikkel Wildner Kildeberg, Emil Allerslev Schledermann, Nicolaj Larsen, Rob van der Goot
Abstract: The best performing transformer-based language models use subword tokenization techniques, such as Byte-Pair-Encoding (BPE). However, these approaches often overlook linguistic principles, such as morphological segmentation, which we believe is fundamental for understanding language-specific word structure. In this study, we leverage an annotated Danish morphological dataset to train a semisupervised model for morphological segmentation, enabling the development of tokenizers optimized for Danish morphology. We evaluate four distinct tokenizers, including two custom morphological tokenizers, by analyzing their performance in morphologically segmenting Danish words. Additionally, we train two generative transformer models, \textit{CerebrasGPT-111M} and \textit{LLaMA-3.2 1B}, using these tokenizers and evaluate their downstream performance. Our findings reveal that our custom-developed tokenizers substantially enhance morphological segmentation, achieving an F1 score of 58.84, compared to 39.28 achieved by a Danish BPE tokenizer. In downstream tasks, models trained with our morphological tokenizers outperform those using BPE tokenizers across different evaluation metrics. These results highlight that incorporating Danish morphological segmentation strategies into tokenizers leads to improved performance in generative transformer models on Danish language
Authors: Amanda Myntti, Erik Henriksson, Veronika Laippala, Sampo Pyysalo
Abstract: Pretraining data curation is a cornerstone in Large Language Model (LLM) development, leading to growing research on quality filtering of large web corpora. From statistical quality flags to LLM-based labeling systems, datasets are divided into categories, frequently reducing to a binary: those passing the filters deemed as valuable examples, others discarded as useless or detrimental. However, a more detailed understanding of the contribution of different kinds of texts to model performance is still largely lacking. In this article, we present the first study utilizing registers (also known as genres) - a widely used standard in corpus linguistics to model linguistic variation - to curate pretraining datasets and investigate the effect of register on the performance of LLMs. We perform comparative studies by training models with register classified data and evaluating them using standard benchmarks, and show that the register of pretraining data substantially affects model performance. We uncover surprising relationships between the pretraining material and the resulting models: using the News register results in subpar performance, and on the contrary, including the Opinion class, covering texts such as reviews and opinion blogs, is highly beneficial. While a model trained on the entire unfiltered dataset outperforms those trained on datasets limited to a single register, combining well-performing registers like How-to-Instructions, Informational Description, and Opinion leads to major improvements. Furthermore, analysis of individual benchmark results reveals key differences in the strengths and drawbacks of specific register classes as pretraining data. These findings show that register is an important explainer of model variation and can facilitate more deliberate future data selection practices.
Authors: Cedric Lothritz, Jordi Cabot
Abstract: Large Language Models (LLMs) have become an increasingly important tool in research and society at large. While LLMs are regularly used all over the world by experts and lay-people alike, they are predominantly developed with English-speaking users in mind, performing well in English and other wide-spread languages while less-resourced languages such as Luxembourgish are seen as a lower priority. This lack of attention is also reflected in the sparsity of available evaluation tools and datasets. In this study, we investigate the viability of language proficiency exams as such evaluation tools for the Luxembourgish language. We find that large models such as ChatGPT, Claude and DeepSeek-R1 typically achieve high scores, while smaller models show weak performances. We also find that the performances in such language exams can be used to predict performances in other NLP tasks.
Authors: Yi-Long Lu, Chunhui Zhang, Jiajun Song, Lifeng Fan, Wei Wang
Abstract: Recent advancements in rule-based reinforcement learning (RL), applied during the post-training phase of large language models (LLMs), have significantly enhanced their capabilities in structured reasoning tasks such as mathematics and logical inference. However, the effectiveness of RL in social reasoning, particularly in Theory of Mind (ToM), the ability to infer others' mental states, remains largely unexplored. In this study, we demonstrate that RL methods effectively unlock ToM reasoning capabilities even in small-scale LLMs (0.5B to 7B parameters). Using a modest dataset comprising 3200 questions across diverse scenarios, our RL-trained 7B model achieves 84.50\% accuracy on the Hi-ToM benchmark, surpassing models like GPT-4o and DeepSeek-v3 despite significantly fewer parameters. While smaller models ($\leq$3B parameters) suffer from reasoning collapse, larger models (7B parameters) maintain stable performance through consistent belief tracking. Additionally, our RL-based models demonstrate robust generalization to higher-order, out-of-distribution ToM problems, novel textual presentations, and previously unseen datasets. These findings highlight RL's potential to enhance social cognitive reasoning, bridging the gap between structured problem-solving and nuanced social inference in LLMs.
Authors: Bowen Cao, Deng Cai, Wai Lam
Abstract: In-context learning (ICL) is critical for large language models (LLMs), but its effectiveness is constrained by finite context windows, particularly in ultra-long contexts. To overcome this, we introduce InfiniteICL, a framework that parallels context and parameters in LLMs with short- and long-term memory in human cognitive systems, focusing on transforming temporary context knowledge into permanent parameter updates. This approach significantly reduces memory usage, maintains robust performance across varying input lengths, and theoretically enables infinite context integration through the principles of context knowledge elicitation, selection, and consolidation. Evaluations demonstrate that our method reduces context length by 90% while achieving 103% average performance of full-context prompting across fact recall, grounded reasoning, and skill acquisition tasks. When conducting sequential multi-turn transformations on complex, real-world contexts (with length up to 2M tokens), our approach surpasses full-context prompting while using only 0.4% of the original contexts. These findings highlight InfiniteICL's potential to enhance the scalability and efficiency of LLMs by breaking the limitations of conventional context window sizes.
Authors: Philip Lippmann, Jie Yang
Abstract: Specialized reasoning language models (RLMs) have demonstrated that scaling test-time computation through detailed reasoning traces significantly enhances performance. Although these traces effectively facilitate knowledge distillation into smaller, instruction-tuned models, the precise nature of transferred reasoning remains unclear. In this study, we investigate to what extent distilled models internalize replicated stylistic patterns during reasoning. To this end, we systematically analyze reasoning traces, identifying structural and lexical patterns that characterize successful reasoning. We then introduce two new datasets -- a dataset of emergent reasoning traces and a synthetic dataset explicitly constructed to replicate these stylistic patterns -- to precisely examine their influence on distilled models' reasoning capabilities. We find that models trained on the synthetic traces achieve comparable performance, indicating that distilled reasoning abilities rely significantly on surface-level patterns. Surprisingly, we observe an increase in performance even when the synthetic traces are altered to lead to the wrong answer. Our findings highlight how stylistic patterns can be leveraged to efficiently enhance LM reasoning across diverse model families.
Authors: Sumeth Yuenyong, Thodsaporn Chay-intr, Kobkrit Viriyayudhakorn
Abstract: We present OpenThaiGPT 1.6 and R1 (OTG-1.6 and OTG-R1), Thai-centric Large Language Models (LLMs) developed through distinct methodologies to enhance generalization and reasoning capabilities. OTG-1.6 employs Task Arithmetic model merging for broad generalization, while OTG-R1 integrates multi-stage training with the Less-Is-More Reasoning Hypothesis (LIMO) for advanced reasoning. Benchmark evaluations demonstrate superior performance across Thai language tasks, achieving competitive results against larger-scale open-source Thai LLMs. This paper details the proposed models, training processes, benchmarks, and results, highlighting improvements over previous models and establishing new performance standards for Thai-centric LLMs.
Authors: Zhijun Wang, Jiahuan Li, Hao Zhou, Rongxiang Weng, Jingang Wang, Xin Huang, Xue Han, Junlan Feng, Chao Deng, Shujian Huang
Abstract: Large language models (LLMs) exhibit remarkable multilingual capabilities despite the extreme language imbalance in the pre-training data. In this paper, we closely examine the reasons behind this phenomenon, focusing on the pre-training corpus. We find that the existence of code-switching, alternating between different languages within a context, is key to multilingual capabilities. We conduct an analysis to investigate code-switching in the pre-training corpus, examining its presence and categorizing it into four types within two quadrants. We then assess its impact on multilingual performance. These types of code-switching data are unbalanced in proportions and demonstrate different effects on facilitating language transfer. To better explore the power of code-switching for language alignment during pre-training, we investigate the strategy of synthetic code-switching. We continuously scale up the synthetic code-switching data and observe remarkable improvements in both benchmarks and representation space. Extensive experiments indicate that incorporating synthetic code-switching data enables better language alignment and generalizes well to high, medium, and low-resource languages with pre-training corpora of varying qualities.
Authors: Sumuk Shashidhar, Cl\'ementine Fourrier, Alina Lozovskia, Thomas Wolf, Gokhan Tur, Dilek Hakkani-T\"ur
Abstract: Evaluating large language models (LLMs) effectively remains a critical bottleneck, as traditional static benchmarks suffer from saturation and contamination, while human evaluations are costly and slow. This hinders timely or domain-specific assessment, crucial for real-world applications. We introduce YourBench, a novel, open-source framework that addresses these limitations by enabling dynamic, automated generation of reliable, up-to-date, and domain-tailored benchmarks cheaply and without manual annotation, directly from user-provided documents. We demonstrate its efficacy by replicating 7 diverse MMLU subsets using minimal source text, achieving this for under 15 USD in total inference costs while perfectly preserving the relative model performance rankings (Spearman Rho = 1) observed on the original benchmark. To ensure that YourBench generates data grounded in provided input instead of relying on posterior parametric knowledge in models, we also introduce Tempora-0325, a novel dataset of over 7K diverse documents, published exclusively after March 2025. Our comprehensive analysis spans 26 SoTA models from 7 major families across varying scales (3-671B parameters) to validate the quality of generated evaluations through rigorous algorithmic checks (e.g., citation grounding) and human assessments. We release the YourBench library, the Tempora-0325 dataset, 150k+ question answer pairs based on Tempora and all evaluation and inference traces to facilitate reproducible research and empower the community to generate bespoke benchmarks on demand, fostering more relevant and trustworthy LLM evaluation.
Authors: Minhu Park, Hongseok Oh, Eunkyung Choi, Wonseok Hwang
Abstract: Recently, building retrieval-augmented generation (RAG) systems to enhance the capability of large language models (LLMs) has become a common practice. Especially in the legal domain, previous judicial decisions play a significant role under the doctrine of stare decisis which emphasizes the importance of making decisions based on (retrieved) prior documents. However, the overall performance of RAG system depends on many components: (1) retrieval corpora, (2) retrieval algorithms, (3) rerankers, (4) LLM backbones, and (5) evaluation metrics. Here we propose LRAGE, an open-source tool for holistic evaluation of RAG systems focusing on the legal domain. LRAGE provides GUI and CLI interfaces to facilitate seamless experiments and investigate how changes in the aforementioned five components affect the overall accuracy. We validated LRAGE using multilingual legal benches including Korean (KBL), English (LegalBench), and Chinese (LawBench) by demonstrating how the overall accuracy changes when varying the five components mentioned above. The source code is available at https://github.com/hoorangyee/LRAGE.
Authors: Zhiwei Yu, Tuo Li, Changhong Wang, Hui Chen, Lang Zhou
Abstract: Chain-of-thought (CoT) has emerged as a critical mechanism for enhancing reasoning capabilities in large language models (LLMs), with self-consistency demonstrating notable promise in boosting performance. However, inherent linguistic biases in multilingual training corpora frequently cause semantic drift and logical inconsistencies, especially in sub-10B parameter LLMs handling complex inference tasks. To overcome these constraints, we propose the Cross-Lingual Consistency (CLC) framework, an innovative inference paradigm that integrates multilingual reasoning paths through majority voting to elevate LLMs' reasoning capabilities. Empirical evaluations on the CMATH dataset reveal CLC's superiority over the conventional self-consistency method, delivering 9.5%, 6.5%, and 6.0% absolute accuracy gains for DeepSeek-Math-7B-Instruct, Qwen2.5-Math-7B-Instruct, and Gemma2-9B-Instruct respectively. Expanding CLC's linguistic scope to 11 diverse languages implies two synergistic benefits: 1) neutralizing linguistic biases in multilingual training corpora through multilingual ensemble voting, 2) escaping monolingual reasoning traps by exploring the broader multilingual solution space. This dual benefits empirically enables more globally optimal reasoning paths compared to monolingual self-consistency baselines, as evidenced by the 4.1%-18.5% accuracy gains using Gemma2-9B-Instruct on the MGSM dataset.
Authors: Abhilash Shankarampeta, Harsh Mahajan, Tushar Kataria, Dan Roth, Vivek Gupta
Abstract: Humans continuously make new discoveries, and understanding temporal sequence of events leading to these breakthroughs is essential for advancing science and society. This ability to reason over time allows us to identify future steps and understand the effects of financial and political decisions on our lives. However, large language models (LLMs) are typically trained on static datasets, limiting their ability to perform effective temporal reasoning. To assess the temporal reasoning capabilities of LLMs, we present the TRANSIENTTABLES dataset, which comprises 3,971 questions derived from over 14,000 tables, spanning 1,238 entities across multiple time periods. We introduce a template-based question-generation pipeline that harnesses LLMs to refine both templates and questions. Additionally, we establish baseline results using state-of-the-art LLMs to create a benchmark. We also introduce novel modeling strategies centered around task decomposition, enhancing LLM performance.
Authors: C\'elia Nouri, Jean-Philippe Cointet, Chlo\'e Clavel
Abstract: Detecting abusive language in social media conversations poses significant challenges, as identifying abusiveness often depends on the conversational context, characterized by the content and topology of preceding comments. Traditional Abusive Language Detection (ALD) models often overlook this context, which can lead to unreliable performance metrics. Recent Natural Language Processing (NLP) methods that integrate conversational context often depend on limited and simplified representations, and report inconsistent results. In this paper, we propose a novel approach that utilize graph neural networks (GNNs) to model social media conversations as graphs, where nodes represent comments, and edges capture reply structures. We systematically investigate various graph representations and context windows to identify the optimal configuration for ALD. Our GNN model outperform both context-agnostic baselines and linear context-aware methods, achieving significant improvements in F1 scores. These findings demonstrate the critical role of structured conversational context and establish GNNs as a robust framework for advancing context-aware abusive language detection.
Authors: Zijun Wang, Haoqin Tu, Yuhan Wang, Juncheng Wu, Jieru Mei, Brian R. Bartoldson, Bhavya Kailkhura, Cihang Xie
Abstract: This paper introduces STAR-1, a high-quality, just-1k-scale safety dataset specifically designed for large reasoning models (LRMs) like DeepSeek-R1. Built on three core principles -- diversity, deliberative reasoning, and rigorous filtering -- STAR-1 aims to address the critical needs for safety alignment in LRMs. Specifically, we begin by integrating existing open-source safety datasets from diverse sources. Then, we curate safety policies to generate policy-grounded deliberative reasoning samples. Lastly, we apply a GPT-4o-based safety scoring system to select training examples aligned with best practices. Experimental results show that fine-tuning LRMs with STAR-1 leads to an average 40% improvement in safety performance across four benchmarks, while only incurring a marginal decrease (e.g., an average of 1.1%) in reasoning ability measured across five reasoning tasks. Extensive ablation studies further validate the importance of our design principles in constructing STAR-1 and analyze its efficacy across both LRMs and traditional LLMs. Our project page is https://ucsc-vlaa.github.io/STAR-1.
Authors: Baban Gain, Dibyanayan Bandyopadhyay, Asif Ekbal
Abstract: The advent of Large Language Models (LLMs) has significantly reshaped the landscape of machine translation (MT), particularly for low-resource languages and domains that lack sufficient parallel corpora, linguistic tools, and computational infrastructure. This survey presents a comprehensive overview of recent progress in leveraging LLMs for MT. We analyze techniques such as few-shot prompting, cross-lingual transfer, and parameter-efficient fine-tuning that enable effective adaptation to under-resourced settings. The paper also explores synthetic data generation strategies using LLMs, including back-translation and lexical augmentation. Additionally, we compare LLM-based translation with traditional encoder-decoder models across diverse language pairs, highlighting the strengths and limitations of each. We discuss persistent challenges such as hallucinations, evaluation inconsistencies, and inherited biases while also evaluating emerging LLM-driven metrics for translation quality. This survey offers practical insights and outlines future directions for building robust, inclusive, and scalable MT systems in the era of large-scale generative models.
Authors: Boshi Wang, Huan Sun
Abstract: Despite their impressive capabilities, LLMs exhibit a basic generalization failure known as the Reversal Curse, where they struggle to learn reversible factual associations. Understanding why this occurs could help identify weaknesses in current models and advance their generalization and robustness. In this paper, we conjecture that the Reversal Curse in LLMs is a manifestation of the long-standing binding problem in cognitive science, neuroscience and AI. Specifically, we identify two primary causes of the Reversal Curse stemming from transformers' limitations in conceptual binding: the inconsistency and entanglements of concept representations. We perform a series of experiments that support these conjectures. Our exploration leads to a model design based on JEPA (Joint-Embedding Predictive Architecture) that for the first time breaks the Reversal Curse without side-stepping it with specialized data augmentation or non-causal masking, and moreover, generalization could be further improved by incorporating special memory layers that support disentangled concept representations. We demonstrate that the skill of reversal unlocks a new kind of memory integration that enables models to solve large-scale arithmetic reasoning problems via parametric forward-chaining, outperforming frontier LLMs based on non-parametric memory and prolonged explicit reasoning.
Authors: Washington Cunha, Leonardo Rocha, Marcos Andr\'e Gon\c{c}alves
Abstract: Automatic text classification (ATC) has experienced remarkable advancements in the past decade, best exemplified by recent small and large language models (SLMs and LLMs), leveraged by Transformer architectures. Despite recent effectiveness improvements, a comprehensive cost-benefit analysis investigating whether the effectiveness gains of these recent approaches compensate their much higher costs when compared to more traditional text classification approaches such as SVMs and Logistic Regression is still missing in the literature. In this context, this work's main contributions are twofold: (i) we provide a scientifically sound comparative analysis of the cost-benefit of twelve traditional and recent ATC solutions including five open LLMs, and (ii) a large benchmark comprising {22 datasets}, including sentiment analysis and topic classification, with their (train-validation-test) partitions based on folded cross-validation procedures, along with documentation, and code. The release of code, data, and documentation enables the community to replicate experiments and advance the field in a more scientifically sound manner. Our comparative experimental results indicate that LLMs outperform traditional approaches (up to 26%-7.1% on average) and SLMs (up to 4.9%-1.9% on average) in terms of effectiveness. However, LLMs incur significantly higher computational costs due to fine-tuning, being, on average 590x and 8.5x slower than traditional methods and SLMs, respectively. Results suggests the following recommendations: (1) LLMs for applications that require the best possible effectiveness and can afford the costs; (2) traditional methods such as Logistic Regression and SVM for resource-limited applications or those that cannot afford the cost of tuning large LLMs; and (3) SLMs like Roberta for near-optimal effectiveness-efficiency trade-off.
Authors: Souradip Chakraborty, Mohammadreza Pourreza, Ruoxi Sun, Yiwen Song, Nino Scherrer, Jindong Gu, Furong Huang, Amrit Singh Bedi, Ahmad Beirami, Hamid Palangi, Tomas Pfister
Abstract: While AI agents have shown remarkable performance at various tasks, they still struggle with complex multi-modal applications, structured generation and strategic planning. Improvements via standard fine-tuning is often impractical, as solving agentic tasks usually relies on black box API access without control over model parameters. Inference-time methods such as Best-of-N (BON) sampling offer a simple yet effective alternative to improve performance. However, BON lacks iterative feedback integration mechanism. Hence, we propose Iterative Agent Decoding (IAD) which combines iterative refinement with dynamic candidate evaluation and selection guided by a verifier. IAD differs in how feedback is designed and integrated, specifically optimized to extract maximal signal from reward scores. We conduct a detailed comparison of baselines across key metrics on Sketch2Code, Text2SQL, and Webshop where IAD consistently outperforms baselines, achieving 3--6% absolute gains on Sketch2Code and Text2SQL (with and without LLM judges) and 8--10% gains on Webshop across multiple metrics. To better understand the source of IAD's gains, we perform controlled experiments to disentangle the effect of adaptive feedback from stochastic sampling, and find that IAD's improvements are primarily driven by verifier-guided refinement, not merely sampling diversity. We also show that both IAD and BON exhibit inference-time scaling with increased compute when guided by an optimal verifier. Our analysis highlights the critical role of verifier quality in effective inference-time optimization and examines the impact of noisy and sparse rewards on scaling behavior. Together, these findings offer key insights into the trade-offs and principles of effective inference-time optimization.
Authors: Wasi Uddin Ahmad, Sean Narenthiran, Somshubra Majumdar, Aleksander Ficek, Siddhartha Jain, Jocelyn Huang, Vahid Noroozi, Boris Ginsburg
Abstract: Since the advent of reasoning-based large language models, many have found great success from distilling reasoning capabilities into student models. Such techniques have significantly bridged the gap between reasoning and standard LLMs on coding tasks. Despite this, much of the progress on distilling reasoning models remains locked behind proprietary datasets or lacks details on data curation, filtering and subsequent training. To address this, we construct a superior supervised fine-tuning (SFT) dataset that we use to achieve state-of-the-art coding capability results in models of various sizes. Our distilled models use only SFT to achieve 61.8% on LiveCodeBench and 24.6% on CodeContests, surpassing alternatives trained with reinforcement learning. We then perform analysis on the data sources used to construct our dataset, the impact of code execution filtering, and the importance of instruction/solution diversity. We observe that execution filtering negatively affected benchmark accuracy, leading us to prioritize instruction diversity over solution correctness. Finally, we also analyze the token efficiency and reasoning patterns utilized by these models. We will open-source these datasets and distilled models to the community.
Authors: Anket Mehra, Malte Prie{\ss}, Marian Himstedt
Abstract: This paper shows that further evaluation metrics during model training are needed to decide about its applicability in inference. As an example, a LayoutLM-based model is trained for token classification in documents. The documents are German receipts. We show that conventional classification metrics, represented by the F1-Score in our experiments, are insufficient for evaluating the applicability of machine learning models in practice. To address this problem, we introduce a novel metric, Document Integrity Precision (DIP), as a solution for visual document understanding and the token classification task. To the best of our knowledge, nothing comparable has been introduced in this context. DIP is a rigorous metric, describing how many documents of the test dataset require manual interventions. It enables AI researchers and software developers to conduct an in-depth investigation of the level of process automation in business software. In order to validate DIP, we conduct experiments with our created models to highlight and analyze the impact and relevance of DIP to evaluate if the model should be deployed or not in different training settings. Our results demonstrate that existing metrics barely change for isolated model impairments, whereas DIP indicates that the model requires substantial human interventions in deployment. The larger the set of entities being predicted, the less sensitive conventional metrics are, entailing poor automation quality. DIP, in contrast, remains a single value to be interpreted for entire entity sets. This highlights the importance of having metrics that focus on the business task for model training in production. Since DIP is created for the token classification task, more research is needed to find suitable metrics for other training tasks.
Authors: Wenjun Zeng, Dana Kurniawan, Ryan Mullins, Yuchi Liu, Tamoghna Saha, Dirichi Ike-Njoku, Jindong Gu, Yiwen Song, Cai Xu, Jingjing Zhou, Aparna Joshi, Shravan Dheep, Mani Malek, Hamid Palangi, Joon Baek, Rick Pereira, Karthik Narasimhan
Abstract: We introduce ShieldGemma 2, a 4B parameter image content moderation model built on Gemma 3. This model provides robust safety risk predictions across the following key harm categories: Sexually Explicit, Violence \& Gore, and Dangerous Content for synthetic images (e.g. output of any image generation model) and natural images (e.g. any image input to a Vision-Language Model). We evaluated on both internal and external benchmarks to demonstrate state-of-the-art performance compared to LlavaGuard \citep{helff2024llavaguard}, GPT-4o mini \citep{hurst2024gpt}, and the base Gemma 3 model \citep{gemma_2025} based on our policies. Additionally, we present a novel adversarial data generation pipeline which enables a controlled, diverse, and robust image generation. ShieldGemma 2 provides an open image moderation tool to advance multimodal safety and responsible AI development.
Authors: Jaechul Roh, Virat Shejwalkar, Amir Houmansadr
Abstract: Large Audio Language Models (LALMs) have significantly advanced audio understanding but introduce critical security risks, particularly through audio jailbreaks. While prior work has focused on English-centric attacks, we expose a far more severe vulnerability: adversarial multilingual and multi-accent audio jailbreaks, where linguistic and acoustic variations dramatically amplify attack success. In this paper, we introduce Multi-AudioJail, the first systematic framework to exploit these vulnerabilities through (1) a novel dataset of adversarially perturbed multilingual/multi-accent audio jailbreaking prompts, and (2) a hierarchical evaluation pipeline revealing that how acoustic perturbations (e.g., reverberation, echo, and whisper effects) interacts with cross-lingual phonetics to cause jailbreak success rates (JSRs) to surge by up to +57.25 percentage points (e.g., reverberated Kenyan-accented attack on MERaLiON). Crucially, our work further reveals that multimodal LLMs are inherently more vulnerable than unimodal systems: attackers need only exploit the weakest link (e.g., non-English audio inputs) to compromise the entire model, which we empirically show by multilingual audio-only attacks achieving 3.1x higher success rates than text-only attacks. We plan to release our dataset to spur research into cross-modal defenses, urging the community to address this expanding attack surface in multimodality as LALMs evolve.
Authors: Nicholas Clark, Hua Shen, Bill Howe, Tanushree Mitra
Abstract: LLMs increasingly serve as tools for knowledge acquisition, yet users cannot effectively specify how they want information presented. When users request that LLMs "cite reputable sources," "express appropriate uncertainty," or "include multiple perspectives," they discover that current interfaces provide no structured way to articulate these preferences. The result is prompt sharing folklore: community-specific copied prompts passed through trust relationships rather than based on measured efficacy. We propose the Epistemic Alignment Framework, a set of ten challenges in knowledge transmission derived from the philosophical literature of epistemology, concerning issues such as evidence quality assessment and calibration of testimonial reliance. The framework serves as a structured intermediary between user needs and system capabilities, creating a common vocabulary to bridge the gap between what users want and what systems deliver. Through a thematic analysis of custom prompts and personalization strategies shared on online communities where these issues are actively discussed, we find users develop elaborate workarounds to address each of the challenges. We then apply our framework to two prominent model providers, OpenAI and Anthropic, through content analysis of their documented policies and product features. Our analysis shows that while these providers have partially addressed the challenges we identified, they fail to establish adequate mechanisms for specifying epistemic preferences, lack transparency about how preferences are implemented, and offer no verification tools to confirm whether preferences were followed. For AI developers, the Epistemic Alignment Framework offers concrete guidance for supporting diverse approaches to knowledge; for users, it works toward information delivery that aligns with their specific needs rather than defaulting to one-size-fits-all approaches.
Authors: Sakhinana Sagar Srinivas, Venkataramana Runkana
Abstract: We present a comprehensive framework for enhancing Retrieval-Augmented Generation (RAG) systems through dynamic retrieval strategies and reinforcement fine-tuning. This approach significantly improves large language models on knowledge-intensive tasks, including opendomain question answering and complex reasoning. Our framework integrates two complementary techniques: Policy-Optimized RetrievalAugmented Generation (PORAG), which optimizes the use of retrieved information, and Adaptive Token-Layer Attention Scoring (ATLAS), which dynamically determines retrieval timing and content based on contextual needs. Together, these techniques enhance both the utilization and relevance of retrieved content, improving factual accuracy and response quality. Designed as a lightweight solution compatible with any Transformer-based LLM without requiring additional training, our framework excels in knowledge-intensive tasks, boosting output accuracy in RAG settings. We further propose CRITIC, a novel method to selectively compress key-value caches by token importance, mitigating memory bottlenecks in long-context applications. The framework also incorporates test-time scaling techniques to dynamically balance reasoning depth and computational resources, alongside optimized decoding strategies for faster inference. Experiments on benchmark datasets show that our framework reduces hallucinations, strengthens domain-specific reasoning, and achieves significant efficiency and scalability gains over traditional RAG systems. This integrated approach advances the development of robust, efficient, and scalable RAG systems across diverse applications.
Authors: Ke Zhu, Yu Wang, Jiangjiang Liu, Qunyi Xie, Shanshan Liu, Gang Zhang
Abstract: This paper is a pioneering work attempting to address abstract visual reasoning (AVR) problems for large vision-language models (VLMs). We make a common LLaVA-NeXT 7B model capable of perceiving and reasoning about specific AVR problems, surpassing both open-sourced (e.g., Qwen-2-VL-72B) and closed-sourced powerful VLMs (e.g., GPT-4o) with significant margin. This is a great breakthrough since almost all previous VLMs fail or show nearly random performance on representative AVR benchmarks. Our key success is our innovative data synthesis and post-training process, aiming to fully relieve the task difficulty and elicit the model to learn, step by step. Our 7B model is also shown to be behave well on AVR without sacrificing common multimodal comprehension abilities. We hope our paper could serve as an early effort in this area and would inspire further research in abstract visual reasoning.
Authors: Mohan Zhang, Pingzhi Li, Jie Peng, Mufan Qiu, Tianlong Chen
Abstract: Mixture-of-Experts (MoE) has successfully scaled up models while maintaining nearly constant computing costs. By employing a gating network to route input tokens, it selectively activates a subset of expert networks to process the corresponding token embeddings. However, in practice, the efficiency of MoE is challenging to achieve due to two key reasons: imbalanced expert activation, which leads to substantial idle time during model or expert parallelism, and insufficient capacity utilization; massive communication overhead, induced by numerous expert routing combinations in expert parallelism at the system level. Previous works typically formulate it as the load imbalance issue characterized by the gating network favoring certain experts over others or attribute it to static execution which fails to adapt to the dynamic expert workload at runtime. In this paper, we exploit it from a brand new perspective, a higher-order view and analysis of MoE routing policies: expert collaboration and specialization where some experts tend to activate broadly with others (collaborative), while others are more likely to activate only with a specific subset of experts (specialized). Our experiments reveal that most experts tend to be overly collaborative, leading to increased communication overhead from repeatedly sending tokens to different accelerators. To this end, we propose a novel collaboration-constrained routing (C2R) strategy to encourage more specialized expert groups, as well as to improve expert utilization, and present an efficient implementation of MoE that further leverages expert specialization. We achieve an average performance improvement of 0.51% and 0.33% on LLaMA-MoE and Qwen-MoE respectively across ten downstream NLP benchmarks, and reduce the all2all communication costs between GPUs, bringing an extra 20%-30% total running time savings on top of the existing SoTA, i.e. MegaBlocks.
Authors: Tianci Xue, Weijian Qi, Tianneng Shi, Chan Hee Song, Boyu Gou, Dawn Song, Huan Sun, Yu Su
Abstract: As digitalization and cloud technologies evolve, the web is becoming increasingly important in the modern society. Autonomous web agents based on large language models (LLMs) hold a great potential in work automation. It is therefore important to accurately measure and monitor the progression of their capabilities. In this work, we conduct a comprehensive and rigorous assessment of the current state of web agents. Our results depict a very different picture of the competency of current agents, suggesting over-optimism in previously reported results. This gap can be attributed to shortcomings in existing benchmarks. We introduce Online-Mind2Web, an online evaluation benchmark consisting of 300 diverse and realistic tasks spanning 136 websites. It enables us to evaluate web agents under a setting that approximates how real users use these agents. To facilitate more scalable evaluation and development, we also develop a novel LLM-as-a-Judge automatic evaluation method and show that it can achieve around 85% agreement with human judgment, substantially higher than existing methods. Finally, we present the first comprehensive comparative analysis of current web agents, highlighting both their strengths and limitations to inspire future research.
Authors: Ming Pang, Chunyuan Yuan, Xiaoyu He, Zheng Fang, Donghao Xie, Fanyi Qu, Xue Jiang, Changping Peng, Zhangang Lin, Zheng Luo, Jingping Shao
Abstract: Traditional sparse and dense retrieval methods struggle to leverage general world knowledge and often fail to capture the nuanced features of queries and products. With the advent of large language models (LLMs), industrial search systems have started to employ LLMs to generate identifiers for product retrieval. Commonly used identifiers include (1) static/semantic IDs and (2) product term sets. The first approach requires creating a product ID system from scratch, missing out on the world knowledge embedded within LLMs. While the second approach leverages this general knowledge, the significant difference in word distribution between queries and products means that product-based identifiers often do not align well with user search queries, leading to missed product recalls. Furthermore, when queries contain numerous attributes, these algorithms generate a large number of identifiers, making it difficult to assess their quality, which results in low overall recall efficiency. To address these challenges, this paper introduces a novel e-commerce retrieval paradigm: the Generative Retrieval and Alignment Model (GRAM). GRAM employs joint training on text information from both queries and products to generate shared text identifier codes, effectively bridging the gap between queries and products. This approach not only enhances the connection between queries and products but also improves inference efficiency. The model uses a co-alignment strategy to generate codes optimized for maximizing retrieval efficiency. Additionally, it introduces a query-product scoring mechanism to compare product values across different codes, further boosting retrieval efficiency. Extensive offline and online A/B testing demonstrates that GRAM significantly outperforms traditional models and the latest generative retrieval models, confirming its effectiveness and practicality.
Authors: Runlong Zhou, Yi Zhang
Abstract: Language models often struggle with cross-mode knowledge retrieval -- the ability to access knowledge learned in one format (mode) when queried in another. We demonstrate that models trained on multiple data sources (e.g., Wikipedia and TinyStories) exhibit significantly reduced accuracy when retrieving knowledge in a format different from its original training mode. This paper quantitatively investigates this phenomenon through a controlled study of random token sequence memorization across different modes. We first explore dataset rewriting as a solution, revealing that effective cross-mode retrieval requires prohibitively extensive rewriting efforts that follow a sigmoid-like relationship. As an alternative, we propose CASCADE, a novel pretraining algorithm that uses cascading datasets with varying sequence lengths to capture knowledge at different scales. Our experiments demonstrate that CASCADE outperforms dataset rewriting approaches, even when compressed into a single model with a unified loss function. This work provides both qualitative evidence of cross-mode retrieval limitations and a practical solution to enhance language models' ability to access knowledge independently of its presentational format.
Authors: Silvia Fernandez-Sabido, Laura Peniche-Sabido
Abstract: In this paper, we offer an overview of indigenous languages, identifying the causes of their devaluation and the need for legislation on language rights. We review the technologies used to revitalize these languages, finding that when they come from outside, they often have the opposite effect to what they seek; however, when developed from within communities, they become powerful instruments of expression. We propose that the inclusion of Indigenous knowledge in large language models (LLMs) will enrich the technological landscape, but must be done in a participatory environment that encourages the exchange of knowledge.
Authors: Ashkan Yousefpour, Taeheon Kim, Ryan S. Kwon, Seungbeen Lee, Wonje Jeung, Seungju Han, Alvin Wan, Harrison Ngan, Youngjae Yu, Jonghyun Choi
Abstract: Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.
Authors: Lena Schmidt, Oshin Sharma, Chris Marshall, Sonia Garcia Gonzalez Moral
Abstract: Introduction: Horizon scanning in healthcare assesses early signals of innovation, crucial for timely adoption. Current horizon scanning faces challenges in efficient information retrieval and analysis, especially from unstructured sources like news, presenting a need for innovative tools. Methodology: The study introduces SCANAR and AIDOC, open-source Python-based tools designed to improve horizon scanning. SCANAR automates the retrieval and processing of news articles, offering functionalities such as de-duplication and unsupervised relevancy ranking. AIDOC aids filtration by leveraging AI to reorder textual data based on relevancy, employing neural networks for semantic similarity, and subsequently prioritizing likely relevant entries for human review. Results: Twelve internal datasets from horizon scans and four external benchmarking datasets were used. SCANAR improved retrieval efficiency by automating processes previously dependent on manual labour. AIDOC displayed work-saving potential, achieving around 62% reduction in manual review efforts at 95% recall. Comparative analysis with benchmarking data showed AIDOC's performance was similar to existing systematic review automation tools, though performance varied depending on dataset characteristics. A smaller case-study on our news datasets shows the potential of ensembling large language models within the active-learning process for faster detection of relevant articles across news datasets. Conclusion: The validation indicates that SCANAR and AIDOC show potential to enhance horizon scanning efficiency by streamlining data retrieval and prioritisation. These tools may alleviate methodological limitations and allow broader, swifter horizon scans. Further studies are suggested to optimize these models and to design new workflows and validation processes that integrate large language models.
Authors: Maelyson R. F. Santos, Marcelo A. F. Gomes
Abstract: This article investigates scaling laws within language families using data from over six thousand languages and analyzing emergent patterns observed in Zipf-like classification graphs. Both macroscopic (based on number of languages by family) and microscopic (based on numbers of speakers by language on a family) aspects of these classifications are examined. Particularly noteworthy is the discovery of a distinct division among the fourteen largest contemporary language families, excluding Afro-Asiatic and Nilo-Saharan languages. These families are found to be distributed across three language family quadruplets, each characterized by significantly different exponents in the Zipf graphs. This finding sheds light on the underlying structure and organization of major language families, revealing intriguing insights into the nature of linguistic diversity and distribution.
Authors: Sean MacAvaney, Antonio Mallia, Nicola Tonellotto
Abstract: Multi-vector retrieval methods, exemplified by the ColBERT architecture, have shown substantial promise for retrieval by providing strong trade-offs in terms of retrieval latency and effectiveness. However, they come at a high cost in terms of storage since a (potentially compressed) vector needs to be stored for every token in the input collection. To overcome this issue, we propose encoding documents to a fixed number of vectors, which are no longer necessarily tied to the input tokens. Beyond reducing the storage costs, our approach has the advantage that document representations become of a fixed size on disk, allowing for better OS paging management. Through experiments using the MSMARCO passage corpus and BEIR with the ColBERT-v2 architecture, a representative multi-vector ranking model architecture, we find that passages can be effectively encoded into a fixed number of vectors while retaining most of the original effectiveness.
Authors: Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidecke, Amelia Glaese, Tejal Patwardhan
Abstract: We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. For objective evaluation, we develop rubrics that hierarchically decompose each replication task into smaller sub-tasks with clear grading criteria. In total, PaperBench contains 8,316 individually gradable tasks. Rubrics are co-developed with the author(s) of each ICML paper for accuracy and realism. To enable scalable evaluation, we also develop an LLM-based judge to automatically grade replication attempts against rubrics, and assess our judge's performance by creating a separate benchmark for judges. We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0\%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that models do not yet outperform the human baseline. We \href{https://github.com/openai/preparedness}{open-source our code} to facilitate future research in understanding the AI engineering capabilities of AI agents.
Authors: Aashiq Muhamed, Mona Diab, Virginia Smith
Abstract: Retrieval-Augmented Generation (RAG) models excel in knowledge-intensive tasks, especially under few-shot learning constraints. We introduce CoRAG, a framework extending RAG to collaborative settings, where clients jointly train a shared model using a collaborative passage store. To evaluate CoRAG, we introduce CRAB, a benchmark for collaborative homogeneous open-domain question answering. Our experiments demonstrate that CoRAG consistently outperforms both parametric collaborative learning methods and locally trained RAG models in low-resource scenarios. Further analysis reveals the critical importance of relevant passages within the shared store, the surprising benefits of incorporating irrelevant passages, and the potential for hard negatives to negatively impact performance. This introduces a novel consideration in collaborative RAG: the trade-off between leveraging a collectively enriched knowledge base and the potential risk of incorporating detrimental passages from other clients. Our findings underscore the viability of CoRAG, while also highlighting key design challenges and promising avenues for future research.
Authors: Haochen Wang, Yucheng Zhao, Tiancai Wang, Haoqiang Fan, Xiangyu Zhang, Zhaoxiang Zhang
Abstract: The rapid development of Large Multimodal Models (LMMs) for 2D images and videos has spurred efforts to adapt these models for interpreting 3D scenes. However, the absence of large-scale 3D vision-language datasets has posed a significant obstacle. To address this issue, typical approaches focus on injecting 3D awareness into 2D LMMs by designing 3D input-level scene representations. This work provides a new perspective. We introduce reconstructive visual instruction tuning with 3D-awareness (Ross3D), which integrates 3D-aware visual supervision into the training procedure. Specifically, it incorporates cross-view and global-view reconstruction. The former requires reconstructing masked views by aggregating overlapping information from other views. The latter aims to aggregate information from all available views to recover Bird's-Eye-View images, contributing to a comprehensive overview of the entire scene. Empirically, Ross3D achieves state-of-the-art performance across various 3D scene understanding benchmarks. More importantly, our semi-supervised experiments demonstrate significant potential in leveraging large amounts of unlabeled 3D vision-only data.
Authors: Yinggan Xu, Hana Kimlee, Yijia Xiao, Di Luo
Abstract: Large Language Models (LLMs) are playing an expanding role in physics research by enhancing reasoning, symbolic manipulation, and numerical computation. However, ensuring the reliability and interpretability of their outputs remains a significant challenge. In our framework, we conceptualize the collaboration between AI and human scientists as a dynamic interplay among three modules: the reasoning module, the interpretation module, and the AI-scientist interaction module. Recognizing that effective physics reasoning demands rigorous logical consistency, quantitative precision, and deep integration with established theoretical models, we introduce the interpretation module to improve the understanding of AI-generated outputs, which is not previously explored in the literature. This module comprises multiple specialized agents, including summarizers, model builders, UI builders, and testers, which collaboratively structure LLM outputs within a physically grounded framework, by constructing a more interpretable science model. A case study demonstrates that our approach enhances transparency, facilitates validation, and strengthens AI-augmented reasoning in scientific discovery.
Authors: Mothilal Asokan, Kebin Wu, Fatima Albreiki
Abstract: As a pioneering vision-language model, CLIP (Contrastive Language-Image Pre-training) has achieved significant success across various domains and a wide range of downstream vision-language tasks. However, the text encoders in popular CLIP models are limited to processing only 77 text tokens, which constrains their ability to effectively handle longer, detail-rich captions. Additionally, CLIP models often struggle to effectively capture detailed visual and textual information, which hampers their performance on tasks that require fine-grained analysis. To address these limitations, we present a novel approach, \textbf{FineLIP}, that extends the capabilities of CLIP. FineLIP enhances cross-modal text-image mapping by incorporating \textbf{Fine}-grained alignment with \textbf{L}onger text input within the CL\textbf{IP}-style framework. FineLIP first extends the positional embeddings to handle longer text, followed by the dynamic aggregation of local image and text tokens. The aggregated results are then used to enforce fine-grained token-to-token cross-modal alignment. We validate our model on datasets with long, detailed captions across two tasks: zero-shot cross-modal retrieval and text-to-image generation. Quantitative and qualitative experimental results demonstrate the effectiveness of FineLIP, outperforming existing state-of-the-art approaches. Furthermore, comprehensive ablation studies validate the benefits of key design elements within FineLIP.
Authors: Massimiliano Luca, Ciro Beneduce, Bruno Lepri, Jacopo Staiano
Abstract: With the wide and cross-domain adoption of Large Language Models, it becomes crucial to assess to which extent the statistical correlations in training data, which underlie their impressive performance, hide subtle and potentially troubling biases. Gender bias in LLMs has been widely investigated from the perspectives of works, hobbies, and emotions typically associated with a specific gender. In this study, we introduce a novel perspective. We investigate whether LLMs can predict an individual's gender based solely on online shopping histories and whether these predictions are influenced by gender biases and stereotypes. Using a dataset of historical online purchases from users in the United States, we evaluate the ability of six LLMs to classify gender and we then analyze their reasoning and products-gender co-occurrences. Results indicate that while models can infer gender with moderate accuracy, their decisions are often rooted in stereotypical associations between product categories and gender. Furthermore, explicit instructions to avoid bias reduce the certainty of model predictions, but do not eliminate stereotypical patterns. Our findings highlight the persistent nature of gender biases in LLMs and emphasize the need for robust bias-mitigation strategies.
Authors: Yue Guo, Tal August, Gondy Leroy, Trevor Cohen, Lucy Lu Wang
Abstract: While there has been significant development of models for Plain Language Summarization (PLS), evaluation remains a challenge. PLS lacks a dedicated assessment metric, and the suitability of text generation evaluation metrics is unclear due to the unique transformations involved (e.g., adding background explanations, removing jargon). To address these questions, our study introduces a granular meta-evaluation testbed, APPLS, designed to evaluate metrics for PLS. We identify four PLS criteria from previous work -- informativeness, simplification, coherence, and faithfulness -- and define a set of perturbations corresponding to these criteria that sensitive metrics should be able to detect. We apply these perturbations to extractive hypotheses for two PLS datasets to form our testbed. Using APPLS, we assess performance of 14 metrics, including automated scores, lexical features, and LLM prompt-based evaluations. Our analysis reveals that while some current metrics show sensitivity to specific criteria, no single method captures all four criteria simultaneously. We therefore recommend a suite of automated metrics be used to capture PLS quality along all relevant criteria. This work contributes the first meta-evaluation testbed for PLS and a comprehensive evaluation of existing metrics. APPLS and our evaluation code is available at https://github.com/LinguisticAnomalies/APPLS.
Authors: Christina Chance, Da Yin, Dakuo Wang, Kai-Wei Chang
Abstract: In this paper, we study whether language models are affected by learned gender stereotypes during the comprehension of stories. Specifically, we investigate how models respond to gender stereotype perturbations through counterfactual data augmentation. Focusing on Question Answering (QA) tasks in fairytales, we modify the FairytaleQA dataset by swapping gendered character information and introducing counterfactual gender stereotypes during training. This allows us to assess model robustness and examine whether learned biases influence story comprehension. Our results show that models exhibit slight performance drops when faced with gender perturbations in the test set, indicating sensitivity to learned stereotypes. However, when fine-tuned on counterfactual training data, models become more robust to anti-stereotypical narratives. Additionally, we conduct a case study demonstrating how incorporating counterfactual anti-stereotype examples can improve inclusivity in downstream applications.
Authors: Chenghao Zhu, Nuo Chen, Yufei Gao, Yunyi Zhang, Prayag Tiwari, Benyou Wang
Abstract: The rapid advancement of Large Language Models (LLMs) has led to the development of benchmarks that consider temporal dynamics, however, there remains a gap in understanding how well these models can generalize across temporal contexts due to the inherent dynamic nature of language and information. This paper introduces the concept of temporal generalization in LLMs, including bias in past and future generalizations. Then we introduce FreshBench, a new evaluation framework that employs fresh text and event prediction for assessing LLMs' temporal adaptability, ensuring the evaluation process free from data leakage and subjective bias. The experiment shows significant temporal biases and a decline in performance over time. Our findings reveal that powerful models, while initially superior, tend to decline more rapidly in future generalization. Additionally, powerful open-source models demonstrate better long-term adaptability compared to their closed-source counterparts. Our code is available at https://github.com/FreedomIntelligence/FreshBench.
Authors: Adithya Bhaskar, Alexander Wettig, Dan Friedman, Danqi Chen
Abstract: The path to interpreting a language model often proceeds via analysis of circuits -- sparse computational subgraphs of the model that capture specific aspects of its behavior. Recent work has automated the task of discovering circuits. Yet, these methods have practical limitations, as they rely either on inefficient search algorithms or inaccurate approximations. In this paper, we frame automated circuit discovery as an optimization problem and propose *Edge Pruning* as an effective and scalable solution. Edge Pruning leverages gradient-based pruning techniques, but instead of removing neurons or components, it prunes the \emph{edges} between components. Our method finds circuits in GPT-2 that use less than half the number of edges compared to circuits found by previous methods while being equally faithful to the full model predictions on standard circuit-finding tasks. Edge Pruning is efficient even with as many as 100K examples, outperforming previous methods in speed and producing substantially better circuits. It also perfectly recovers the ground-truth circuits in two models compiled with Tracr. Thanks to its efficiency, we scale Edge Pruning to CodeLlama-13B, a model over 100x the scale that prior methods operate on. We use this setting for a case study comparing the mechanisms behind instruction prompting and in-context learning. We find two circuits with more than 99.96% sparsity that match the performance of the full model and reveal that the mechanisms in the two settings overlap substantially. Our case study shows that Edge Pruning is a practical and scalable tool for interpretability and sheds light on behaviors that only emerge in large models.
Authors: Joy Crosbie, Ekaterina Shutova
Abstract: Large language models (LLMs) have shown a remarkable ability to learn and perform complex tasks through in-context learning (ICL). However, a comprehensive understanding of its internal mechanisms is still lacking. This paper explores the role of induction heads in a few-shot ICL setting. We analyse two state-of-the-art models, Llama-3-8B and InternLM2-20B on abstract pattern recognition and NLP tasks. Our results show that even a minimal ablation of induction heads leads to ICL performance decreases of up to ~32% for abstract pattern recognition tasks, bringing the performance close to random. For NLP tasks, this ablation substantially decreases the model's ability to benefit from examples, bringing few-shot ICL performance close to that of zero-shot prompts. We further use attention knockout to disable specific induction patterns, and present fine-grained evidence for the role that the induction mechanism plays in ICL.
Authors: Berk Atil, Sarp Aykent, Alexa Chittams, Lisheng Fu, Rebecca J. Passonneau, Evan Radcliffe, Guru Rajan Rajagopal, Adam Sloan, Tomasz Tudrej, Ferhan Ture, Zhe Wu, Lixinyu Xu, Breck Baldwin
Abstract: LLM (large language model) practitioners commonly notice that outputs can vary for the same inputs under settings expected to be deterministic. Yet the questions of how pervasive this is, and with what impact on results, have not to our knowledge been systematically investigated. We investigate non-determinism in five LLMs configured to be deterministic when applied to eight common tasks in across 10 runs, in both zero-shot and few-shot settings. We see accuracy variations up to 15% across naturally occurring runs with a gap of best possible performance to worst possible performance up to 70%. In fact, none of the LLMs consistently delivers repeatable accuracy across all tasks, much less identical output strings. Sharing preliminary results with insiders has revealed that non-determinism perhaps essential to the efficient use of compute resources via co-mingled data in input buffers so this issue is not going away anytime soon. To better quantify our observations, we introduce metrics focused on quantifying determinism, TARr@N for the total agreement rate at N runs over raw output, and TARa@N for total agreement rate of parsed-out answers. Our code and data are publicly available at https://github.com/breckbaldwin/llm-stability.
Authors: Ranajoy Sadhukhan, Jian Chen, Zhuoming Chen, Vashisth Tiwari, Ruihang Lai, Jinyuan Shi, Ian En-Hsu Yen, Avner May, Tianqi Chen, Beidi Chen
Abstract: Large Language Models (LLMs) have become more prevalent in long-context applications such as interactive chatbots, document analysis, and agent workflows, but it is challenging to serve long-context requests with low latency and high throughput. Speculative decoding (SD) is a widely used technique to reduce latency losslessly, but the conventional wisdom suggests that its efficacy is limited to small batch sizes. In MagicDec, we show that surprisingly SD can achieve speedup even for a high throughput inference regime for moderate to long sequences. More interestingly, an intelligent drafting strategy can achieve better speedup with increasing batch size based on our rigorous analysis. MagicDec first identifies the bottleneck shifts with increasing batch size and sequence length, and uses these insights to deploy SD more effectively for high throughput inference. We leverage draft model with sparse KV cache to address the KV bottleneck, which scales with both sequence length and batch size. Additionally, we propose a theoretical model to select the optimal drafting strategy for maximum speedup. Our work highlights the broad applicability of speculative decoding in long-context serving, as it can enhance throughput and reduce latency without compromising accuracy. For moderate to long sequences, we demonstrate up to 2.51x speedup for Llama3.1-8B when serving batch sizes ranging from 32 to 256 on various types of hardware and tasks.
Authors: Jimin Huang, Mengxi Xiao, Dong Li, Zihao Jiang, Yuzhe Yang, Yifei Zhang, Lingfei Qian, Yan Wang, Xueqing Peng, Yang Ren, Ruoyu Xiang, Zhengyu Chen, Xiao Zhang, Yueru He, Weiguang Han, Shunian Chen, Lihang Shen, Daniel Kim, Yangyang Yu, Yupeng Cao, Zhiyang Deng, Haohang Li, Duanyu Feng, Yongfu Dai, VijayaSai Somasundaram, Peng Lu, Guojun Xiong, Zhiwei Liu, Zheheng Luo, Zhiyuan Yao, Ruey-Ling Weng, Meikang Qiu, Kaleb E Smith, Honghai Yu, Yanzhao Lai, Min Peng, Jian-Yun Nie, Jordan W. Suchow, Xiao-Yang Liu, Benyou Wang, Alejandro Lopez-Lira, Qianqian Xie, Sophia Ananiadou, Junichi Tsujii
Abstract: Financial LLMs hold promise for advancing financial tasks and domain-specific applications. However, they are limited by scarce corpora, weak multimodal capabilities, and narrow evaluations, making them less suited for real-world application. To address this, we introduce \textit{Open-FinLLMs}, the first open-source multimodal financial LLMs designed to handle diverse tasks across text, tabular, time-series, and chart data, excelling in zero-shot, few-shot, and fine-tuning settings. The suite includes FinLLaMA, pre-trained on a comprehensive 52-billion-token corpus; FinLLaMA-Instruct, fine-tuned with 573K financial instructions; and FinLLaVA, enhanced with 1.43M multimodal tuning pairs for strong cross-modal reasoning. We comprehensively evaluate Open-FinLLMs across 14 financial tasks, 30 datasets, and 4 multimodal tasks in zero-shot, few-shot, and supervised fine-tuning settings, introducing two new multimodal evaluation datasets. Our results show that Open-FinLLMs outperforms afvanced financial and general LLMs such as GPT-4, across financial NLP, decision-making, and multi-modal tasks, highlighting their potential to tackle real-world challenges. To foster innovation and collaboration across academia and industry, we release all codes (https://anonymous.4open.science/r/PIXIU2-0D70/B1D7/LICENSE) and models under OSI-approved licenses.
URLs: https://anonymous.4open.science/r/PIXIU2-0D70/B1D7/LICENSE)
Authors: Lingwei Meng, Shujie Hu, Jiawen Kang, Zhaoqing Li, Yuejiao Wang, Wenxuan Wu, Xixin Wu, Xunying Liu, Helen Meng
Abstract: Recent advancements in large language models (LLMs) have revolutionized various domains, bringing significant progress and new opportunities. Despite progress in speech-related tasks, LLMs have not been sufficiently explored in multi-talker scenarios. In this work, we present a pioneering effort to investigate the capability of LLMs in transcribing speech in multi-talker environments, following versatile instructions related to multi-talker automatic speech recognition (ASR), target talker ASR, and ASR based on specific talker attributes such as sex, occurrence order, language, and keyword spoken. Our approach utilizes WavLM and Whisper encoder to extract multi-faceted speech representations that are sensitive to speaker characteristics and semantic context. These representations are then fed into an LLM fine-tuned using LoRA, enabling the capabilities for speech comprehension and transcription. Comprehensive experiments reveal the promising performance of our proposed system, MT-LLM, in cocktail party scenarios, highlighting the potential of LLM to handle speech-related tasks based on user instructions in such complex settings. The code, model, and samples are available at https://github.com/cuhealthybrains/MT-LLM.
Authors: Aniket Deroy, Subhankar Maity
Abstract: Code generation by Llama 3.1 models, such as Meta's Llama 3.1 405B, represents a significant advancement in the field of artificial intelligence, particularly in natural language processing and programming automation. This paper explores the capabilities and applications of Llama-driven code generation, highlighting its ability to translate natural language prompts into executable code across multiple programming languages. Key features include contextual awareness, multi-language support, and enhanced debugging and optimization functionalities. By examining these aspects, we illustrate how Llama can serve as a versatile tool for developers of all skill levels, improving productivity and efficiency in software development. The potential implications for education, industry, and the future of coding practices are also discussed, underscoring the transformative impact of AI in programming. Experimentation shows that while Llama 3.1 405B performs well with simple algorithmic and data structure based problems, it still struggles with problems on Quantum Computing, Bioinformatics, and Artificial Intelligence.
Authors: Peiqi Wang, Barbara D. Lam, Yingcheng Liu, Ameneh Asgari-Targhi, Rameswar Panda, William M. Wells, Tina Kapur, Polina Golland
Abstract: We present a novel approach to calibrating linguistic expressions of certainty, e.g., "Maybe" and "Likely". Unlike prior work that assigns a single score to each certainty phrase, we model uncertainty as distributions over the simplex to capture their semantics more accurately. To accommodate this new representation of certainty, we generalize existing measures of miscalibration and introduce a novel post-hoc calibration method. Leveraging these tools, we analyze the calibration of both humans (e.g., radiologists) and computational models (e.g., language models) and provide interpretable suggestions to improve their calibration.
Authors: Xiangru Zhu, Penglei Sun, Yaoxian Song, Yanghua Xiao, Zhixu Li, Chengyu Wang, Jun Huang, Bei Yang, Xiaoxiao Xu
Abstract: Accurate interpretation and visualization of human instructions are crucial for text-to-image (T2I) synthesis. However, current models struggle to capture semantic variations from word order changes, and existing evaluations, relying on indirect metrics like text-image similarity, fail to reliably assess these challenges. This often obscures poor performance on complex or uncommon linguistic patterns by the focus on frequent word combinations. To address these deficiencies, we propose a novel metric called SemVarEffect and a benchmark named SemVarBench, designed to evaluate the causality between semantic variations in inputs and outputs in T2I synthesis. Semantic variations are achieved through two types of linguistic permutations, while avoiding easily predictable literal variations. Experiments reveal that the CogView-3-Plus and Ideogram 2 performed the best, achieving a score of 0.2/1. Semantic variations in object relations are less understood than attributes, scoring 0.07/1 compared to 0.17-0.19/1. We found that cross-modal alignment in UNet or Transformers plays a crucial role in handling semantic variations, a factor previously overlooked by a focus on textual encoders. Our work establishes an effective evaluation framework that advances the T2I synthesis community's exploration of human instruction understanding. Our benchmark and code are available at https://github.com/zhuxiangru/SemVarBench .
Authors: Julie Kallini, Shikhar Murty, Christopher D. Manning, Christopher Potts, R\'obert Csord\'as
Abstract: Models that rely on subword tokenization have significant drawbacks, such as sensitivity to character-level noise like spelling errors and inconsistent compression rates across different languages and scripts. While character- or byte-level models like ByT5 attempt to address these concerns, they have not gained widespread adoption -- processing raw byte streams without tokenization results in significantly longer sequence lengths, making training and inference inefficient. This work introduces MrT5 (MergeT5), a more efficient variant of ByT5 that integrates a token deletion mechanism in its encoder to dynamically shorten the input sequence length. After processing through a fixed number of encoder layers, a learned delete gate determines which tokens are to be removed and which are to be retained for subsequent layers. MrT5 effectively "merges" critical information from deleted tokens into a more compact sequence, leveraging contextual information from the remaining tokens. In continued pre-training experiments, we find that MrT5 can achieve significant gains in inference runtime with minimal effect on performance, as measured by bits-per-byte. Additionally, with multilingual training, MrT5 adapts to the orthographic characteristics of each language, learning language-specific compression rates. Furthermore, MrT5 shows comparable accuracy to ByT5 on downstream evaluations such as XNLI, TyDi QA, and character-level tasks while reducing sequence lengths by up to 75%. Our approach presents a solution to the practical limitations of existing byte-level models.
Authors: Raphael Merx, Ad\'erito Jos\'e Guterres Correia, Hanna Suominen, Ekaterina Vylomova
Abstract: Low-resource machine translation (MT) presents a diversity of community needs and application challenges that remain poorly understood. To complement surveys and focus groups, which tend to rely on small samples of respondents, we propose an observational study on actual usage patterns of tetun$.$org, a specialized MT service for the Tetun language, which is the lingua franca in Timor-Leste. Our analysis of 100,000 translation requests reveals patterns that challenge assumptions based on existing corpora. We find that users, many of them students on mobile devices, typically translate text from a high-resource language into Tetun across diverse domains including science, healthcare, and daily life. This contrasts sharply with available Tetun corpora, which are dominated by news articles covering government and social issues. Our results suggest that MT systems for institutionalized minority languages like Tetun should prioritize accuracy on domains relevant to educational contexts, in the high-resource to low-resource direction. More broadly, this study demonstrates how observational analysis can inform low-resource language technology development, by grounding research in practical community needs.
Authors: Luhang Sun, Varsha Pendyala, Yun-Shiuan Chuang, Shanglin Yang, Jonathan Feldman, Andrew Zhao, Munmun De Choudhury, Sijia Yang, Dhavan Shah
Abstract: This paper leverages large-language models (LLMs) to experimentally determine optimal strategies for scaling up social media content annotation for stance detection on HPV vaccine-related tweets. We examine both conventional fine-tuning and emergent in-context learning methods, systematically varying strategies of prompt engineering across widely used LLMs and their variants (e.g., GPT4, Mistral, and Llama3, etc.). Specifically, we varied prompt template design, shot sampling methods, and shot quantity to detect stance on HPV vaccination. Our findings reveal that 1) in general, in-context learning outperforms fine-tuning in stance detection for HPV vaccine social media content; 2) increasing shot quantity does not necessarily enhance performance across models; and 3) different LLMs and their variants present differing sensitivity to in-context learning conditions. We uncovered that the optimal in-context learning configuration for stance detection on HPV vaccine tweets involves six stratified shots paired with detailed contextual prompts. This study highlights the potential and provides an applicable approach for applying LLMs to research on social media stance and skepticism detection.
Authors: Jiangnan Fang, Cheng-Tse Liu, Jieun Kim, Yash Bhedaru, Ethan Liu, Nikhil Singh, Nedim Lipka, Puneet Mathur, Nesreen K. Ahmed, Franck Dernoncourt, Ryan A. Rossi, Hanieh Deilamsalehy
Abstract: In this work, we propose a Multi-LLM summarization framework, and investigate two different multi-LLM strategies including centralized and decentralized. Our multi-LLM summarization framework has two fundamentally important steps at each round of conversation: generation and evaluation. These steps are different depending on whether our multi-LLM decentralized summarization is used or centralized. In both our multi-LLM decentralized and centralized strategies, we have k different LLMs that generate diverse summaries of the text. However, during evaluation, our multi-LLM centralized summarization approach leverages a single LLM to evaluate the summaries and select the best one whereas k LLMs are used for decentralized multi-LLM summarization. Overall, we find that our multi-LLM summarization approaches significantly outperform the baselines that leverage only a single LLM by up to 3x. These results indicate the effectiveness of multi-LLM approaches for summarization.
Authors: Casey Kennington, Pierre Lison, David Schlangen
Abstract: Efforts towards endowing robots with the ability to speak have benefited from recent advancements in natural language processing, in particular large language models. However, current language models are not fully incremental, as their processing is inherently monotonic and thus lack the ability to revise their interpretations or output in light of newer observations. This monotonicity has important implications for the development of dialogue systems for human--robot interaction. In this paper, we review the literature on interactive systems that operate incrementally (i.e., at the word level or below it). We motivate the need for incremental systems, survey incremental modeling of important aspects of dialogue like speech recognition and language generation. Primary focus is on the part of the system that makes decisions, known as the dialogue manager. We find that there is very little research on incremental dialogue management, offer some requirements for practical incremental dialogue management, and the implications of incremental dialogue for embodied, robotic platforms in the age of large language models.
Authors: Yongkang Du, Jen-tse Huang, Jieyu Zhao, Lu Lin
Abstract: Large language models (LLMs) have been widely deployed in coding tasks, drawing increasing attention to the evaluation of the quality and safety of LLMs' outputs. However, research on bias in code generation remains limited. Existing studies typically identify bias by applying malicious prompts or reusing tasks and dataset originally designed for discriminative models. Given that prior datasets are not fully optimized for code-related tasks, there is a pressing need for benchmarks specifically designed for evaluating code models. In this study, we introduce FairCoder, a novel benchmark for evaluating social bias in code generation. FairCoder explores the bias issue following the pipeline in software development, from function implementation to unit test, with diverse real-world scenarios. Additionally, three metrics are designed to assess fairness performance on this benchmark. We conduct experiments on widely used LLMs and provide a comprehensive analysis of the results. The findings reveal that all tested LLMs exhibit social bias.
Authors: Fabio Barth, Georg Rehm
Abstract: The breakthrough of generative large language models (LLMs) that can solve different tasks through chat interaction has led to a significant increase in the use of general benchmarks to assess the quality or performance of these models beyond individual applications. There is also a need for better methods to evaluate and also to compare models due to the ever increasing number of new models published. However, most of the established benchmarks revolve around the English language. This paper analyses the benefits and limitations of current evaluation datasets, focusing on multilingual European benchmarks. We analyse seven multilingual benchmarks and identify four major challenges. Furthermore, we discuss potential solutions to enhance translation quality and mitigate cultural biases, including human-in-the-loop verification and iterative translation ranking. Our analysis highlights the need for culturally aware and rigorously validated benchmarks to assess the reasoning and question-answering capabilities of multilingual LLMs accurately.
Authors: Jianghao Chen, Zhenlin Wei, Zhenjiang Ren, Ziyong Li, Jiajun Zhang
Abstract: Recent progress in o1-like models has significantly enhanced the reasoning abilities of Large Language Models (LLMs), empowering them to tackle increasingly complex tasks through reflection capabilities, such as making assumptions, backtracking, and self-refinement. However, effectively evaluating such reflection capabilities remains challenging due to the lack of appropriate benchmarks. To bridge this gap, we introduce LR$^2$Bench, a novel benchmark designed to evaluate the Long-chain Reflective Reasoning capabilities of LLMs. LR$^2$Bench comprises 850 samples across six Constraint Satisfaction Problems (CSPs) where reflective reasoning is crucial for deriving solutions that meet all given constraints. Each type of task focuses on distinct constraint patterns, such as knowledge-based, logical, and spatial constraints, providing a comprehensive evaluation of diverse problem-solving scenarios. We conduct extensive evaluation on both conventional models and o1-like models. Our experimental results reveal that even the most advanced reasoning-specific models, such as DeepSeek-R1 and OpenAI o1-preview, struggle with tasks in LR$^2$Bench, achieving an average Exact Match score of only 20.0% and 23.6%, respectively. These findings underscore the significant room for improvement in the reflective reasoning capabilities of current LLMs. The leaderboard of our benchmark is available at https://huggingface.co/spaces/UltraRonin/LR2Bench
Authors: Junsol Kim, James Evans, Aaron Schein
Abstract: Large language models (LLMs) have demonstrated the ability to generate text that realistically reflects a range of different subjective human perspectives. This paper studies how LLMs are seemingly able to reflect more liberal versus more conservative viewpoints among other political perspectives in American politics. We show that LLMs possess linear representations of political perspectives within activation space, wherein more similar perspectives are represented closer together. To do so, we probe the attention heads across the layers of three open transformer-based LLMs (Llama-2-7b-chat, Mistral-7b-instruct, Vicuna-7b). We first prompt models to generate text from the perspectives of different U.S. lawmakers. We then identify sets of attention heads whose activations linearly predict those lawmakers' DW-NOMINATE scores, a widely-used and validated measure of political ideology. We find that highly predictive heads are primarily located in the middle layers, often speculated to encode high-level concepts and tasks. Using probes only trained to predict lawmakers' ideology, we then show that the same probes can predict measures of news outlets' slant from the activations of models prompted to simulate text from those news outlets. These linear probes allow us to visualize, interpret, and monitor ideological stances implicitly adopted by an LLM as it generates open-ended responses. Finally, we demonstrate that by applying linear interventions to these attention heads, we can steer the model outputs toward a more liberal or conservative stance. Overall, our research suggests that LLMs possess a high-level linear representation of American political ideology and that by leveraging recent advances in mechanistic interpretability, we can identify, monitor, and steer the subjective perspective underlying generated text.
Authors: Ben Hutchinson, Celeste Rodr\'iguez Louro, Glenys Collard, Ned Cooper
Abstract: In Australia, post-contact language varieties, including creoles and local varieties of international languages, emerged as a result of forced contact between Indigenous communities and English speakers. These contact varieties are widely used, yet are poorly supported by language technologies. This gap presents barriers to participation in civil and economic society for Indigenous communities using these varieties, and reproduces minoritisation of contemporary Indigenous sociolinguistic identities. This paper concerns three questions regarding this context. First, can speech technologies support speakers of Australian Aboriginal English, a local indigenised variety of English? Second, what risks are inherent in such a project? Third, what technology development practices are appropriate for this context, and how can researchers integrate meaningful community participation in order to mitigate risks? We argue that opportunities do exist -- as well as risks -- and demonstrate this through a case study exploring design practices in a real-world project aiming to improve speech technologies for Australian Aboriginal English. We discuss how we integrated culturally appropriate and participatory processes throughout the project. We call for increased support for languages used by Indigenous communities, including contact varieties, which provide practical economic and socio-cultural benefits, provided that participatory and culturally safe practices are enacted.
Authors: Yijiong Yu
Abstract: Recent advances in reasoning models have demonstrated significant improvements in accuracy, particularly for complex tasks such as mathematical reasoning, by employing detailed and comprehensive reasoning processes. However, generating these lengthy reasoning sequences is computationally expensive and time-consuming. To address this inefficiency, we leverage the inherent parallelizability of certain tasks to accelerate the reasoning process. Specifically, when multiple parallel reasoning branches exist, we decode multiple tokens per step using a specialized attention mask, processing them within a single sequence, avoiding additional memory usage. Experimental results show that our method achieves over 100% speedup in decoding time while maintaining the answer quality.
Authors: Rohitash Chandra, Aryan Chaudhari, Yeshwanth Rayavarapu
Abstract: Large Language models (LLMs) have been prominent for language translation, including low-resource languages. There has been limited study about the assessment of the quality of translations generated by LLMs, including Gemini, GPT and Google Translate. In this study, we address this limitation by using semantic and sentiment analysis of selected LLMs for Indian languages, including Sanskrit, Telugu and Hindi. We select prominent texts that have been well translated by experts and use LLMs to generate their translations to English, and then we provide a comparison with selected expert (human) translations. Our findings suggest that while LLMs have made significant progress in translation accuracy, challenges remain in preserving sentiment and semantic integrity, especially in figurative and philosophical contexts. The sentiment analysis revealed that GPT-4o and GPT-3.5 are better at preserving the sentiments for the Bhagavad Gita (Sanskrit-English) translations when compared to Google Translate. We observed a similar trend for the case of Tamas (Hindi-English) and Maha P (Telugu-English) translations. GPT-4o performs similarly to GPT-3.5 in the translation in terms of sentiments for the three languages. We found that LLMs are generally better at translation for capturing sentiments when compared to Google Translate.
Authors: Alejandro Lozano, Min Woo Sun, James Burgess, Jeffrey J. Nirschl, Christopher Polzak, Yuhui Zhang, Liangyu Chen, Jeffrey Gu, Ivan Lopez, Josiah Aklilu, Anita Rau, Austin Wolfgang Katzer, Collin Chiu, Orr Zohar, Xiaohan Wang, Alfred Seunghoon Song, Chiang Chia-Chun, Robert Tibshirani, Serena Yeung-Levy
Abstract: Despite the excitement behind biomedical artificial intelligence (AI), access to high-quality, diverse, and large-scale data - the foundation for modern AI systems - is still a bottleneck to unlocking its full potential. To address this gap, we introduce Biomedica, an open-source dataset derived from the PubMed Central Open Access subset, containing over 6 million scientific articles and 24 million image-text pairs, along with 27 metadata fields (including expert human annotations). To overcome the challenges of accessing our large-scale dataset, we provide scalable streaming and search APIs through a web server, facilitating seamless integration with AI systems. We demonstrate the utility of the Biomedica dataset by building embedding models, chat-style models, and retrieval-augmented chat agents. Notably, all our AI models surpass previous open systems in their respective categories, underscoring the critical role of diverse, high-quality, and large-scale biomedical data.
Authors: Tongke Ni, Yang Fan, Junru Zhou, Xiangping Wu, Qingcai Chen
Abstract: Text semantic segmentation involves partitioning a document into multiple paragraphs with continuous semantics based on the subject matter, contextual information, and document structure. Traditional approaches have typically relied on preprocessing documents into segments to address input length constraints, resulting in the loss of critical semantic information across segments. To address this, we present CrossFormer, a transformer-based model featuring a novel cross-segment fusion module that dynamically models latent semantic dependencies across document segments, substantially elevating segmentation accuracy. Additionally, CrossFormer can replace rule-based chunk methods within the Retrieval-Augmented Generation (RAG) system, producing more semantically coherent chunks that enhance its efficacy. Comprehensive evaluations confirm CrossFormer's state-of-the-art performance on public text semantic segmentation datasets, alongside considerable gains on RAG benchmarks.
Authors: Zhiming Ma, Peidong Wang, Minhua Huang, Jingpeng Wang, Kai Wu, Xiangzhao Lv, Yachun Pang, Yin Yang, Wenjie Tang, Yuchen Kang
Abstract: The detection of telecom fraud faces significant challenges due to the lack of high-quality multimodal training data that integrates audio signals with reasoning-oriented textual analysis. To address this gap, we present TeleAntiFraud-28k, the first open-source audio-text slow-thinking dataset specifically designed for automated telecom fraud analysis. Our dataset is constructed through three strategies: (1) Privacy-preserved text-truth sample generation using automatically speech recognition (ASR)-transcribed call recordings (with anonymized original audio), ensuring real-world consistency through text-to-speech (TTS) model regeneration; (2) Semantic enhancement via large language model (LLM)-based self-instruction sampling on authentic ASR outputs to expand scenario coverage; (3) Multi-agent adversarial synthesis that simulates emerging fraud tactics through predefined communication scenarios and fraud typologies. The generated dataset contains 28,511 rigorously processed speech-text pairs, complete with detailed annotations for fraud reasoning. The dataset is divided into three tasks: scenario classification, fraud detection, fraud type classification. Furthermore, we construct TeleAntiFraud-Bench, a standardized evaluation benchmark comprising proportionally sampled instances from the dataset, to facilitate systematic testing of model performance on telecom fraud detection tasks. We also contribute a production-optimized supervised fine-tuning (SFT) model trained on hybrid real/synthetic data, while open-sourcing the data processing framework to enable community-driven dataset expansion. This work establishes a foundational framework for multimodal anti-fraud research while addressing critical challenges in data privacy and scenario diversity. The project will be released at https://github.com/JimmyMa99/TeleAntiFraud.
Authors: Weizhi Wang, Yu Tian, Linjie Yang, Heng Wang, Xifeng Yan
Abstract: The reproduction of state-of-the-art multimodal LLM pre-training faces barriers at every stage of the pipeline, including high-quality data filtering, multimodal data mixture strategies, sequence packing techniques, and training frameworks. We introduce Open-Qwen2VL, a fully open-source 2B-parameter Multimodal Large Language Model pre-trained efficiently on 29M image-text pairs using only 220 A100-40G GPU hours. Our approach employs low-to-high dynamic image resolution and multimodal sequence packing to significantly enhance pre-training efficiency. The training dataset was carefully curated using both MLLM-based filtering techniques (e.g., MLM-Filter) and conventional CLIP-based filtering methods, substantially improving data quality and training efficiency. The Open-Qwen2VL pre-training is conducted on academic level 8xA100-40G GPUs at UCSB on 5B packed multimodal tokens, which is 0.36% of 1.4T multimodal pre-training tokens of Qwen2-VL. The final instruction-tuned Open-Qwen2VL outperforms partially-open state-of-the-art MLLM Qwen2-VL-2B on various multimodal benchmarks of MMBench, SEEDBench, MMstar, and MathVista, indicating the remarkable training efficiency of Open-Qwen2VL. We open-source all aspects of our work, including compute-efficient and data-efficient training details, data filtering methods, sequence packing scripts, pre-training data in WebDataset format, FSDP-based training codebase, and both base and instruction-tuned model checkpoints. We redefine "fully open" for multimodal LLMs as the complete release of: 1) the training codebase, 2) detailed data filtering techniques, and 3) all pre-training and supervised fine-tuning data used to develop the model.
Authors: Jianyou Wang, Kaicheng Wang, Xiaoyue Wang, Weili Cao, Ramamohan Paturi, Leon Bergen
Abstract: Effective information retrieval (IR) in settings with limited training data, particularly for complex queries, remains a challenging task. This paper introduces IR2, Information Regularization for Information Retrieval, a technique for reducing overfitting during synthetic data generation. This approach, representing a novel application of regularization techniques in synthetic data creation for IR, is tested on three recent IR tasks characterized by complex queries: DORIS-MAE, ArguAna, and WhatsThatBook. Experimental results indicate that our regularization techniques not only outperform previous synthetic query generation methods on the tasks considered but also reduce cost by up to 50%. Furthermore, this paper categorizes and explores three regularization methods at different stages of the query synthesis pipeline-input, prompt, and output-each offering varying degrees of performance improvement compared to models where no regularization is applied. This provides a systematic approach for optimizing synthetic data generation in data-limited, complex-query IR scenarios. All code, prompts and synthetic data are available at https://github.com/Info-Regularization/Information-Regularization.
URLs: https://github.com/Info-Regularization/Information-Regularization.
Authors: Khai Le-Duc, David Thulke, Hung-Phong Tran, Long Vo-Dang, Khai-Nguyen Nguyen, Truong-Son Hy, Ralf Schl\"uter
Abstract: Spoken Named Entity Recognition (NER) aims to extract named entities from speech and categorise them into types like person, location, organization, etc. In this work, we present VietMed-NER - the first spoken NER dataset in the medical domain. To our knowledge, our Vietnamese real-world dataset is the largest spoken NER dataset in the world regarding the number of entity types, featuring 18 distinct types. Furthermore, we present baseline results using various state-of-the-art pre-trained models: encoder-only and sequence-to-sequence; and conduct quantitative and qualitative error analysis. We found that pre-trained multilingual models generally outperform monolingual models on reference text and ASR output and encoders outperform sequence-to-sequence models in NER tasks. By translating the transcripts, the dataset can also be utilised for text NER in the medical domain in other languages than Vietnamese. All code, data and models are publicly available: https://github.com/leduckhai/MultiMed/tree/master/VietMed-NER.
URLs: https://github.com/leduckhai/MultiMed/tree/master/VietMed-NER.
Authors: Saber Malekmohammadi, Golnoosh Farnadi
Abstract: A significant approach in natural language processing involves large-scale pre-training of models on general domain data followed by their adaptation to specific tasks or domains. As models grow in size, full fine-tuning all of their parameters becomes increasingly impractical. To address this, some methods for low-rank task adaptation of language models have been proposed, e.g., LoRA and FLoRA. These methods keep the pre-trained model weights fixed and incorporate trainable low-rank decomposition matrices into some layers of the transformer architecture, called adapters. This approach significantly reduces the number of trainable parameters required for downstream tasks compared to full fine-tuning all parameters. In this work, we look at low-rank adaptation from the lens of data privacy. We show theoretically that the low-rank adaptation used in LoRA and FLoRA leads to the injection of some random noise into the batch gradients w.r.t the adapter parameters. We quantify the variance of the injected noise and show that the smaller the adaptation rank, the larger the noise variance. By establishing a Berry-Esseen type bound on the total variation distance between distribution of the injected noise and a Gaussian distribution with the same variance, we show that the dynamics of low-rank adaptation is close to that of differentially private fine-tuning of the adapters. Finally, using Johnson-Lindenstrauss lemma, we show that when augmented with gradient scaling, low-rank adaptation is very close to performing DPSGD algorithm with a fixed noise scale to fine-tune the adapters. Suggested by our theoretical findings and approved by our experimental results, we show that low-rank adaptation, besides mitigating the space and computational complexities, implicitly provides a privacy protection w.r.t the fine-tuning data, without inducing the high space complexity of DPSGD.
Authors: Yihong Dong, Ge Li, Yongding Tao, Xue Jiang, Kechi Zhang, Jia Li, Jinliang Deng, Jing Su, Jun Zhang, Jingjing Xu
Abstract: Despite the remarkable successes of general-purpose neural networks, such as MLPs and Transformers, we find that they exhibit notable shortcomings in modeling and reasoning about periodic phenomena, achieving only marginal performance within the training domain and failing to generalize effectively to out-of-domain (OOD) scenarios. Periodicity is ubiquitous throughout nature and science. Therefore, neural networks should be equipped with the essential ability to model and handle periodicity. In this work, we propose FAN, a novel general-purpose neural network that offers broad applicability similar to MLP while effectively addressing periodicity modeling challenges. Periodicity is naturally integrated into FAN's structure and computational processes by introducing the Fourier Principle. Unlike existing Fourier-based networks, which possess particular periodicity modeling abilities but are typically designed for specific tasks, our approach maintains the general-purpose modeling capability. Therefore, FAN can seamlessly replace MLP in various model architectures with fewer parameters and FLOPs. Through extensive experiments, we demonstrate the superiority of FAN in periodicity modeling tasks and the effectiveness and generalizability of FAN across a range of real-world tasks, e.g., symbolic formula representation, time series forecasting, language modeling, and image recognition.
Authors: Alex Iacob, Lorenzo Sani, Meghdad Kurmanji, William F. Shen, Xinchi Qiu, Dongqi Cai, Yan Gao, Nicholas D. Lane
Abstract: Language Model pre-training uses broad data mixtures to enhance performance across domains and languages. However, training on such heterogeneous text corpora requires extensive and expensive efforts. Since these data sources vary significantly in lexical, syntactic, and semantic aspects, they cause negative interference or the ``curse of multilinguality''. To address these challenges we propose a communication-efficient pre-training framework, DEPT. Our method decouples embeddings from the transformer body while simultaneously training the latter on multiple data sources without requiring a shared vocabulary. DEPT can: (1) train robustly and effectively under significant data heterogeneity, (2) minimize token embedding parameters to only what the data source vocabulary requires, while cutting communication costs in direct proportion to both the communication frequency and the reduction in parameters, (3) enhance transformer body plasticity and generalization, improving both average perplexity (up to 20%) and downstream task performance, and (4) enable training with custom optimized vocabularies per data source. We demonstrate DEPT's potential via the first vocabulary-agnostic federated pre-training of billion-scale models, reducing communication costs by orders of magnitude and embedding memory by 4-5x.
Authors: Federico Cocchi, Nicholas Moratelli, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
Abstract: Multimodal LLMs (MLLMs) are the natural extension of large language models to handle multimodal inputs, combining text and image data. They have recently garnered attention due to their capability to address complex tasks involving both modalities. However, their effectiveness is limited to the knowledge acquired during training, which restricts their practical utility. In this work, we introduce a novel method to enhance the adaptability of MLLMs by integrating external knowledge sources. Our proposed model, Reflective LLaVA (ReflectiVA), utilizes reflective tokens to dynamically determine the need for external knowledge and predict the relevance of information retrieved from an external database. Tokens are trained following a two-stage two-model training recipe. This ultimately enables the MLLM to manage external knowledge while preserving fluency and performance on tasks where external knowledge is not needed. Through our experiments, we demonstrate the efficacy of ReflectiVA for knowledge-based visual question answering, highlighting its superior performance compared to existing methods. Source code and trained models are publicly available at https://aimagelab.github.io/ReflectiVA.
Authors: Yuanze Lin, Yunsheng Li, Dongdong Chen, Weijian Xu, Ronald Clark, Philip H. S. Torr
Abstract: We introduce Olympus, a new approach that transforms Multimodal Large Language Models (MLLMs) into a unified framework capable of handling a wide array of computer vision tasks. Utilizing a controller MLLM, Olympus delegates over 20 specialized tasks across images, videos, and 3D objects to dedicated modules. This instruction-based routing enables complex workflows through chained actions without the need for training heavy generative models. Olympus easily integrates with existing MLLMs, expanding their capabilities with comparable performance. Experimental results demonstrate that Olympus achieves an average routing accuracy of 94.75% across 20 tasks and precision of 91.82% in chained action scenarios, showcasing its effectiveness as a universal task router that can solve a diverse range of computer vision tasks. Project page: http://yuanze-lin.me/Olympus_page/
Authors: Alejandro Lozano, Min Woo Sun, James Burgess, Liangyu Chen, Jeffrey J Nirschl, Jeffrey Gu, Ivan Lopez, Josiah Aklilu, Austin Wolfgang Katzer, Collin Chiu, Anita Rau, Xiaohan Wang, Yuhui Zhang, Alfred Seunghoon Song, Robert Tibshirani, Serena Yeung-Levy
Abstract: The development of vision-language models (VLMs) is driven by large-scale and diverse multimodal datasets. However, progress toward generalist biomedical VLMs is limited by the lack of annotated, publicly accessible datasets across biology and medicine. Existing efforts are restricted to narrow domains, missing the full diversity of biomedical knowledge encoded in scientific literature. To address this gap, we introduce BIOMEDICA, a scalable, open-source framework to extract, annotate, and serialize the entirety of the PubMed Central Open Access subset into an easy-to-use, publicly accessible dataset. Our framework produces a comprehensive archive with over 24 million unique image-text pairs from over 6 million articles. Metadata and expert-guided annotations are also provided. We demonstrate the utility and accessibility of our resource by releasing BMCA-CLIP, a suite of CLIP-style models continuously pre-trained on the BIOMEDICA dataset via streaming, eliminating the need to download 27 TB of data locally. On average, our models achieve state-of-the-art performance across 40 tasks - spanning pathology, radiology, ophthalmology, dermatology, surgery, molecular biology, parasitology, and cell biology - excelling in zero-shot classification with a 6.56% average improvement (as high as 29.8% and 17.5% in dermatology and ophthalmology, respectively), and stronger image-text retrieval, all while using 10x less compute. To foster reproducibility and collaboration, we release our codebase and dataset for the broader research community.
Authors: Samuel Soo, Chen Guang, Wesley Teng, Chandrasekaran Balaganesh, Tan Guoxian, Yan Ming
Abstract: Effective and reliable control over large language model (LLM) behavior is a significant challenge. While activation steering methods, which add steering vectors to a model's hidden states, are a promising approach, existing techniques often lack precision and interpretability in how they influence model outputs. We introduce Feature Guided Activation Additions (FGAA), a novel activation steering method that leverages insights from Contrastive Activation Addition (CAA) and Sparse Autoencoder-Targeted Steering (SAE-TS). By operating in the latent space of a Sparse Autoencoder (SAE) and employing optimization techniques to select desired SAE features, FGAA constructs precise steering vectors that provide better steering effects while maintaining coherence of steered model outputs. In this regard, evaluations on Gemma-2-2B and Gemma-2-9B models across various steering tasks demonstrate that FGAA outperforms existing steering methods of CAA, SAE decoder steering, and SAE-TS. Our results also highlight important trade-offs between steering scale and general model capabilities that are consistent across all tested steering methods.
Authors: Aleksander Ficek, Somshubra Majumdar, Vahid Noroozi, Boris Ginsburg
Abstract: Synthetic verification techniques such as generating test cases and reward modelling are common ways to enhance the coding capabilities of large language models (LLM) beyond predefined tests. Additionally, code verification has recently found great success as a critical component in improving reasoning capability of LLMs via reinforcement learning. In this paper, we propose a an approach which can transform existing coding benchmarks into scoring and ranking datasets to evaluate the effectiveness of synthetic verifiers. We also propose multiple metrics to measure different aspects of the synthetic verifiers with the proposed benchmarks. By employing the proposed approach, we release four new benchmarks (HE-R, HE-R+, MBPP-R, and MBPP-R+), and analyzed synthetic verification methods with standard, reasoning-based, and reward-based LLMs. Our experiments show that reasoning can significantly improve test case generation and that scaling the number of test cases enhances the verification accuracy.
Authors: Abe Bohan Hou, Hongru Du, Yichen Wang, Jingyu Zhang, Zixiao Wang, Paul Pu Liang, Daniel Khashabi, Lauren Gardner, Tianxing He
Abstract: Can we simulate a sandbox society with generative agents to model human behavior, thereby reducing the over-reliance on real human trials for assessing public policies? In this work, we investigate the feasibility of simulating health-related decision-making, using vaccine hesitancy, defined as the delay in acceptance or refusal of vaccines despite the availability of vaccination services (MacDonald, 2015), as a case study. To this end, we introduce the VacSim framework with 100 generative agents powered by Large Language Models (LLMs). VacSim simulates vaccine policy outcomes with the following steps: 1) instantiate a population of agents with demographics based on census data; 2) connect the agents via a social network and model vaccine attitudes as a function of social dynamics and disease-related information; 3) design and evaluate various public health interventions aimed at mitigating vaccine hesitancy. To align with real-world results, we also introduce simulation warmup and attitude modulation to adjust agents' attitudes. We propose a series of evaluations to assess the reliability of various LLM simulations. Experiments indicate that models like Llama and Qwen can simulate aspects of human behavior but also highlight real-world alignment challenges, such as inconsistent responses with demographic profiles. This early exploration of LLM-driven simulations is not meant to serve as definitive policy guidance; instead, it serves as a call for action to examine social simulation for policy development.
Authors: Boyi Ma, Yanguang Zhao, Jie Wang, Guankun Wang, Kun Yuan, Tong Chen, Long Bai, Hongliang Ren
Abstract: DeepSeek series have demonstrated outstanding performance in general scene understanding, question-answering (QA), and text generation tasks, owing to its efficient training paradigm and strong reasoning capabilities. In this study, we investigate the dialogue capabilities of the DeepSeek model in robotic surgery scenarios, focusing on tasks such as Single Phrase QA, Visual QA, and Detailed Description. The Single Phrase QA tasks further include sub-tasks such as surgical instrument recognition, action understanding, and spatial position analysis. We conduct extensive evaluations using publicly available datasets, including EndoVis18 and CholecT50, along with their corresponding dialogue data. Our comprehensive evaluation results indicate that, when provided with specific prompts, DeepSeek-V3 performs well in surgical instrument and tissue recognition tasks However, DeepSeek-V3 exhibits significant limitations in spatial position analysis and struggles to understand surgical actions accurately. Additionally, our findings reveal that, under general prompts, DeepSeek-V3 lacks the ability to effectively analyze global surgical concepts and fails to provide detailed insights into surgical scenarios. Based on our observations, we argue that the DeepSeek-V3 is not ready for vision-language tasks in surgical contexts without fine-tuning on surgery-specific datasets.
Authors: Mattia Opper, Roland Fernandez, Paul Smolensky, Jianfeng Gao
Abstract: Transformers struggle with length generalisation, displaying poor performance even on basic tasks. We test whether these limitations can be explained through two key failures of the self-attention mechanism. The first is the inability to fully remove irrelevant information. The second is tied to position, even if the dot product between a key and query is highly negative (i.e. an irrelevant key) learned positional biases may unintentionally up-weight such information - dangerous when distances become out of distribution. Put together, these two failure cases lead to compounding generalisation difficulties. We test whether they can be mitigated through the combination of a) selective sparsity - completely removing irrelevant keys from the attention softmax and b) contextualised relative distance - distance is only considered as between the query and the keys that matter. We show how refactoring the attention mechanism with these two mitigations in place can substantially improve generalisation capabilities of decoder only transformers.
Authors: Neil Mallinar, A. Ali Heydari, Xin Liu, Anthony Z. Faranesh, Brent Winslow, Nova Hammerquist, Benjamin Graef, Cathy Speed, Mark Malhotra, Shwetak Patel, Javier L. Prieto, Daniel McDuff, Ahmed A. Metwally
Abstract: Large language models (LLMs) have emerged as powerful tools for analyzing complex datasets. Recent studies demonstrate their potential to generate useful, personalized responses when provided with patient-specific health information that encompasses lifestyle, biomarkers, and context. As LLM-driven health applications are increasingly adopted, rigorous and efficient one-sided evaluation methodologies are crucial to ensure response quality across multiple dimensions, including accuracy, personalization and safety. Current evaluation practices for open-ended text responses heavily rely on human experts. This approach introduces human factors and is often cost-prohibitive, labor-intensive, and hinders scalability, especially in complex domains like healthcare where response assessment necessitates domain expertise and considers multifaceted patient data. In this work, we introduce Adaptive Precise Boolean rubrics: an evaluation framework that streamlines human and automated evaluation of open-ended questions by identifying gaps in model responses using a minimal set of targeted rubrics questions. Our approach is based on recent work in more general evaluation settings that contrasts a smaller set of complex evaluation targets with a larger set of more precise, granular targets answerable with simple boolean responses. We validate this approach in metabolic health, a domain encompassing diabetes, cardiovascular disease, and obesity. Our results demonstrate that Adaptive Precise Boolean rubrics yield higher inter-rater agreement among expert and non-expert human evaluators, and in automated assessments, compared to traditional Likert scales, while requiring approximately half the evaluation time of Likert-based methods. This enhanced efficiency, particularly in automated evaluation and non-expert contributions, paves the way for more extensive and cost-effective evaluation of LLMs in health.
Authors: Jie Ma, Zhitao Gao, Qi Chai, Jun Liu, Pinghui Wang, Jing Tao, Zhou Su
Abstract: Audio-Visual Question Answering (AVQA) is a challenging multimodal reasoning task requiring intelligent systems to answer natural language queries based on paired audio-video inputs accurately. However, existing AVQA approaches often suffer from overfitting to dataset biases, leading to poor robustness. Moreover, current datasets may not effectively diagnose these methods. To address these challenges, we first introduce a novel dataset, FortisAVQA, constructed in two stages: (1) rephrasing questions in the test split of the public MUSIC-AVQA dataset and (2) introducing distribution shifts across questions. The first stage expands the test space with greater diversity, while the second enables a refined robustness evaluation across rare, frequent, and overall question distributions. Second, we introduce a robust Multimodal Audio-Visual Epistemic Network (MAVEN) that leverages a multifaceted cycle collaborative debiasing strategy to mitigate bias learning. Experimental results demonstrate that our architecture achieves state-of-the-art performance on FortisAVQA, with a notable improvement of 7.81\%. Extensive ablation studies on both datasets validate the effectiveness of our debiasing components. Additionally, our evaluation reveals the limited robustness of existing multimodal QA methods. We also verify the plug-and-play capability of our strategy by integrating it with various baseline models across both datasets. Our dataset and code are available at https://github.com/reml-group/fortisavqa.