Multi-agent debates have been introduced to improve the accuracy of Large Language Models (LLMs) by having multiple agents discuss solutions to a problem over several rounds of debate. However, models often generate incorrect yet confident-sounding responses, which can mislead the others. This issue arises partly because agents do not consider how confident their peers are. To address this, we propose DebUnc, a debate framework that uses uncertainty metrics to assess agent confidence. Confidence is then conveyed through textual prompts or via a modified attention mechanism that adjusts token weights. Evaluations across benchmarks show that attention-based methods are particularly effective and that performance continues to improve as uncertainty estimation becomes more reliable. The code is available at https://github.com/lukeyoffe/debunc.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Large language models (LLMs) often fail to scale their performance on long-context tasks performance in line with the context lengths they support. This gap is commonly attributed to retrieval failures—the models’ inability to identify information in the long inputs that is relevant to the task they are solving. Accordingly, recent efforts often focus on evaluating and improving LLMs’ retrieval performance: if retrieval is perfect, a model should, in principle, perform just as well on a long input as it does on a short one—or should it? This paper presents findings that the answer to this question may be negative. Our systematic experiments across 5 open- and closed-source LLMs on math, question answering, and coding tasks reveal that, even when models can perfectly retrieve all relevant information, their performance still degrades substantially (13.9%–85%) as input length increases but remains well within their claimed context lengths. This failure occurs even when the irrelevant tokens are replaced with minimally distracting whitespace, and, more surprisingly, when they are all masked and the models are forced to attend only to the relevant tokens. A similar performance drop is observed when all relevant evidence is placed immediately before the question. Our findings reveal a previously-unrealized limitation: the sheer length of the input alone can hurt LLM performance, independent of retrieval quality and without any distraction. They motivate our simple, model-agnostic mitigation strategy that transforms a long-context task into a short-context one by prompting the model to recite the retrieved evidence before attempting to solve the problem. On RULER, we observe a consistent improvement of GPT-4o up to 4% on an already strong baseline.
The reliability of large language models (LLMs) is greatly compromised by their tendency to hallucinate, underscoring the need for precise identification of knowledge gaps within LLMs. Various methods for probing such gaps exist, ranging from calibration-based to prompting-based methods. To evaluate these probing methods, in this paper, we propose a new process based on using input variations and quantitative metrics. Through this, we expose two dimensions of inconsistency in knowledge gap probing. (1) **Intra-method inconsistency:** Minimal non-semantic perturbations in prompts lead to considerable variance in detected knowledge gaps within the same probing method; e.g., the simple variation of shuffling answer options can decrease agreement to around 40%. (2) **Cross-method inconsistency:** Probing methods contradict each other on whether a model knows the answer. Methods are highly inconsistent – with decision consistency across methods being as low as 7% – even though the model, dataset, and prompt are all the same. These findings challenge existing probing methods and highlight the urgent need for perturbation-robust probing frameworks.
The Spiral of Silence (SoS) theory holds that individuals with minority views often refrain from speaking out for fear of social isolation, enabling majority positions to dominate public discourse. When the “agents” are large language models (LLMs), however, the classical psychological explanation is not directly applicable, since SoS was developed for human societies. This raises a central question: can SoS-like dynamics nevertheless emerge from purely statistical language generation in LLM collectives? We propose an evaluation framework for examining SoS in LLM agents. Specifically, we consider four controlled conditions that systematically vary the availability of “History” and “Persona” signals. Opinion dynamics are assessed using trend tests such as Mann–Kendall and Spearman’s rank, along with concentration measures including kurtosis and interquartile range. Experiments across open-source and closed-source models show that history and persona together produce strong majority dominance and replicate SoS patterns; history signals alone induce strong anchoring; and persona signals alone foster diverse but uncorrelated opinions, indicating that without historical anchoring, SoS dynamics cannot emerge. The work bridges computational sociology and responsible AI design, highlighting the need to monitor and mitigate emergent conformity in LLM-agent systems.
The Prompt Makes the Person(a): A Systematic Evaluation of Sociodemographic Persona Prompting for Large Language Models
PDF ↗Persona prompting is increasingly used in large language models (LLMs) to simulate views of various sociodemographic groups. However, how a persona prompt is formulated can significantly affect outcomes, raising concerns about the fidelity of such simulations. Using five open-source LLMs, we systematically examine how different persona prompt strategies, specifically role adoption formats and demographic priming strategies, influence LLM simulations across 15 intersectional demographic groups in both open- and closed-ended tasks. Our findings show that LLMs struggle to simulate marginalized groups but that the choice of demographic priming and role adoption strategy significantly impacts their portrayal. Specifically, we find that prompting in an interview-style format and name-based priming can help reduce stereotyping and improve alignment. Surprisingly, smaller models like OLMo-2-7B outperform larger ones such as Llama-3.3-70B.Our findings offer actionable guidance for designing sociodemographic persona prompts in LLM-based simulation studies.
Many applications that modern large language models (LLMs) are deployed on are retrieval tasks: the answer can be recovered from context and success is a matter of learning generalizable features from data. However, this is easier said than done. Overparametrized models trained on cross-entropy loss can overfit on noise. We argue that such overfitting is prone to happen when the model can identify mechanisms that rapidly drive down the loss of certain tokens early on in training. Fitting some tokens early reduce gradient signals in later iterations, as such, remaining tokens are more vulnerable to noise overfitting. We dub this phenomenon unequal learning and show that LLMs with longer contexts or larger embedding sizes are prone to this failure mode. In this work, we argue that learning training samples at an equal rate helps counter such biases. We highlight two mechanisms that promote equal learning: (i) loss functions that regularize uniform margins across training samples, (ii) small learning rates (e.g. by warming up) at the start of training. We demonstrate these approaches on various synthetic and natural language datasets.
Using language models to scalably approximate human preferences on text quality (LLM-as-a-judge) has become a standard practice applicable to many tasks. A judgment is often extracted from the judge’s textual output alone, typically with greedy decoding. However, LLM judges naturally provide distributions over judgment tokens, inviting a breadth of inference methods for extracting fine-grained preferences. We find that taking the mean of the judgment distribution consistently outperforms taking the mode (i.e. greedy decoding) in all evaluation settings (i.e. pointwise, pairwise, and listwise). We further explore novel methods of deriving preferences from judgment distributions, and find that methods incorporating risk aversion often improve performance. Lastly, we analyze LLM-as-a-judge paired with chain-of-thought (CoT) prompting, showing that CoT can collapse the spread of the judgment distribution, often harming performance. Our findings show that leveraging distributional output improves LLM-as-a-judge, as opposed to using the text interface alone.
There emerges a critical security risk of LLM agents: indirect prompt injection, a sophisticated attack vector that compromises thecore of these agents, the LLM, by manipulating contextual information rather than direct user prompts. In this work, we propose a generic black-box optimization framework, AGENTVIGIL, designed to automatically discover and exploit indirect prompt injection vulnerabilities across diverse LLM agents. Our approach starts by constructing a high-quality initial seed corpus, then employs a seed selectionalgorithm based on Monte Carlo Tree Search (MCTS) to iteratively refine inputs, therebymaximizing the likelihood of uncovering agent weaknesses. We evaluate AGENTVIGIL on twopublic benchmarks, AgentDojo and VWA-adv, where it achieves 71% and 70% success rates against agents based on o3-mini and GPT-4o, respectively, nearly doubling the performance of handcrafted baseline attacks. Moreover, AGENTVIGIL exhibits strong transferability across unseen tasks and internal LLMs, as well as promising results against defenses. Beyondbenchmark evaluations, we apply our attacks in real-world environments, successfully misleading agents to navigate to arbitrary URLs,including malicious sites.
Visual metaphors are a complex vision–language phenomenon that requires both perceptual and conceptual reasoning to understand. They provide a valuable test of a model’s ability to interpret visual input and reason about it with creativity and coherence. We introduce ImageMet, a visual metaphor dataset, featuring 2177 synthetic and 350 human-annotated images. We benchmark several SOTA VLMs on two tasks: Visual Metaphor Captioning (VMC) and Visual Metaphor VQA (VM-VQA). We establish strong baselines by fine-tuning on ImageMet, which yields substantial performance gains in VMC (+4.67% SBERT-Similarity, +4.84% task-specific metric) and VM-VQA (+9.3% Accuracy on average). Additionally, we introduce a task-specific CoT prompting strategy that outperforms standard few-shot baselines (+1.99% in VMC, +5.21% in VM-VQA). We observe that despite strong performance on the VMC task, VLMs still significantly lag behind humans in understanding visual metaphors, indicating that their success often relies on learned associations rather than genuine analytical reasoning. We note that this gap is often obscured in metaphor captioning tasks where the automatic metrics correlate only moderately at best with human judgment (Pearson r < 0.6), highlighting the need for careful, holistic evaluation of the visual metaphor understanding of the models.
MAKIEval: A Multilingual Automatic WiKidata-based Framework for Cultural Awareness Evaluation for LLMs
PDF ↗Large language models (LLMs) are used globally across many languages, but their English-centric pretraining raises concerns about cross-lingual disparities for cultural awareness, often resulting in biased outputs. However, comprehensive multilingual evaluation remains challenging due to limited benchmarks and questionable translation quality. To better assess these disparities, we introduce MAKIEval, an automatic multilingual framework for evaluating cultural awareness in LLMs across languages, regions, and topics. MAKIEval evaluates open-ended text generation, capturing how models express culturally grounded knowledge in natural language. Leveraging Wikidata’s multilingual structure as a cross-lingual anchor, it automatically identifies cultural entities in model outputs and links them to structured knowledge, enabling scalable, language-agnostic evaluation without manual annotation or translation. We then introduce four metrics that capture complementary dimensions of cultural awareness: granularity, diversity, cultural specificity, and consensus across languages. We assess 7 LLMs developed from different parts of the world, encompassing both open-source and proprietary systems, across 13 languages, 19 countries and regions, and 6 culturally salient topics (e.g., food, clothing). Notably, we find that models tend to exhibit stronger cultural awareness in English, suggesting that English prompts more effectively activate culturally grounded knowledge. We publicly release our code and data.
Protein-specific large language models (ProteinLLMs) are revolutionizing protein science by enabling more efficient protein structure prediction, function annotation, and design. While existing surveys focus on specific aspects or applications, this work provides the first comprehensive overview of ProteinLLMs, covering their architectures, training datasets, evaluation metrics, and diverse applications. Through a systematic analysis of over 100 articles, we propose a structured taxonomy of state-of-the-art ProteinLLMs, analyze how they leverage large-scale protein sequence data for improved accuracy, and explore their potential in advancing protein engineering and biomedical research. Additionally, we discuss key challenges and future directions, positioning ProteinLLMs as essential tools for scientific discovery in protein science. Resources are maintained at https://github.com/Yijia-Xiao/Protein-LLM-Survey.
Mainstream large vision-language models (LVLMs) inherently encode cultural biases, highlighting the need for diverse multimodal datasets. To address this gap, we introduce PEARL, a large-scale Arabic multimodal dataset and benchmark explicitly designed for cultural understanding. Constructed through advanced agentic workflows and extensive human-in-the-loop annotations by 37 annotators from across the Arab world, PEARL comprises over 309K multimodal examples spanning ten culturally significant domains covering all Arab countries. We further provide two robust evaluation benchmarks (PEARL and PEARL-LITE) along with a specialized subset (PEARL-X) explicitly developed to assess nuanced cultural variations. Comprehensive evaluations on state-of-the-art open and proprietary LVLMs demonstrate that reasoning-centric instruction alignment substantially improves models’ cultural grounding compared to conventional scaling methods. PEARL establishes a foundational resource for advancing culturally-informed multimodal modeling research. All datasets and benchmarks are publicly available.
Backdoor attacks are powerful and effective, but distributing LLMs without a proven track record like ‘meta-llama‘ or ‘qwen‘ rarely gains community traction. We identify LoRA sharing as a unique scenario where users are more willing to try unendorsed assets, since such shared LoRAs allow them to enjoy personalized LLMs with negligible investment. However, this convenient share-and-play ecosystem also introduces a new attack surface, where attackers can distribute malicious LoRAs to an undefended community. Despite the high-risk potential, no prior art has comprehensively explored LoRA’s attack surface under the downstream-enhancing share-and-play context. In this paper, we investigate how backdoors can be injected into task-enhancing LoRAs and examine the mechanisms of such infections. We find that with a simple, efficient, yet specific recipe, **a backdoor LoRA can be trained once and then seamlessly merged (in a training-free fashion) with multiple task-enhancing LoRAs, retaining both its malicious backdoor and benign downstream capabilities.** This allows attackers to scale the distribution of compromised LoRAs with minimal effort by leveraging the rich pool of existing shared LoRA assets. We note that such merged LoRAs are particularly *infectious* — because their malicious intent is cleverly concealed behind improved downstream capabilities, creating a strong incentive for voluntary download — and *dangerous* — because under local deployment, no safety measures exist to intervene when things go wrong. Our work is among the first to study this new threat model of training-free distribution of downstream-capable-yet-backdoor-injected LoRAs, highlighting the urgent need for heightened security awareness in the LoRA ecosystem. **Warning: This paper contains offensive content and involves a real-life tragedy.**
Understanding human social behavior such as recognizing emotions and the social dynamics causing them is an important and challenging problem. While LLMs have made remarkable advances, they are limited to the textual domain and cannot account for the major role that non-verbal cues play in understanding social situations. Vision Language Models (VLMs) can potentially account for this gap, however their ability to make correct inferences over such social cues has received little attention. In this paper, we explore the capabilities of VLMs at social reasoning. We identify a previously overlooked limitation in VLMs: the Visual Social-Pragmatic Inference gap. To target this gap, we propose a new task for VLMs: Visual Social-Pragmatic Inference. We construct a high quality dataset to test the abilities of a VLM for this task and benchmark the performance of several VLMs on it.
Evaluating Fairness in Large Vision-Language Models Across Diverse Demographic Attributes and Prompts
PDF ↗Large vision-language models (LVLMs) have recently achieved significant progress, demonstrating strong capabilities in open-world visual understanding. However, it is not yet clear how LVLMs address demographic biases in real life, especially the disparities across attributes such as gender, skin tone, age and race. In this paper, We empirically investigate visual fairness in several mainstream LVLMs by auditing their performance disparities across demographic attributes using public fairness benchmark datasets (e.g., FACET, UTKFace). Our fairness evaluation framework employs direct and single-choice question prompt on visual question-answering/classification tasks. Despite advancements in visual understanding, our zero-shot prompting results show that both open-source and closed-source LVLMs continue to exhibit fairness issues across different prompts and demographic groups. Furthermore, we propose a potential multi-modal Chain-of-thought (CoT) based strategy for unfairness mitigation, applicable to both open-source and closed-source LVLMs. This approach enhances transparency and offers a scalable solution for addressing fairness, providing a solid foundation for future research and practical efforts in unfairness mitigation. The dataset and code used in this study are publicly available at this GitHub Repository.
How Can Input Reformulation Improve Tool Usage Accuracy in a Complex Dynamic Environment? A Study on tau-bench
PDF ↗Recent advances in reasoning and planning capabilities of large language models (LLMs) have enabled their potential as autonomous agents capable of tool use in dynamic environments. However, in multi-turn conversational environments like \tau‐bench, these agents often struggle with consistent reasoning, adherence to domain-specific policies, and extracting correct information over a long horizon of tool-calls and conversation. To capture and mitigate these failures, we conduct a comprehensive manual analysis of the common errors occurring in the conversation trajectories. We then experiment with reformulations of inputs to the tool-calling agent for improvement in agent decision making. Finally, we propose the Input-Reformulation Multi-Agent (IRMA) framework, which automatically reformulates user queries augmented with relevant domain rules and tool suggestions for the tool-calling agent to focus on. The results show that IRMA significantly outperforms ReAct, Function Calling, and Self-Reflection by 16.1%, 12.7%, and 19.1%, respectively, in overall pass^5 scores. These findings highlight the superior reliability and consistency of IRMA compared to other methods in dynamic environments.
Recent advancements in Large Language Models (LLMs) have significantly enhanced their code generation capabilities. However, their robustness against adversarial misuse, particularly through multi-turn malicious coding prompts, remains underexplored. In this work, we introduce code decomposition attacks, where a malicious coding task is broken down into a series of seemingly benign subtasks across multiple conversational turns to evade safety filters. To facilitate systematic evaluation, we introduce MOCHA, a large-scale benchmark designed to evaluate the robustness of code LLMs against both single-turn and multi-turn malicious prompts. Empirical results across open- and closed-source models reveal persistent vulnerabilities, especially under multi-turn scenarios. Fine-tuning on MOCHA improves rejection rates while preserving coding ability, and importantly, enhances robustness on external adversarial datasets with up to 32.4% increase in rejection rates without any additional supervision.
In collaborative creation tasks, people steer artifacts towards specific goals by _refining_ them with _multimodal_ communication over multiple rounds of interaction. In contrast, generative AI excels at creating artifacts in a single turn but can struggle to make precise refinements that match our design intent. To close this gap, we present mrCAD, a dataset of multi-turn interactions in which pairs of humans iteratively created and refined computer-aided designs (CADs). In each game, a _Designer sent instructions to a _Maker_, explaining how to create and subsequently refine a CAD to match a target design that only the _Designer_ could see. mrCAD consists of 6,082 communication games, 15,163 instruction-execution rounds, played between 1,092 pairs of human players. Crucially, _Designers_ had access to two communication modalities – text and drawing. Analysis finds that players relied more on text in refinement than in initial generation instructions, and used different linguistic elements for refinement than for generation. We also find that state-of-the-art VLMs are better at following generation instructions than refinement instructions. These results lay the foundation for modeling multi-turn, multimodal communication not captured in prior datasets.
A Comparison of Independent and Joint Fine-tuning Strategies for Retrieval-Augmented Generation
PDF ↗Retrieval augmented generation (RAG) is a popular framework for question answering that is powered by two large language models (LLMs): an embedding model that retrieves context documents from a database that are relevant to a given question, and a generator model that uses the retrieved context to generate an answer to the question. Both the embedding and generator models can be fine-tuned to increase performance of a RAG pipeline on a new task, but multiple fine-tuning strategies exist with different costs and benefits. In this paper, we evaluate and compare several RAG fine-tuning strategies, including independent, joint, and two-phase fine-tuning. In our experiments, we observe that all of these strategies achieve about equal improvement in EM and F1 generation quality metrics, although they have significantly different computational costs. We conclude the optimal fine-tuning strategy to use depends on whether the training dataset includes context labels and whether a grid search over the learning rates for the embedding and generator models is required.
When training large language models (LLMs), it is common practice to track downstream task performance throughout the training process and select the checkpoint with the highest validation score.However, downstream metrics often exhibit substantial fluctuations, making it difficult to identify the checkpoint that truly represents the best-performing model.In this study, we empirically analyze the stability of downstream task performance in an LLM trained on diverse web-scale corpora.We find that task scores frequently fluctuate throughout training, both at the aggregate and example levels.To address this instability, we investigate two post-hoc checkpoint integration methods: checkpoint averaging and ensemble, motivated by the hypothesis that aggregating neighboring checkpoints can reduce performance volatility.We demonstrate both empirically and theoretically that these methods improve downstream performance stability without requiring any changes to the training procedure.