In this paper, our objective is to develop a multi-agent financial system that incorporates simulated trading, a technique extensively utilized by financial professionals. While current LLM-based agent models demonstrate competitive performance, they still exhibit significant deviations from real-world fund companies. A critical distinction lies in the agents’ reliance on “post-reflection”, particularly in response to adverse outcomes, but lack a distinctly human capability: long-term prediction of future trends. Therefore, we introduce QuantAgents, a multi-agent system integrating simulated trading, to comprehensively evaluate various investment strategies and market scenarios without assuming actual risks. Specifically, QuantAgents comprises four agents: a simulated trading analyst, a risk control analyst, a market news analyst, and a manager, who collaborate through several meetings. Moreover, our system incentivizes agents to receive feedback on two fronts: performance in real-world markets and predictive accuracy in simulated trading. Extensive experiments demonstrate that our framework excels across all metrics, yielding an overall return of nearly 300% over the three years (https://quantagents.github.io).
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
Multimodal Large Language Models (MLLMs) show promising results for embodied agents in operating meaningfully in complex, human-centered environments. Yet, evaluating their capacity for nuanced, human-like reasoning and decision-making remains challenging. In this work, we introduce VIVA+, a cognitively grounded benchmark for evaluating the reasoning and decision-making of MLLMs in human-centered situations. VIVA+ consists of 1,317 real-world situations paired with 6,373 multiple-choice questions, targeting three core abilities for decision-making: (1) Foundational Situation Comprehension, (2) Context-Driven Action Justification, and (3) Reflective Reasoning. Together, these dimensions provide a systematic framework for assessing a model’s ability to perceive, reason, and act in socially meaningful ways. We evaluate the latest commercial and open-source models on VIVA+, where we reveal distinct performance patterns and highlight significant challenges. We further explore targeted training and multi-step reasoning strategies, which yield consistent performance improvements. Finally, our in-depth analysis highlights current model limitations and provides actionable insights for advancing MLLMs toward more robust, context-aware, and socially adept decision-making in real-world settings.
One challenge of quantizing a large language model (LLM) is the presence of outliers. Outliers often make uniform quantization schemes less effective, particularly in extreme cases such as 4-bit quantization. We introduce KurTail, a new post-training quantization (PTQ) scheme that leverages Kurtosis-based rotation to mitigate outliers in the activations of LLMs. Our method optimizes Kurtosis as a measure of tailedness. This approach enables the quantization of weights, activations, and the KV cache in 4 bits. We utilize layer-wise optimization, ensuring memory efficiency. KurTail outperforms existing quantization methods, offering a 13.3% boost in MMLU accuracy and a 15.5% boost in Wiki perplexity compared to QuaRot. It also outperforms SpinQuant with a 2.6% MMLU gain and reduces perplexity by 2.9%, all while reducing the training cost. For comparison, learning the rotation using SpinQuant for Llama3-70B requires at least four NVIDIA H100 80GB GPUs, whereas our method requires only a single GPU, making it more accessible.
Investigating the Impact of Conceptual Metaphors on LLM-based NLI through Shapley Interactions
PDF ↗Metaphorical language is prevalent in everyday communication, often used unconsciously, as in “rising crime.” While LLMs excel at identifying metaphors in text, they struggle with downstream tasks that implicitly require correct metaphor interpretation, such as natural language inference (NLI). This work explores how LLMs perform on NLI with metaphorical input. Particularly, we investigate whether incorporating conceptual metaphors (source and target domains) enhances performance in zero-shot and few-shot settings. Our contributions are two-fold: (1) we extend metaphorical texts in an existing NLI dataset by source and target domains, and (2) we conduct an ablation study using Shapley values and interactions to assess the extent to which LLMs interpret metaphorical language correctly in NLI. Our results indicate that incorporating conceptual metaphors often improves task performance.
Judging with Many Minds: Do More Perspectives Mean Less Prejudice? On Bias Amplification and Resistance in Multi-Agent Based LLM-as-Judge
PDF ↗LLM-as-Judge has emerged as a scalable alternative to human evaluation, enabling large language models (LLMs) to provide reward signals in trainings. While recent work has explored multi-agent extensions such as multi-agent debate and meta-judging to enhance evaluation quality, the question of how intrinsic biases manifest in these settings remains underexplored. In this study, we conduct a systematic analysis of four diverse bias types: position bias, verbosity bias, chain-of-thought bias, and bandwagon bias. We evaluate these biases across two widely adopted multi-agent LLM-as-Judge frameworks: Multi-Agent-Debate and LLM-as-Meta-Judge. Our results show that debate framework amplifies biases sharply after the initial debate, and this increased bias is sustained in subsequent rounds, while meta-judge approaches exhibit greater resistance. We further investigate the incorporation of PINE, a leading single-agent debiasing method, as a bias-free agent within these systems. The results reveal that this bias-free agent effectively reduces biases in debate settings but provides less benefit in meta-judge scenarios. Our work provides a comprehensive study of bias behavior in multi-agent LLM-as-Judge systems and highlights the need for targeted bias mitigation strategies in collaborative evaluation settings.
Optimising Factual Consistency in Summarisation via Preference Learning from Multiple Imperfect Metrics
PDF ↗Reinforcement learning with evaluation metrics as rewards is widely used to enhance specific capabilities of language models. However, for tasks such as factually consistent summarisation, existing metrics remain underdeveloped, limiting their effectiveness as signals for shaping model behaviour.While individual factuality metrics are unreliable, their combination can more effectively capture diverse factual errors. We leverage this insight to introduce an automated training pipeline that improves factual consistency in summaries by aggregating scores from different weak metrics. Our approach avoids the need for complex reward shaping by mapping scores to preferences and filtering out cases with high disagreement between metrics. For each source document, we generate lexically similar summary pairs by varying decoding strategies, enabling the model to learn from factual differences caused by subtle lexical differences. This approach constructs a high-quality preference dataset using only source documents.Experiments demonstrate consistent factuality gains across models, ranging from early encoder-decoder architectures to modern large language models, with smaller models reaching comparable factuality to larger ones.
Choosing a Model, Shaping a Future: Comparing LLM Perspectives on Sustainability and its Relationship with AI
PDF ↗As organizations increasingly rely on AI systems for decision support in sustainability contexts, it becomes critical to understand the inherent biases and perspectives embedded in Large Language Models (LLMs). This study systematically investigates how five state-of-the-art LLMs – Claude, DeepSeek, GPT, LLaMA, and Mistral – conceptualize sustainability and its relationship with AI. We administered validated, psychometric sustainability-related questionnaires – each 100 times per model – to capture response patterns and variability. Our findings revealed significant inter-model differences: For example, GPT exhibited skepticism about the compatibility of AI and sustainability, whereas LLaMA demonstrated extreme techno-optimism with perfect scores for several Sustainable Development Goals (SDGs). Models also diverged in attributing institutional responsibility for AI and sustainability integration, a results that holds implications for technology governance approaches. Our results demonstrate that model selection could substantially influence organizational sustainability strategies, highlighting the need for awareness of model-specific biases when deploying LLMs for sustainability-related decision-making.
The rapid advancement of large language models (LLMs) significantly enhances long-context Retrieval-Augmented Generation (RAG), yet existing benchmarks focus primarily on English. This leaves low-resource languages without comprehensive evaluation frameworks, limiting their progress in retrieval-based tasks. To bridge this gap, we introduce Ko-LongRAG, the first Korean long-context RAG benchmark. Unlike conventional benchmarks that depend on external retrievers, Ko-LongRAG adopts a retrieval-free approach designed around Specialized Content Knowledge (SCK), enabling controlled and high-quality QA pair generation without the need for an extensive retrieval infrastructure. Our evaluation shows that o1 model achieves the highest performance among proprietary models, while EXAONE 3.5 leads among open-sourced models. Additionally, various findings confirm Ko-LongRAG as a reliable benchmark for assessing Korean long-context RAG capabilities and highlight its potential for advancing multilingual RAG research. The dataset and source code will be released publicly.
PhysicsArena: The First Multimodal Physics Reasoning Benchmark Exploring Variable, Process, and Solution Dimensions
PDF ↗Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in diverse reasoning tasks, yet their application to complex physics reasoning remains underexplored. Physics reasoning presents unique challenges, requiring grounding in physical conditions and the interpretation of multimodal information. Current physics benchmarks are limited, often focusing on text-only inputs or solely on problem-solving, thereby overlooking the critical intermediate steps of variable identification and process formulation. To address these limitations, we introduce **PhysicsArena, the first multimodal physics reasoning benchmark designed to holistically evaluate MLLMs across three critical dimensions: variable identification, physical process formulation, and solution derivation.** PhysicsArena aims to provide a comprehensive platform for assessing and advancing the multimodal physics reasoning abilities of MLLMs.
Diving into Mitigating Hallucinations from a Vision Perspective for Large Vision-Language Models
PDF ↗Object hallucinations in Large Vision-Language Models (LVLMs) significantly impede their real-world applicability. As the primary component for accurately interpreting visual information, the choice of visual encoder is pivotal. We hypothesize that the diverse training paradigms employed by different visual encoders instill them with distinct inductive biases, which leads to their diverse hallucination performances. Existing benchmarks typically focus on coarse-grained hallucination detection and fail to capture the diverse hallucinations elaborated in our hypothesis. To systematically analyze these effects, we introduce VHBench-10, a comprehensive benchmark for evaluating LVLMs across ten fine-grained hallucination categories. Our evaluations confirm encoders exhibit unique hallucination characteristics. Building on these insights and the suboptimality of simple feature fusion, we propose VisionWeaver, a novel Context-Aware Routing Network. It employs global visual features to generate routing signals, dynamically aggregating visual features from multiple specialized experts. Comprehensive experiments confirm the effectiveness of VisionWeaver in significantly reducing hallucinations and improving overall model performance. Our code and benchmark are available at https://github.com/whwangovo/VisionWeaver.
As large language models (LLMs) increasingly assist in subjective decision-making (e.g., moral reasoning, advice), it is critical to understand whose preferences they align with—and why. While prior work uses aggregate human judgments, demographic variation and its linguistic drivers remain underexplored. We present a comprehensive analysis of how demographic background and empathy level correlate with preferences for LLM-generated dilemma responses, alongside a systematic study of predictive linguistic features (e.g., agency, emotional tone). Our findings reveal significant demographic divides and identify markers (e.g., power verbs, tentative phrasing) that predict group-level differences. These results underscore the need for demographically informed LLM evaluation.
Large Language Models (LLMs) have shown promising results in coreference resolution, especially after fine-tuning. However, recent generative approaches face a critical issue: hallucinations—where the model generates content not present in the original input. These hallucinations make evaluation difficult and decrease overall performance. To address this issue, we analyze the underlying causes of hallucinations and propose a low-hallucination and efficient solution. Specifically, we introduce Efficient Constrained Decoding for Coreference Resolution, which maintains strong robustness while significantly improving computational efficiency. On the English OntoNotes development set, our approach achieved slightly better performance than previous state-of-the-art methods, while requiring substantially fewer parameters.
Recent advances in large language models (LLMs) have driven significant progress in end-to-end spoken dialogue models (SDMs). In contrast to text-based LLMs, the evaluation framework for SDMs should encompass both cognitive dimensions (e.g., logical reasoning, knowledge) and speech-related aspects (e.g., paralinguistic cues, audio quality). However, there is still a lack of comprehensive evaluations for SDMs in speech-to-speech (S2S) scenarios. To address this gap, we propose **URO-Bench**, an extensive benchmark for SDMs. Notably, URO-Bench is the first S2S benchmark that covers evaluations about multilingualism, multi-round dialogues, and paralinguistics. Our benchmark is divided into two difficulty levels: basic track and pro track, each comprising 20 test sets, evaluating the spoken dialogue model’s abilities in **U**nderstanding, **R**easoning, and **O**ral conversation. Evaluations on our proposed benchmark reveal that current open-source SDMs perform rather well in daily QA tasks, but lag behind their backbone LLMs in terms of instruction-following ability and also suffer from catastrophic forgetting. Their performance in advanced evaluations of paralinguistic information and audio understanding remains subpar, highlighting the need for further research in this direction. We hope that URO-Bench can facilitate the development of spoken dialogue models by providing a multifaceted evaluation of existing models and helping to track progress in this area.
Large Language Models are increasingly used in applications requiring objective assessment, which could be compromised by political bias. Many studies found preferences for left-leaning positions in LLMs, but downstream effects on tasks like fact-checking remain underexplored. In this study, we systematically investigate political bias through exchanging words with euphemisms or dysphemisms in German claims. We construct minimal pairs of factually equivalent claims that differ in political connotation, to assess the consistency of LLMs in classifying them as true or false. We evaluate six LLMs and find that, more than political leaning, the presence of judgmental words significantly influences truthfulness assessment. While a few models show tendencies of political bias, this is not mitigated by explicitly calling for objectivism in prompts. Warning: This paper contains content that may be offensive or upsetting.
Tracing Training Footprints: A Calibration Approach for Membership Inference Attacks Against Multimodal Large Language Models
PDF ↗With the increasing scale of training data for Multimodal Large Language Models (MLLMs) and the lack of data details, there is growing concern about privacy breaches and data security issues. Under black-box access, exploring effective Membership Inference Attacks (MIA) has garnered increasing attention. In real-world applications, where most samples are non-members, the issue of non-members being over-represented in the data manifold, leading to misclassification as member samples, becomes more prominent. This has motivated recent work to focus on developing effective difficulty calibration strategies, producing promising results. However, these methods only consider text-only input during calibration, and their effectiveness is diminished when migrated to MLLMs due to the presence of visual embeddings. To address the above problem, we propose PC-MMIA, focusing on visual instruction fine-tuning data. PC-MMIA is based on the idea that tokens located in poorly generalized local manifolds can better reflect traces of member samples that have been trained. By employing bidirectional perturbation of image embeddings to capture tokens critical to MIA and assigning them different weights, we achieve difficulty calibration. Experimental results demonstrate that our proposed method surpasses existing methods.
Recent advances in large language models (LLMs) enable effective in-context learning (ICL) with many-shot examples, but at the cost of high computational demand due to longer input tokens. To address this, we propose cheat-sheet ICL, which distills the information from many-shot ICL into a concise textual summary (cheat sheet) used as the context at inference time. Experiments on challenging reasoning tasks show that cheat-sheet ICL achieves comparable or better performance than many-shot ICL with far fewer tokens, and matches retrieval-based ICL without requiring test-time retrieval. These findings demonstrate that cheat-sheet ICL is a practical alternative for leveraging LLMs in downstream tasks.
Multi-turn jailbreak attacks simulate real-world human interactions by engaging large language models (LLMs) in iterative dialogues, exposing critical safety vulnerabilities. However, existing methods often struggle to balance semantic coherence with attack effectiveness, resulting in either benign semantic drift or ineffective detection evasion. To address this challenge, we propose Reasoning-Augmented Conversation (RACE), a novel multi-turn jailbreak framework that reformulates harmful queries into benign reasoning tasks and leverages LLMs’ strong reasoning capabilities to compromise safety alignment. Specifically, we introduce an attack state machine framework to systematically model problem translation and iterative reasoning, ensuring coherent query generation across multiple turns. Building on this framework, we design gain-guided exploration, self-play, and rejection feedback modules to preserve attack semantics, enhance effectiveness, and sustain reasoning-driven attack progression. Extensive experiments on multiple LLMs demonstrate that RACE achieves state-of-the-art attack effectiveness in complex conversational scenarios, with attack success rates (ASRs) increasing by up to 96%. Notably, our approach achieves average ASR of 83.3% against leading commercial models, including Gemini 2.0 Flashing Thinking and OpenAI o1, underscoring its potency.
Test-Time Scaling (TTS) has proven effective in improving the performance of Large Language Models (LLMs) during inference. However, existing research has overlooked the efficiency of TTS from a latency-sensitive perspective. Through a latency-aware evaluation of representative TTS methods, we demonstrate that a compute-optimal TTS does not always result in the lowest latency in scenarios where latency is critical. To address this gap and achieve latency-optimal TTS, we propose two key approaches by optimizing the concurrency configurations: (1) branch-wise parallelism, which leverages multiple concurrent inference branches, and (2) sequence-wise parallelism, enabled by speculative decoding. By integrating these two approaches and allocating computational resources properly to each, our latency-optimal TTS enables a 32B model to reach 82.3% accuracy on MATH-500 within 1 minute and a smaller 3B model to achieve 72.4% within 10 seconds. Our work emphasizes the importance of latency-aware TTS and demonstrates its ability to deliver both speed and accuracy in latency-sensitive scenarios.
Large language models excel at complex tasks by breaking down problems into structured reasoning steps. However, reasoning traces often extend beyond reaching a correct answer, causing wasted computation, reduced readability, and hallucinations. To address this, we introduce a novel hyperparameter-free conciseness score used as a reward signal within a reinforcement learning framework to guide models toward generating correct and concise reasoning traces. This score is evaluated by a large language model acting as a judge, enabling dynamic, context-aware feedback beyond simple token length. Our method achieves state-of-the-art efficiency–accuracy trade-offs on the MATH dataset, reducing token usage by up to 31x on simple problems while improving accuracy by 7%, and on the hardest problems, it outperforms full reasoning by +7.5% accuracy with up to 3.6x fewer tokens. On TheoremQA, our method improves accuracy by +2.2% using 12.5x fewer tokens. We also conduct ablation studies on the judge model, reward composition, and problem difficulty, showing that our method dynamically adapts reasoning length based on problem difficulty and benefits significantly from stronger judges. The code, model weights, and datasets are open-sourced at https://github.com/RazvanDu/ConciseRL.
Logical reasoning is a core capability for large language models (LLMs), yet existing benchmarks that rely solely on final-answer accuracy fail to capture the quality of the reasoning process. To address this, we introduce FineLogic, a fine-grained evaluation framework that assesses logical reasoning across three dimensions: overall accuracy, stepwise soundness, and representation-level probing. Leveraging this framework, we conduct a comprehensive study on how different supervision formats in fine-tuning shape reasoning abilities. We fine-tune LLMs on four supervision styles—one in natural language and three symbolic variants—and find a key trade-off: natural language supervision excels at generalization to out-of-distribution and long-chain problems, whereas symbolic supervision is superior at instilling structurally sound, atomic reasoning steps. Furthermore, our probing analysis indicates that fine-tuning primarily refines the model’s step-by-step generation process, rather than improving its ability to converge on an answer early. Together, our framework and analysis provide a more rigorous lens for evaluating and improving logical reasoning in LLMs. The code is available at https://github.com/YujunZhou/FineLogic.