NeurIPS 2025 Papers

Large Language Models (LLMs) have demonstrated notable capabilities across financial tasks, including financial report summarization, earnings call transcript analysis, and asset classification. However, their real-world effectiveness in managing complex fund investment remains inadequately assessed. A fundamental limitation of existing benchmarks for evaluating LLM-driven trading strategies is their reliance on historical back-testing, inadvertently enabling LLMs to "time travel"—leveraging future information embedded in their training corpora, thus resulting in possible information leakage and overly optimistic performance estimates. To address this issue, we introduce DeepFund, a live fund benchmark tool designed to rigorously evaluate LLM in real-time market conditions. Utilizing a multi-agent architecture, DeepFund connects directly with real-time stock market data—specifically data published after each model’s pretraining cutoff—to ensure fair and leakage-free evaluations. Empirical tests on nine flagship LLMs from leading global institutions across multiple investment dimensions—including ticker-level analysis, investment decision-making, portfolio management, and risk control—reveal significant practical challenges. Notably, even cutting-edge models such as DeepSeek-V3 and Claude-3.7-Sonnet incur net trading losses within DeepFund real-time evaluation environment, underscoring the present limitations of LLMs for active fund management. Our code is available at https://github.com/HKUSTDial/DeepFund.

The success of Graph Neural Networks (GNNs) leverages the homophily principle, where connected nodes share similar features and labels. However, this assumption breaks down in heterophilic graphs, where same-class nodes are often distributed across distant neighborhoods rather than immediate connections. Recent attempts expand the receptive field through multi-hop aggregation schemes that explicitly preserve intermediate representations from each hop distance. While effective at capturing heterophilic patterns, these methods require separate weight matrices per hop and feature concatenation, causing parameters to scale linearly with hop count. This leads to high computational complexity and GPU memory consumption. We propose Gated Multi-hop Message Passing (GAMMA), where nodes assess how relevant the aggregated information is from their k-hop neighbors. This assessment occurs through multiple refinement steps where the node compares each hop's embedding with its current representation, allowing it to focus on the most informative hops. During the forward pass, GAMMA finds the optimal mix of multi-hop information local to each node using a single feature vector without needing separate representations for each hop, thereby maintaining dimensionality comparable to single hop GNNs. In addition, we propose a weight sharing scheme that leverages a unified transformation for aggregated features from multiple hops so the global heterophilic patterns specific to each hop are learned during training. As such, GAMMA captures both global (per-hop) and local (per-node) heterophily patterns without high computation and memory overhead. Experiments show GAMMA matches or exceeds state-of-the-art heterophilic GNN accuracy, achieving up to $\approx20\times$ faster inference. Our code is publicly available at \url{https://github.com/amir-ghz/GAMMA}.

Large language models (LLMs) are inherently designed to support multi-turn interactions, which opens up new possibilities for jailbreak attacks that unfold gradually and potentially bypass safety mechanisms more effectively than single-turn attacks. However, current multi-turn jailbreak methods are still in their early stages and suffer from two key limitations. First, they all inherently require inserting sensitive phrases into the context, which makes the dialogue appear suspicious and increases the likelihood of rejection, undermining the effectiveness of the attack. Second, even when harmful content is generated, the response often fails to align with the malicious prompt due to semantic drift, where the conversation slowly moves away from its intended goal. To address these challenges, we propose an analogy-based black-box multi-turn jailbreak framework that constructs fully benign contexts to improve attack success rate while ensuring semantic alignment with the malicious intent. The method first guides the model through safe tasks that mirror the response structure of the malicious prompt, enabling it to internalize the format without exposure to sensitive content. A controlled semantic shift is then introduced in the final turn, substituting benign elements with malicious ones while preserving structural coherence. Experiments on six commercial and open-source LLMs, two benchmark datasets show that our method significantly improves attack performance, achieving an average attack success rate of 93.3\% and outperforming five competitive baselines. Our code is released at https://anonymous.4open.science/r/AMA-E1C4

Large Language Model (LLM) agents are capable of task execution across various domains by autonomously interacting with environments and refining LLM responses based on feedback. However, existing model serving systems are not optimized for the unique demands of serving agents. Compared to classic model serving, agent serving has different characteristics: predictable request pattern, increasing quality requirement, and unique prompt formatting. We identify a key problem for agent serving: LLM serving systems lack session-awareness. They neither perform effective KV cache management nor precisely select the cheapest yet competent model in each round. This leads to a cost-quality tradeoff, and we identify an opportunity to surpass it in an agent serving system. To this end, we introduce AgServe for AGile AGent SERVing. AgServe features a session-aware server that boosts KV cache reuse via Estimated-Time-of-Arrival-based eviction and in-place positional embedding calibration, a quality-aware client that performs session-aware model cascading through real-time quality assessment, and a dynamic resource scheduler that maximizes GPU utilization. With AgServe, we allow agents to select and upgrade models during the session lifetime, and to achieve similar quality at much lower costs, effectively transcending the tradeoff. Extensive experiments on real testbeds demonstrate that AgServe (1) achieves comparable response quality to GPT-4o at a 16.5\% cost. (2) delivers 1.8$\times$ improvement in quality relative to the tradeoff curve.

Learning from multi-agent expert demonstrations, known as Multi-Agent Imitation Learning (MAIL), provides a promising approach to sequential decision-making. However, existing MAIL methods including Behavior Cloning (BC) and Adversarial Imitation Learning (AIL) face significant challenges: BC suffers from the compounding error issue, while the very nature of adversarial optimization makes AIL prone to instability. In this work, we propose \textbf{M}ulti-\textbf{A}gent imitation by learning and sampling from \textbf{F}actor\textbf{I}zed \textbf{S}oft Q-function (MAFIS), a novel method that addresses these limitations for both online and offline MAIL settings. Built upon the single-agent IQ-Learn framework, MAFIS introduces the value decomposition network to factorize the imitation objective at agent level, thus enabling scalable training for multi-agent systems. Moreover, we observe that the soft Q-function implicitly defines the optimal policy as an energy-based model, from which we can sample actions via stochastic gradient Langevin dynamics. This allows us to estimate the gradient of the factorized optimization objective for continuous control tasks, avoiding the adversarial optimization between the soft Q-function and the policy required by prior work. By doing so, we obtain a tractable and \emph{non-adversarial} objective for both discrete and continuous multi-agent control. Experiments on common benchmarks including the discrete control tasks StarCraft Multi-Agent Challenge v2 (SMACv2), Gold Miner, and Multi Particle Environments (MPE), as well as the continuous control task Multi-Agent MuJoCo (MaMuJoCo), demonstrate that MAFIS achieves superior performance compared with baselines. Our code is available at https://github.com/LAMDA-RL/MAFIS.

We study the online multi-class selection problem with group fairness guarantees, where limited resources must be allocated to sequentially arriving agents. Our work addresses two key limitations in the existing literature. First, we introduce a novel lossless rounding scheme that ensures the integral algorithm achieves the same expected performance as any fractional solution. Second, we explicitly address the challenges introduced by agents who belong to multiple classes. To this end, we develop a randomized algorithm based on a relax-and-round framework. The algorithm first computes a fractional solution using a resource reservation approach---referred to as the *set-aside* mechanism---to enforce fairness across classes. The subsequent rounding step preserves these fairness guarantees without degrading performance. Additionally, we propose a learning-augmented variant that incorporates untrusted machine-learned predictions to better balance fairness and efficiency in practical settings.

Speculative decoding is an effective and lossless method for Large Language Model (LLM) inference acceleration. It employs a smaller model to generate a draft token sequence, which is then verified by the original base model. In multi-GPU systems, inference latency can be further reduced through tensor parallelism (TP), while the optimal TP size of the draft model is typically smaller than that of the base model, leading to GPU idling during the drafting stage. We observe that such inefficiency stems from the sequential execution of layers, which is seemingly natural but actually unnecessary. Therefore, we propose EasySpec, a layer-parallel speculation strategy that optimizes the efficiency of multi-GPU utilization. EasySpec breaks the inter-layer data dependencies in the draft model, enabling multiple layers to run simultaneously across multiple devices as ``fuzzy'' speculation. After each drafting-and-verification iteration, the draft model’s key-value cache is calibrated in a single forward pass, preventing long-term fuzzy-error accumulation at minimal additional latency. EasySpec is a training-free and plug-in method. We evaluated EasySpec on several mainstream open-source LLMs, using smaller versions of models from the same series as drafters. The results demonstrate that EasySpec can achieve a peak speedup of 4.17x compared to vanilla decoding, while preserving the original distributions of the base LLMs. Specifically, the drafting stage can be accelerated by up to 1.62x with a maximum speculation accuracy drop of only 7\%. The code is available at https://github.com/Yize-Wu/EasySpec.

Emergency traffic control presents an increasingly critical challenge, requiring seamless coordination among emergency vehicles, regular vehicles, and traffic lights to ensure efficient passage for all vehicles. Existing models primarily only focus on traffic light control, leaving emergency and regular vehicles prone to delay due to the lack of navigation strategies. To address this issue, we propose the ***R*ole-aware *M*ulti-agent *T*raffic *C*ontrol (RMTC)** framework, which dynamically assigns appropriate roles to traffic components for better cooperation by considering their relations with emergency vehicles and adaptively adjusting their policies. Specifically, RMTC introduces a *Heterogeneous Temporal Traffic Graph (HTTG)* to model the spatial and temporal relationships among all traffic components (traffic lights, regular and emergency vehicles) at each time step. Furthermore, we develop a *Dynamic Role Learning* model to infer the evolving roles of traffic lights and regular vehicles based on HTTG. Finally, we present a *Role-aware Multi-agent Reinforcement Learning* approach that learns traffic policies conditioned on the dynamically roles. Extensive experiments across four public traffic scenarios show that RMTC outperforms existing traffic light control methods by significantly reducing emergency vehicle travel time, while effectively preserving traffic efficiency for regular vehicles. The code is released at [https://anonymous.4open.science/r/RMTC-5E28](https://anonymous.4open.science/r/RMTC-5E28).

Recent efforts to employ large language models (LLMs) as agents have demonstrated promising results in a wide range of multi-step agent tasks. However, existing agents lack an effective experience reuse approach to leverage historical completed tasks. In this paper, we propose a novel experience reuse framework MetaFlowLLM, which constructs a hierarchical experience tree from historically completed tasks. Each node in this experience tree is presented as a MetaFlow which contains static execution workflow and subtask required by agents to complete dynamically. Then, we propose a Hierarchical MetaFlow Merging algorithm to construct the hierarchical experience tree. When accomplishing a new task, MetaFlowLLM can first retrieve the most relevant MetaFlow node from the experience tree and then execute it accordingly. To effectively generate valid MetaFlows from historical data, we further propose a reinforcement learning pipeline to train the MetaFlowGen. Extensive experimental results on AppWorld and WorkBench demonstrate that integrating with MetaFlowLLM, existing agents (e.g., ReAct, Reflexion) can gain substantial performance improvement with reducing execution costs. Notably, MetaFlowLLM achieves an average success rate improvement of 32.3% on AppWorld and 6.2% on WorkBench, respectively.

In real-word scenarios, person re-identification (ReID) expects to identify a person-of-interest via the descriptive query, regardless of whether the query is a single modality or a combination of multiple modalities. However, existing methods and datasets remain constrained to limited modalities, failing to meet this requirement. Therefore, we investigate a new challenging problem called Omni Multi-modal Person Re-identification (OM-ReID), which aims to achieve effective retrieval with varying multi-modal queries. To address dataset scarcity, we construct ORBench, the first high-quality multi-modal dataset comprising 1,000 unique identities across five modalities: RGB, infrared, color pencil, sketch, and textual description. This dataset also has significant superiority in terms of diversity, such as the painting perspectives and textual information. It could serve as an ideal platform for follow-up investigations in OM-ReID. Moreover, we propose ReID5o, a novel multi-modal learning framework for person ReID. It enables synergistic fusion and cross-modal alignment of arbitrary modality combinations in a single model, with a unified encoding and multi-expert routing mechanism proposed. Extensive experiments verify the advancement and practicality of our ORBench. A range of models have been compared on it, and our proposed ReID5o gives the best performance.

Recent advances in group-based reinforcement learning (RL) have driven frontier large language models (LLMs) in single-turn tasks like mathematical reasoning. However, their scalability to multi-turn LLM agent training remains limited. Unlike static tasks, agent-environment interactions unfold over many steps and often yield sparse or delayed rewards, making credit assignment across individual steps significantly more challenging. In this work, we propose Group-in-Group Policy Optimization (GiGPO), a novel RL algorithm that achieves fine-grained credit assignment for LLM agents while preserving the appealing properties of group-based RL: critic-free, low memory, and stable convergence. GiGPO introduces a two-level structure for estimating relative advantage: (i) At the episode-level, GiGPO computes macro relative advantages based on groups of complete trajectories; (ii) At the step-level, GiGPO introduces an anchor state grouping mechanism that retroactively constructs step-level groups by identifying repeated environment states across trajectories. Actions stemming from the same state are grouped together, enabling micro relative advantage estimation. This hierarchical structure effectively captures both global trajectory quality and local step effectiveness without relying on auxiliary models or additional rollouts. We evaluate GiGPO on challenging agent benchmarks, including ALFWorld and WebShop, as well as tool-integrated reasoning on search-augmented QA tasks, using Qwen2.5-1.5B/3B/7B-Instruct. Crucially, GiGPO delivers fine-grained per-step credit signals, achieves performance gains of > 12\% on ALFWorld and > 9\% on WebShop over GRPO, and obtains superior performance on QA tasks (42.1\% on 3B and 47.2\% on 7B): all while maintaining the same GPU memory overhead, identical LLM rollout, and incurring little to no additional time cost.

This paper provides the first expert sample complexity characterization for learning a Nash equilibrium from expert data in Markov Games. We show that a new quantity named the *single policy deviation concentrability coefficient* is unavoidable in the non-interactive imitation learning setting, and we provide an upper bound for behavioral cloning (BC) featuring such coefficient. BC exhibits substantial regret in games with high concentrability coefficient, leading us to utilize expert queries to develop and introduce two novel solution algorithms: MAIL-BRO and MURMAIL. The former employs a best response oracle and learns an $\varepsilon$-Nash equilibrium with $\mathcal{O}(\varepsilon^{-4})$ expert and oracle queries. The latter bypasses completely the best response oracle at the cost of a worse expert query complexity of order $\mathcal{O}(\varepsilon^{-8})$. Finally, we provide numerical evidence, confirming our theoretical findings.

Despite their outstanding performance in numerous applications, large language models (LLMs) remain prone to hallucinations, generating content inconsistent with their pretraining corpora. Currently, almost all contrastive decoding approaches alleviate hallucinations by introducing a model susceptible to hallucinations and appropriately widening the contrastive logits gap between hallucinatory tokens and target tokens. However, although existing contrastive decoding methods mitigate hallucinations, they lack enough confidence in the factual accuracy of the generated content. In this work, we propose Multi-Model Contrastive Decoding (MCD), which integrates a pretrained language model with an evil model and a truthful model for contrastive decoding. Intuitively, a token is assigned a high probability only when deemed potentially hallucinatory by the evil model while being considered factual by the truthful model. This decoding strategy significantly enhances the model’s confidence in its generated responses and reduces potential hallucinations. Furthermore, we introduce a dynamic hallucination detection mechanism that facilitates token-by-token identification of hallucinations during generation and a tree-based revision mechanism to diminish hallucinations further. Extensive experimental evaluations demonstrate that our MCD strategy effectively reduces hallucinations in LLMs and outperforms state-of-the-art methods across various benchmarks.

Automated Theorem Proving (ATP) in formal languages remains a formidable challenge in AI, demanding rigorous logical deduction and navigating vast search spaces. While large language models (LLMs) have shown promising performance, existing stepwise provers often suffer from biased search guidance, leading to inefficiencies and suboptimal proof strategies. This paper introduces the Multi-Perspective Search Prover (MPS-Prover), a novel stepwise ATP system designed to overcome these limitations. MPS-Prover incorporates two key innovations: a highly effective post-training data curation strategy that prunes approximately 40\% of redundant training data without sacrificing performance, and a multi-perspective tree search mechanism. This search integrates a learned critic model with strategically designed heuristic rules to diversify tactic selection, prevent getting trapped in unproductive states, and enhance search robustness. Extensive evaluations demonstrate that MPS-Prover achieves state-of-the-art performance on multiple challenging benchmarks, including miniF2F and ProofNet, outperforming prior 7B parameter models. Furthermore, our analyses reveal that MPS-Prover generates significantly shorter and more diverse proofs compared to existing stepwise and whole-proof methods, highlighting its efficiency and efficacy. Our work advances the capabilities of LLM-based formal reasoning and offers a robust framework and a comprehensive analysis for developing more powerful theorem provers.

Multi-modal learning has emerged as a key technique for improving performance across domains such as autonomous driving, robotics, and reasoning. However, in certain scenarios, particularly in resource-constrained environments, some modalities available during training may be absent during inference. While existing frameworks effectively utilize multiple data sources during training and enable inference with reduced modalities, they are primarily designed for single-agent settings. This poses a critical limitation in dynamic environments such as connected autonomous vehicles (CAV), where incomplete data coverage can lead to decision-making blind spots. Conversely, some works explore multi-agent collaboration but without addressing missing modality at test time. To overcome these limitations, we propose Collaborative Auxiliary Modality Learning (CAML), a novel multi-modal multi-agent framework that enables agents to collaborate and share multi-modal data during training, while allowing inference with reduced modalities during testing. Experimental results in collaborative decision-making for CAV in accident-prone scenarios demonstrate that CAML achieves up to a 58.1% improvement in accident detection. Additionally, we validate CAML on real-world aerial-ground robot data for collaborative semantic segmentation, achieving up to a 10.6% improvement in mIoU.

We present new fast-rate PAC-Bayesian generalization bounds for multi-task and meta-learning in the unbalanced setting, i.e. when the tasks have training sets of different sizes, as is typically the case in real-world scenarios. Previously, only standard-rate bounds were known for this situation, while fast-rate bounds were limited to the setting where all training sets are of equal size. Our new bounds are numerically computable as well as interpretable, and we demonstrate their flexibility in handling a number of cases where they give stronger guarantees than previous bounds. Besides the bounds themselves, we also make conceptual contributions: we demonstrate that the unbalanced multi-task setting has different statistical properties than the balanced situation, specifically that proofs from the balanced situation do not carry over to the unbalanced setting. Additionally, we shed light on the fact that the unbalanced situation allows two meaningful definitions of multi-task risk, depending on whether all tasks should be considered equally important or if sample-rich tasks should receive more weight than sample-poor ones.

Multi-modal large language models (MLLMs) have emerged as a transformative approach for aligning visual and textual understanding. They typically require extremely high computational resources (e.g., thousands of GPUs) for training to achieve cross-modal alignment at multi-granularity levels. We argue that a key source of this inefficiency lies in the vision encoders they widely equip with, e.g., CLIP and SAM, which lack the alignment with language at multi-granularity levels. To address this issue, in this paper, we leverage hyperbolic space, which inherently models hierarchical levels and thus provides a principled framework for bridging the granularity gap between visual and textual modalities at an arbitrary granularity level. Concretely, we propose an efficient training paradigm for MLLMs, dubbed as \blg, which can optimize visual representations to align with their textual counterparts at an arbitrary granularity level through dynamic hyperbolic radius adjustment in hyperbolic space. \alg employs learnable matrices with M\"{o}bius multiplication operations, implemented via three effective configurations: diagonal scaling matrices, block-diagonal matrices, and banded matrices, providing a flexible yet efficient parametrization strategy. Comprehensive experiments across multiple MLLM benchmarks demonstrate that \alg consistently improves both existing pre-training and fine-tuning MLLMs clearly with less than 1\% additional parameters.

Multi-frame video enhancement tasks aim to improve the spatial and temporal resolution and quality of video sequences by leveraging temporal information from multiple frames, which are widely used in streaming video processing, surveillance, and generation. Although numerous Transformer-based enhancement methods have achieved impressive performance, their computational and memory demands hinder deployment on edge devices. Quantization offers a practical solution by reducing the bit-width of weights and activations to improve efficiency. However, directly applying existing quantization methods to video enhancement tasks often leads to significant performance degradation and loss of fine details. This stems from two limitations: (a) inability to allocate varying representational capacity across frames, which results in suboptimal dynamic range adaptation; (b) over-reliance on full-precision teachers, which limits the learning of low-bit student models. To tackle these challenges, we propose a novel quantization method for video enhancement: Progressive Multi-Frame Quantization for Video Enhancement (PMQ-VE). This framework features a coarse-to-fine two-stage process: Backtracking-based Multi-Frame Quantization (BMFQ) and Progressive Multi-Teacher Distillation (PMTD). BMFQ utilizes a percentile-based initialization and iterative search with pruning and backtracking for robust clipping bounds. PMTD employs a progressive distillation strategy with both full-precision and multiple high-bit (INT) teachers to enhance low-bit models' capacity and quality. Extensive experiments demonstrate that our method outperforms existing approaches, achieving state-of-the-art performance across multiple tasks and benchmarks. The code will be made publicly available.

LLM-based agent systems are emerging as a new software paradigm and have been widely adopted across diverse domains such as medicine, robotics, and programming. However, maintaining these systems requires substantial effort, as they are inevitably prone to bugs and continually evolve to meet changing external requirements. Therefore, automatically resolving agent issues (i.e.,bug reports or feature requests) is a crucial and challenging task. While recent software engineering (SE) agents (e.g., SWE-agent) have shown promise in addressing issues in traditional software systems, it remains unclear how effectively they can resolve real-world issues in agent systems, which differ significantly from traditional software. To fill this gap, we first manually analyze 201 real-world agent issues and identify common categories of agent issues. We then spend 500 person-hours constructing AgentIssue-bench, a reproducible benchmark comprising 50 agent issue resolution tasks (each with an executable environment and failure-triggering tests). We further evaluate state-of-the-art SE agents on AgentIssue-bench and reveal their limited effectiveness (\.e., with only 0.67% - 4.67% resolution rates). These results underscore the unique challenges of maintaining agent systems compared to traditional software, highlighting the need for further research to develop advanced SE agents for resolving agent issues.