Yexin Li

Yuanhao Zeng, Ao Lu, Lufei Li et al.

Generating diverse responses is crucial for test-time scaling of large language models (LLMs), yet standard stochastic sampling mostly yields surface-level lexical variation, limiting semantic exploration. In this paper, we propose Exploratory Sampling (ESamp), a decoding approach that explicitly encourages semantic diversity during generation. ESamp is motivated by the well-known observation that neural networks tend to make lower-error predictions on inputs similar to those encountered before, and incur higher prediction error on novel ones. Building on this property, we train a lightweight Distiller at test time to predict deep-layer hidden representations of the LLM from its shallow-layer representations to model the LLM's depth-wise representation transitions. During decoding, the Distiller continuously adapts to the mappings induced by the current generation context. ESamp uses the prediction error as a novelty signal to reweight candidate token extensions conditioned on the current prefix, thereby biasing decoding toward less-explored semantic patterns. ESamp is implemented with an asynchronous training--inference pipeline, with less than 5% worst case overhead (1.2% in the optimized release). Empirical results show that ESamp significantly boosts the Pass@k efficiency of reasoning models, showing superior or comparable performance to strong stochastic and heuristic baselines. Notably, ESamp achieves robust generalization across mathematics, science, and code generation benchmarks and breaks the trade-off between diversity and coherence in creative writing. Our code has released at: https://github.com/LinesHogan/tLLM.

Siyuan Qi, Shuo Chen, Yexin Li et al.

The generalization of decision-making agents encompasses two fundamental elements: learning from past experiences and reasoning in novel contexts. However, the predominant emphasis in most interactive environments is on learning, often at the expense of complexity in reasoning. In this paper, we introduce CivRealm, an environment inspired by the Civilization game. Civilization's profound alignment with human history and society necessitates sophisticated learning, while its ever-changing situations demand strong reasoning to generalize. Particularly, CivRealm sets up an imperfect-information general-sum game with a changing number of players; it presents a plethora of complex features, challenging the agent to deal with open-ended stochastic environments that require diplomacy and negotiation skills. Within CivRealm, we provide interfaces for two typical agent types: tensor-based agents that focus on learning, and language-based agents that emphasize reasoning. To catalyze further research, we present initial results for both paradigms. The canonical RL-based agents exhibit reasonable performance in mini-games, whereas both RL- and LLM-based agents struggle to make substantial progress in the full game. Overall, CivRealm stands as a unique learning and reasoning challenge for decision-making agents. The code is available at https://github.com/bigai-ai/civrealm.

Honglei Guo, Yuhan Zhao, Yexin Li

Reinforcement learning has increasingly been used to study economic decision-making, such as taxation, public spending, and labour supply. However, most existing RL-based economic models focus on a single government--household group, thereby overlooking the strategic interactions that arise when multiple governments compete while managing their own populations. In practice, many economic systems (e.g., taxation) exhibit a multi-group structure, where each government must optimize its fiscal policy in response not only to household behaviour within its jurisdiction, but also to the policies of other competing governments. To capture this structure, we formulate taxation as a hierarchical multi-group game. Within each group, the interaction between the government and households is modelled as a leader--follower game; across groups, governments are modelled as players in a competitive game. This results in a hybrid hierarchical game that is difficult to solve using standard multi-agent reinforcement learning algorithms. We therefore propose a bi-level training framework built on multi-agent reinforcement learning, together with \textit{ Curriculum Learning} and a \textit{ Closed-Loop Sequential Update} strategy, to stabilize training and promote convergence. We instantiate this framework in a taxation game simulation environment grounded in classical economic models. The environment supports the evaluation of different taxation algorithms and provides multiple economic indicators for assessing policy performance. Experiments show that our approach can learn stable tax policies that benefit all participating groups. Compared with a two-group baseline without the proposed update mechanisms, our method avoids premature game collapse, extends the effective game duration by 60.92\%, produces more sustainable and robust tax policies, and reduces GDP disparities among governments by 44.12\%.

Zheng Zhang, Ao Lu, Yuanhao Zeng et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has catalyzed significant breakthroughs in complex LLM reasoning within verifiable domains, such as mathematics and programming. Recent efforts have sought to extend this paradigm to open-ended tasks by employing LLMs-as-a-Judge to provide sequence-level rewards for policy optimization. However, these rewards are inherently sparse, failing to provide the fine-grained supervision necessary for generating complex, long-form trajectories. Furthermore, current work treats the Judge as a black-box oracle, discarding the rich intermediate feedback signals encoded in it. To address these limitations, we introduce Grad2Reward, a novel framework that extracts dense process rewards directly from the Judge's model inference process via a single backward pass. By leveraging gradient-based attribution, Grad2Reward enables precise token-level credit assignment, substantially enhancing training efficiency and reasoning quality. Additionally, Grad2Reward introduces a self-judging mechanism, allowing the policy to improve through its own evaluative signals without training specialized reward models or reliance on superior external Judges. The experiments demonstrate that policies optimized with Grad2Reward achieve outstanding performance across diverse open-ended tasks, affirming its effectiveness and broad generalizability.

Yexin Li, Pring Wong, Hanfang Zhang et al.

Exploration remains a critical challenge in reinforcement learning, as many existing methods either lack theoretical guarantees or fall short of practical effectiveness. In this paper, we introduce CAE, a lightweight algorithm that repurposes the value networks in standard deep RL algorithms to drive exploration without introducing additional parameters. CAE utilizes any linear multi-armed bandit technique and incorporates an appropriate scaling strategy, enabling efficient exploration with provable sub-linear regret bounds and practical stability. Notably, it is simple to implement, requiring only around 10 lines of code. In complex tasks where learning an effective value network proves challenging, we propose CAE+, an extension of CAE that incorporates an auxiliary network. This extension increases the parameter count by less than 1% while maintaining implementation simplicity, adding only about 10 additional lines of code. Experiments on MuJoCo and MiniHack show that both CAE and CAE+ outperform state-of-the-art baselines, bridging the gap between theoretical rigor and practical efficiency.

Yipeng Kang, Junqi Wang, Yexin Li et al.

As large language models (LLMs) become increasingly integrated into critical applications, aligning their behavior with human values presents significant challenges. Current methods, such as Reinforcement Learning from Human Feedback (RLHF), typically focus on a limited set of coarse-grained values and are resource-intensive. Moreover, the correlations between these values remain implicit, leading to unclear explanations for value-steering outcomes. Our work argues that a latent causal value graph underlies the value dimensions of LLMs and that, despite alignment training, this structure remains significantly different from human value systems. We leverage these causal value graphs to guide two lightweight value-steering methods: role-based prompting and sparse autoencoder (SAE) steering, effectively mitigating unexpected side effects. Furthermore, SAE provides a more fine-grained approach to value steering. Experiments on Gemma-2B-IT and Llama3-8B-IT demonstrate the effectiveness and controllability of our methods.

Jinjin Guo, Yexin Li, Zhichao Huang et al.

Large-scale multimodal contrastive learning has recently achieved impressive success in learning rich and transferable representations, yet it remains fundamentally limited by the uniform treatment of feature dimensions and the neglect of the intrinsic spectral structure of the learned features. Empirical evidence indicates that high-dimensional embeddings tend to collapse into narrow cones, concentrating task-relevant semantics in a small subspace, while the majority of dimensions remain occupied by noise and spurious correlations. Such spectral imbalance and entanglement undermine model generalization. We propose Spectral Disentanglement and Enhancement (SDE), a novel framework that bridges the gap between the geometry of the embedded spaces and their spectral properties. Our approach leverages singular value decomposition to adaptively partition feature dimensions into strong signals that capture task-critical semantics, weak signals that reflect ancillary correlations, and noise representing irrelevant perturbations. A curriculum-based spectral enhancement strategy is then applied, selectively amplifying informative components with theoretical guarantees on training stability. Building upon the enhanced features, we further introduce a dual-domain contrastive loss that jointly optimizes alignment in both the feature and spectral spaces, effectively integrating spectral regularization into the training process and encouraging richer, more robust representations. Extensive experiments on large-scale multimodal benchmarks demonstrate that SDE consistently improves representation robustness and generalization, outperforming state-of-the-art methods. SDE integrates seamlessly with existing contrastive pipelines, offering an effective solution for multimodal representation learning.

Zheng Zhang, Ziwei Shan, Kaitao Song et al.

Process Reward Models (PRMs) have emerged as a promising approach to enhance the reasoning capabilities of large language models (LLMs) by guiding their step-by-step reasoning toward a final answer. However, existing PRMs either treat each reasoning step in isolation, failing to capture inter-step dependencies, or struggle to align process rewards with the final outcome. Consequently, the reward signal fails to respect temporal causality in sequential reasoning and faces ambiguous credit assignment. These limitations make downstream models vulnerable to reward hacking and lead to suboptimal performance. In this work, we propose Conditional Reward Modeling (CRM) that frames LLM reasoning as a temporal process leading to a correct answer. The reward of each reasoning step is not only conditioned on the preceding steps but also explicitly linked to the final outcome of the reasoning trajectory. By enforcing conditional probability rules, our design captures the causal relationships among reasoning steps, with the link to the outcome allowing precise attribution of each intermediate step, thereby resolving credit assignment ambiguity. Further, through this consistent probabilistic modeling, the rewards produced by CRM enable more reliable cross-sample comparison. Experiments across Best-of-N sampling, beam search and reinforcement learning demonstrate that CRM consistently outperforms existing reward models, offering a principled framework for enhancing LLM reasoning. In particular, CRM is more robust to reward hacking and delivers stable downstream improvements without relying on verifiable rewards derived from ground truth.

Jiulin Li, Ping Huang, Yexin Li et al.

Real-world multimodal applications often require any-to-any capabilities, enabling both understanding and generation across modalities including text, image, audio, and video. However, integrating the strengths of autoregressive language models (LLMs) for reasoning and diffusion models for high-fidelity generation remains challenging. Existing approaches rely on rigid pipelines or tightly coupled architectures, limiting flexibility and scalability. We propose MAGUS (Multi-Agent Guided Unified Multimodal System), a modular framework that unifies multimodal understanding and generation via two decoupled phases: Cognition and Deliberation. MAGUS enables symbolic multi-agent collaboration within a shared textual workspace. In the Cognition phase, three role-conditioned multimodal LLM agents - Perceiver, Planner, and Reflector - engage in collaborative dialogue to perform structured understanding and planning. The Deliberation phase incorporates a Growth-Aware Search mechanism that orchestrates LLM-based reasoning and diffusion-based generation in a mutually reinforcing manner. MAGUS supports plug-and-play extensibility, scalable any-to-any modality conversion, and semantic alignment - all without the need for joint training. Experiments across multiple benchmarks, including image, video, and audio generation, as well as cross-modal instruction following, demonstrate that MAGUS outperforms strong baselines and state-of-the-art systems. Notably, on the MME benchmark, MAGUS surpasses the powerful closed-source model GPT-4o.

Juewen Hu, Yexin Li, Jiulin Li et al.

Emotion recognition plays a vital role in enhancing human-computer interaction. In this study, we tackle the MER-SEMI challenge of the MER2025 competition by proposing a novel multimodal emotion recognition framework. To address the issue of data scarcity, we leverage large-scale pre-trained models to extract informative features from visual, audio, and textual modalities. Specifically, for the visual modality, we design a dual-branch visual encoder that captures both global frame-level features and localized facial representations. For the textual modality, we introduce a context-enriched method that employs large language models to enrich emotional cues within the input text. To effectively integrate these multimodal features, we propose a fusion strategy comprising two key components, i.e., self-attention mechanisms for dynamic modality weighting, and residual connections to preserve original representations. Beyond architectural design, we further refine noisy labels in the training set by a multi-source labeling strategy. Our approach achieves a substantial performance improvement over the official baseline on the MER2025-SEMI dataset, attaining a weighted F-score of 87.49% compared to 78.63%, thereby validating the effectiveness of the proposed framework.

Haoyuan Gao, Zicong Zhang, Yuqi Wei et al.

Vision-Language Models (VLMs) represent a breakthrough in artificial intelligence by integrating visual and textual modalities to achieve impressive zero-shot capabilities. However, VLMs are susceptible to catastrophic forgetting when sequentially fine-tuned on multiple downstream tasks. Existing continual learning methods for VLMs often rely heavily on additional reference datasets, compromise zero-shot performance, or are limited to parameter-efficient fine-tuning scenarios. In this paper, we propose Continual Decoupling-Unifying (ConDU), a novel approach, by introducing model fusion into continual learning for VLMs. ConDU maintains a unified model along with task triggers and prototype sets, employing an iterative process of decoupling task-specific models for previous tasks and unifying them with the model for the newly learned task. Additionally, we introduce an inference strategy for zero-shot scenarios by aggregating predictions from multiple decoupled task-specific models. Extensive experiments across various settings show that ConDU achieves up to a 2\% improvement in average performance across all seen tasks compared to state-of-the-art baselines, while also enhancing zero-shot capabilities relative to the original VLM.

Yexin Li, Zhancun Mu, Siyuan Qi

Learning effective negotiation strategies poses two key challenges: the exploration-exploitation dilemma and dealing with large action spaces. However, there is an absence of learning-based approaches that effectively address these challenges in negotiation. This paper introduces a comprehensive formulation to tackle various negotiation problems. Our approach leverages contextual combinatorial multi-armed bandits, with the bandits resolving the exploration-exploitation dilemma, and the combinatorial nature handles large action spaces. Building upon this formulation, we introduce NegUCB, a novel method that also handles common issues such as partial observations and complex reward functions in negotiation. NegUCB is contextual and tailored for full-bandit feedback without constraints on the reward functions. Under mild assumptions, it ensures a sub-linear regret upper bound. Experiments conducted on three negotiation tasks demonstrate the superiority of our approach.

Tianfu He, Jie Bao, Yexin Li et al.

Illegal vehicle parking is a common urban problem faced by major cities in the world, as it incurs traffic jams, which lead to air pollution and traffic accidents. The government highly relies on active human efforts to detect illegal parking events. However, such an approach is extremely ineffective to cover a large city since the police have to patrol over the entire city roads. The massive and high-quality sharing bike trajectories from Mobike offer us a unique opportunity to design a ubiquitous illegal parking detection approach, as most of the illegal parking events happen at curbsides and have significant impact on the bike users. The detection result can guide the patrol schedule, i.e. send the patrol policemen to the region with higher illegal parking risks, and further improve the patrol efficiency. Inspired by this idea, three main components are employed in the proposed framework: 1)~{\em trajectory pre-processing}, which filters outlier GPS points, performs map-matching, and builds trajectory indexes; 2)~{\em illegal parking detection}, which models the normal trajectories, extracts features from the evaluation trajectories, and utilizes a distribution test-based method to discover the illegal parking events; and 3)~{\em patrol scheduling}, which leverages the detection result as reference context, and models the scheduling task as a multi-agent reinforcement learning problem to guide the patrol police. Finally, extensive experiments are presented to validate the effectiveness of illegal parking detection, as well as the improvement of patrol efficiency.

Tianfu He, Guochun Chen, Chuishi Meng et al.

People often refer to a place of interest (POI) by an alias. In e-commerce scenarios, the POI alias problem affects the quality of the delivery address of online orders, bringing substantial challenges to intelligent logistics systems and market decision-making. Labeling the aliases of POIs involves heavy human labor, which is inefficient and expensive. Inspired by the observation that the users' GPS locations are highly related to their delivery address, we propose a ubiquitous alias discovery framework. Firstly, for each POI name in delivery addresses, the location data of its associated users, namely Mobility Profile are extracted. Then, we identify the alias relationship by modeling the similarity of mobility profiles. Comprehensive experiments on the large-scale location data and delivery address data from JD logistics validate the effectiveness.