Group Relative Policy Optimization (GRPO) enhances LLM reasoning but often induces overconfidence, where incorrect responses yield lower perplexity than correct ones, degrading relative calibration as described by the Area Under the Curve (AUC).Existing approaches either yield limited improvements in calibration or sacrifice gains in reasoning accuracy.We first prove that this degradation in GRPO-style algorithms stems from their uncertainty-agnostic advantage estimation, which inevitably misaligns optimization gradients with calibration.This leads to improved accuracy at the expense of degraded calibration.We then propose Calibration-Aware Policy Optimization (CAPO).It adopts a logistic AUC surrogate loss that is theoretically consistent and admits regret bound, enabling uncertainty-aware advantage estimation.By further incorporating a noise masking mechanism, CAPO achieves stable learning dynamics that jointly optimize calibration and accuracy.Experiments on multiple mathematical reasoning benchmarks show that CAPO-1.5Bsignificantly improves calibration by up to 15% while achieving accuracy comparable to or better than GRPO, and further boosts accuracy on downstream inference-time scaling tasks by up to 5%.Moreover, when allowed to abstain under low-confidence conditions, CAPO achieves a Pareto-optimal precision-coverage trade-off, highlighting its practical value for hallucination mitigation.

Human preference alignment is critical in building powerful and reliable large language models (LLMs). However, current methods either ignore the multi-dimensionality of human preferences (e.g. helpfulness and harmlessness) or struggle with the complexity of managing multiple reward models. To address these issues, we propose Sequential Preference Optimization (SPO), a method that sequentially fine-tunes LLMs to align with multiple dimensions of human preferences. SPO avoids explicit reward modeling, directly optimizing the models to align with nuanced human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Gradient analysis is conducted to show how SPO manages to fine-tune the LLMs while maintaining alignment on previously optimized dimensions. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences and significantly outperforms the baselines.

Chain of thought reasoning has demonstrated remarkable success in large language models, yet its adaptation to vision-language reasoning remains an open challenge with unclear best practices. Existing attempts typically employ reasoning chains at a coarse-grained level, which struggles to perform fine-grained structured reasoning and, more importantly, are difficult to evaluate the reward and quality of intermediate reasoning. In this work, we delve into chain of step reasoning for vision-language models, enabling assessing reasoning step quality accurately and leading to effective reinforcement learning and inference-time scaling with fine-grained rewards. We present a simple, effective, and fully transparent framework, including the step-level reasoning data, process reward model (PRM), and reinforcement learning training. With the proposed approaches, our models set strong baselines with consistent improvements on challenging vision-language benchmarks. More importantly, we conduct a thorough empirical analysis and ablation study, unveiling the impact of each component and several intriguing properties of inference-time scaling. We believe this paper serves as a baseline for vision-language models and offers insights into more complex multimodal reasoning. Our dataset, PRM, and code at https://github.com/baaivision/CoS.

Multi-Agent Policy Gradient (MAPG) has made significant progress in recent years. However, centralized critics in state-of-the-art MAPG methods still face the centralized-decentralized mismatch (CDM) issue, which means sub-optimal actions by some agents will affect other agent's policy learning. While using individual critics for policy updates can avoid this issue, they severely limit cooperation among agents. To address this issue, we propose an agent topology framework, which decides whether other agents should be considered in policy gradient and achieves compromise between facilitating cooperation and alleviating the CDM issue. The agent topology allows agents to use coalition utility as learning objective instead of global utility by centralized critics or local utility by individual critics. To constitute the agent topology, various models are studied. We propose Topology-based multi-Agent Policy gradiEnt (TAPE) for both stochastic and deterministic MAPG methods. We prove the policy improvement theorem for stochastic TAPE and give a theoretical explanation for the improved cooperation among agents. Experiment results on several benchmarks show the agent topology is able to facilitate agent cooperation and alleviate CDM issue respectively to improve performance of TAPE. Finally, multiple ablation studies and a heuristic graph search algorithm are devised to show the efficacy of the agent topology.

Decision making demands intricate interplay between perception, memory, and reasoning to discern optimal policies. Conventional approaches to decision making face challenges related to low sample efficiency and poor generalization. In contrast, foundation models in language and vision have showcased rapid adaptation to diverse new tasks. Therefore, we advocate for the construction of foundation agents as a transformative shift in the learning paradigm of agents. This proposal is underpinned by the formulation of foundation agents with their fundamental characteristics and challenges motivated by the success of large language models (LLMs). Moreover, we specify the roadmap of foundation agents from large interactive data collection or generation, to self-supervised pretraining and adaptation, and knowledge and value alignment with LLMs. Lastly, we pinpoint critical research questions derived from the formulation and delineate trends for foundation agents supported by real-world use cases, addressing both technical and theoretical aspects to propel the field towards a more comprehensive and impactful future.

State-of-the-art multi-agent policy gradient (MAPG) methods have demonstrated convincing capability in many cooperative games. However, the exponentially growing joint-action space severely challenges the critic's value evaluation and hinders performance of MAPG methods. To address this issue, we augment Central-Q policy gradient with a joint-action embedding function and propose Mutual-information Maximization MAPG (M3APG). The joint-action embedding function makes joint-actions contain information of state transitions, which will improve the critic's generalization over the joint-action space by allowing it to infer joint-actions' outcomes. We theoretically prove that with a fixed joint-action embedding function, the convergence of M3APG is guaranteed. Experiment results on the StarCraft Multi-Agent Challenge (SMAC) demonstrate that M3APG gives evaluation results with better accuracy and outperform other MAPG basic models across various maps of multiple difficulty levels. We empirically show that our joint-action embedding model can be extended to value-based multi-agent reinforcement learning methods and state-of-the-art MAPG methods. Finally, we run ablation study to show that the usage of mutual information in our method is necessary and effective.

The goal of zero-shot human-AI coordination is to develop an agent that can collaborate with humans without relying on human data. Prevailing two-stage population-based methods require a diverse population of mutually distinct policies to simulate diverse human behaviors. The necessity of such populations severely limits their computational efficiency. To address this issue, we propose E3T, an **E**fficient **E**nd-to-**E**nd **T**raining approach for zero-shot human-AI coordination. E3T employs a mixture of ego policy and random policy to construct the partner policy, making it both coordination-skilled and diverse. In this way, the ego agent is end-to-end trained with this mixture policy without the need of a pre-trained population, thus significantly improving the training efficiency. In addition, a partner modeling module is proposed to predict the partner's action from historical information. With the predicted partner's action, the ego policy is able to adapt its policy and take actions accordingly when collaborating with humans of different behavior patterns. Empirical results on the Overcooked environment show that our method significantly improves the training efficiency while preserving comparable or superior performance than the population-based baselines. Demo videos are available at https://sites.google.com/view/e3t-overcooked.