Theory of Mind (ToM), the ability to infer others’ mental states from behavior, is pivotal for developing machines with human-level social intelligence. Existing methods endowing LLMs with ToM fall into two paradigms: training-free methods and those repurposing ToM evaluation benchmarks as training data for RL-based fine-tuning. However, training-free methods fail to internalize the augmented ToM into the LLMs. Meanwhile, using evaluation benchmarks as training sources is conceptually problematic and, in practice, results in narrow in-domain overfitting rather than robust ToM. To address the lack of training resources within the ToM community and to empower LLMs with robust ToM, we introduce ToM-Synth, a factorial combinatorial synthesis framework of 6912 social units. This framework enables the systematic synthesis of ToM data, yielding a training dataset of 27,648 instances, termed ToM-Synth-27K. Utilizing ToM-Synth-27K for RL fine-tuning, experimental results demonstrate consistent and significant improvements across models of varying families and scales on ToM, Emotional Intelligence, and Social Commonsense benchmarks. Furthermore, we observe concurrent enhancements in IQ-related tasks (math, science, logic) and effective performance scaling with increasing data scale.

Currently, most reinforcement learning tasks focus on domains like mathematics and programming, where verification is relatively straightforward. However, in subjective tasks such as role-playing, alignment techniques struggle to make progress, primarily because subjective reward modeling using the Bradley-Terry model faces significant challenges when dealing with ambiguous preferences. To improve reward modeling in subjective tasks, this paper proposes AAM (Act-Adaptive Margin), which enhances reward modeling by dynamically calibrating preference margins using the model’s internal parameter knowledge. We design two versions of AAM that efficiently generate contextually-appropriate preference gaps without additional human annotation. This approach fundamentally improves how reward models handle subjective rewards by better integrating generative understanding with preference scoring. To validate AAM’s effectiveness in subjective reward modeling, we conduct evaluations on RewardBench, JudgeBench, and challenging role-playing tasks. Results show that AAM significantly improves subjective reward modeling performance, enhancing Bradley-Terry reward models by 2.95% in general tasks and 4.85% in subjective role-playing tasks. Furthermore, reward models trained with AAM can help downstream alignment tasks achieve better results. Our test results show that applying rewards generated by AAM-Augmented RM to preference learning techniques (e.g., GRPO) achieves state-of-the-art results on CharacterEval and Charm. The code and dataset will be released upon acceptance.

Role-playing agents (RPAs) require balancing multiple objectives, such as instruction following, persona consistency, and stylistic fidelity, which are not always perfectly aligned across different dimensions. While prior work has primarily relied on supervised fine-tuning or reinforcement learning with scalarized rewards, these approaches do not explicitly address the coordination of multiple reward dimensions during optimization. We present MOA (Multi-Objective Alignment), a reinforcement-learning framework that enables multi-dimensional, fine-grained rubric optimization for general RPAs. MOA introduces a novel multi-objective optimization strategy that trains simultaneously on multiple fine-grained rubrics to boost optimization performance. Besides, to address the issues of model output diversity and quality, we have also employed thought-augmented rollout with off-policy guidance. Experiments on PersonaGym and RoleMRC show that MOA consistently improves multi-dimensional role-playing performance over supervised and standard RL baselines. Under identical evaluation protocols, an 8B model trained with MOA reaches performance competitive with strong closed-source models across multiple evaluation dimensions. These results suggest that MOA provides a practical framework for training more capable general-purpose role-playing agents.

Reinforcement Learning Fine-Tuning (RLFT) has achieved notable success in tasks with objectively verifiable answers (e.g., code generation, mathematical reasoning), yet struggles with open-ended subjective tasks like role-playing dialogue. Traditional reward modeling approaches, which rely on independent sample-wise scoring, face dual challenges: subjective evaluation criteria and unstable reward signals. Motivated by the insight that human evaluation inherently combines explicit criteria with implicit comparative judgments, we propose Comparative Policy Optimization (CPO). CPO redefines the reward evaluation paradigm by shifting from sample-wise scoring to comparative group-wise scoring. Building on the same principle, we introduce the CharacterArena evaluation framework, which comprises two stages: (1) Contextualized Multi-turn Role-playing Simulation, and (2) Trajectory-level Comparative Evaluation. By operationalizing subjective scoring via objective trajectory comparisons, CharacterArena minimizes contextual bias and enables more robust and fair performance evaluation. Empirical results on CharacterEval, CharacterBench, and CharacterArena confirm that CPO effectively mitigates reward ambiguity and leads to substantial improvements in dialogue quality.

Instruction following (IF) is a critical capability for large language models (LLMs). However, handling complex instructions with multiple constraints remains challenging. Previous methods typically select preference pairs based on the number of constraints they satisfy, introducing noise where chosen examples may fail to follow some constraints and rejected examples may excel in certain respects over the chosen ones. To address the challenge of aligning with multiple preferences, we propose a simple yet effective method called Reverse Preference Optimization (RPO). It mitigates noise in preference pairs by dynamically reversing the constraints within the instruction to ensure the chosen response is perfect, alleviating the burden of extensive sampling and filtering to collect perfect responses. Besides, reversal also enlarges the gap between chosen and rejected responses, thereby clarifying the optimization direction and making it more robust to noise. We evaluate RPO on two multi-turn IF benchmarks, Sysbench and Multi-IF, demonstrating average improvements over the DPO baseline of 4.6 and 2.5 points (on Llama-3.1 8B), respectively. Moreover, RPO scales effectively across model sizes (8B to 70B parameters), with the 70B RPO model surpassing GPT-4o.

The development of Multimodal Large Language Models (MLLMs) has seen significant progress, driven by increasing demands across various fields (e.g., multimodal agents, embodied intelligence). While model-driven approaches aim to enhance MLLM capabilities through diverse architectures, their performance gains have become increasingly marginal. In contrast, data-driven methods, which scale up image-text instruction datasets, have proven more effective but face challenges related to limited data diversity and complexity. The absence of high-quality instruction data remains a major bottleneck in MLLM development. To address this issue, we propose , a novel multimodal instruction data evolution framework. This framework iteratively enhances data quality through a refined combination of fine-grained perception, cognitive reasoning, and interaction evolution, generating a more complex and diverse image-text instruction dataset that significantly improves MLLM capabilities. Starting with an initial dataset, SEED-163K, we employ to systematically expand instruction diversity, extend visual reasoning steps to improve cognitive abilities, and extract fine-grained visual details to enhance understanding and robustness. To rigorously evaluate our approach, we conduct extensive qualitative analysis and quantitative experiments across 13 vision-language tasks. Compared to baseline models trained on the original seed dataset, our method achieves an average accuracy improvement of 3.1 percentage points. Moreover, our approach attains state-of-the-art (SOTA) performance in nine tasks while using significantly less data than existing state-of-the-art models.

Large Language Models (LLMs) have shown impressive reasoning capabilities in well-defined problems with clear solutions, such as mathematics and coding. However, they still struggle with complex real-world scenarios like business negotiations, which require strategic reasoning—an ability to navigate dynamic environments and align long-term goals amidst uncertainty.Existing methods for strategic reasoning face challenges in adaptability, scalability, and transferring strategies to new contexts.To address these issues, we propose explicit policy optimization (EPO) for strategic reasoning, featuring an LLM that provides strategies in open-ended action space and can be plugged into arbitrary LLM agents to motivate goal-directed behavior.To improve adaptability and policy transferability, we train the strategic reasoning model via multi-turn reinforcement learning (RL), utilizing process rewards and iterative self-play.Experiments across social and physical domains demonstrate EPO’s ability of long-term goal alignment through enhanced strategic reasoning, achieving state-of-the-art performance on social dialogue and web navigation tasks. Our findings reveal various collaborative reasoning mechanisms emergent in EPO and its effectiveness in generating novel strategies, underscoring its potential for strategic reasoning in real-world applications. Code and data are available at https://github.com/lxqpku/EPO.

Social agents powered by large language models (LLMs) can simulate human social behaviors but fall short in handling complex social dialogues. Direct Preference Optimization (DPO) has proven effective in aligning LLM behavior with human preferences across various agent tasks. However, standard DPO focuses solely on individual turns, which limits its effectiveness in multi-turn social interactions. Several DPO-based multi-turn alignment methods with session-level data have shown potential in addressing this problem. While these methods consider multiple turns across entire sessions, they are often overly coarse-grained, introducing training noise, and lack robust theoretical support. To resolve these limitations, we propose Segment-Level Direct Preference Optimization (SDPO), which dynamically select key segments within interactions to optimize multi-turn agent behavior. SDPO minimizes training noise and is grounded in a rigorous theoretical framework. Evaluations on the SOTOPIA benchmark demonstrate that SDPO-tuned agents consistently outperform both existing DPO-based methods and proprietary LLMs like GPT-4o, underscoring SDPO’s potential to advance the social intelligence of LLM-based agents. We release our code and data at https://anonymous.4open.science/r/SDPO-CE8F.

Role-Playing Agents (RPAs), benefiting from large language models, is an emerging interactive AI system that simulates roles or characters with diverse personalities. However, existing methods primarily focus on mimicking dialogues among roles in textual form, neglecting the role’s voice traits (e.g., voice style and emotions) as playing a crucial effect in interaction, which tends to be more immersive experiences in realistic scenarios. Towards this goal, we propose OmniCharacter, a first seamless speech-language personality interaction model to achieve immersive RPAs with low latency. Specifically, OmniCharacter enables agents to consistently exhibit role-specific personality traits and vocal traits throughout the interaction, enabling a mixture of speech and language responses. To align the model with speech-language scenarios, we construct a dataset named OmniCharacter-10K, which involves more distinctive characters (20), richly contextualized multi-round dialogue (10K), and dynamic speech response (135K). Experimental results showcase that our method yields better responses in terms of both content and style compared to existing RPAs and mainstream speech-language models, with a response latency as low as 289ms.