Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Social-R1: Towards Human-like Social Reasoning in LLMs

Jincenzi Wu, Yuxuan Lei, Jianxun Lian, Yitian Huang, Lexin Zhou, Haotian Li, Xing Xie, Helen Meng

TL;DR

This work introduces ToMBench-Hard, an adversarial benchmark designed to provide hard training examples for social reasoning and proposes Social-R1, a reinforcement learning framework that aligns model reasoning with human cognition through multi-dimensional rewards.

Abstract

While large language models demonstrate remarkable capabilities across numerous domains, social intelligence - the capacity to perceive social cues, infer mental states, and generate appropriate responses - remains a critical challenge, particularly for enabling effective human-AI collaboration and developing AI that truly serves human needs. Current models often rely on superficial patterns rather than genuine social reasoning. We argue that cultivating human-like social intelligence requires training with challenging cases that resist shortcut solutions. To this end, we introduce ToMBench-Hard, an adversarial benchmark designed to provide hard training examples for social reasoning. Building on this, we propose Social-R1, a reinforcement learning framework that aligns model reasoning with human cognition through multi-dimensional rewards. Unlike outcome-based RL, Social-R1 supervises the entire reasoning process, enforcing structural alignment, logical integrity, and information density. Results show that our approach enables a 4B parameter model to surpass much larger counterparts and generalize robustly across eight diverse benchmarks. These findings demonstrate that challenging training cases with trajectory-level alignment offer a path toward efficient and reliable social intelligence.

Social-R1: Towards Human-like Social Reasoning in LLMs

TL;DR

This work introduces ToMBench-Hard, an adversarial benchmark designed to provide hard training examples for social reasoning and proposes Social-R1, a reinforcement learning framework that aligns model reasoning with human cognition through multi-dimensional rewards.

Abstract

While large language models demonstrate remarkable capabilities across numerous domains, social intelligence - the capacity to perceive social cues, infer mental states, and generate appropriate responses - remains a critical challenge, particularly for enabling effective human-AI collaboration and developing AI that truly serves human needs. Current models often rely on superficial patterns rather than genuine social reasoning. We argue that cultivating human-like social intelligence requires training with challenging cases that resist shortcut solutions. To this end, we introduce ToMBench-Hard, an adversarial benchmark designed to provide hard training examples for social reasoning. Building on this, we propose Social-R1, a reinforcement learning framework that aligns model reasoning with human cognition through multi-dimensional rewards. Unlike outcome-based RL, Social-R1 supervises the entire reasoning process, enforcing structural alignment, logical integrity, and information density. Results show that our approach enables a 4B parameter model to surpass much larger counterparts and generalize robustly across eight diverse benchmarks. These findings demonstrate that challenging training cases with trajectory-level alignment offer a path toward efficient and reliable social intelligence.
Paper Structure (43 sections, 4 equations, 12 figures, 18 tables)

This paper contains 43 sections, 4 equations, 12 figures, 18 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Theory-of-Mind Analysis in LLMs
  4. RL for LLM General Reasoning
  5. The ToMBench-Hard Benchmark
  6. Adversarial Data Construction
  7. Diagnostic Analysis: Exposing the Shortcut Illusion
  8. The Social-R1 Framework
  9. Multi-Dimensional Reward Design
  10. Verifiable Format Alignment ($R_{{fmt}}$)
  11. Reward Synthesis and Learning
  12. Experiment
  13. Experiment Setting
  14. SIP Content Integrity Reward Model ($RM_{content}$).
  15. Implementation Details
  16. ...and 28 more sections

Figures (12)

  • Figure 1: Social-R1 for Human-like and Efficient Social Reasoning. By integrating SIP-guided rewards into reinforcement learning, Social-R1 mitigates reasoning shortcuts and enforces structured human-like social inference, improving both accuracy and efficiency across model scales. Detailed cases are in Appendix \ref{['app:0mainfigure']}.
  • Figure 2: Option-Mention Density across SIP reasoning stages.
  • Figure 3: Stage-wise SIP accuracy across models.
  • Figure 4: Case study highlighting the Interpretation Bottleneck. Detailed cases are provided in Appendix \ref{['app:5optionmatching']}.
  • Figure 5: Robustness study under story-consistent distractors.
  • ...and 7 more figures