Research map / Multi-agent systems and games / LLM agents
CollabBench: Benchmarking and Unleashing Collaborative Ability of LLMs with Diverse Players via Proactive Engagement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CollabBench, a cooperative-game benchmark with simulated diverse personality-driven players, and a GiGPO-based agentic RL training method with hybrid efficiency and affective rewards that improves Qwen2.5-7B-Instruct on CWAH and Overcooked-based collaboration.
Problem
LLM agents struggle to collaborate with diverse human-like partners in grounded cooperative games, and existing evaluations focus on efficiency (game score) while neglecting affective qualities like helpfulness, trustfulness, and empathy.
Approach
Simulates diverse players via Big Five personality prompts filtered by an LLM-judge consistency check, then trains a collaborating LLM with GiGPO using a trajectory-level efficiency reward plus a step-level affective reward from an LLM judge, within a single-pass think/message/action rollout.
Key results
- Trained Qwen2.5-7B-Instruct reduces CWAH completion steps by 15.2% (Agent 1) and 29.3% (Agent 2) relative to the base 7B model.
- Affective scores improve over the base Qwen2.5-7B-Instruct, e.g. CWAH helpfulness +17.2% and +39.4% for the two agent roles.
- In a 15-participant user study, the fully trained model significantly reduced completion steps and scored higher on affective Likert items than the base model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| collaborative agentic training with affective and efficiency rewards | Qwen2.5-7B-Instruct · CWAH-MultiPlayer | Step count (Agent 1 and 2, lower better): 84.51 -> 71.64 (Agent 1); 90.03 -> 63.65 (Agent 2) | Helped |
| collaborative agentic training with affective and efficiency rewards | Qwen2.5-7B-Instruct · CWAH-MultiPlayer | Affective average (Helpfulness, Trustfulness, Empathy): Helpfulness +17.2%/+39.4%; Empathy +33.5%/+31.5% | Helped |
| collaborative agentic training | Qwen2.5-7B-Instruct · Cook-MultiPlayer | Game score: 86.93 -> 99.20 (Agent 1); 85.87 -> 102.40 (Agent 2) | Helped |
| collaborative agentic training | Qwen2.5-7B-Instruct · Cook-MultiPlayer | Std. of score: 33.30 -> 34.03 (Agent 1); 33.73 -> 34.76 (Agent 2) | No clear effect |
| removing personality-driven simulated players (self-play partner) during training (ablation) | Qwen2.5-7B-Instruct trained agent · CWAH-MultiPlayer | Efficiency and affective average: affective performance drops sharply; efficiency improves | Hurt |
| removing affective reward (efficiency-only reward) (ablation) | Qwen2.5-7B-Instruct trained agent · CWAH-MultiPlayer | Affective average score: affective drops sharply when removed | Helped |
| profile filtering via interaction-based consistency scoring (ablation) | player profile set · CWAH-MultiPlayer | k-NN label consistency LC^k: 0.4511 -> 0.5470 | Helped |
| profile filtering via interaction-based consistency scoring (ablation) | player profile set · Cook-MultiPlayer | k-NN label consistency LC^k: 0.5729 -> 0.6760 | Helped |
| adding a user study comparison of base vs affective-reward-free vs full training | Qwen2.5-7B-Instruct · Human user study CWAH (15 participants) | Completion steps and Likert affective ratings: both ablation and full model reduced steps vs base; full model higher on all affective dimensions | Helped |
Surprising findings
- Larger proprietary and open models (GPT-5.2, DeepSeek-V3.1, Qwen2.5-72B) show weak helpfulness and empathy despite similar game scores, and increased token generation does not improve affective performance.
- Ablation without affective reward causes affective performance to drop sharply as agents lose communication and effectively play as single-player.
Limitations
- Affective metrics rely on an LLM judge (DeepSeek-V3.1) with residual subjectivity; human calibration is costly and limited.
- Player simulator is an LLM (DeepSeek-V3.1), so simulated players may not reflect real human behavior; user study has only 15 participants.
- Evaluation covers a limited set of models due to resource constraints; agents may still exhibit reward hacking in affective optimization.
Experimental setup
Datasets: CWAH-MultiPlayer Cook-MultiPlayer
Benchmarks: CWAH-MultiPlayer Cook-MultiPlayer
Metrics: steps game score standard deviation tokens per step helpfulness trustfulness empathy Pearson correlation
Baselines: CoELA with GPT-5.2 CoELA with DeepSeek-V3.1 CoELA with Qwen2.5-72B-Instruct ProAgent with GPT-5.2 ProAgent with DeepSeek-V3.1 ProAgent with Qwen2.5-72B-Instruct MHP SP PBT FCP MEP COLE
Builds on: GiGPO GRPO CoELA ProAgent verl-agent
Closest papers by meaning
- ACC-Collab: An Actor-Critic Approach to Multi-Agent LLM Collaboration (ICLR 2025 · similarity 0.89)
- Learning “Partner-Aware” Collaborators in Multi-Party Collaboration (NeurIPS 2025 · similarity 0.88)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
- Shapley-Coop: Credit Assignment for Emergent Cooperation in Self-Interested LLM Agents (NeurIPS 2025 · similarity 0.87)
- CollabLLM: From Passive Responders to Active Collaborators (ICML 2025 · similarity 0.87)
- MARSHAL: Incentivizing Multi-Agent Reasoning via Self-Play with Strategic LLMs (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.