Paper: solo leaderboards can't pick good collaborators; Shapley-based RL fixes it
shangbinfeng · x · 2026-10-22
New work from CMU researchers argues that LLMs are usually trained and evaluated alone, and leaderboards built to rank soloists miss what makes a good team player—solo RL can even make an LLM better alone but worse together. The paper proposes Compositional RL: training policies with GRPO where the reward is the model's Shapley marginal contribution to the multi-LLM system, plus methods to evaluate collaborators, including with frontier models.
More from Research
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- CMU PhD talk: patching full-attention transformers into hybrid architectures late in training — Cohere · 2026-10-23
- Ultrafast waves flush brain waste in minutes: Science Advances mouse study on Alzheimer's — NickPassig · 2026-10-23
- arXiv paper gives StopGrads theoretical grounding: convergence without an underlying loss — srush_nlp · 2026-10-23
- PEARL demo: an LLM agent that checks experiment results, revises plans, and re-tests — KevinKaichuang · 2026-10-23