Awesome Credit Assignment in LLM RL
A curated research survey covering how reinforcement learning for large language models assigns credit to individual tokens, steps, turns or agents when the only feedback is a single outcome reward at the end of a long trajectory.
This repository accompanies a survey paper on credit assignment in reinforcement learning for large language models, organizing 47 methods from the original survey snapshot, plus 29 newer entries added since, by two axes: granularity, from token-level up through multi-agent, and methodology, such as Monte Carlo, temporal difference, LLM-as-critic, game-theoretic and information-theoretic approaches. It separates methods developed for single-chain reasoning RL, where credit is distributed across tokens and steps within one long chain of thought, from methods for agentic RL, where a multi-turn environment introduces stochastic transitions and much longer horizons.
The living catalog includes foundational and background material on classical RL and reinforcement learning algorithms for language models, dedicated sections on reasoning-RL and agentic-RL methods, a section on multi-agent credit assignment, and a list of benchmarks used to evaluate these methods, alongside guides meant to help a reader quickly find the method decision path or a machine-readable version of the catalog.