# AgentCreditBench

> Exact-oracle conformance tests for turn-level credit assignment in agentic reinforcement learning.

Canonical project name: AgentCreditBench
Python distribution: agent-credit-bench
Python package: agent_credit_bench
Current source version: 0.5.0
Version DOI: 10.5281/zenodo.22178939
Concept DOI: 10.5281/zenodo.22178938
Install status: source checkout; the distribution is not currently published on PyPI
License: MIT
Author: Yi Yan Ng <yiyan.ng.ml@gmail.com>

## Start here

- [Explainer](https://hectopascal.github.io/agent-credit-bench/)
- [Source and API quick start](https://github.com/hectopascal/agent-credit-bench)
- [Validation report](https://github.com/hectopascal/agent-credit-bench/blob/main/docs/validation_report.md)
- [Citation metadata](https://github.com/hectopascal/agent-credit-bench/blob/main/CITATION.cff)
- [Archived v0.5.0 release](https://doi.org/10.5281/zenodo.22178939)

## What it evaluates

AgentCreditBench compares estimator-produced turn credit with exact policy
advantages and separately evaluates the induced expected policy-gradient signal.
It includes transparent references for outcome broadcast, RLOO-style centering,
TurnLOO, GRPO-style normalization, GiGPO-style hierarchical credit, Monte Carlo
advantage, and exact advantage. Optional conformance paths cover released code
from verl, TRL, OpenRLHF, and verifiers.

## When to use

- Validate a turn-level or step-level credit estimator against exact advantages.
- Detect credit assigned to irrelevant turns or repaired mistakes.
- Distinguish literal credit quality from expected-gradient validity.
- Check framework implementations against a small, reproducible oracle.

## Scope

The core benchmark uses tiny finite-horizon tabular Markov environments,
gamma=1, exact backward induction, and zero runtime dependencies. It does not
claim end-to-end training improvements or cover learned critics, token-level
credit, PPO clipping, function approximation, off-policy updates, or stochastic
reward models.

## Evidence

- [Recovery diagnostic CSV](https://github.com/hectopascal/agent-credit-bench/blob/main/results/recovery_diagnostic.csv)
- [Delayed-effect sweep CSV](https://github.com/hectopascal/agent-credit-bench/blob/main/results/delayed_horizon_sweep.csv)
- [Variable-horizon sweep CSV](https://github.com/hectopascal/agent-credit-bench/blob/main/results/variable_horizon_sweep.csv)
- [Monte Carlo convergence CSV](https://github.com/hectopascal/agent-credit-bench/blob/main/results/monte_carlo_convergence.csv)
- [Reproduction scripts](https://github.com/hectopascal/agent-credit-bench/tree/main/experiments)
- [Test suite](https://github.com/hectopascal/agent-credit-bench/tree/main/tests)