AgentCreditBench v0.5.0

AgentCreditBench is a CPU-first conformance-test suite for turn-level credit assignment in agentic reinforcement learning. This is the first public pre-release; APIs and experiment protocols may still change.

Archive and citation

• Version DOI: 10.5281/zenodo.22178939
• Concept DOI for all versions: 10.5281/zenodo.22178938

Highlights

• Exact policy-advantage oracles on three tiny finite-horizon MDPs: delayed effects, recoverable mistakes, and variable horizons.
• Separate identification metrics from expected policy-gradient metrics so state-dependent baselines are scored correctly.
• Reference estimators for exact advantage, outcome broadcast, group centering, TurnLOO, GRPO-style normalization, GiGPO-style hierarchical credit, and Monte Carlo advantage.
• Optional conformance paths for released advantage code in verl 0.9.0, TRL 1.12.0, OpenRLHF 0.11.0, and verifiers 0.1.14.
• Hardened tests for episode termination, non-uniform policies, dense rewards, Spearman ranking and ties, population variance, Monte Carlo early termination, and GiGPO's episode-level contribution and normalization.
• Committed CSV evidence, reproduction scripts, a scoped validation report, and an exact-oracle explainer.

Install from source

The distribution is not yet published on PyPI.

git clone https://github.com/hectopascal/agent-credit-bench.git
cd agent-credit-bench
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
pytest

Validation

• 118 tests pass in the dependency-free local suite; optional framework tests skip when their pinned extras are absent.
• GitHub CI runs Python 3.10 and 3.12 plus dedicated verl, TRL, OpenRLHF, and verifiers integration jobs.
• CITATION.cff validates against CFF 1.2.0.

See the interactive explainer, validation report, and committed results.