TRCA: Transition-wise Rubric Credit Assignment for Long-horizon LLM Agents¶
Deterministic core credit mechanism, not a full external environment.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | No affiliation disclosed; first author Huan Zhang. |
| First public release | 2026-08-17 (arXiv v1) |
| Author code | No public author code found as of 2026-08-24. |
| Local adapter or method key | trca |
| Local reproduction code | src/auto_research/agent_research/historical_b10_b11.py |
Paper method and evidence¶
Without successful anchors, derive base and breakthrough rewards for each transition from evidence, execution and invalidity rubrics.
flowchart LR
A[State transition] --> B[Apply evidence execution and invalidity rubrics]
B[Apply evidence execution and invalidity rubrics] --> C[Base and breakthrough reward]Original paper figure¶
Figure shows transition-wise credit construction. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
The paper reports +6.0%–12.6% on WebShop and mean +1.9%–18.3% on SearchQA versus baselines.
Local reproduction¶
PlanBench mini 120 episodes: TRCA joint success 1.0000, mean cost 0.7100; long context 1.0000, 64.5000. Metrics, Batch index. With saturated accuracy, inspect traces, provenance and method-specific counters rather than claiming gains. No real browser, AppWorld, SWE-bench or external services. Raw outputs stay under runs/agent-research/; only metrics are committed.