EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents¶
Core harness-policy operator mini-suite, not a trained paper-scale agent.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | University of Illinois Urbana–Champaign / Meta AI. |
| First public release | 2026-08-05 (arXiv v1) |
| Author code | No public author code found as of 2026-08-09. |
| Local adapter or method key | evoharness-rl |
| Local reproduction code | src/auto_research/agent_research/latest_20260809.py |
Paper method and evidence¶
Expose belief, progress and experience as editable external state. SFT teaches harness actions; cost-aware GRPO learns when to read, write or merge.
flowchart LR
A[External belief progress and experience] --> B[Learn harness actions]
B[Learn harness actions] --> C[Optimize cost-aware RL]Original paper figure¶
The original figure shows policy-managed external state. From the original paper; copyright remains with the authors.
Core equation¶
\(R=R_{task}-\lambda C_{harness}\).
Reported results¶
Qwen3-8B reports 96.9% ALFWorld success with harness annealing and evolution.
Local reproduction¶
Local metrics. Same mini-suite and mechanism-disabled comparator; no author model, proprietary environment or expensive judge. Evolve can discover the operator, but genome composition is constrained by public evaluation and budget. Checkpoints and raw runs are not committed.