Harness-S rescues failed RL rollouts with sibling harnesses for generalizable agents
SJUT1 · hf · 2026-11-06
A paper from SJTU, Xiaohongshu, CMU and others introduces Harness-S, a selective harness-rescue training method. Sibling harness variants from a common base exhibit locality, complementarity, and nonstationarity; the pool is invoked only when all attempts under the base harness fail, replacing failed RL rollout groups with trajectories regenerated under selected auxiliary harnesses. An adaptive router picks Top-k variants via behavior profiles and multi-LLM voting. Across AppWorld, SWE-Gym, and DBBench, the trained policy substantially outperforms policy-only training under the base harness, and reusing the router for Rank-1 inference adds gains at the same rollout count.
More from coding & agent
- Scale AI ships 3,300+ RL environments; some tasks take humans 3 weeks and 500+ tool calls — alexandr_wang · 2026-11-06
- Agent self-analysis shows late-night AI frustration comes from long sessions, not fatigue — dbreunig · 2026-11-06
- Elastic's VP of AI: Treat search as test-time compute, latency is a currency to spend — gaganghotra_ · 2026-11-06
- EvoMemState: State-Centric Long-Term Memory for Personal Agents — Zhishang Xiang · 2026-11-06
- OpsBridge: a workbench for investigating blocked invoices with versioned evidence and human approval — Fit_Fortune953 · 2026-11-06
- Why hasn't anyone built the 'Tab, Tab, Tab' predictive experience for building apps? — johnlindquist · 2026-11-06