EvoHeir paper: fresh evaluation cuts harmful inherited updates in self-evolving LLM loops by 68–83%
MBZUAI · hf · 2026-10-19
A MBZUAI paper, 'Selection Is Not Learning,' identifies a selection-to-update gap in self-evolving LLM loops (R-Zero, Agent0, Skill Self-Pay and peers): tasks that look learnable often don't yield useful updates, yet these loops inherit every resulting change unverified.
Two fixes are proposed:
- HeirCommit freezes the updated model and uses fresh evaluation questions to decide whether to inherit it or keep the current model;
- EvoHeir separates parameter inheritance from experience inheritance, using current and archived experience to repair failures and preserve successes before final confirmation.
The analysis decomposes selection-induced optimism into selection, absorption, and confirmation errors. HeirCommit cuts the harmful share of accepted updates by 68–83% relative to random inheritance at the same acceptance rate, and EvoHeir achieves the best math and general-reasoning averages across four backbones.
More from Research
- Dream2Discern: Self-Generated Hard Examples Predict Multi-Agent Failures With 90%+ Accuracy — RutgersU · 2026-10-19
- BridgeWAM: latent bridge queries decouple video and action experts, 2.47x faster robot inference — Fudan-University · 2026-10-19
- NVIDIA's FC-SWE: Failure-Conditioned RL Boosts Multi-Attempt Resolution for SWE Agents — nvidia · 2026-10-19
- AVC-Reward: an event-centric framework for evaluating audio-video generation consistency — ZhejiangUniversity · 2026-10-19
- PremiseCheck Fact-Checks Peer Reviews at GPT-5.5-Level Accuracy for Half the Cost — Dingjie Song · 2026-10-19
- SciΠvot benchmark: frontier models prefer flawed research plans and lag human scientific judgment — Youran Sun · 2026-10-19