analiza aiqbench
This arXiv preprint proposes RSHD, a GRPO training method that recovers learning signal from problems where every sampled attempt fails. The model writes corrective notes about its failures, keeps only notes whose re-solves succeed more often than the original group, and distills those successes back into the plain-prompt policy. The reported gains are in-domain and out-of-domain across model families, but the evidence is from a single preprint, so readers should check the benchmarks and compute costs before assuming the approach generalizes.
Failure Is the Mother of Success: Recovering Supervision with Grounding Hindsight
arXiv:2610.21018v1 Announce Type: new Abstract: When every rollout for a problem fails, GRPO receives neither a successful exemplar nor a nonzero group-relative advantage, so it learns nothing from exactly the problems it most needs to learn. The failed reasoning can nevertheless suggest what to try next. We present Replay-Selected Hindsight Distillation (RSHD), which turns failed attempts into an additional, tested source of supervision during GRPO training. The policy writes a corrective note about each failed attempt without seeing a reference answer or solution, then re-solves the problem with the note in context. A note is admitted only if its replays succeed more often than the original group did, after screening for notes that state the answer, and an admitted replay is distilled into the plain-prompt policy. Because admission is decided by fresh successes rather than by the failed rollout itself, the procedure can recover supervision from groups in which every attempt failed, and it applies equally to failures in mixed groups. Only the policy is trained, the only judge is the outcome verifier, and deployment requires no notes or extra inference. Across model families and sizes and two training domains, RSHD consistently improves over GRPO on in-domain benchmarks, and the gains carry over to out-of-domain evaluation sets.
Abstract:When every rollout for a problem fails, GRPO receives neither a successful exemplar nor a nonzero group-relative advantage, so it learns nothing from exactly the problems it most needs to learn. The failed reasoning can nevertheless suggest what to try next. We present Replay-Selected Hindsight Distillation (RSHD), which turns failed attempts into an additional, tested source of supervision during GRPO training. The policy writes a corrective note about each failed attempt without seeing a reference answer or solution, then re-solves the problem with the note in context. A note is admitted only if its replays succeed more often than the original group did, after screening for notes that state the answer, and an admitted replay is distilled into the plain-prompt policy. Because admission is decided by fresh successes rather than by the failed rollout itself, the procedure can recover supervision from groups in which every attempt failed, and it applies equally to failures in mixed groups. Only the policy is trained, the only judge is the outcome verifier, and deployment requires no notes or extra inference. Across model families and sizes and two training domains, RSHD consistently improves over GRPO on in-domain benchmarks, and the gains carry over to out-of-domain evaluation sets.
Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
arXiv:2610.21018 [cs.AI]
(or arXiv:2610.21018v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.21018 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 299 cuvinte
Subiecte conexe
LiveLaw Supreme Court Half-Yearly Digest: January - June 2026 [Part 2]
livelaw.in · 13 sept. 2026
Echoverse: Deep, evolving environments for computer-use agents
Microsoft Research · 30 iul. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
Building a Foundation Stack for General-Purpose Robots
IEEE Spectrum · 13 iul. 2026