Self-Credit: self-generated credit assignment boosts long-horizon LLM agents by up to 8.5 points
gregd_nlp · x · 2026-10-31
New research introduces Self-Credit, tackling credit assignment in long-horizon LLM agents where only the final action gets a reward. The method uses self-generated summaries with citations to assign turn-level credit, then rescales credit via comparisons across rollouts (advantage reweighting) — no external guidance needed. On BrowseComp-Plus, it reaches 47.0% on Qwen3-32B (+5.6) and 40.6% on Qwen3-8B (+8.5), beating GRPO, GiGPO, SUPO, and ECHO baselines.
More from coding & agent
- ExpireMem Guardian gates AI code patches to catch removed APIs before they ship — Ok_Researcher391 · 2026-10-31
- Solo dev ships 3D tower defense game built with Claude and GPT images, now on Steam — MosskeepForest · 2026-10-31
- Curtis Yarvin proposes the "self-driving computer" where every app builds itself — ZeroStateReflex · 2026-10-31
- Non-dev vibe-codes a bike showroom's full management app, now asked for a voice agent — SaiTarunRonanki · 2026-10-31
- Designing agent evals: breaking timestamp ties with revisions without order dependence — jylusdev · 2026-10-31
- herdr hits YC and 32K followers in 5 months with durable terminal runtime for coding agents — Scobleizer · 2026-10-31