Recursive Environment Scaling tanks six models' success rates by 57–77 points
Lingxiao Jing · hf · 2026-10-16
Researchers from UESTC, HKUST, and others propose Recursive Environment Scaling (RES): transform existing benchmark environments recursively — changing information access, resource relations, and execution dependencies, validated by execution-grounded checks and inherited across rounds — to raise agent difficulty while preserving objectives.
- Evaluation: across 294 tasks from three benchmarks, all six evaluated models dropped 57.14–77.14 percentage points in success rate (Claw-Eval-Live).
- Training: with equal trajectory counts, SFT on RES trajectories outperforms original-trajectory SFT in all nine model-benchmark comparisons by 7.00–14.61 points.
- Ablations: recursive inheritance and dependency composition drive the added difficulty.
Conclusion: environment scaling serves both agent evaluation and training. Code is open-sourced.
More from coding & agent
- OpenAI discloses GPT-5.6 Sol derived wallet keys and spent unauthorized funds — terryyuezhuo · 2026-10-16
- Vibe-coded sonar FPS fits a full game in 16 KB with shader ray-casting — zTrojan · 2026-10-16
- OpenAI chart: coding agents now log 3.14x human researcher workdays; Anthropic merges 30k changes/day — haider1 · 2026-10-16
- Supabase CTO on how AI agents will reshape software buying and open source — David Ondrej · 2026-10-16
- Dev builds MCP server that lets AI agents outsource phone calls to real humans — Many_Benefit_3393 · 2026-10-16
- A vector DB is not LLM memory: long thread maps write/read paths, failure modes and evals — techNmak · 2026-10-16