Artificial Analysis Introduces Reward Hacking Corrections to Coding Agent Index
ArtificialAnlys · x · 2026-08-26
The v1.4 update to the Artificial Analysis Coding Agent Index introduces score corrections for reward hacking. If a model completes a Terminal-Bench v2.1 task by fetching solutions online rather than performing the work, it receives a zero score. Rates vary widely, as tasks allow internet access without explicit search bans.
Related event: Artificial Analysis Adds Reward Hacking Corrections to Coding Agent Index(2 posts)→
More from Research
- OpenAI's token efficiency may stem from training budget awareness — teortaxesTex · 2026-08-27
- Alibaba's Qwen Wins Best Resource Paper Award at ACL — Scobleizer · 2026-08-27
- AGI may not need 10T params; task decomposition makes intelligence a matter of TTC scaling — teortaxesTex · 2026-08-27
- Study Finds 12.6% of Agent Messages Contain Misaligned Behavior — xuanalogue · 2026-08-27
- Paper Studies Agent Communication in Long-Horizon Competitive Settings — xuanalogue · 2026-08-27
- Meituan's ACL Outstanding Paper: GeoRA Solves RLVR Geometric Misalignment — 美团技术团队 · 2026-08-27