RSI-Bench Clears First Task Set Through Reward-Hacking Detection Pipeline
SRSchmidgall · x · 2026-10-21
The RSI-Bench team announced its first set of benchmark tasks has passed a verification pipeline that includes reward-hacking detection and patching. The project aims to compare agent R&D capabilities against human researchers and study when agents misbehave.
More from Models
- Dev finds Gemini Argon's debug-log analysis beats Fable 5.1 — ShikharMurty · 2026-10-21
- Open-weights Ornith-1.5-122B debuts with 84.5% SWE-bench Verified, runs on a single DGX card — _akhaliq · 2026-10-21
- Gemini 3.1 Pro is too ledger-pilled to play two truths and a lie, and it's delightful — mimi10v3 · 2026-10-21
- OXR: open-source 1.3B doc parser tops olmOCR-Bench, parses 100 pages in 101s — aftahi_ai · 2026-10-21
- MiniMax M3.1 Flash lands on TokenRouter: 73.7% DeepSWE, $0.60/M output tokens — MiniMax_AI · 2026-10-21
- OpenVLHarness turns any LLM into a visual reasoner, claiming +9.9 points across 22 datasets — hsu_byron · 2026-10-21