New "Discovery Episode" Framework Measures AI Scientists by Real Research Cycles
量子位 · wechat · 2026-08-24
A joint paper by DeepPrinciple, Microsoft, and Stanford introduces the "discovery episode" evaluation framework, shifting from closed-book exams like HLE to assessing AI's full research cycle. The system tracks decision-making trajectories through hypothesis, design, execution, and interpretation, valuing failed data as core assets. DeepPrinciple's MIRA platform, a real-world application of this framework, ranks first on the ResearchClawBenchmark and ScienceAgentArena, demonstrating its ability to close the loop between dry and wet labs.
More from Research
- Two code-LLM papers accepted at EMNLP2026: behavior benchmark and self-verifying codegen — algo_diver · 2026-10-25
- "Do LLMs Feel?" paper maps localized emotion circuits inside LLMs — mikeflache · 2026-10-25
- Researchers push back on hype that AI papers done in 2–3 days make peer review obsolete — ChongZzZhang · 2026-10-25
- EC2-Bench: edge-cloud collaboration hits 73% of cloud-only agent performance at 15% of the cost — zibuyu9 · 2026-10-25
- Paper: additive "score centering" term stabilizes LLM RL under training-inference mismatch — finbarrtimbers · 2026-10-25
- Score centering cancels RL training-inference drift via a bias term — finbarrtimbers · 2026-10-25