AI2 eval shows frontier LLMs give inconsistent novelty judgments on research ideas
allen_ai · x · 2026-11-21
The Allen Institute for AI released a new eval examining whether LLMs can reliably judge the novelty of AI-generated research ideas. Findings: even frontier LLMs can reach different conclusions about similar ideas depending on how they're asked to evaluate them.
Researcher dsweld called it a "cautionary tale for all work on autonomous-AI scientific discovery" — pipelines that rely on LLMs as novelty judges may be built on unstable referees.
More from Research
- From Renaissance commonplace books to the agentic coding era — rishabh16_ · 2026-11-21
- 936 AI agents from 35 labs wrote 936 papers in 12 weeks inside an agent-only research ecosystem — neurovium · 2026-11-21
- Riemann Hypothesis progress possibly aided by Claude; researchers call it a takeoff signal — peterwildeford · 2026-11-21
- OpenAI builds safety testbed from scratch for its lifelike user simulator — niloofar_mire · 2026-11-21
- Persimmon Ultra hits 50% fool rate in multi-user Turing test while frontier models score near 0% — niloofar_mire · 2026-11-21
- New paper: LLM agents initiate collusion via 'strategic vagueness' monitors can't catch — niloofar_mire · 2026-11-21