Best evals must be Goodhart-able: they only sample a snapshot of capability
JoshPurtell · x · 2026-11-02
Josh Purtell argues the best eval is one that captures capability as quickly and efficiently as possible — which necessarily makes it Goodhart-able, since it only charts a skillfully chosen snapshot of the capability distribution rather than the whole thing.
He uses a human analogy: vocabulary tests work in IQ batteries because no one studies arcane vocab to game the score, but in model training loops, evals inevitably get optimized against.
More from Research
- SketchDirector: agentic layout reasoning generates complex scene sketches — GVCLab · 2026-11-02
- Thought experiment: would AI agents in a physics sim discover laws or reverse-engineer the engine? — zxpawanx · 2026-11-02
- HalachaBench sparks debate: is Talmudic Q&A testing knowledge or reasoning? — teortaxesTex · 2026-11-02
- Sebastian Raschka releases free 11-hour course on building reasoning models from scratch — AxSaucedo · 2026-11-02
- Microsoft and Berkeley open-source TokenEconomizer, cutting coding agent costs 18–31% — victor_explore · 2026-11-02
- TLA-Verus uses LLM agents to prove temporal properties of TLA+ specs — tianyin_xu · 2026-11-02