Evals should be neutral, negative, or positive, argues an AI researcher
secemp9 · x · 2026-07-22
A simpler eval rule: neutral, negative, and positive evals
The author argues that eval design should be thought of in three buckets:
- Neutral evals: tasks that can be done naturally without the model being aware it is in an evaluation.
- Negative evals: setups where the model becomes eval-aware.
- Positive evals: eval-obliviousness, where the harness avoids giving the model hints that it is being tested.
The main point is that good evals should reproduce the task as it would happen in the real world, while avoiding prompts or harness design that telegraph the evaluation setting to the model. The thread uses Anthropic-related examples to argue that this issue is visible directly in the model's thinking traces.
Related event: Researchers Propose New Framework for AI Evaluation Design(3 posts)→
More from Research
- arXiv moderator's tip: accept the classifier's primary category or risk a hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27