# Three New Evaluations Show That Test Format Changes What Models Score on Safety and Social Reasoning

A mechanistic study finds that multiple-choice prompts move model activations about three times farther from the refusal direction than open-ended prompts, and a $20,000 deception tournament puts OpenAI's GPT-6 Astra 264 Elo points ahead of Anthropic's best model.

- Published: 2026-11-09T06:17:39.329Z
- Canonical: https://polylog.news/ai/2026-11-09/three-new-evaluations-show-that-test-format-changes-what-mod
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [arXiv cs.CL](https://arxiv.org/abs/2611.08639), [arXiv cs.AI](https://arxiv.org/abs/2611.08613)

Earlier studies found that safety benchmarks in multiple-choice question (MCQ) format draw fewer refusals. A paper posted November 5, "Forcing a Letter Suppresses Refusal", traces that effect inside the models. Stepan Epifantsev and four co…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-11-09/three-new-evaluations-show-that-test-format-changes-what-mod (subscription information: https://polylog.news/pricing).