New benchmark probes LLM self-modeling: RL lifts open models but counterfactual errors persist
dair_ai · x · 2026-09-07
A paper dair-ai highlights tests whether models can answer verifiable questions about their own behavior, e.g. whether a specific prompt edit would change their final answer — avoiding unverifiable introspection claims.
- A new benchmark covers diverse self-modeling question types; current models show real but limited skill and consistently fail simple counterfactuals about themselves.
- A scalable synthetic-data pipeline plus reinforcement learning improves aggregate scores across three open-source model families, with some transfer to held-out tasks.
- The authors decline the tempting interpretation: the gains may not come from privileged access to the model's internal decision process.
More from Models
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
- supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
- Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26
- AI search APIs ranked: top 12 within 7 points, yet prices differ nearly 7x — FinanceYF5 · 2026-10-26