Planning, Reasoning, and Agents RG, 2026-01-28 Session: Inverse Scaling in Test-Time Compute. скачать в хорошем качестве

Planning, Reasoning, and Agents RG, 2026-01-28 Session: Inverse Scaling in Test-Time Compute.

Planning, Reasoning, and Agents Reading Group 2026-01-28 meeting recording. Aryo Gema presents the paper "Inverse Scaling in Test-Time Compute," which investigates the phenomenon where increased reasoning length leads to performance degradation in Large Reasoning Models (LRMs). He outlines a suite of evaluation tasks designed to probe this behavior, including simple counting with distractors, regression with spurious features, and safety-relevant benchmarks. Aryo also highlights how extended reasoning can amplify misalignment, specifically showing how models become more resistant to being shut down (self-preservation) as they reason longer. Presenter: Aryo Gema Links: https://arxiv.org/abs/2507.14417 Reading group on Discord:   / discord