# New work examines whether reasoning distillation losses differ in weight space, not just accuracy

A paper asks whether offline RL objectives that transfer reasoning from large teachers to small students produce genuinely different models, beyond their downstream scores.

- Published: 2026-06-24T10:42:52.470Z
- Canonical: https://polylog.news/ai/2026-06-24/new-work-examines-whether-reasoning-distillation-losses-diff
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv (Weight-Space Geometry)](https://arxiv.org/abs/2606.23740), [arXiv (EXPO-SQL)](https://arxiv.org/abs/2606.23693)

An arXiv paper on the weight-space geometry of offline reasoning training takes up a question the field usually skips. Offline reinforcement-learning (RL) objectives such as RFT, DFT, offline GRPO and DPO are widely used to distill reasonin…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-06-24/new-work-examines-whether-reasoning-distillation-losses-diff (subscription information: https://polylog.news/pricing).