← The Polylog AI Intelligence Brief
Morning Edition · Monday, October 5, 2026Published at 2:00 AM EDT · New York
Two papers find supervised fine-tuning teaches surface style and can widen a falsification gap that reinforcement learning closes
One study explains why supervised fine-tuning (SFT) causes more forgetting than reinforcement fine-tuning (RFT), even when every demonstration is correct. The other shows that imitation makes models worse at disproving false theorems.
Two papers posted on Monday reach the same conclusion about post-training from different directions. Imitation learning transfers surface form. Reinforcement against a verifier transfers the target behavior. The first, Learning Style, Forge…
Continue the AI Intelligence Brief
Track frontier labs, chips, export controls, model releases, regulation, and AI infrastructure.
- 5 AI intelligence signals a day
- Frontier labs, compute, and chips
- Model releases and AI infrastructure
- Source-grounded analysis with confidence labels
The Global Intelligence Brief stays free.
Part of a tracked trend
Reinforcement Learning Displaces Imitation in Post-Training
Post-training budgets will keep moving from supervised imitation to reinforcement learning against verifiers, because imitation copies surface style and can erode knowledge, which makes verifier construction a recurring competitive asset.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Anthropic and OpenAI price their near-flagship models at $2 and $10 per million tokens while open-weight Qwen runs on a gaming GPU
- Hinton says Trump is "badly advised" on AI after the White House replaced rules with a voluntary self-policing accord
- New papers show jailbreaks and GraphRAG poisoning work by composing benign parts, a pattern OpenAI says it saw in a reasoning-extraction campaign
- Sonnet 5.5's jump from 10.3% to 70.6% on Terminal-Bench 4.0 tests how far vendor agent evals can be trusted
- MIRROR paper defends multi-agent systems against agent-in-the-middle tampering as harnesses hand routing and injection checks to fast single-pass models
- Anthropic's 2030 scenarios put US output up to 32.4% above a no-AI path while labor's share of income falls to 45.2%
- Meta sells its closed-weight Muse Spark 1.3 at $1.25 and $4.25 per million tokens through an OpenAI-compatible API
Comments
0No comments yet.