Untuned models solved agent tasks their polished versions could not.
A paper posted on 1 October compared 14 pairs of models, each a raw version and its tuned twin. With enough attempts, raw models cracked tasks the tuned ones never solved at all. The authors call the gap a sharpening tax and offer a sampling fix for it.
Why it mattersFine-tuning makes a model more reliable on the first try and narrower overall. If your agent keeps failing a task, a plainer model may still get there.
Read in the app — free, in 9 languages
Related stories
Meta open-sourced the firmware for building your own Muse gadget.2026-10-04Apple is locking down the Mac permission that AI agents ask for.
2026-10-04Pretraining on everyday photos got a model to 70% on a reasoning test.
2026-10-04DeepSeek put its agent on your desktop and opened the code.
2026-10-03Cloudflare gave away two models that only choose, never chat.
2026-10-03