Trajectory researcher identifies 'but wait' model collapse in self-distillation

Scaling policy self-distillation to trajectories with 100+ tool calls causes models to collapse into hedging. The model begins favoring tokens like 'wait', 'but', and 'maybe', leading to a degradation in performance known as the 'but wait' problem.
Featured · Ronak Malde
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- OpenAI-backed Thrive Holdings raises $2B to bring AI to the enterprise
- Open Instruct tutorial covers LLM post-training with SFT, DPO, GRPO
- Twitch streamers can now opt out from training Amazon's AI
- OpenWALDO project launches to create shared, open-source AI training dataset
- MIT Technology Review report: Legacy data systems limit AI agents