The Pitfalls of Next-token Prediction
summary
Argues that next-token prediction splits into distinct phases—autoregressive inference and teacher-forcing training—and that teacher‑forcing can provably fail to learn even simple planning tasks. Builds a minimal path‑star sequencing task (adjacency lists with randomized IDs) and shows both Transformer and Mamba architectures trained with standard next‑token objectives hit trivial accuracy (~1/degree) on held‑out in‑distribution examples, while reversing targets or training to predict multiple tokens in advance often recovers 100% accuracy. Presents empirical experiments and a proposed multi-token/teacherless fix.
AI-written summary · about
chapters (12)
- 0:00Next token prediction debate
- 2:33What both sides of the debate are missing
- 8:43Auto-regression vs teacher-forcing
- 9:05Planning task failure
- 17:46Clever Hans cheat
- 24:12First token hardest
- 33:49In-distribution failure evidence
- 34:12Teacherless multi-token training
- 43:09Reversed targets and autonomy
- 46:50Broader implications
- 51:32Key takeaways
- 52:14Q&A
related talks
40mAn Occam's Razor Principle for Transformers?
45mFrom Next Token Prediction to Reasoning and Beyond
46mEvaluating Patterns in Gen AI Outputs
01h 00mInside GPT – Large Language Models Demystified
39mComputational Thinking on Learning Models
08mLeveraging large-scale models for efficient learning
01h 05mCan you trust your (large language) model?
32mBeyond Algorithms – on Human(oid) Intelligence