Guiding machine design using insights from simple sandboxes
summary
Research uses minimal finite-automaton sandboxes to map which sequential tasks Transformers can implement exactly, proving solvable automata decompose into two base types: mod counters (one-layer solution with uniform attention) and memory units (implemented with sparse attention). It shows number of Transformer layers needed is bounded by layers ≤ Q^2 independent of sequence length, but finds practical failings: OOD robustness breaks unless OOD examples are included in training, RNNs perfectly solve some cases Transformers struggle with, and intermediate-checkpoint distillation speeds training.
AI-written summary · about
chapters (10)
- 0:00Bridging theory and practice
- 1:06Simple sandboxes defined
- 2:29Finite state automata intro
- 4:16Transformers for reasoning
- 6:24Efficient automata decomposition
- 7:24Transformer failure diagnosis
- 9:30RNNs vs Transformers
- 10:55Sparse parity insights
- 11:38Training acceleration
- 12:21Q&A
related talks
01h 01mUnderstanding Transformers through the Lens of Logic and Automata
38mTalk by Pranjal Awasthi (Google)
40mAn Occam's Razor Principle for Transformers?
26mRethinking the Transformer: Toward Native Multimodal Architectures
46mEvaluating Patterns in Gen AI Outputs
17mCode World Model: Building World Models for Computation
39mComputational Thinking on Learning Models
28mEverything you need to know about Transformer Models