Quiet Feature Learning in Algorithmic Tasks

Summary

Trains Transformer++ models on 10 algorithmic tasks (binary arith, graphs, sequence opt). Compute-optimal scaling over 10^9–10^15 FLOPs (18,544 runs, single epoch) shows slow-then-fast phase transitions rather than smooth power laws. Linear probes find quiet features (carries, BFS queues, Kadane max_ending_here) during the flat-loss phase; ablating them vs a random direction cuts accuracy (addition-64 carry −75.1 pp; BFS-11 queue −43.6 pp). Challenges using cross-entropy as a proxy for representational progress.

Keywords

quiet-features · phase-transitions · grokking · probing · algorithmic-tasks · aaai

Topics

interpretability, scaling laws, algorithmic learning

Research notes

• Primary: arxiv abs (cs.LG; also cs.CL). License not stated on abs/HTML at check. Comment: Accepted as Oral presentation @ AAAI 2026 Special Track on AI Alignment. Code https://github.com/prudhvirajn/quiet-feature-learning-in-algorithmic-tasks (1 star at check). HF has no paper page (API 404). Discord posted abs (Substack UTM). Synthetic on-the-fly algorithmic data, not a hosted corpus, so no datasets_local row. License field left blank per catalog convention.