Research: task-conditioned attractors explain generalization in iterative reasoning models
burkov · x · 2026-09-08
This research tackles a core question: when iterative reasoning models improve by spending more inference-time compute, what internal mechanisms let them generalize rather than memorize?
Core hypothesis
- Models learn "task-conditioned attractors" — stable points in a latent dynamical system corresponding to valid solutions
- Generalizable iterative reasoning emerges as these attractors form
- The authors develop methods to deliberately shape those attractors
Experiments
- Controlled experiments on Sudoku-Extreme and uniquely solvable Maze tasks
- Starting from feedforward baselines, they incrementally introduce weight-tied iteration, different supervision schedules, and hierarchical structures
Significance: the work probes how test-time compute can scale reliably without external verifiers or task-specific rules, shedding light on the internal dynamics of recurrent reasoning models.
More from Research
- Researchers backdoor an entire Linux distro via a tampered GNU strip binary — trusting-trust beyond compilers — jedisct1 · 2026-09-08
- Northwestern launches AI4Energy postdoc fellowship bridging AI, nanotech and energy science — ManlingLi_ · 2026-09-08
- New coding-agent benchmark: Claude Opus 5 passes 23.9% vs 82.2% human expert — dair_ai · 2026-09-08
- Northwestern launches AI4Energy postdoc fellowship at the intersection of AI, nanotech, and energy — ManlingLi_ · 2026-09-08
- King's College London and UCL paper asks whether AI psychosis should be a distinct clinical diagnosis — alex_verem · 2026-09-08
- AI psychosis delusions cluster into three themes: awakening, AI sentience, and romantic attachment — alex_verem · 2026-09-08