Energy-Based Reasoners: 99.9% on Sudoku-Extreme with 345× less inference compute
burny_tech · x · 2026-11-11
A Duke/UIUC/Harvard paper introduces Energy-Based Reasoners (EBRs), which define an explicit energy function over latent reasoning states and perform looped updates as gradient descent on that energy surface. The energy gradient guides thought refinement while the scalar energy value provides an intrinsic stopping signal — fixing the "blind looping" of prior recurrent reasoners like HRM/TRM. With exploration-based training, EBRs match or beat prior recurrent reasoning models across four benchmarks, hit 99.9% on Sudoku-Extreme using 345× less inference compute, and achieve SOTA distinct-solution coverage on multi-solution tasks. Code is open-sourced.
More from Research
- Four matrix factorizations — LU, QR, SVD, Cholesky — explained in one chart — burny_tech · 2026-11-11
- Generative AI editing other models' weights: tiny neuron edits do improve performance, but modestly — zakkohane · 2026-11-11
- Reverse-engineering Apple's on-device personal knowledge graph: 181 entity classes, 278 predicates — JeremyCMorgan · 2026-11-11
- SAGE: Experience-Learning Agent Solves 84/88 NPU Kernel Synthesis Tasks on Ascend 910C — jiqizhixin · 2026-11-11
- xkcd-artist: a 1.72MB browser model plans stick-figure comics in 0.55ms — init0 · 2026-11-11
- Researchers argue in Nature Human Behaviour that AI 'slop' is a social judgement, not a natural category — manoelribeiro · 2026-11-11