Energy-Based Reasoners: recurrent reasoner claims 345x less inference compute than the next-best modelImage: energy-based-reasoners.github.io
Cooling

Energy-Based Reasoners: recurrent reasoner claims 345x less inference compute than the next-best model

Heat 50 · #46 in Tech · Cooling3 reports · 3 sources · First reported Nov 6 · Updated Nov 20

The story

A paper by Duke researchers proposes Energy-Based Reasoners, which learn an energy function over latent reasoning states and refine their answers through looped updates that perform gradient descent on it. On Sudoku-Extreme, the authors report 99.9% accuracy with 345 times less inference compute than the next-best model, from a model with 1.8 million parameters. Some headlines put the compute figure at 385 times; the paper's own number is 345. The results cover algorithmic puzzle benchmarks rather than open-ended language reasoning, and the claims rest on the paper itself, with no independent verification yet.
Latest development
cooling cleared, status active (round 35 revival precedent)

How it unfolded

  1. cooling cleared, status active (round 35 revival precedent)
  2. back in HN top 30 after cooling in round 188
  3. Lobsters #19 same project page (energy-based-reasoners.github.io), NEW platform for this row, cross +1
  4. same post, same pts/cmt as round 162
  5. single-platform entry on HN score alone per Switcher/JEPA/Lonewolf precedent
  6. byteseu/peacknews cover a different concurrent paper arXiv:2611.17309 that only cites EBR)
  7. Nov 18
    search 2026-11-18 found no independent media coverage (arxivb.org and latent-implicit-thinking.github.io are arXiv aggregator mirrors, not a second platform
  8. HN title says 385x, the paper's own number is 345x - discrepancy flagged
  9. 99.9% accuracy on Sudoku-Extreme with 345x less inference compute than the next-best model, 1.8M parameters
  10. Nov 6
    paper arXiv:2611.08977 (Chen et al., Duke, submitted 2026-11-06): Energy-Based Reasoners learn an energy function over latent reasoning states and do looped updates as gradient descent on it

The angle

A tiny model claims 99.9% on Sudoku-Extreme with 345x less compute - what energy-based reasoning is, and why the HN title's 385x doesn't match the paper's own 345xSuggested format: Short video · for AI followers, developers, ML researchers

Heat

Now 50 · Peak 50 · 64 recorded checks since Nov 20
Full ranking →