25launches

The first launch of Energy-Based Reasoners.

Energy-Based Reasoners page

Energy-Based Reasoners

Launching today

A 1.8M-parameter model that cracks the hardest Sudoku

Hunted by IlyasHunter
Visit energy-based-reasoners.github.io (opens in a new tab)

Energy-Based Reasoners (EBRs) are a research paper with code from Duke, UIUC and Harvard. Instead of looping a network a fixed number of times, an EBR learns an energy score for its current reasoning state, refines it by gradient descent, stops when it settles and keeps the lowest-energy answer. With 1.8M parameters it solves 99.4% of Sudoku-Extreme in one try and 99.9% with a few samples, using 345x less inference compute than the next-best model. Code is MIT; training needs a CUDA GPU.

Meet the team

Hunted by Ilyas. The makers are not on 25launches yet.

H (Hunter)
  • Energy-Based Reasoners screenshot 1
  • Energy-Based Reasoners screenshot 2
  • Energy-Based Reasoners screenshot 3

Discussion 1

Ask questions, give feedback. Comments don't count in the ranking.

Sign in to join the discussion.

  • IlyasHunter Pinned

    Small looped models now beat large LLMs on puzzles like Sudoku, but they have no sense of when to stop thinking. This paper gives them an energy score for the current guess, which also tells them when to stop and which of several answers to keep. It's research, not a tool: the code and checkpoints are on GitHub (87 stars), and ARC-AGI results are announced for later.