Energy-Based Reasoners: 385x more efficient inference
- ID
- 48151
- Status
- summarized
- Published
- 18 Nov 2026, 5:29 PM
- Fetched
- 19 Nov 2026, 9:23 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://energy-based-reasoners.github.io/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 6.5
- Created
- 19 Nov 2026, 9:23 AM
- Tags
- Audience
- ai_ml_learnersdevelopersai_agent_users
What happened
A Duke/UIUC/Harvard paper introduces Energy-Based Reasoners (EBRs), which learn an energy function over latent reasoning states and treat looped recurrent updates as gradient descent on that energy — so the energy both scores candidate answers and its gradient norm acts as the stop signal. With 1.8M parameters, EBRs hit 99.9% accuracy on Sudoku-Extreme using 345× less inference compute than the next-best model, and reach state-of-the-art coverage on multi-solution benchmarks (N-Queens, Graph Coloring) while keeping near-100% single-sample accuracy. Note the headline says 385× while the abstract and body say 345×. The Hacker News thread drew 298 points and 52 comments.
Why it matters
The concrete takeaway is about stopping criteria, not Sudoku: EBRs remove the external verifier and the fixed loop count by using gradient magnitude as the 'am I done' signal, and pick among candidates by lowest energy. If you build agent loops that retry until a checker passes, that's the pattern worth stealing — one learned score replacing separate verifier calls. Caveat before you get excited: every benchmark here is algorithmic (Sudoku, N-Queens, Graph Coloring) at 1.8M parameters, so nothing in the text shows this transfers to open-ended text tasks or to the model sizes you'd actually deploy.
Discussion angle
Ask whether the gradient-as-stopping-criterion idea survives outside puzzle domains — and why the 385×/345× numbers disagree in the same paper, since that's the kind of detail that decides whether you trust a benchmark table.