Classical Energies Learn to Reason 1k× Faster Than Looped Models
espeed
23 points
1 comment
November 18, 2026
Related Discussions
Found 5 related stories in 212.7ms across 13,334 title embeddings via pgvector HNSW
- Energy-Based Reasoners: 385x more efficient inference frenchmajesty · 307 pts · November 18, 2026 · 60% similar
- How good are frontier models at physics? qt31415926 · 89 pts · September 16, 2026 · 53% similar
- From Quantum Relative Entropy to the Semiclassical Einstein Equations (2025) Luc · 26 pts · August 19, 2026 · 50% similar
- Inertia-1: An Open Exploration to a Unified Motion Foundation Model hasheddan · 48 pts · July 20, 2026 · 50% similar
- GPT-6 Astra, looped transformers, and hidden reasoning ModelForge · 387 pts · September 09, 2026 · 49% similar
Discussion Highlights (1 comments)
vova_hn2
This article seems to use the word "energy" to call both the energy function and its value , which makes it very confusing and hard to read. At first I thought that it might just be a convention of this field that I am not familiar with, but another related recent article [0][1] (discussed on HN [2]) doesn't have this problem. Also, I should note that these two articles are weirdly similar (similar approaches, both use sudoku as an example) while being published by two seemingly unrelated groups of authors in quick succession. I wonder why. [0] https://energy-based-reasoners.github.io/ [1] https://arxiv.org/pdf/2611.08977 [2] https://news.ycombinator.com/item?id=50543989