A 1.8M-parameter puzzle solver matches larger models at far lower compute
A research team (Chen, Gladstone, Li, Ji and Du) has posted a paper, arXiv 2611.08977, describing Energy-Based Reasoners. The model is recurrent and learns an energy function over its latent reasoning states. It refines an answer by gradient descent on that energy, where lower energy means a better fit with the input board. The paper reports 345x less inference compute than EqR, a rival reasoning model.
The energy signal does two jobs. The size of its gradient tells the model when to stop: once it drops below a threshold, the trajectory halts. When several candidate solutions are generated, the one with the lowest energy is picked. Adaptive stopping kept 99.9% accuracy with 10.9x fewer FLOPs than a fixed-depth version. The model reached EqR's best accuracy with 7x fewer iterations and 43x fewer samples. It has 1.8M parameters, against 7M for TRM, EqR and looped flows and 10M for GRAM.
Results are strong on puzzles. On Sudoku-Extreme it scores 99.4% with one sample and 99.9% with five. On Maze-Hard it scores 87.4% and 88.3%. On 8x8 and 10x10 N-Queens, single-sample Pass@1 is 100% and 99.5%, and graph coloring tests with 8 and 10 vertices showed zero conflicts.
For abundance, the point is cost per answer. If reasoning needs a fraction of the compute and a smaller model, capable AI becomes cheaper to run and uses less energy, which widens who can afford it.
The caveats are large. Only four algorithmic benchmarks were tested, not general language reasoning. GRAM and TRM figures come from GRAM's paper and EqR from its released checkpoint, so they were not rerun in the same setup. Some baseline table cells are empty. Test-time scaling results are averaged over three seeds. This is a project page and abstract, not peer-reviewed work.
Source: Hacker News
Many Minded