Hacker News AIT2·Models
78 pts
Researchers introduce Energy-Based Reasoners (EBRs), which learn an energy function over latent reasoning states and cast looped updates as gradient descent on that energy, using the energy value as a stopping signal. Across four algorithmic reasoning benchmarks, EBRs match or beat prior recurrent reasoning models, reaching 99.9% accuracy on Sudoku-Extreme with 345x less inference compute and achieving state-of-the-art coverage of distinct valid solutions on multi-solution tasks.