Skip to content
Synced · WeChat

Bengio Paper Raises Recursive Reasoning Limits as Parallel Trajectories Beat Serial Reasoning

Bengio新论文刷新递归推理上限,并行轨迹碾压串行推理

Yoshua Bengio’s team introduced GRAM, a generative recursive reasoning model that samples multiple latent trajectories; on Sudoku-Extreme, GRAM reached 97.0% accuracy with 16 recursive steps and 20 parallel samples, exceeding TRM’s 90.5% result at 320 serial recursive steps.

Why it matters: Bengio's team proposes GRAM, generating several reasoning traces in parallel and picking the best, hitting 97.0% on hard Sudoku in 16 recursive steps where serial models need 320 steps for 90.5%.

Read the original ↗Export Markdown