Transcription
Samsung AI research team introduces tiny recursive model TRM, a 7 million parameter two-layer solver that replaces hierarchical reasoning models HRM's dual module hierarchy with a single recurrent core.
At each outer step, it first drafts a full answer, then runs approximately six latent scratchpad updates, z = f(xyz), think before refining the solution. Y = G(YZ) act.
The think act block is unrolled up to 16 times with deep supervision and a halting head during training and gradients are propagated through the recursion rather than using fixed point approximations.
On public ARC evaluations, TRM reports 45% test accuracy on ARC AGI1 and 8% two try on ARC AGI 2, improving over HRM's 40% and 5% while using far fewer parameters. It also reaches 87.4% on Sudoku Extreme and 85.3% on Maze Hard.
The paper sites lower scores for several much larger LLMs under its protocol. For example, Gemini 2.5 Pro, 03 Mini High, DeepSeek R1.
Practically, this research inclines in spending inference FLOPs on recursive refinement over parameter count for structured reasoning. The pattern is a tiny recurrent core with deep supervision across up to 16 outer steps, approximately six latent updates per step, EMA stabilized training, and full backrop through the loop.
[Music]