Research
CosmicFish-HRM
A study of adaptive reasoning in compact language models. A hierarchical recurrent mechanism lets the model decide how much computation each input deserves, continuing or halting its reasoning rather than spending a fixed amount on every request.
The model
82.77M parameters · transformer · GQA · RoPE · SwiGLU · RMSNorm
Adaptive depth with a learned halt/continue mechanism, up to 16 reasoning steps.
Adaptive computation
Reported mean reasoning steps by task:
| Simple factual completion | 2.78 |
| Scientific concepts | 4.77 |
| Logical reasoning | 7.03 |
| Multi-step arithmetic | 8.40 |
Paper