Most of the reasoning gains coming out of the big labs are still tied to scale. More params, more compute, better reasoning. That's been the play for a while.
Ran into TwIL-LM2 which flips the script for narrow tasks. PEFT LoRA adapter on SmolLM2-1.7B, specialized purely for formal logic translation. On strict-7 scoring (no partial credit, exact-format required) it hits 0.2386 - ahead of Qwen3-8B at 0.2093 and Gemma-4-26B at 0.2050. On the loose-match six-lane average it's a different story (Qwen3-8B still wins there) but for the "actually usable formal output" measurement, the 1.7B leads.
Makes me wonder how much of the "we need bigger models for reasoning" narrative is actually about complex multi-step reasoning vs. just having enough capacity to hold multiple approaches. If you can specialize hard on one reasoning task and lead 8B+ models on the strictest scoring at 1.7B, that's real efficiency.
Kind of hoping this becomes a trend. A pipeline of narrow specialists on 1-3B models sounds a lot more practical than routing everything through a 70B.
Non-commercial license, worth flagging. Anyone doing something similar with narrow fine-tunes? What tasks have you found respond well to this approach?
[link] [comments]