1.7B Logic-Tuned Model Outperforms Qwen3-8B on Formal Logic

TwIL-LM2, a 72M-parameter LoRA adapter on SmolLM2-1.7B, translates natural language into verifiable first-order logic. It scored 0.2386 on Strict-7, beating Qwen3-8B and even a 26B model, showing specialized small models can outperform far larger general ones.

2026-08-15 ~ 2026-08-17 · 2 related posts

• 1.7B TwIL-LM2 Specializes in Formal Logic, Beats 26B Giants — itsalreadyhardenough · 2026-08-15
• 1.7B specialized model outperforms Qwen3-8B on strict logic benchmarks — Creative-Fig522 · 2026-08-17