HKUDS/LightReasoner
A post-training framework where small language models generate reasoning traces to teach large language models, improving their reasoning ability.
Velocity · 7d
+2.4
★ / day
Trend
→steady
star history
LightReasoner is an ACL 2026 Oral paper exploring knowledge distillation for reasoning in LLMs. The framework uses smaller models to generate and label reasoning data that is then used to fine-tune larger models, with a focus on token efficiency. It releases trained checkpoints and evaluates against reasoning benchmarks like Qwen2.5-Math baselines.