Hong-Lin Lin

Failed medical reasoning trajectories, when intelligently repaired using knowledge navigation, unlock substantial accuracy gains beyond standard distillation.

Recursive self-improvement on ECG models boosts accuracy by over 20 percentage points by dynamically evolving supervision from direct reinforcement to confusion-focused contrasts.

Chasing large token-level teacher-student gaps in agent distillation is largely misguided: local divergence fails to predict task completion, whereas unpenalized minor deviations often trigger catastrophic multi-turn failure.

LLM datasets aren't independent islands: tracing their lineage reveals hidden redundancy, benchmark contamination, and opportunities for more diverse training data.

An 8B model can now generate scientific graphics code that rivals or surpasses the output of much larger proprietary models, thanks to a new dataset, benchmark, and reinforcement learning approach.

Forget scaling laws, targeted data engineering—specifically multi-stage distillation and difficulty-aware sampling—allows an 8B model to outperform larger open-source financial LLMs.

Forget simplistic synthetic data: ChartVerse generates complex charts and reliable reasoning data from scratch, enabling an 8B model to outperform its 30B teacher in chart reasoning.