HORIZON LLM
GitHub
THE HORIZON NOTEBOOK
Research, ideas, and progress toward more capable intelligence.
4 articles
From searching more to refining better. An agentic verifier turns failed solutions into actionable feedback, guiding competitive-programming agents toward the IOI gold-medal line.
Higher scores are only part of the story. Harness-Delta Attribution separates gains from overfitting, additional test-time compute, and improvements that generalize.
Learning even when success is rare. RESD turns feedback from failed attempts into reflections and reusable lessons, giving a model richer supervision through self-distillation.
What makes Kimi’s policy updates work? An analysis of PMD reveals implicit mixed KL–χ² regularization, explaining more conservative updates and robustness when rollout samples are limited.
Try a different keyword or explore all posts.
