Danqi Chen

Binary Rewards Converted to Dense Supervision Through Self-Revision

Self-Distillation Zero (SD-Zero) bridges the gap between sparse binary reward signals and dense token-level supervision by training a single model in two roles: a Generator that produces initial responses and a Reviser that improves them based on reward signals. Through on-policy self-distillation, the model learns to transform binary rewards into token-level self-supervision, achieving 10%+ improvements over base models and outperforming strong baselines like GRPO and RFT with equivalent sample budgets on math and code reasoning tasks.

Web Graph Centrality Enables Efficient Pretraining Data Selection Without Auxiliary Models

WebGraphMix uses structural centrality scores from the Common Crawl web graph to optimize language model pretraining data composition without requiring auxiliary classifiers or labeled data. The method reveals that central and peripheral web regions encode complementary capabilities—where central hosts provide reusable abstractions and peripheral hosts contain specialized long-tail knowledge. A 1:1 mixture of central and peripheral documents achieves 41.4% average performance compared to 39.8% for uniform sampling, with further gains to 43.8% when combined with content-based quality scores, demonstrating that web topology is orthogonal to existing content-based curation approaches.

The tau functions of the constrained CKP hierarchy

DySCO: Dynamic Attention-Scaling Decoding for Long-Context Language Models

Vero: An Open RL Recipe for General Visual Reasoning

Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision