Self-Distillation Zero: New Method Converts Binary Rewards into Dense Training Supervision
Researchers propose Self-Distillation Zero (SD-Zero), a training method that enables a single model to learn from binary rewards without requiring external teachers or high-quality demonstrations. The approach trains a model in two roles—generator and reviser—then distills the reviser's improvements back into the generator using token-level supervision. On math and code benchmarks, SD-Zero achieves at least 10% improvement over base models and outperforms existing methods like GRPO and Rejection Fine-Tuning.
Self-Distillation Zero addresses a key limitation in current post-training methods: binary reward signals provide sparse supervision, while dense supervision typically requires costly external teachers or high-quality demonstrations. The method trains a single model to function as both a Generator (producing initial responses) and a Reviser (improving responses based on binary rewards). Through on-policy self-distillation, the reviser's token distributions are used to supervise the generator, effectively converting sparse binary rewards into dense token-level supervision. Testing on Qwen3-4B-Instruct and Olmo-3-7B-Instruct models shows consistent improvements of at least 10% over baselines, with the method outperforming Rejection Fine-Tuning, GRPO, and Self-Distillation Fine-Tuning under equivalent training budgets. The approach exhibits two key properties: token-level self-localization (identifying which tokens need revision) and iterative self-evolution (improving revision ability through regular teacher synchronization).
What's missing
The study does not discuss computational costs or training time comparisons relative to baseline methods, nor does it address potential limitations when scaling to larger models or more complex reasoning tasks beyond math and code.
What different sources said
- arXiv cs.CLCenter
Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.