OC-OPD improves on-policy distillation via observation-coordinated supervision
OC-OPD uses reverse and forward KL divergences with a four-observation matrix to coordinate gradients, achieving strong gains in mathematical reasoning and OOD transfer.
Stories per day · 30 days
2 RSS
Relatedllm-training2qwen32arabic1benchmark1criticopd1dialectal-arabic1distillation1himath1kl-divergence1llm-eval1model-training1observation-coordination1on-policy1on-policy-distillation1qwen1reinforcement-learning1reward-design1self-distillation1
14 stories, newest first · one card per story
Stories tagged #mathematical-reasoning
MTA-OPSD lets each token pick its own teacher in self-distillation
On Qwen3 math tasks, the method beats prior on-policy self-distillation, with gains of up to 5.9 over the base model.
H2R uses hierarchical hints as rewards for math reasoning
Qwen2.5-Math-7B rises from 30.4% to 39.1% on HiMath with hierarchical hints, and to 45.6% with oracle hints.
DAR-MGSM benchmark tests math reasoning in Arabic dialects
Eleven models generally scored worse on culturally localized problems than on translated-only versions.
Recovering Reliable Supervision via Critique-Guided On-Policy Distillation
CriticOPD uses teacher critiques to restore recoverability in on-policy distillation, lifting two Qwen3 students on math benchmarks.
Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVR
EDEN adds anchored embedding noise to RLVR rollouts, improving exploration and solution coverage over GRPO.
When Tools Lie: Reliability of Mathematical Agents Under Corrupted Tool Feedback
Corrupted tool feedback drops math-agent accuracy from 100% to 72.4%; mandatory reflection restores 100%.
The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models
A new benchmark and self-distillation method diagnose and repair weak mathematical reasoning in LLMs.
The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models
A primitive-based diagnosis finds Discovery is the main bottleneck in LLM mathematical reasoning and shows it can be repaired.
Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation
Neighborhood OPSD lifts Qwen3 math scores by routing student prefixes to nearby expert teachers.
Smaller Models, Better Rejects: Preference Distillation Scaling
Smaller frozen models produce more useful preference-distillation rejects than a student's own failures on code and math.
Order-Invariant Answers, Order-Sensitive Representations in Mathematical Reasoning
Across 16 models from 1B to 8B, higher math accuracy tracks more distinct internal encodings of equivalent rule orders.
When and Where to Trust the Teacher: Unifying On-Policy Distillation and GRPO through Entropy-Calibrated Credit Assignment
UECR-GRPO blends teacher and verifier credit, slightly improving Qwen3 math-reasoning accuracy.
An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics
Open post-training pipeline turns Nemotron 3 Ultra checkpoints into an IMO 2026 gold-medal system, scoring 30/42 without formal provers or external tools.