News without the noise
Benchmark
Reported in 102 papers on Zotpaper. The latest, each as the paper states it:
15 results for “GSM8K”
Research · Reinforcement learning
RLVR Sharpens the Pencil, Not the Mind
Tracking one small model through arithmetic training, the authors find modest accuracy gains but no measurable improvement in internal signals distinguishing flawed reasoning.
GSM8K: 3.3 pts pass@1 improvement across training (as reported)
Indian Institute of Technology Roorkee
Research · Language models
Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining
Diagnostic phrases in failed reasoning traces retrieved a compact training corpus that improved subsequent reinforcement learning across model families.
GSM8K: 2.4 pts accuracy gain over matched control (as reported)
Princeton Language and Intelligence
Research · Reinforcement learning
Think Less, Locally: Local Preference Learning for Reasoning Compression
A new method teaches large language models when to conclude reasoning rather than continue, cutting output length by up to 45% on math and 36% on code without losing accuracy.
AIME24: 54.17% Accuracy (as reported)
Xi’an Jiaotong University, Tencent Inc.
Research · Evaluation & benchmarks
What pass@k Cannot Measure: Evaluating Diversity and Capability Retention after Post-Training
GRPO and rejection-sampling fine-tuning produced sharply different output distributions despite often similar pass@k scores.
GSM8K: 0.51 distinct-3 (verifier-correct completions) (as reported)
Accenture Strategy & Consulting, Vizuara AI Labs
Research · Language models
TeachBack: Learning to Teach via Bidirectional Preference Optimization for On-Policy Self-Distillation
The TeachBack framework uses bidirectional preference optimization to encourage teachers to produce exploration-rich reasoning trajectories that students can learn from.
AMC23/AIME24/AIME25: 74.21% mean@4 (as reported)
Huazhong University of Science and Technology, Tsinghua University, Harbin Institute of Technology
Research · Language models
Structural Chain-of-Thought: Reasoning with Executable Computational Graphs
The Structural Chain-of-Thought method trains models to generate executable directed acyclic graphs, improving accuracy, efficiency, and faithfulness across multiple benchmarks.
GSM8k+: 0.35 accuracy (as reported)
MIPT, AIRI
Research · Language models
Learning from Think-Mode Advantage via On-Policy Distillation
Authors propose ThinkOPD, which ranks student responses by outcome benefit and trace-response compatibility, outperforming uniform distillation and baselines on math and code tasks.
AIME24: 35.2% avg@16 (as reported)
Research · Reinforcement learning
ReFlow: Replay-Driven Decoupling for Efficient RL Post-Training
ReFlow, a decoupled and adaptive system, achieves up to 2.06× throughput improvement over existing methods while maintaining model quality.
DAPO-Math-17K: 15.8% trainer idle ratio (as reported)
The University of Hong Kong, University of Science and Technology of China, Alibaba Group
Research · Language models
Compatibility Matters: A Framework for Token-Level Teacher Selection in On-Policy Distillation
A theoretical framework and new routing scheme yield up to 10 percentage point gains in pass@k across math, code, and reasoning tasks.
AIME 2025 × KernelBench L2: 29.3% joint fast0 pass@16 (as reported)
Princeton University, Amazon Annapurna Labs
Research · Language models
Learning to Revise Reasoning with Segment-wise On-Policy Distillation
Replacing student reasoning segments with teacher redrafts boosts accuracy by 5.22% over baselines.
AIME25: 59.7% accuracy (as reported)
The Hong Kong University of Science and Technology (Guangzhou), University of Science and Technology of China, Alibaba Group
Research · Reinforcement learning
Residual Advantage: Student-Relative Teacher Guidance for RL with Verifiable Rewards
Chen and Pang show that centering the teacher's probability residual within each response allows step-level credit redistribution while preserving verifier outcome labels, improving reasoning performance across multiple benchmarks.
Macro: 53.02% Pass@8 (as reported)
Harbin Engineering University, Tencent, Harbin Institute of Technology
Research · Reinforcement learning
Lexicographic Multi-Objective On-Policy Distillation
Method routes student rollouts to highest-priority unmet specialist and projects corrections to avoid interfering with higher-priority experts.
AIME 2024, AIME 2025, HMMT February 2025: 28% retained gain, composite conciseness (as reported)
Cohere, Mila, Université de Montréal
Research · Reinforcement learning
Outcome-Guided On-Policy Self-Distillation
The authors propose OG-OPSD, which selects forward or reverse KL divergence based on answer correctness and shortens distillation on incorrect trajectories using teacher entropy, outperforming vanilla on-policy self-distillation.
AIME 2024: 77.5% Avg@12 (as reported)
Southeast University, Huawei Noah’s Ark Lab, Zhongguancun Academy
Research · Reinforcement learning
Privy to the Foil: Recasting Value Estimation with a Self-Privileged Critic for RLVR
πPPO reuses verified same-prompt rollouts as contrastive evidence, boosting value quality and policy performance on math reasoning.
AIME 2024: 73.5% Pass@1 (as reported)
Peking University, Tencent
Research · Reinforcement learning
Mitigating the Length-Scaling Tax with Online Distillation
Length Self-Distillation reduces excess response length on easy queries without sacrificing accuracy on hard ones.
BrowseComp-Plus: 30.72% Pass@1 (as reported)
ByteDance Seed, Tongji University, Peking University