math-reasoning

GPT-6.1 Sol claims the top spot on MathArena with 86.3% accuracy at a cost of just $0.94, outperforming Astra's 81.9% achieved at $2.26. The result highlights a significant advance in math reasoning performance and cost-efficiency among frontier models.

The paper introduces Neighborhood On-Policy Self-Distillation (N-OPSD), which exploits local parameter perturbations of a privileged teacher to pool complementary reference-aligned corrections and route them to student-visited states, improving math reasoning benchmarks (AIME 2024/2025, HMMT Feb 2025) across Qwen3 1.7B/4B/8B models.

This paper analyzes on-policy self-distillation (OPSD) for LLM reasoning and shows that unverified student scaffolds create an imitation gap that worsens with model scale, motivating OASIS, a method that supervises mostly verified by-label on-policy trajectories using model-generated attempts as teacher context, improving Qwen3 1.7B/4B/8B by 3.2-3.8 points on AIME/HMMT benchmarks.

This paper proposes SCALE (Selective Control of Adaptation via Local Entropy), an entropy-guided method that freezes the pretrained model and SFT delta while learning bounded token- and module-specific gates to suppress, reverse, or extrapolate learned features during supervised fine-tuning. The approach achieves strong math-reasoning and code-generation results on Qwen2.5-Math and Qwen3-4B-Base models, outperforming existing token-reweighting baselines.

The paper introduces Allspark, a training and inference framework for transferring reasoning improvements from a weak model to a strong model via alternating chains of thought, with experiments demonstrating accuracy gains in math, reasoning, and ARC-AGI-2 tasks.

This paper investigates whether large language models internally organize mathematical reasoning by topic or approach, finding evidence that approach is the key organizing principle, challenging traditional benchmarking methods.

Kyutai Labs releases Voice of Reason, a speech-native model that solves spoken math problems without transcription or text LLM, achieving 77.1% on GSM8K, up from GLM-4-Voice's 27.3%.

This paper addresses advantage scale calibration in group-relative optimization under low-variance rewards, proposing methods like the Reward-Resolution Protocol and MaxNorm-AC to filter sub-resolution jitter and provide bounded recovery for credible gaps.

This paper proposes a four-stage mechanistic decomposition of how large language models solve math word problems, localizing fragility from irrelevant clauses to the Operation Planning stage via specific attention heads.

Fine-tuning small AI models on logic puzzles derived from zebra puzzles yields significant improvements in math benchmarks like MATH-500, with open-source resources and notebooks for reproduction.

ZGCM-1 is a 7B open foundation model trained from scratch with extreme efficiency, combining internal reasoning and external tool use for math and agentic search tasks, achieving competitive performance with much larger models like Qwen3-235B-A22B and GLM-5.1.

FlowBalance introduces a verifier-grounded self-improvement technique that improves math reasoning performance by an average of 2.12 over GRPO on the Qwen3-8B model, offering faster training, enhanced stability, and greater solution diversity.

This paper compares prior injection methods for sparse-reward reinforcement learning in vision-language math reasoning, finding that effectiveness depends on delivery and that certain evaluation slices can mislead generalization assessments.

This paper presents SimpleOPD, a method for on-policy distillation from long-context reasoning teachers to short-context students, overcoming tokenizer mismatch and training instability to enhance mathematical reasoning capabilities.

This paper introduces SFS-DPO, a reinforcement learning two-stage framework for step-level self-verification and self-correction in LLMs, with a teacher-assisted variant SFS-DPO-R. It demonstrates improvements in self-correction effectiveness across multiple LLMs with less training data than prior approaches.

Introduces U-OPSD, an unsupervised on-policy self-distillation method that uses internal consistency and majority-vote pseudo-solutions to improve LLMs without external supervision, matching or exceeding supervised methods on math benchmarks.

Introduces AuroOFT, a method that enhances quantized orthogonal fine-tuning (QOFT) with a zero-start gated low-rank nonlinear residual, improving mathematical reasoning accuracy over QOFT and QLoRA while using fewer trainable parameters on low-bit language models.

Proposes Woodpecker Distillation, a weak-to-strong training framework that uses weak probe models to identify and repair local reasoning bugs in stronger models via contrastive local interventions, improving performance on math reasoning benchmarks.

This paper studies On-Policy Delta Distillation (OPD^2) for multilingual math reasoning in English, Korean, and Japanese, showing consistent improvements over standard OPD and narrowing language gaps.

This paper proposes LSPO (LoRA Scaffolded Policy Optimization) to solve the vanishing-gradient problem of GRPO on zero-reward "cliff problems": a temporary LoRA adapter is used for a short SFT, successful trajectories are sampled, corrected via importance sampling, and then fed back into the RL batch; ultimately, only the base model is updated. Experiments on DeepMath-103K show it outperforms the DAPO baseline by an average of 3.8 points.