Tag
Introduces Weak-to-Strong On-Policy Distillation (W2S-OPD), a framework that improves a strong language model by distilling from multiple weaker models using contrast pairs in logit space, consistently outperforming standard on-policy distillation on math and code benchmarks.
This paper investigates multi-agent math reasoning systems, finding that a reviewer's precision in identifying errors does not guarantee that the solver will act on those critiques, revealing a gap between detection and effective uptake that limits overall performance gains.
CADENCE introduces a unified framework addressing cold-start collapse, state-agnostic scheduling, and reward sparsity in on-policy knowledge distillation for reasoning, achieving strong results on GSM8K and MATH-500 with compact student models on a single Mac Studio.
The AIMO Interpretability Challenge is a competition aimed at distinguishing robust from spurious reasoning in frontier mathematical language models using interpretability methods, providing new problems, model access, and computing infrastructure.
Presents an agentic framework using general coding LLMs to autoformalize research-level mathematics into Lean 4 code, evaluated on Putnam problems and STOC conference papers.
HuiHui, an ablated model variant, surprisingly outperforms the baseline 3.6-35B-a3b model on math and code tasks.
This paper introduces approach-level diversity for LLM math reasoning, showing that surface-level diversity metrics are unreliable proxies and that directly optimizing for approach diversity remains an open problem.
SHAPE is a framework that analyzes chain-of-thought reasoning in large language models using semantic spaces and heuristics to diagnose and improve mathematical reasoning through post-training.
Weibo AI releases VibeThinker-3B, a 3B parameter open-source reasoning model with MIT license, achieving competitive results on math, coding, and STEM reasoning benchmarks.
SuperThoughts compresses consecutive chain-of-thought tokens into latent representations and decodes two tokens per step, achieving ~20–30% CoT length reduction with minimal accuracy loss on math reasoning benchmarks, while doubling inference throughput.
ExpRL is a new RL-based mid-training method that uses human-written reference solutions as dense reward scaffolds (never shown to the policy) to improve LLM reasoning, achieving significant gains on hard math benchmarks like AIME-2026.
Epoch AI released a v2 update to the FrontierMath benchmark, correcting errors in 42% of problems and increasing scores across all models, though rankings remained largely unchanged; Tiers 1-4 are approaching saturation.
This paper presents a mathematical forum platform that integrates an image-to-LaTeX conversion pipeline directly into the posting interface, reducing friction for users. The system is designed to generate a community-validated dataset of math problems and solutions for training AI reasoning systems.
This paper introduces SWARR, a two-stage recipe using supervised fine-tuning and reinforcement learning to adapt sliding-window attention models for mathematical reasoning, showing that RL can narrow the performance gap with self-attention while maintaining efficiency.
该论文推出KCSAT-ML基准，包含十年韩国高考数学题及全国考生错误率，并提出难度对齐推理增益（DRG）指标，揭示模型错误与人类难度的对齐模式，展现相同准确率下截然不同的推理行为。
This tweet thread introduces research showing that training models to verify their own work can nearly double accuracy on hard math problems and improve scientific reasoning by 14x.
Microsoft AI introduces MAI-Thinking-1, a 35B-active parameter reasoning model trained from scratch without distillation, achieving strong performance on software engineering and math benchmarks while emphasizing clean data and self-sufficiency.
DenseSteer is a training-free inference-time framework that improves small language models' math reasoning by steering their internal representations towards dense reasoning patterns, achieving accuracy gains without increasing token-level negative log-likelihood.
Trust-Region behavior Blending (TRB) improves on-policy distillation by replacing poor early student rollouts with teacher-like behavior within a KL trust region during warmup, achieving stronger results on math-reasoning tasks.
This paper evaluates three approaches (pure chain-of-thought reasoning, single-shot code execution, and iterative code execution) on 1,000 GSM-Symbolic problems using Claude Haiku 4.5, finding that chain-of-thought is the most robust to perturbation, while code execution does not improve reasoning robustness on grade-school math problems.