Tagged articles
Mathematical Reasoning
17 articles · Page 1 of 1
Tsinghua researchers show that training diffusion language models with left-to-right token unmasking and GRPO significantly improves math and code reasoning performance, reducing entropy at logical branching points and outperforming flexible decoding strategies on GSM8K and MATH-500 benchmarks.
Entropy ReductionGRPOICML 2026
0 likes · 11 min read

Zhejiang University and Alibaba propose Relay-OPD, an online distillation method that detects student reasoning errors via teacher-student token divergence and briefly hands control to the teacher for targeted correction, achieving +5.73% accuracy on math benchmarks while halving training trajectory length.
Mathematical ReasoningPrefix FailureRelay-OPD
0 likes · 17 min read

Microsoft and Shanghai Jiao Tong University open-sourced Argus, an evidence-driven autonomous research agent that ran 1,548 hours across 27 campaigns with 95–98% duty cycle, using a multi-agent loop and core-vertical decoupling to minimize human intervention.
AI for ScienceAgent RuntimeArgus
0 likes · 10 min read

The paper identifies the broadcast‑style credit assignment of GRPO as a bottleneck for RL‑LLM math reasoning, proposes the Outcome‑Grounded Advantage Reshaping (OAR) framework with token‑importance estimation, and demonstrates that its two variants, OAR‑P and OAR‑G, consistently improve accuracy, training efficiency, and stability across multiple math benchmarks.
Credit AssignmentGRPOLLM
0 likes · 15 min read

The paper “What Really Improves Mathematical Reasoning: Structured Reasoning Signals Beyond Pure Code” (ICML 2026) shows that while pure code data clearly enhances programming ability, it does not consistently improve complex mathematical reasoning and can even compete with math data, whereas structured reasoning signals—cognitive scaffolds—significantly lift difficult math benchmarks without harming coding performance.
Mathematical Reasoningcode datalarge language models
0 likes · 13 min read

MathForge tackles the long‑standing question of which math problems deserve focus in reinforcement‑learning‑based training, introducing a difficulty‑aware optimizer (DGPO) and multi‑aspect question reformulation (MQR) that together prioritize harder‑but‑learnable questions, yielding consistent performance gains across model sizes and modalities.
DGPODifficulty‑Aware OptimizationMQR
0 likes · 11 min read

The SHAPE framework (Stage‑aware Hierarchical Advantage via Potential Estimation) adds a milestone‑based “reasoning tax” to large language model inference, providing step‑wise correctness signals and penalizing verbosity, which yields an average 3% accuracy gain and a 30% reduction in token consumption across multiple math‑reasoning benchmarks.
ACL 2026LLMMathematical Reasoning
0 likes · 10 min read

MathForge tackles the overlooked issue of training large language models on mathematically challenging yet learnable problems by introducing a difficulty‑aware group policy optimization (DGPO) and multi‑aspect question reformulation (MQR), achieving consistent gains across model sizes and modalities.
DGPODifficulty‑Aware OptimizationMQR
0 likes · 13 min read

SePT (Self‑evolving Post‑Training) shows that a large language model can improve its mathematical reasoning ability by about ten percentage points using a reward‑free online self‑training loop that decouples generation temperature from standard SFT, matching or surpassing RL‑based methods without harming general performance.
LLMMathematical ReasoningSFT
0 likes · 9 min read

DeepSeek-Math-V2, an open‑source math‑reasoning model from DeepSeek, introduces a self‑verification mechanism, strong theorem‑proving ability, closed‑loop evolution, and record‑breaking competition scores, offering researchers, educators, and engineers a reliable tool for rigorous mathematical AI tasks.
AI MathMathematical Reasoningproof assistant
0 likes · 13 min read

DeepSeek-Math-V2, an open‑source math reasoning model from DeepSeek, introduces a self‑verification mechanism that ensures step‑by‑step logical correctness, achieving gold‑medal scores in IMO 2025, CMO 2024 and near‑perfect results in the Putnam 2024 competition, while offering free, extensible deployment for research, training, and scientific computation.
AI MathDeepSeekMathematical Reasoning
0 likes · 13 min read

This roundup surveys five cutting‑edge AI papers—DeepSeekMath‑V2’s self‑verifiable mathematical reasoning, MedSAM‑3’s promptable medical image and video segmentation, SAM 3D’s single‑image 3D reconstruction, Qwen3‑VL’s high‑capacity vision‑language model, and the M² memory‑mesh transformer for image captioning—highlighting their key methods, benchmarks, and code links.
3D reconstructionImage CaptioningMathematical Reasoning
0 likes · 6 min read

The LightReasoner paper from Hong Kong University shows that small language models can guide large models on critical reasoning steps, achieving up to 90% faster inference and significant accuracy gains across multiple math benchmarks.
Contrastive DecodingKL DivergenceMathematical Reasoning
0 likes · 9 min read

This paper introduces the Thought Leap Bridge task and the CoT‑Bridge model, which detect and fill missing intermediate steps in chain‑of‑thought reasoning, dramatically improving large language model performance on mathematical and logical benchmarks and enhancing downstream distillation and reinforcement‑learning pipelines.
Chain-of-ThoughtCoT-BridgeLLM
0 likes · 8 min read


DevOpsMay 5, 2025 · Artificial Intelligence
DeepSeek has quietly open‑sourced a new mathematics‑focused large language model, DeepSeek‑Prover‑V2 (available in 671B and 7B variants), achieving 88.9% on MiniF2F and strong results on PutnamBench, alongside the high‑quality ProverBench dataset and a novel recursive theorem‑proving pipeline.
AIDeepSeekMathematical Reasoning
0 likes · 4 min read

Researchers from Nanjing University of Aeronautics, Nanjing University of Technology and Oxford show that high‑instruction prompts dramatically boost large language models' mathematical reasoning, enabling DeepSeek‑R1 and Qwen2.5 to solve complex polynomial tasks and even produce a new counterexample to Hilbert's 17th problem.
AIDeepSeekMathematical Reasoning
0 likes · 6 min read

Recent research by Apple’s Iman Mirzadeh team introduces the GSM‑Symbolic benchmark, revealing that large language models, despite high scores on GSM8K, exhibit significant performance drops when problem numbers, names, or extra clauses change, indicating a lack of true formal reasoning ability.
AI safetyGSM‑SymbolicMathematical Reasoning
0 likes · 9 min read
