A new arXiv paper introduces Func-R1, an approach aimed at strengthening mathematical function reasoning in multimodal large language models. The work targets the challenge of combining visual perception with symbolic logic when solving math problems from images. The abstract frames deliberate mathematical reasoning in visual settings as an indicator of advanced multimodal model capability.
Researchers posted an arXiv preprint describing Stellar Colosseum, a model-agnostic framework that coordinates multiple language-model agents on extended research problems in mathematics and theoretical computer science. The authors argue that while models can generate convincing short proofs, they remain unreliable when progress requires many uncertain, interdependent decisions in sequence. The harness is presented as a way to structure such long-horizon work rather than a single model release.
A new arXiv paper introduces HISPO (Hierarchical Importance-Sampling Policy Optimization with Entropy-Derived Segments), a reinforcement learning approach aimed at reinforcement learning with verifiable rewards. The method addresses the difficulty of assigning credit across long solution traces by splitting them into segments derived from entropy, so that different parts of a model's mathematical reasoning receive appropriate weight during training. The work is a cross-listed submission on arXiv's machine learning category.
Researchers released ZGCM-1, a 7-billion-parameter dense foundation model trained from scratch with a focus on data, system, and algorithmic efficiency. The work argues that smaller models should not try to memorize the open web, but instead be optimized for targeted capabilities such as mathematical reasoning and agentic search. It is presented as a fully open release.
A new arXiv paper addresses a known failure mode in reinforcement learning with verifiable rewards (RLVR), where training problems that are too hard for a model produce uniformly failed attempts and yield no learning signal. The authors propose a teacher-guided curriculum that sequences training data so the model encounters problems it can actually solve, making the process more data-efficient. The work targets mathematical reasoning in large language models and falls within the cs.CL area.
A new arXiv preprint examines the kinds of mistakes large language models make while working through mathematics problems, grouping them into distinct error categories. The authors use that taxonomy of reasoning failures to target improvements in the models' mathematical problem-solving. The work aims to give a clearer picture of where current LLM reasoning breaks down and how to address it.
A revised arXiv paper introduces NeuroProlog, a neurosymbolic approach that pairs language models with symbolic reasoning to improve mathematical problem solving. The authors use multi-task fine-tuning and describe a "cocktail effect," where combining several training tasks yields better results than training on them individually. The work targets a known weakness in LLMs, which often produce fluent but logically inconsistent math solutions.
A new arXiv paper examines how post-training choices and test-time inference setups influence a model's ability to write natural-language proofs for difficult olympiad problems. Using Nemotron 3 Ultra as a base, the authors produce two specialist checkpoints via supervised fine-tuning and reinforcement learning, and release the training approach publicly.
A new arXiv paper introduces Magenta, a method that connects informal natural-language mathematical reasoning by large language models with the formal proof assistant Lean. The approach aims to let models generate reasoning in ordinary language while Lean checks correctness, closing the gap between informal and formally verified mathematics.
A revised arXiv preprint introduces a method that uses automatic verification to track how well language models reason about unsolved mathematical problems. The author notes that although large language models now handle sophisticated math and science reasoning, whether they can contribute genuinely new research remains contested and thinly studied. The work aims to give a measurable way to assess progress on that question.
Researchers have posted an arXiv paper presenting a framework that helps language models solve plane geometry problems by separating visual perception of geometric symbols from logical deduction. The approach aims to reduce reliance on computationally heavy large multimodal models by using symbolic representations to guide text-only reasoning. The work addresses a long-standing AI challenge that requires combining perception with rigorous mathematical reasoning.
A new arXiv paper argues that standard supervised fine-tuning applies its loss uniformly across all tokens, even though some are already mastered and others carry far more useful learning signal for mathematical reasoning. The authors introduce a token-trimming perspective that prioritizes which tokens a model should actually learn during fine-tuning, aiming to avoid over-sharpening well-understood tokens while strengthening the ones that matter most.