mathematical-reasoning

The paper investigates the contribution of privileged information in on-policy self-distillation for language models, finding that reference-free distillation accounts for most improvements, with limited additional benefits from privileged references.

UniCAR-RL introduces a reinforcement learning framework that decouples perception and reasoning to improve multimodal large language models' visual mathematical reasoning without annotation.

The paper introduces a teacher-guided curriculum learning approach for Reinforcement Learning with Verifiable Rewards (RLVR) to efficiently train language models on initially unsolvable mathematical problems, achieving significant data efficiency and expanding reasoning boundaries.

Bellman Policy Optimization (BPO) is a critic-free reinforcement learning method that reformulates Policy Mirror Descent using the Bellman equation for autoregressive generation with terminal rewards, improving mathematical reasoning in large language models.

This paper introduces TrimSFT, a token-level reweighting method for supervised fine-tuning that enhances mathematical reasoning by focusing on intermediate logit-gap tokens, demonstrating significant performance gains over standard SFT.

This paper introduces Negative Self-Distillation (NSD), a framework for improving large language model reasoning by diverging from flawed reasoning instead of imitating privileged solutions, showing consistent gains over existing methods on mathematical benchmarks.

This paper investigates the structural sensitivity of multilingual large language models to semantics-preserving perturbations in Hindi and Malayalam, showing significant degradation in mathematical reasoning performance and introducing the IndicReStruct benchmark for evaluation.

RePro integrates Lean-oriented neural automated theorem provers into benchmark rewriting to ensure problem validity and answer correctness for reliable evaluation of LLMs in mathematical problem solving.

This paper investigates how reinforcement learning with verifiable rewards (RLVR) narrows the solution space in LLM reasoning by analyzing where diversity is lost, finding it concentrated at the 'entrance' of trajectories. It demonstrates interventions to recover breadth without compromising accuracy.

MathAdv is a diagnostic benchmark for formal theorem proving in mathematics, covering 13 domains with auxiliary tasks to evaluate knowledge, reasoning, and robustness. The study reveals formalization bottlenecks and performance variations across models.

TTPO introduces a label-free method for test-time training in mathematical reasoning that matches supervised performance by asymmetrically handling agreeing and disagreeing rollouts.

DirEAG is a method for calibrating verbalized confidence in large language models during mathematical reasoning by aggregating Dirichlet evidence from multiple prompts, showing improved calibration on benchmarks like GSM8K.

The paper reveals that exact-match verifiers in reinforcement learning with verifiable rewards introduce language-dependent biases in multilingual mathematical reasoning, proposes a protocol for auditing, and identifies a cross-lingual selection bottleneck.

The paper introduces ERPO, a method that moves regularization from the action-side to the input-side by controlling query distribution, addressing the stability-exploration dilemma in LLM policy optimization, and showing improvements on mathematical reasoning benchmarks.

This report investigates reinforcement fine-tuning strategies for adapting the Qwen2.5-3B-Base model to signal mathematical problems, achieving a more than threefold improvement in accuracy over the base model.

This paper proposes Adaptive Entropy Distillation (AED), a method that dynamically calibrates token-level imitation strength in knowledge distillation using teacher entropy, achieving superior performance on instruction-following and mathematical reasoning benchmarks.

MathForm introduces a framework for mathematical autoformalization using knowledge retrieval and verification-guided refinement, yielding the FormalVerse dataset and an 8B model that outperforms specialized baselines.

Introduces TREAT, a benchmark for evaluating whether large language models can recover known theorem identities from equivalence-preserving transformations of mathematical formulas. The best tested model achieves only 60.73% accuracy, showing that theorem knowledge is fragile under representation changes.

Elon Musk praises Grok's image generation after it correctly visualizes a mathematician drinking from a genus-1 object, while ChatGPT produces a genus-2 donut-mug.

FutureBridge introduces a token reranker for collaborative decoding that ranks LLM-SLM candidates based on how well the SLM can continue reasoning from them, improving the Qwen3-1.7B SLM's math accuracy by 35.1% over greedy decoding.