Tag
The paper introduces Success-Referenced On-Policy Distillation (SR-OPD), which uses a student's own successful rollouts as references to selectively allocate teacher supervision, achieving comparable reasoning performance on math benchmarks using only 3.46–5.02% of the teacher-input tokens required by vanilla on-policy distillation.
This paper shows that when selecting math training data for LLM fine-tuning, sharing a reasoning approach (e.g., invariant reasoning or double counting) yields significantly greater transfer to held-out targets than sharing a mathematical topic, with advantages of 8–16 percentage points across five base models in counterbalanced 2x2 designs.
The paper introduces Multilingual GSM-Symbolic, a symbolic-template dataset covering 15 languages and 30k item-matched QA pairs, to quantify what determines cross-lingual capability transfer in LLMs. It finds model size, language resource level, reasoning, and typological distance are key drivers, and shows a framework explaining 92% of between-language variation while predicting unseen-language performance within 6.0pp.
This workshop paper audits how a cumulative token cap falling inside a mathematical derivation affects test-time controllers, comparing strict (stop at cap) versus advisory (finish current attempt) stopping across 19,200 archived traces. It finds that budget curves must jointly report cap, realized cost, eligible candidates, and selector information, since candidate coverage and nominal budgets alone can mislead.
A workshop paper presenting a single-instance case study in which an AI agent is asked to rediscover a hidden invariant of degree-four Blaschke curves from numerical data, yielding a homogeneous cubic that predicts unseen configurations with machine-precision residuals, while honestly noting limitations such as a deterministic polynomial-fitting baseline matching the agent's result.
The paper introduces the concept of 'Mathematical Primitive' to probe structural mathematical understanding in LLMs, presents a benchmark evaluating reasoning along Discovery, Generation, Digestion, and Execution, and proposes a primitive-privileged self-distillation framework that improves mathematical reasoning across model scales.
The paper introduces Advantage-Based Control Variates (ABC), an advantage–value formulation of trajectory-level control variates for RL with verifiable rewards, combined with Direct Advantage Estimation into a single actor-critic algorithm that matches GRPO's performance on mathematical reasoning with far fewer online optimization steps.
This paper proposes entropy-regularized cross-entropy (ER-CE) to address the misalignment between cross-entropy loss and verifier objectives in verifiable domains, showing consistent improvements in mathematical reasoning and code generation benchmarks.
This paper proposes GRAFT, an off-policy-aware framework for reinforcement learning with verifiable rewards (RLVR) that replaces all-fail groups with peer trajectories to improve performance in mathematical reasoning benchmarks.
SAKI introduces a supervision allocation method for on-policy distillation that uses KL-constrained teacher-guided rollouts and maximal coupling to route token-level supervision, improving performance on mathematical reasoning benchmarks for small models.
The paper introduces "The Endless Exam," a benchmark where models automatically construct and verify mathematical objects like sets, graphs, and codes, earning uncapped continuous scores even after surpassing published references. Current frontier models evaluated across 17 configurations fail to beat the published frontier, highlighting a persistent gap toward superintelligence-level mathematical capability.
This paper introduces Decoupled Credit Self-Distillation (DCSD) to separate credit direction and magnitude for self-distillation, achieving improvements of 8.45 points in mathematical reasoning and 7.01 points in multimodal reasoning across 11 benchmarks.
Introduces Least Square Policy Distillation (LSPD), an RL-based method for sample-efficient distillation in large language models, improving performance on mathematical reasoning benchmarks.
This paper introduces calibrated importance sampling to address the training-inference mismatch in reinforcement learning for large language models, improving policy updates and performance on mathematical reasoning benchmarks.
ELF-REG scales continuous diffusion language models to reasoning tasks by using representation alignment and entanglement, achieving improved performance on benchmarks like GSM8K and HumanEval compared to other diffusion models.
This paper introduces Planned Test-Time Scaling (PTTS), a method that coordinates reasoning branches to enhance performance on challenging tasks, achieving significant gains over repeated sampling in mathematical reasoning benchmarks.
The paper introduces RECAP, a redundancy-aware learning method that improves the efficiency of large reasoning models by assigning credit to steps based on their structural role and efficacy, enhancing accuracy while reducing token usage.
This paper studies the interaction between prompt breadth and rollout refresh in on-policy distillation for mathematical reasoning, revealing that prompt efficiency depends on both the refresh rate of student policies and the inference budget.
The paper introduces Calibrated On-Policy Distillation (Cal-OPD), a method that estimates the teacher's self-deviation to calibrate teacher-student discrepancies, improving on-policy distillation for mathematical reasoning tasks.
This paper investigates how on-policy distillation can cause length inflation due to EOS token mismatches between student and teacher models, and proposes a correction method by aggregating EOS probabilities to reduce response length.