The Quality-Utility Paradox: Why High-Reward Data Impairs Small Model Mathematical Reasoning

• URL: http://arxiv.org/abs/2606.16152v1
• Date: Mon, 15 Jun 2026 03:13:07 GMT
• Title: The Quality-Utility Paradox: Why High-Reward Data Impairs Small Model Mathematical Reasoning
• Abstract summary: Data refined or synthesized by a stronger Oracle obtains higher perceived quality according to reward models.<n>Our analysis shows that Oracle refinement couples logical repair with distributional drift away from the SLM's native reasoning distribution.<n>These findings suggest that effective mathematical reasoning distillation should jointly optimize perceived solution quality and learner-data compatibility.
• Score: 54.477658074293885
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Knowledge distillation from powerful reasoning models is widely used to improve Small Language Models (SLMs) on mathematical reasoning, often assuming that traces with higher reward model scores provide more useful supervision. We identify a counterintuitive \textbf{Quality-Utility Paradox} in mathematical reasoning distillation. Data refined or synthesized by a stronger Oracle obtains higher perceived quality according to reward models, yet consistently underperforms traces generated by the SLM itself and selected through rejection sampling across Qwen2.5, LLaMA-3, and DeepSeek families. Our analysis shows that Oracle refinement couples logical repair with distributional drift away from the SLM's native reasoning distribution. This drift increases the learner's adaptation cost and can outweigh the benefit of improved reasoning logic. To test this mechanism, we introduce \textbf{Style-Aligned Refinement}, which preserves the native trajectory of the SLM while retaining logical repair from the Oracle. This intervention lowers adaptation cost and restores downstream utility. These findings suggest that effective mathematical reasoning distillation should jointly optimize perceived solution quality and learner-data compatibility, rather than relying solely on reward-model scores. The datasets and code are available at https://github.com/Dracoqhl/Quality-Utility-Paradox.

Related papers

• Can Recommender Systems Teach Themselves? A Recursive Self-Improving Framework with Fidelity Control [82.30868101940068]
  We propose a paradigm in which a model bootstraps its own performance without reliance on external data or teacher models.<n>Our theoretical analysis shows that RSIR acts as a data-driven implicit regularizer, smoothing the optimization landscape.<n>We show that even smaller models benefit, and weak models can generate effective training curricula for stronger ones.
  arXiv  Detail & Related papers  (2026-02-17T15:31:32Z)
• Generating Data-Driven Reasoning Rubrics for Domain-Adaptive Reward Modeling [21.45871501724415]
  We propose a data-driven approach to automatically construct highly granular reasoning model error.<n>Rubrics can be used to build stronger LLM-as-judge reward functions.<n>This extension opens the door for teaching models to solve complex technical problems without a full dataset of gold labels.
  arXiv  Detail & Related papers  (2026-02-06T15:51:52Z)
• Budget-Aware Anytime Reasoning with LLM-Synthesized Preference Data [57.996437077411315]
  We study the reasoning behavior of large language models (LLMs) under limited computation budgets.<n>We introduce an anytime reasoning framework and the Anytime Index, a metric that quantifies how effectively solution quality improves as reasoning tokens increase.<n> Experiments on NaturalPlan (Trip), AIME, and GPQA datasets show consistent gains across Grok-3, GPT-oss, GPT-4.1/4o, and LLaMA models.
  arXiv  Detail & Related papers  (2026-01-16T07:09:30Z)
• Reasoning Pattern Alignment Merging for Adaptive Reasoning [48.347817456299104]
  Reasoning Pattern Alignment Merging (RPAM)<n>RPAM is a layer-wise model merging framework based on feature alignment to facilitate query-adaptive reasoning.<n> Experiments on seven widely used reasoning benchmarks show that RPAM substantially reduces inference cost while maintaining strong performance.
  arXiv  Detail & Related papers  (2026-01-07T01:36:39Z)
• Learning to Reason in LLMs by Expectation Maximization [55.721496945401846]
  We formalize reasoning as a latent variable model and derive an expectation-maximization objective for learning to reason.<n>This view connects EM and modern reward-based optimization, and shows that the main challenge lies in designing a sampling distribution that generates rationales that justify correct answers.
  arXiv  Detail & Related papers  (2025-12-23T08:56:49Z)
• Making Mathematical Reasoning Adaptive [61.45161826629692]
  We propose the AdaR framework to enable adaptive reasoning in large language models (LLMs)<n>AdaR synthesizes logically equivalent queries by varying variable values, and trains models with RLVR on these data to penalize spurious logic.<n> Experimental results demonstrate that AdaR improves robustness and generalization, achieving substantial improvement in mathematical reasoning.
  arXiv  Detail & Related papers  (2025-10-06T09:30:05Z)
• Mitigating Spurious Correlations Between Question and Answer via Chain-of-Thought Correctness Perception Distillation [25.195244084313114]
  Chain-of-Thought Correctness Perception Distillation (CoPeD) aims to improve the reasoning quality of the student model.<n>CoPeD encourages the student model to predict answers based on correct rationales and revise them when they are incorrect.
  arXiv  Detail & Related papers  (2025-09-06T05:33:17Z)
• From Answers to Rationales: Self-Aligning Multimodal Reasoning with Answer-Oriented Chain-of-Thought [43.07899102255169]
  Current methods primarily focus on positive rationales, typically relying on manual annotations or complex systems.<n>We propose a novel framework: textbfSelf-Aligning textbfMultimodal Reasoning with textbfAnswertextbfriented Chain-of-textbfThought.
  arXiv  Detail & Related papers  (2025-07-01T08:24:51Z)
• Scaling Relationship on Learning Mathematical Reasoning with Large Language Models [75.29595679428105]
  We investigate how the pre-training loss, supervised data amount, and augmented data amount influence the reasoning performances of a supervised LLM. We find that rejection samples from multiple models push LLaMA-7B to an accuracy of 49.3% on GSM8K which outperforms the supervised fine-tuning (SFT) accuracy of 35.9% significantly.
  arXiv  Detail & Related papers  (2023-08-03T15:34:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.