Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination

• URL: http://arxiv.org/abs/2507.10532v2
• Date: Tue, 05 Aug 2025 14:47:50 GMT
• Title: Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination
• Authors: Mingqi Wu, Zhihao Zhang, Qiaole Dong, Zhiheng Xi, Jun Zhao, Senjie Jin, Xiaoran Fan, Yuhao Zhou, Huijie Lv, Ming Zhang, Yanwei Fu, Qin Liu, Songyang Zhang, Qi Zhang,
• Abstract summary: Pre-training on massive web-scale corpora leaves Qwen2.5 susceptible to data contamination in widely used benchmarks.<n>We introduce a generator that creates fully clean arithmetic problems of arbitrary length and difficulty, dubbed RandomCalculation.<n>We show that only accurate reward signals yield steady improvements that surpass the base model's performance boundary.
• Score: 67.67725938962798
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reasoning in large language models has long been a central research focus, and recent studies employing reinforcement learning (RL) have introduced diverse methods that yield substantial performance gains with minimal or even no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance performance. However, these breakthroughs are predominantly observed for the mathematically strong Qwen2.5 series on benchmarks such as MATH-500, AMC, and AIME, and seldom transfer to models like Llama, which warrants a more in-depth investigation. In this work, our empirical analysis reveals that pre-training on massive web-scale corpora leaves Qwen2.5 susceptible to data contamination in widely used benchmarks. Consequently, conclusions derived from contaminated benchmarks on Qwen2.5 series may be unreliable. To obtain trustworthy evaluation results, we introduce a generator that creates fully clean arithmetic problems of arbitrary length and difficulty, dubbed RandomCalculation. Using this leakage-free dataset, we show that only accurate reward signals yield steady improvements that surpass the base model's performance boundary in mathematical reasoning, whereas random or incorrect rewards do not. Moreover, we conduct more fine-grained analyses to elucidate the factors underlying the different performance observed on the MATH-500 and RandomCalculation benchmarks. Consequently, we recommend that future studies evaluate models on uncontaminated benchmarks and, when feasible, test various model series to ensure trustworthy conclusions about RL and related methods.

Related papers

• Inverse Scaling in Test-Time Compute [51.16323216811257]
  Extending the reasoning length of Large Reasoning Models (LRMs) deteriorates performance.<n>We identify five distinct failure modes when models reason for longer.<n>These findings suggest that while test-time compute scaling remains promising for improving model capabilities, it may inadvertently reinforce problematic reasoning patterns.
  arXiv  Detail & Related papers  (2025-07-19T00:06:13Z)
• VAR-MATH: Probing True Mathematical Reasoning in Large Language Models via Symbolic Multi-Instance Benchmarks [25.295071827427677]
  emphbenchmark contamination arises from the public availability of test problems.<n>emphevaluation fragility stems from the reliance on single-instance assessments.<n>IME-MATH is a symbolic evaluation framework designed to probe genuine reasoning ability.
  arXiv  Detail & Related papers  (2025-07-17T08:10:55Z)
• REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once [33.049237516125146]
  We present REST (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that exposes Large Reasoning Models to multiple problems simultaneously.<n>Our evaluation reveals several striking findings: Even state-of-the-art (SOTA) models like DeepSeek-R1 exhibit substantial performance degradation under stress testing.
  arXiv  Detail & Related papers  (2025-07-14T17:58:47Z)
• ReasonFlux-PRM: Trajectory-Aware PRMs for Long Chain-of-Thought Reasoning in LLMs [56.32212611983997]
  We introduce ReasonFlux-PRM, a novel trajectory-aware PRM to evaluate trajectory-response type of reasoning traces.<n>ReasonFlux-PRM incorporates both step-level and trajectory-level supervision, enabling fine-grained reward assignment aligned with structured chain-of-thought data.<n>Our derived ReasonFlux-PRM-7B yields consistent performance improvements, achieving average gains of 12.1% in supervised fine-tuning, 4.5% in reinforcement learning, and 6.3% in test-time scaling.
  arXiv  Detail & Related papers  (2025-06-23T17:59:02Z)
• Model-agnostic Mitigation Strategies of Data Imbalance for Regression [0.0]
  Data imbalance persists as a pervasive challenge in regression tasks, introducing bias in model performance and undermining predictive reliability.<n>We present advanced mitigation techniques, which build upon and improve existing sampling methods.<n>We demonstrate that constructing an ensemble of models -- one trained with imbalance mitigation and another without -- can significantly reduce these negative effects.
  arXiv  Detail & Related papers  (2025-06-02T09:46:08Z)
• TinyV: Reducing False Negatives in Verification Improves RL for LLM Reasoning [11.573904453859098]
  Reinforcement Learning (RL) has become a powerful tool for enhancing the reasoning abilities of large language models (LLMs)<n>Yet, RL's success relies on the reliability of rewards, which are provided by verifiers.<n>In this paper, we expose and analyze a widespread problem--false negatives--where verifiers wrongly reject correct model outputs.<n>We propose tinyV, a lightweight LLM-based verifier that augments existing rule-based methods.
  arXiv  Detail & Related papers  (2025-05-20T17:16:44Z)
• Model Utility Law: Evaluating LLMs beyond Performance through Mechanism Interpretable Metric [99.56567010306807]
  Large Language Models (LLMs) have become indispensable across academia, industry, and daily applications.<n>One core challenge of evaluation in the large language model (LLM) era is the generalization issue.<n>We propose Model Utilization Index (MUI), a mechanism interpretability enhanced metric that complements traditional performance scores.
  arXiv  Detail & Related papers  (2025-04-10T04:09:47Z)
• Do Theory of Mind Benchmarks Need Explicit Human-like Reasoning in Language Models? [14.29992535286614]
  Theory of Mind (ToM) is the ability to attribute mental states to others.<n>Recent advancements in Large Language Models have shown promising performance on ToM benchmarks.<n>Do these benchmarks necessitate explicit human-like reasoning processes, or can models succeed through alternative strategies?
  arXiv  Detail & Related papers  (2025-04-02T12:58:42Z)
• Proof or Bluff? Evaluating LLMs on 2025 USA Math Olympiad [0.8030359871216614]
  We evaluate state-of-the-art reasoning models on six problems from the 2025 USAMO.<n>Only Gemini-2.5-Pro achieves a non-trivial score of 25%.<n>Our results suggest that current LLMs are inadequate for rigorous mathematical reasoning tasks.
  arXiv  Detail & Related papers  (2025-03-27T19:21:05Z)
• Exploring the Limit of Outcome Reward for Learning Mathematical Reasoning [65.2421542320293]
  Reasoning abilities are crucial components of general intelligence.<n>Recent advances by proprietary companies, such as o-series models of OpenAI, have made remarkable progress on reasoning tasks.<n>This paper proposes a new RL framework, termed OREAL, to pursue the performance limit that can be achieved through textbfOutcome textbfREwtextbfArd-based reinforcement textbfLearning for mathematical reasoning tasks.
  arXiv  Detail & Related papers  (2025-02-10T18:57:29Z)
• MR-Ben: A Meta-Reasoning Benchmark for Evaluating System-2 Thinking in LLMs [55.20845457594977]
  Large language models (LLMs) have shown increasing capability in problem-solving and decision-making.<n>We present a process-based benchmark MR-Ben that demands a meta-reasoning skill.<n>Our meta-reasoning paradigm is especially suited for system-2 slow thinking.
  arXiv  Detail & Related papers  (2024-06-20T03:50:23Z)
• Sources of Gain: Decomposing Performance in Conditional Average Dose Response Estimation [0.9332308328407303]
  Estimating conditional average dose responses (CADR) is an important but challenging problem. Our paper analyses this practice and shows that using popular benchmark datasets without further analysis is insufficient to judge model performance. We propose a novel decomposition scheme that allows the evaluation of the impact of five distinct components contributing to CADR estimator performance.
  arXiv  Detail & Related papers  (2024-06-12T13:39:32Z)
• Masked Thought: Simply Masking Partial Reasoning Steps Can Improve Mathematical Reasoning Learning of Language Models [102.72940700598055]
  In reasoning tasks, even a minor error can cascade into inaccurate results. We develop a method that avoids introducing external resources, relying instead on perturbations to the input. Our training approach randomly masks certain tokens within the chain of thought, a technique we found to be particularly effective for reasoning tasks.
  arXiv  Detail & Related papers  (2024-03-04T16:21:54Z)
• Rethinking Benchmark and Contamination for Language Models with Rephrased Samples [49.18977581962162]
  Large language models are increasingly trained on all the data ever produced by humans. Many have raised concerns about the trustworthiness of public benchmarks due to potential contamination in pre-training or fine-tuning datasets.
  arXiv  Detail & Related papers  (2023-11-08T17:35:20Z)
• Don't Make Your LLM an Evaluation Benchmark Cheater [142.24553056600627]
  Large language models(LLMs) have greatly advanced the frontiers of artificial intelligence, attaining remarkable improvement in model capacity. To assess the model performance, a typical approach is to construct evaluation benchmarks for measuring the ability level of LLMs. We discuss the potential risk and impact of inappropriately using evaluation benchmarks and misleadingly interpreting the evaluation results.
  arXiv  Detail & Related papers  (2023-11-03T14:59:54Z)
• 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.