When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models

• URL: http://arxiv.org/abs/2504.02010v2
• Date: Thu, 02 Oct 2025 02:39:53 GMT
• Title: When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models
• Abstract summary: We investigate how the reasoning capabilities of large reasoning models are compromised during compression.<n>We benchmark quantized, distilled, and pruned DeepSeek-R1 models on four reasoning datasets.<n>Overall, we find dynamically quantized 2.51-bit R1 reaches close-to-R1 performance.
• Score: 18.37440002326957
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Compression methods, including quantization, distillation, and pruning, improve the computational efficiency of large reasoning models (LRMs). However, existing studies either fail to sufficiently compare all three compression methods on LRMs or lack in-depth interpretation analysis. In this paper, we investigate how the reasoning capabilities of LRMs are compromised during compression, through performance benchmarking and mechanistic interpretation. To uncover the effects of compression on reasoning performance, we benchmark quantized, distilled, and pruned DeepSeek-R1 models on four reasoning datasets (AIME 2024, FOLIO, Temporal Sequences, and MuSiQue). To precisely locate compression effects on model weights, we adapt difference of means and attribution patching techniques, focusing on the activation of every linear component in compressed LRMs, to interpret fine-grained causal relationships between weights and various reasoning capabilities. This fine-grained interpretation addresses a fundamental question of compression: which weights are the most important for reasoning? Overall, we find dynamically quantized 2.51-bit R1 reaches close-to-R1 performance. With empirical verification, we present three main findings that generalize across both Llama and Qwen: (1) Weight count has a greater impact on LRMs' knowledge memorization than reasoning, highlighting the risks of pruning and distillation; (2) The MLP up projection in the final layer of distilled LRMs is one of the most important components, offering a new perspective on locating critical weights - a fundamental problem in model compression; and (3) Current quantization methods overly compress the final-layer modules and MLP gate projections, so protecting just 2% of all weights that are excessively compressed can raise average accuracy by 6.57%, greatly surpassing the state-of-the-art.

Related papers

• The Asymmetric Harms of LLM Compression [9.45803162325743]
  Large language models (LLMs) compression reduces deployment costs, but standard aggregate metrics like perplexity and accuracy often mask underlying behavioral shifts.<n>We systematically evaluate 3 LLMs across 11 compression methods to investigate the effects of compression on knowledge retention, model confidence, and social bias.
  arXiv  Detail & Related papers  (2026-08-20T06:06:14Z)
• Arbitrary Ratio Feature Compression via Next Token Prediction [52.10426317889982]
  Arbitrary Ratio Feature Compression (ARFC) framework supports any compression ratio with a single model.<n>ARC is an auto-regressive model that performs compression via next-gressive prediction.<n>MoS module refines the compressed tokens by utilizing multiple compression results.<n>ERGC is integrated into the training process to preserve semantic and structural relationships during compression.
  arXiv  Detail & Related papers  (2026-02-12T02:38:57Z)
• Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression [55.63153956934198]
  Chain-of-Thought (CoT) reasoning successfully enhances the reasoning capabilities of Large Language Models (LLMs)<n>Existing CoT compression methods often suffer from a critical loss of logical fidelity at high compression ratios.<n>We propose a novel EXTreme-RAtio Chain-of-Thought Compression framework, termed Extra-CoT, which aggressively reduces the token budget while preserving answer accuracy.
  arXiv  Detail & Related papers  (2026-02-09T06:57:15Z)
• A Systematic Study of Compression Ordering for Large Language Models [0.5926203312586109]
  This study systematically examines how knowledge distillation, structured pruning, and low-bit quantization perform when applied to the Qwen2.5 3B model.<n>Experiments show that quantization provides the greatest standalone compression, while pruning introduces moderate quality degradation.
  arXiv  Detail & Related papers  (2025-11-23T12:46:56Z)
• MGAA: Multi-Granular Adaptive Allocation fof Low-Rank Compression of LLMs [9.244526043014098]
  Multi-Granular Adaptive Allocation (MGAA) method can adaptively allocate parameters between and within sublayers without task-specific evaluations in the compression process.<n> Comprehensive evaluations of MGAA across multiple LLMs backbone models and benchmark datasets demonstrate its superior performance.
  arXiv  Detail & Related papers  (2025-07-04T04:54:01Z)
• Fractional Reasoning via Latent Steering Vectors Improves Inference Time Compute [57.16286134405821]
  We propose Fractional Reasoning, a framework that enables continuous control over reasoning intensity at inference time.<n>Our method operates by extracting the latent steering vector associated with deeper reasoning and reapplying it with a tunable scaling factor.<n> Experiments on GSM8K, MATH500, and GPQA demonstrate that Fractional Reasoning consistently improves performance across diverse reasoning tasks and models.
  arXiv  Detail & Related papers  (2025-06-18T21:15:59Z)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity [16.266145641151375]
  Large Reasoning Models generate detailed thinking processes before providing answers.<n>We show that LRMs face a complete accuracy collapse beyond certain complexities.<n>We also investigate the reasoning traces in more depth, studying the patterns of explored solutions.
  arXiv  Detail & Related papers  (2025-06-07T22:42:29Z)
• TL;DR: Too Long, Do Re-weighting for Efficient LLM Reasoning Compression [55.37723860832064]
  We propose a dynamic ratio-based training pipeline that does not rely on sophisticated data annotations.<n>We validate our approach across models on DeepSeek-R1-Distill-7B and DeepSeek-R1-Distill-14B and on a diverse set of benchmarks with varying difficulty levels.
  arXiv  Detail & Related papers  (2025-06-03T09:23:41Z)
• Compression Hacking: A Supplementary Perspective on Informatics Properties of Language Models from Geometric Distortion [56.12939353271623]
  From a geometric standpoint, the word representation space of highly compressed LMs tends to degenerate into a highly anisotropic state.<n>We find this synchronicity is essentially the Compression Hacking'' in LM representations.<n>We propose three refined compression metrics by incorporating geometric distortion analysis and integrate them into a self-evaluation pipeline.
  arXiv  Detail & Related papers  (2025-05-23T12:11:03Z)
• TrimR: Verifier-based Training-Free Thinking Compression for Efficient Test-Time Scaling [20.980976778470247]
  Large Reasoning Models (LRMs) demonstrate exceptional capability in tackling complex mathematical, logical, and coding tasks.<n>We propose TrimR, a verifier-based, training-free, efficient framework for dynamic Chain-of-Thought (CoT) compression.
  arXiv  Detail & Related papers  (2025-05-22T12:23:30Z)
• Semantic Retention and Extreme Compression in LLMs: Can We Have Both? [0.0]
  Large Language Model (LLM) deployment has intensified the need for efficient model compression techniques.<n>We show how strategically combining pruning and quantization could yield superior performance-to-compression ratios.<n>We introduce the Semantic Retention Compression Rate (SrCr), a novel metric that quantifies the trade-off between model compression and semantic preservation.
  arXiv  Detail & Related papers  (2025-05-12T07:23:19Z)
• SplitReason: Learning To Offload Reasoning [7.016347390223799]
  Reasoning in large language models (LLMs) tends to produce substantially longer token generation sequences than simpler language modeling tasks. We leverage this by offloading only the most challenging parts of the reasoning process to a larger, more capable model. This approach improves AIME24 reasoning accuracy by 24% and 28.3% while offloading 1.35% and 5% of the generated tokens respectively.
  arXiv  Detail & Related papers  (2025-04-23T03:00:02Z)
• M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models [72.75501495786297]
  We introduce a novel hybrid linear RNN reasoning model, M1, built on the Mamba architecture. Experimental results show that M1 not only outperforms previous linear RNN models but also matches the performance of state-of-the-art DeepSeek R1 distilled reasoning models.
  arXiv  Detail & Related papers  (2025-04-14T17:38:25Z)
• SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning [14.020244011380063]
  SpecReason is a system that accelerates LRM inference by using a lightweight model to (speculatively) carry out simpler intermediate reasoning steps. It achieves 1.5-2.5$times$ speedup over vanilla LRM inference while improving accuracy by 1.0-9.9%. Compared to speculative decoding without SpecReason, their combination yields an additional 19.4-44.2% latency reduction.
  arXiv  Detail & Related papers  (2025-04-10T16:05:19Z)
• Revisiting Prompt Optimization with Large Reasoning Models-A Case Study on Event Extraction [8.88001387249786]
  Large Reasoning Models (LRMs) such as DeepSeek-R1 and OpenAI o1 have demonstrated remarkable capabilities in various reasoning tasks. Their strong capability to generate and reason over intermediate thoughts has led to arguments that they may no longer require extensive prompt engineering or optimization to interpret human instructions. In this work, we aim to systematically study this open question, using the structured task of event extraction for a case study.
  arXiv  Detail & Related papers  (2025-04-10T00:53:59Z)
• START: Self-taught Reasoner with Tools [51.38785489790888]
  We introduce START (Self-Taught Reasoner with Tools), a tool-integrated long Chain-of-thought (CoT) reasoning LLM.<n> START is capable of performing complex computations, self-checking, exploring diverse methods, and self-ging.<n>It significantly outperforms the base QwQ-32B and achieves performance comparable to the state-of-the-art open-weight model R1-Distill-Qwen-32B.
  arXiv  Detail & Related papers  (2025-03-06T17:11:51Z)
• Compression Scaling Laws:Unifying Sparsity and Quantization [65.05818215339498]
  We investigate how different compression techniques affect the scaling behavior of large language models (LLMs) during pretraining.<n>We show that weight-only quantization achieves strong parameter efficiency multipliers, while full quantization of both weights and activations shows diminishing returns at lower bitwidths.<n>Our results suggest that different compression techniques can be unified under a common scaling law framework.
  arXiv  Detail & Related papers  (2025-02-23T04:47:36Z)
• Optimizing Singular Spectrum for Large Language Model Compression [95.7621116637755]
  We introduce SoCo, a novel compression framework that learns to rescale the decomposed components of SVD in a data-driven manner.<n>Thanks to the learnable singular spectrum, SoCo adaptively prunes components according to the sparsified importance scores.<n> Experimental evaluations across multiple LLMs and benchmarks demonstrate that SoCo surpasses the state-of-the-art methods in model compression.
  arXiv  Detail & Related papers  (2025-02-20T23:18:39Z)
• Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs [76.43407125275202]
  o1-like models can emulate human-like long-time thinking during inference.<n>This paper presents the first comprehensive study on the prevalent issue of overthinking in these models.<n>We propose strategies to mitigate overthinking, streamlining reasoning processes without compromising accuracy.
  arXiv  Detail & Related papers  (2024-12-30T18:55:12Z)
• LLMCBench: Benchmarking Large Language Model Compression for Efficient Deployment [36.958867918858296]
  Large language models (LLMs) have demonstrated their strong intelligence ability, but high demand for computation and storage hinders their practical application. We present the Large Language Model Compression Benchmark (LLMCBench), a rigorously designed benchmark with an in-depth analysis for LLM compression algorithms.
  arXiv  Detail & Related papers  (2024-10-28T14:45:01Z)
• Mixture of Parrots: Experts improve memorization more than reasoning [72.445819694797]
  We show that as we increase the number of experts, the memorization performance consistently increases while the reasoning capabilities saturate. We find that increasing the number of experts helps solve knowledge-intensive tasks, but fails to yield the same benefits for reasoning tasks.
  arXiv  Detail & Related papers  (2024-10-24T17:54:41Z)
• A Comparative Study on Reasoning Patterns of OpenAI's o1 Model [69.08287909042421]
  We show that OpenAI's o1 model has achieved the best performance on most datasets. We also provide a detailed analysis on several reasoning benchmarks.
  arXiv  Detail & Related papers  (2024-10-17T15:09:03Z)
• Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression [33.45167213570976]
  We investigate the impact of model compression on four dimensions: (1) degeneration harm, i.e., bias and toxicity in generation; (2) representational harm, i.e., biases in discriminative tasks; (3) dialect bias; and(4) language modeling and downstream task performance. Our analysis reveals that compression can lead to unexpected consequences.
  arXiv  Detail & Related papers  (2024-07-06T05:56:22Z)
• 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)
• LLMC: Benchmarking Large Language Model Quantization with a Versatile Compression Toolkit [55.73370804397226]
  Quantization, a key compression technique, can effectively mitigate these demands by compressing and accelerating large language models. We present LLMC, a plug-and-play compression toolkit, to fairly and systematically explore the impact of quantization. Powered by this versatile toolkit, our benchmark covers three key aspects: calibration data, algorithms (three strategies), and data formats.
  arXiv  Detail & Related papers  (2024-05-09T11:49:05Z)
• The Cost of Compression: Investigating the Impact of Compression on Parametric Knowledge in Language Models [11.156816338995503]
  Large language models (LLMs) provide faster inference, smaller memory footprints, and enables local deployment. Two standard compression techniques are pruning and quantization, with the former eliminating redundant connections in model layers and the latter representing model parameters with fewer bits. Existing research on LLM compression primarily focuses on performance in terms of general metrics like perplexity or downstream task accuracy. More fine-grained metrics, such as those measuring parametric knowledge, remain significantly underexplored.
  arXiv  Detail & Related papers  (2023-12-01T22:27:12Z)
• What do Compressed Large Language Models Forget? Robustness Challenges in Model Compression [68.82486784654817]
  We study two popular model compression techniques including knowledge distillation and pruning. We show that compressed models are significantly less robust than their PLM counterparts on adversarial test sets. We develop a regularization strategy for model compression based on sample uncertainty.
  arXiv  Detail & Related papers  (2021-10-16T00:20:04Z)

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.