mlsanityData to 14 Oct 2026

Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction

arXiv 2509.12464 ↗First posted 15 Sep 2025cs.AIcited by 4

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

Reasoning language models such as DeepSeek-R1 produce long chain-of-thought traces during inference time which make them costly to deploy at scale. We show that using compression techniques such as neural network pruning produces greater performance loss than in typical language modeling tasks, and in some cases can make the model slower since they cause the model to produce more thinking tokens but with worse performance. We show that this is partly due to the fact that standard LLM pruning methods often focus on input reconstruction, whereas reasoning is a decode-dominated task. We introduce a simple, drop-in fix: during pruning we jointly reconstruct activations from the input and the model's on-policy chain-of-thought traces. This "Reasoning-Aware Compression" (RAC) integrates seamlessly into existing pruning workflows such as SparseGPT, and boosts their performance significantly. Code reproducing the results in the paper can be found at: https://github.com/RyanLucas3/RAC

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 10 of its 21 references that are in the map. First posted on arXiv 15 Sep 2025.

Builds on

10 of 21 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models2 Apr
  2. 2023Tree of Thoughts: Deliberate Problem Solving with Large Language Models17 May

Other references · 8

  1. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  2. 2023On-Policy Distillation of Language Models: Learning from Self-Generated Mistakescited by 939
  3. 2023A Simple and Effective Pruning Approach for Large Language Modelscited by 800
  4. 2021Accelerating Sparse Deep Neural Networkscited by 223
  5. 2025CodeElo: Benchmarking Competition-level Code Generation of LLMs with Human-comparable Elo Ratingscited by 115
  6. 2024Is C4 Dataset Optimal for Pruning? An Investigation of Calibration Data for LLM Pruningcited by 20
  7. 2024ALPS: Improved Optimization for Highly Sparse One-Shot Pruning for Large Language Modelscited by 19
  8. 2025PPC-GPT: Federated Task-Specific Compression of Large Language Models via Pruning and Chain-of-Thought Distillationcited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 4.

4 papers.

  1. 2026OBC-Prune: Outcome-Based Calibration for Large Reasoning Model Pruningcited by 1
  2. 2026Hardware-Native Joint Sparse-Quantization for Trillion-Scale Mixture-of-Expertscited by 0
  3. 2026Dynamics as Code: On Model Compression via Dynamic Systemcited by 0
  4. 2026Unlocking Structured Pruning for Reasoning Language Models via Attention-Preserving MLP Trimmingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Think Before You Prune: Self-Reflective Structured Pruning for Reasoning Language Models
  2. 2026Gradually Compacting Large Language Models for Reasoning Like a Boiling Frog
  3. 2025Pruning the Unsurprising: Efficient LLM Reasoning via First-Token Surprisal
  4. 2025Can Pruning Improve Reasoning? Revisiting Long-CoT Compression with Capability in Mind for Better Reasoning
  5. 2026Memory Augmentation Unlocks Efficient Chain-of-Thought Reasoning
  6. 2026On the Limits of Layer Pruning for Generative Reasoning in Large Language Models
  7. 2026Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression
  8. 2025Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy