Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning
On the map: LLM reasoning → Test-time compute and efficient thinking
Abstract (arXiv)
Parallel reasoning enhances Large Reasoning Models (LRMs) but incurs prohibitive costs due to futile paths caused by early errors. To mitigate this, path pruning at the prefix level is essential, yet existing research remains fragmented without a standardized framework. In this work, we propose the first systematic taxonomy of path pruning, categorizing methods by their signal source (internal vs. external) and learnability (learnable vs. non-learnable). This classification reveals the unexplored potential of learnable internal methods, motivating our proposal of STOP (Super TOken for Pruning). Extensive evaluations across LRMs ranging from 1.5B to 20B parameters demonstrate that STOP achieves superior effectiveness and efficiency compared to existing baselines. Furthermore, we rigorously validate the scalability of STOP under varying compute budgets - for instance, boosting GPT-OSS-20B accuracy on AIME25 from 84% to nearly 90% under fixed compute budgets. Finally, we distill our findings into formalized empirical guidelines to facilitate optimal real-world deployment. Code, data and models are available at https://bijiaxihh.github.io/STOP
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 19 of its 29 references that are in the map. First posted on arXiv 17 Apr 2026.
Builds on
19 of 29 references.
Closest ancestors · 11
Same direction, newest first
- 2026One-Token Verification for Reasoning Correctness Estimation1 Mar
- 2025A Survey on Parallel Reasoning14 Oct
- 2025DeepPrune: Parallel Scaling without Inter-trace Redundancy9 Oct
- 2025Think Just Enough: Sequence-Level Entropy as a Confidence Signal for LLM Reasoning9 Oct
- 2025Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency17 Sep
- 2025Deep Think with Confidence21 Aug
- 2025Lost at the Beginning of Reasoning27 Jun
- 2025Don't Overthink it. Preferring Shorter Thinking Chains for Improved LLM Reasoning23 May
- 2025The Energy Cost of Reasoning: Analyzing Energy Usage in LLMs with Test-time Compute20 May
- 2025Learning from Peers in Reasoning Models12 May
- 2025Multi-Agent Verification: Scaling Test-Time Compute with Multiple Verifiers27 Feb
Other references · 8
- 2020Language Models are Few-Shot Learnerscited by 32,199
- 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
- 2022Self-Consistency Improves Chain of Thought Reasoning in Language Modelscited by 5,601
- 2024Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Headscited by 821
- 2023Chameleon: Plug-and-Play Compositional Reasoning with Large Language Modelscited by 352
- 2025Process Reward Models That Thinkcited by 118
- 2025GenPRM: Scaling Test-Time Compute of Process Reward Models via Generative Reasoningcited by 94
- 2025AdaDec: A Uncertainty-Guided Lookahead Decoding Framework for LLM-Based Code Generationcited by 2
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2025Can Pruning Improve Reasoning? Revisiting Long-CoT Compression with Capability in Mind for Better Reasoning
- 2026Hidden States as Early Signals: Step-level Trace Evaluation and Pruning for Efficient Test-Time Scaling
- 2026Revisiting the Effectiveness of LLM Pruning for Test-Time Scaling
- 2025When Fewer Layers Break More Chains: Layer Pruning Harms Test-Time Scaling in LLMs
- 2025Pruning the Unsurprising: Efficient LLM Reasoning via First-Token Surprisal
- 2026OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping
- 2025Think Before You Prune: Self-Reflective Structured Pruning for Reasoning Language Models
- 2026From LLMs to LRMs: Rethinking Pruning for Reasoning-Centric Models