Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning

2026-04-17

AI Summary

This work addresses the high computational cost in parallel inference caused by early errors that propagate through numerous invalid reasoning paths. The paper introduces the first systematic taxonomy for path pruning and proposes STOP, a learnable internal-signal-based pruning method grounded in this framework. STOP leverages a Super Token mechanism to dynamically model internal confidence at the prefix level, enabling effective elimination of unpromising paths during inference. Evaluated across models ranging from 1.5B to 20B parameters, STOP consistently outperforms existing baselines. Notably, under fixed computational budgets, it boosts the accuracy of GPT-OSS-20B on AIME25 from 84% to nearly 90%, while also offering practical deployment guidelines for real-world applications.

Technology Category

• Search and Optimization: Learning to Search
• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

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

Problem

• path pruning
• parallel reasoning
• Large Reasoning Models
• compute efficiency
• futile paths

Innovation

• path pruning
• learnable internal signals
• STOP
• parallel reasoning
• Large Reasoning Models