Message Passing Enables Efficient Reasoning

2026-07-01

AI Summary

This work addresses the high computational cost of chain-of-thought (CoT) reasoning in large language models and the limited scalability of existing parallel approaches—such as fork-join—due to inefficient inter-thread communication. To overcome these bottlenecks, the authors propose the Message-Passing Language Model (MPLM) framework, which introduces lightweight point-to-point message-passing primitives to enable direct multi-thread communication. MPLM further incorporates context deduplication and a partial-information-based preemptive termination mechanism to enhance inference efficiency. Experimental results demonstrate that MPLM substantially reduces context length while successfully solving challenging 25×25 Sudoku puzzles, improves reasoning efficiency via preemption on 3-SAT problems, and achieves competitive performance against state-of-the-art fork-join methods in long-context question answering.

Abstract

While inference-time scaling has improved the reasoning abilities of large language models (LLMs), the need to generate long chains-of-thought (CoTs) is a computational bottleneck. Thus, in contrast to sequential scaling methods like CoT, recent parallel scaling techniques instead use fork and join (FJ) primitives to divide work across multiple LLM threads. However, in the fork-join paradigm, threads are typically transient and do not communicate pointwise with one another which limits scalability. To tackle this, we introduce Message Passing Language Models (MPLMs), a framework for LLM reasoning in which threads communicate directly via lightweight send and receive primitives. MPLMs enable efficient scaling through two key mechanisms: (1) reduced communication costs, achieved by avoiding redundant context sharing, and (2) preemption, which allows threads to terminate early based on partial information from their peers. We demonstrate the promise of MPLMs on 3 classes of tasks. First, on Sudoku puzzles, we show that MPLMs require an asymptotically smaller context than both serial CoT and parallel FJ. We then fine-tune a single model to solve 25 x 25 puzzles that remain challenging for standard CoT and FJ approaches, as well as frontier reasoning models without tools. Second, on 3-SAT puzzles, the capability of preemption allows termination of unpromising branches, which results in improved efficiency. Finally, we show that appropriately prompted large pre-trained models follow the MPLM protocol, achieving competitive results on long-context question answering relative to popular fork-join approaches.

Problem

• chain-of-thought
• fork-join
• reasoning efficiency
• message passing
• large language models

Innovation

• Message Passing
• Parallel Reasoning
• Preemption
• Fork-Join
• Context Efficiency