Paper 2601.05593
PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 16
- GitHub
- 338 stars
01 In brief
Summary
PaCoRe (Parallel Coordinated Reasoning) is a training-and-inference framework that scales test-time compute (TTC) beyond sequential reasoning limits by using parallel exploration coordinated via message passing.
It runs multiple rounds: each round generates parallel reasoning trajectories, compacts them into context-bounded messages, and synthesizes these to guide the next round, ultimately producing a final answer.
Trained with large-scale, outcome-based reinforcement learning, the model learns synthesis abilities, overcoming 'Reasoning Solipsism' where models ignore parallel insights.
PaCoRe-8B achieves 94.5% on HMMT 2025, surpassing GPT-5's 93.2%, by scaling effective TTC to ~2 million tokens.
It also improves on LiveCodeBench (78.2%), IMOAnswerBench (78.4%), and Apex (2.3%), outperforming RLVR-8B baseline.
Ablations show parallel scaling is more effective than sequential, and message passing is essential for unbounded scaling.
Training data is filtered to require genuine synthesis, and the model shows emergent cross-checking and correctness even from all-incorrect messages.
The framework generalizes to SWE-Verified (34.0% vs 29.8%) and MultiChallenge (48.0% vs 33.3%).
Resources are open-sourced.
02 From the paper
Abstract
We introduce Parallel Coordinated Reasoning (PaCoRe), a training-and-inference framework designed to overcome a central limitation of contemporary language models: their inability to scale test-time compute (TTC) far beyond sequential reasoning under a fixed context window. PaCoRe departs from the traditional sequential paradigm by driving TTC through massive parallel exploration coordinated via a message-passing architecture in multiple rounds. Each round launches many parallel reasoning trajectories, compacts their findings into context-bounded messages, and synthesizes these messages to guide the next round and ultimately produce the final answer. Trained end-to-end with large-scale, outcome-based reinforcement learning, the model masters the synthesis abilities required by PaCoRe and scales to multi-million-token effective TTC without exceeding context limits. The approach yields strong improvements across diverse domains, and notably pushes reasoning beyond frontier systems in mathematics: an 8B model reaches 94.5% on HMMT 2025, surpassing GPT-5's 93.2% by scaling effective TTC to roughly two million tokens. We open-source model checkpoints, training data, and the full inference pipeline to accelerate follow-up work.