stepfun-ai/PaCoRe

stepfun-ai/PaCoRe

Description: PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning

Language: Python

Stars: 334

Forks: 15

Open issues: 0

Created: 2025-12-09T10:43:09Z

Pushed: 2026-02-05T01:27:48Z

Default branch: main

Fork: no

Archived: no

README:

PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning

📖 Overview

We introduce PaCoRe (Parallel Coordinated Reasoning), a framework that shifts the driver of inference from sequential depth to coordinated parallel breadth, breaking the model context limitation and massively scaling test time compute:

• Think in Parallel: PaCoRe launches massive parallel exploration trajectories.
• Coordinate in Multi-rounds: It employs a message-passing architecture to compact these thoughts into concise messages and synthesize them to guide the next round.

Trained via large-scale, outcome-based reinforcement learning, PaCoRe masters the Reasoning Synthesis capabilities required to reconcile diverse parallel insights.

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!

------

*Figure 1 | Parallel Coordinated Reasoning (PaCoRe) performance. Left: On HMMT 2025, PaCoRe-8B demonstrates remarkable test-time scaling, yielding steady gains and ultimately surpassing GPT-5. Right: On LiveCodeBench, the RLVR-8B model fails to leverage increased test-time compute, while PaCoRe-8B model effectively unlocks substantial gains as the test-time compute increases.*

*Figure 2 | PaCoRe Training dynamics. Left panels: The Training Reward and Response Length steadily increase, demonstrating the training stability and effectiveness. Right panels: Evaluation on HMMT 2025 and LiveCodeBench (2408-2505). Performance is reported using single round coordinated reasoning in PaCoRe inference setting with $\vec{K} = [16]$.*

🔥 Releases

[2026/02/03] 🚀 PaCoRe Server is now open source!

• 🔗 Effortless to play PaCoRe with any LLM endpoint you have!
• 🍻 Even better—we've added first-class support for **Step 3.5 Flash** with **OpenRouter free provider**, StepFun's blazing-fast flagship model!
• 🎁 Check out the [Inference Pipeline](#inference-pipeline) section to get started!

[2025/12/09] We are excited to release the PaCoRe-8B ecosystem:

• 📝 In-depth Technical Report: **PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning.**
• 🤖 Model:
• PaCoRe-8B: Our final PaCoRe-trained model checkpoint!
• RLVR-8B-0926: The initial checkpoint of our study, conducted strong reasoning-oriented post-trained on Qwen3-8B-Base.
• 📚 Data: PaCoRe-Train-8k The high-quality training corpus, including opensource_math, public_mathcontest, synthetic_math and code:
• 🤗 Stage1-3k: PaCoRe-Train-Stage1-3k
• 🤗 Stage2-5k: PaCoRe-Train-Stage2-5k

🔍 Experiments

AIME 2025 HMMT 2025 IMO AnswerBench Apex LiveCodeBench HLEtext MultiChallenge

GPT-5 93.5 (13k) 93.2 (16k) 72.9 (26k) 1.0 (33k) 83.5 (13k) 26.0 (14k) 71.1 (5.0k)

Qwen3-235B-Thinking 91.6 (26k) 82.3 (32k) 71.7 (34k) 3.3 (46k) 74.5 (21k) 18.2 (23k) 60.3 (1.6k)

GLM-4.6 92.3 (20k) 88.7 (25k) 73.5 (37k) 0.7 (53k) 79.5 (19k) 17.2 (21k) 54.9 (2.2k)

DeepSeek-v3.1* 90.2 (16k) 86.1 (20k) 63.0 (27k) 1.4 (36k) 74.9 (11k) 19.3 (18k) 54.4 (1.1k)

Kimi-K2-Thinking 95.3 (25k) 86.5 (33k) 76.5 (44k) 0.8 (60k) 79.2 (25k) 23.9 (29k) 66.4 (1.6k)

RLVR-8B 84.1 (50k) 75.4 (48k) 64.6 (56k) 0.0 (65k) 70.6 (34k) 9.3 (35k) 33.3 (1.7k)

PaCoRe-8B (low) 89.7 (255k) 88.1 (243k) 76.1 (306k) 0.7 (362k) 75.8 (188k) 13.0 (196k) 41.8 (13k)

PaCoRe-8B (medium) 92.5 (908k) 92.9 (869k) 77.3 (1080k) 1.4 (1280k) 76.7 (659k) 14.6 (694k) 45.7 (45k)

PaCoRe-8B (high) 93.7 (1873k) 94.5 (1796k) 78.4 (2258k) 2.3 (2679k) 78.2 (1391k) 16.0 (1451k) 48.0 (95.3k)

*Table 1 | For each benchmark, we report accuracy together with total TTC (in thousands). For *Low*, *Medium*, and *High*, we apply the inference trajectory configuration as $\vec{K}=[4]$, $[16]$, and $[32, 4]$ separately.\* DeepSeek-V3.1 refers to the Terminus version.*

Key Findings

• Message Passing Unlocks Scaling. Without compaction, performance flatlines at the context limit. PaCoRe breaks the memory barrier and lets reasoning scale freely.
• Breadth > Depth. All compute is not equal. Coordinated parallel reasoning delivers far higher returns than extending a single chain.
• Data as a Force Multiplier. The PaCoRe corpus provides exceptionally valuable supervision—even baseline models see substantial gains when trained on it.

Getting Started 🚀

Data

The data is provided as a list[dict], where each entry represents a training instance:

• conversation: The original problem/prompt messages.
• responses: A list of cached generated responses (trajectories). These serve as the input messages ($M$) used during PaCoRe training.
• ground_truth: The verifiable answer used for correctness evaluation.

Model Serving

You can directly use vllm serve to serve the model! More inference details of PaCoRe will be handled in Inference Pipeline.

Inference Pipeline


*Figure 3 | Inference pipeline of PaCoRe. Each round launches broad parallel exploration, compacts the resulting trajectories into compacted messages, and feeds these messages together with the question forward to coordinate the next round. Repeating this process $\hat{R}$ times yields multi-million-token effective TTC while respecting fixed context limits, with the final compacted message serving as the system’s answer.*

We will explain the PaCoRe inference pipeline in this section.

PaCoRe Server Mode…