Parallel-Probe: Towards Efficient Parallel Thinking via 2D Probing

February 3, 2026 · View on GitHub

Overview of Parallel-Probe framework

🚀 Parallel-Probe is a training-free controller for efficient parallel reasoning in large language models.

Using 2D Probing, we reveal global width–depth dynamics of parallel trajectories, uncovering key inefficiencies such as non-monotonic scaling, heterogeneous branch lengths, and early consensus.

Built on these insights, Parallel-Probe adaptively prunes branches and enables early stopping, achieving superior accuracy–efficiency trade-offs with reduced latency and token cost.

Observations

Key Contributions

• 2D Probing Interface
A structured width × depth probing matrix that reveals global parallel reasoning dynamics during decoding.

• Dynamics Analysis
We uncover three fundamental properties:

  • Non-monotonic width–depth scaling under fixed token budgets
  • Long-tailed heterogeneous branch lengths
  • Early stabilization of global consensus

• Parallel-Probe Controller
A principled training-free policy that:

  • Prunes divergent branches to reduce width
  • Stops generation once global consensus stabilizes to reduce depth

• SCOUT Testbed
An offline evaluation framework that decouples trajectory generation from control for fair and efficient test-time scaling research.

SCOUT Online

Results

Across multiple models and challenging reasoning benchmarks, Parallel-Probe consistently achieves superior accuracy–efficiency trade-offs, reducing:

  • Sequential tokens (latency proxy) by over 30%
  • Total token cost by over 20%

while maintaining competitive accuracy compared to standard self-consistency baselines.

Base Model: Qwen3-0.6B

MethodTypeAIME24 AccAIME24 SeqTokAIME24 TokAIME25 AccAIME25 SeqTokAIME25 TokHMMT25 AccHMMT25 SeqTokHMMT25 TokAvg AccAvg SeqTokAvg Tok
SC@64Parallel21.432.7k1008.6k28.931.1k890.5k18.131.0k937.8k22.831.6k945.7k
ASCSeq21.4805.5k805.5k28.9653.8k653.8k18.1580.8k580.8k22.8680.0k680.0k
ESCHybrid21.4192.9k986.7k28.9171.8k868.8k18.1179.5k923.9k22.8181.4k926.5k
SC@64+SACParallel19.526.8k820.7k25.427.2k819.4k17.426.3k808.2k20.726.8k816.1k
Parallel-ProbeParallel21.820.8k773.8k29.719.6k697.8k18.520.5k734.5k23.320.3k735.3k

Base Model: Qwen3-1.7B

MethodTypeAIME24 AccAIME24 SeqTokAIME24 TokAIME25 AccAIME25 SeqTokAIME25 TokHMMT25 AccHMMT25 SeqTokHMMT25 TokAvg AccAvg SeqTokAvg Tok
SC@64Parallel72.531.4k1025.8k44.430.0k1054.1k24.232.4k1132.9k47.031.3k1070.9k
ASCSeq72.3482.6k482.6k44.4600.9k600.9k24.2586.3k586.3k47.0556.6k556.6k
ESCHybrid72.5170.4k909.2k44.4160.6k913.8k24.2174.9k1014.2k47.0168.6k945.7k
SC@64+SACParallel64.527.3k868.2k40.026.4k909.0k21.426.9k889.1k42.026.9k888.8k
Parallel-ProbeParallel68.120.5k748.5k44.721.3k775.8k22.622.8k860.2k45.121.5k794.8k

Base Model: Qwen3-4B

MethodTypeAIME24 AccAIME24 SeqTokAIME24 TokAIME25 AccAIME25 SeqTokAIME25 TokHMMT25 AccHMMT25 SeqTokHMMT25 TokAvg AccAvg SeqTokAvg Tok
SC@64Parallel80.029.3k886.8k76.630.5k1088.1k43.633.9k1168.3k66.831.2k1047.7k
ASCSeq80.0214.2k214.2k76.6325.1k325.1k43.6487.3k487.3k66.7342.2k342.2k
ESCHybrid80.098.9k528.9k76.6137.0k793.3k43.6174.0k990.2k66.8136.6k770.8k
SC@64+SACParallel80.024.8k782.2k73.327.9k995.4k41.927.1k863.0k65.126.6k880.2k
Parallel-ProbeParallel79.719.2k688.9k76.122.2k806.0k44.721.5k872.3k66.820.9k789.0k

Base Model: Qwen3-8B

MethodTypeAIME24 AccAIME24 SeqTokAIME24 TokAIME25 AccAIME25 SeqTokAIME25 TokHMMT25 AccHMMT25 SeqTokHMMT25 TokAvg AccAvg SeqTokAvg Tok
SC@64Parallel80.430.1k910.8k76.730.7k1124.4k48.934.8k1267.0k68.631.9k1100.7k
ASCSeq80.4226.0k226.0k76.7406.2k406.2k48.8565.1k565.1k68.6399.1k399.1k
ESCHybrid80.484.7k459.4k76.7132.4k793.1k48.9184.5k1062.1k68.6133.9k771.5k
SC@64+SACParallel76.725.6k773.4k70.228.1k998.5k42.728.5k896.8k63.227.4k889.5k
Parallel-ProbeParallel81.520.3k730.8k76.921.9k846.7k47.122.4k897.2k68.521.6k824.9k

Quick Start

This repository includes a partial release of the trajectory data for local evaluation. For complete benchmarks and large-scale experiments, please use the SCOUT Online Platform (We will support soon).

pip install pandas & pip install scipy

cd evaluation

python evaluation_main_table.py