Stream of Belief: Eliciting Collective Intelligence through Latent Belief Dynamics
Organizations: MIT · UCLA · Harvard
Abstract
In this work, we view multi-agent reasoning as a process of seeking a collective equilibrium of implicit belief dynamics and introduce Stream of Belief (SoB), where agents maintain continuous latent beliefs and iteratively refine them through communication and reflection. Our method seeks a collective belief state of a multi-agent system where further communication induces no meaningful information propagation. Through extensive experiments, we demonstrate the generalization of SoB across tasks and model families. More Broadly, our results suggest a new perspective on scaling multi-agent reasoning from belief dynamics.
Figures & tables
| Model | Method | Math and science | Knowledge, logic, and multi-hop | ||||||
|---|---|---|---|---|---|---|---|---|---|
| GSM8K | MATH-500 | MMLU-Pro | GPQA-D | ARC-C | MedQA | BBH | HotpotQA | ||
| Qwen3-8B | Multi-agent debate | 96.5 1.3 | 86.4 2.2 | 75.3 2.5 | 53.0 3.4 | 94.5 1.6 | 73.7 2.6 | 93.5 1.2 | 70.7 3.6 |
| LatentMAS | 87.5 2.3 | 65.6 3.1 | 62.7 2.7 | 34.3 3.4 | 91.0 2.1 | 64.0 2.8 | 86.5 1.6 | 60.0 3.9 | |
| ThoughtComm | 43.5 3.5 | 5.2 1.4 | 16.0 2.2 | 13.1 2.5 | 69.5 3.3 | 33.3 2.7 | 24.8 2.0 | 48.7 4.2 | |
| Interlat | 84.5 2.5 | 47.2 3.2 | 25.7 2.4 | 13.1 2.4 | 85.0 2.5 | 52.7 3.0 | 56.8 2.4 | 59.3 4.0 | |
| SwiReasoning | 23.0 2.9 | 25.6 2.6 | 21.0 2.3 | 7.6 1.9 | 19.5 2.7 | 18.3 2.2 | 25.8 2.2 | 16.7 3.1 | |
| Model | Method | MATH-500 | GPQA-D | MedQA | BBH |
|---|---|---|---|---|---|
| Qwen3-0.6B | Multi-agent debate | 57.2 3.2 | 23.2 2.9 | 34.0 2.9 | 35.2 2.3 |
| LatentMAS | 14.4 2.1 | 14.1 2.4 | 16.3 2.1 | 43.5 2.5 | |
| SwiReasoning | 39.6 3.1 | 10.6 2.2 | 14.7 2.1 | 26.7 2.2 | |
| Soft Thinking | 26.0 2.8 | 3.5 1.3 | 12.0 2.0 | 24.0 2.1 | |
| Ours | 49.6 3.1 | 25.3 3.2 | 34.7 2.9 | 43.0 2.5 | |
| Qwen3-1.7B | Multi-agent debate | 77.6 2.6 | 34.8 3.4 | 47.7 2.8 | 72.5 2.3 |
Multi-Agent Reasoning
Multi-Agent Reasoning studies how multiple AI agents share evidence, model one another, and coordinate decisions on complex tasks.
See all 247 papers