Research map / Large language models / Retrieval and knowledge
Behavioral Probes for Information Flow in LLM Swarms
Junhui Chen, Yu Zhang, Lantian Li, Dong Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies LLM swarm interactions through an information-flow lens by introducing behavioral probes that quantify how information is processed, integrated, and relied upon.
Problem
Existing LLM swarm optimization is utility-driven, focusing on final task utility and leaving how information is processed, integrated, and relied upon within individual interactions largely implicit.
Approach
Introduces behavioral probes that quantify information flow patterns in LLM swarms, repurposing External Context (abstract truncated, so the specific probe design is not stated).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine ECS and PKS behavioral profiles (Probe-Joint) with utility for node pruning | GPT Swarm-style DAG pruning · Random-DAG pruning, Qwen-2.5-32B | redundant LA/BA weight: 2.943 -> 0.997 | Improved |
| combine ECS and PKS behavioral profiles (Probe-Joint) with utility for node pruning | GPT Swarm-style DAG pruning · Random-DAG pruning, LLaMA-3-70B | redundant LA/BA weight: 2.228 -> 0.498 | Improved |
| Probe-Joint instead of GPT Swarm: token consumption | Node pruning · Random-DAG pruning, Qwen-2.5-32B | tokens per question: 2294 -> 1589 | Improved |
| Probe-Joint instead of utility-only baselines: accuracy | Node pruning · Random-DAG pruning, Qwen-2.5-32B | accuracy (%): 80.8 -> 80.0 | No clear effect |
| use ECS-only probe instead of joint ECS-PKS (ablation) | Probe-guided node pruning · Random-DAG pruning, Qwen-2.5-32B | redundant LA/BA weight: 0.997 -> 2.201 | Hurt |
| use PKS-only probe instead of joint ECS-PKS (ablation) | Probe-guided node pruning · Random-DAG pruning, LLaMA-3-70B | redundant LA/BA weight: 0.498 -> 2.174 | Hurt |
| add PKS-based convergence reward to utility after 50 warm-up iterations (ablation) | Reasoning-path search · Taxonomic path search, three warm-starts | critical-edge weight growth | Improved |
| remove WebSearch outputs at CombineAnswer stage (ablation) | GAIA ToT pipeline · GAIA Level-1 | PKS: paired mean difference 0.214 | Improved |
| remove an entire evidence branch (WebSearch or FileAnalyse) (ablation) | GAIA ToT pipeline · GAIA Level-1 | accuracy (%): 21.2 -> 15.4 | Hurt |
| prune only low-ECS evidence (ablation) | GAIA ToT pipeline · GAIA Level-1 | accuracy (%): 21.2 -> 21.2 | No clear effect |
| prune high-ECS evidence (ablation) | GAIA ToT pipeline · GAIA Level-1 | accuracy (%): 21.2 -> 15.4 | Hurt |
| increase context relevance (bubble to answer) | ECS probe, node-level · Kinship questions, Qwen2.5-32B | ECS (mean adjacent Cliff's delta): 0.624 | Improved |
Closest papers by meaning
- Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence (ICML 2025 · similarity 0.87)
- Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems (NeurIPS 2025 · similarity 0.86)
- Recognize Your Orchestrator: An Entropy Dynamics Perspective for LLM Multi-Agent Systems (ICML 2026 · similarity 0.85)
- FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.85)
- JADE: Bridging the Strategic-Operational Gap in Dynamic Agentic RAG (ICML 2026 · similarity 0.84)
- LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
