Research map / Multi-agent systems and games / LLM agents
GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling
Read the paperCheck your idea against 108,967 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.
Proposes GUARDIAN, an unsupervised temporal attributed graph encoder-decoder with information-bottleneck graph abstraction that detects and removes anomalous agents/edges to defend LLM multi-agent debates against hallucination amplification and injected errors.
Problem
Multi-agent LLM collaboration suffers from hallucination amplification and error injection/propagation (agent-targeted and communication-targeted), and existing defenses either check outputs in isolation or assume simplified agent dependencies and often need base-model modifications.
Approach
Model the multi-agent collaboration as a discrete-time temporal attributed graph (agents as nodes with BERT-embedded responses, communications as edges), then use an unsupervised GCN+Transformer encoder with dual attribute/structure reconstruction decoders, IB-based graph abstraction, and incremental training to score and prune anomalous nodes each round, leaving the LLMs unchanged.
Key results
- On MATH with GPT-3.5-turbo under hallucination amplification, GUARDIAN reaches 56.2% accuracy versus the best baseline (DyLAN 40.8%), reported as up to 15.4% improvement.
- Averaged across hallucination amplification scenarios, GUARDIAN reports a 4.2% average gain over state-of-the-art baselines; on MATH under agent-targeted attacks it reports 4.3% over the best baseline.
- GUARDIAN's average API calls are lowest among compared methods in most settings (e.g., 4.48 vs 8.00 for LLM Debate on MMLU with GPT-3.5-turbo).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply GUARDIAN instead of DyLAN | LLM multi-agent collaboration · MMLU, hallucination amplification, GPT-3.5-turbo | accuracy: 56.3 -> 57.2 | Helped |
| apply GUARDIAN instead of DyLAN | LLM multi-agent collaboration · MATH, hallucination amplification, GPT-3.5-turbo | accuracy: 40.8 -> 56.2 | Helped |
| apply GUARDIAN instead of DyLAN | LLM multi-agent collaboration · FEVER, hallucination amplification, GPT-3.5-turbo | accuracy: 32.3 -> 34.5 | Helped |
| apply GUARDIAN instead of DyLAN | LLM multi-agent collaboration · MMLU, agent-targeted error injection, GPT-3.5-turbo | accuracy: 55.2 -> 57.3 | Helped |
| apply GUARDIAN instead of DyLAN | LLM multi-agent collaboration · MATH, agent-targeted error injection, GPT-3.5-turbo | accuracy: 43.6 -> 52.2 | Helped |
| apply GUARDIAN instead of Inspector | LLM multi-agent collaboration · FEVER, agent-targeted error injection, GPT-3.5-turbo | accuracy: 32.1 -> 33.3 | Helped |
| apply GUARDIAN instead of DyLAN | LLM multi-agent collaboration · MMLU, communication-targeted error injection, GPT-3.5-turbo | accuracy: 52.6 -> 60.1 | Helped |
| apply GUARDIAN instead of Challenger | LLM multi-agent collaboration · MATH, communication-targeted error injection, GPT-3.5-turbo | accuracy: 45.2 -> 53.9 | Helped |
| apply GUARDIAN instead of DyLAN | LLM multi-agent collaboration · FEVER, communication-targeted error injection, GPT-3.5-turbo | accuracy: 34.1 -> 35.3 | Helped |
| add historical graph information rather than using only the current timestep graph (ablation) | GUARDIAN · MMLU, hallucination amplification, GPT-3.5-turbo | accuracy: 56.2 -> 57.2 | Helped |
| add historical graph information rather than using only the current timestep graph (ablation) | GUARDIAN · MATH, hallucination amplification, GPT-3.5-turbo | accuracy: 49.3 -> 56.2 | Helped |
| add historical graph information rather than using only the current timestep graph (ablation) | GUARDIAN · FEVER, hallucination amplification, GPT-3.5-turbo | accuracy: 34.1 -> 34.5 | Helped |
Surprising findings
- The SelfCheckGPT baseline performs very poorly on several datasets with GPT-3.5-turbo (e.g., 7.4% on MATH and 3.3% on FEVER), far below plain LLM Debate.
- Reported FDR stays low (8.32% to 30.67%), and the paper argues that wrongly removed correct agents have minimal impact on system performance.
Limitations
- Each dataset gets its own trained model, and evaluation is in-distribution only with 100 sampled questions per dataset, over three iterations.
- Accuracy gains are not uniform: GUARDIAN underperforms GUARDIAN.s or baselines in some cases, such as GPT-4o on MMLU under hallucination amplification.
- Attacks are simulated under fixed rules (random single-agent or edge perturbation), and the claimed theoretical bounds are loosely connected to the empirical method.
Experimental setup
Datasets: MMLU MATH FEVER Biographies
Benchmarks: MMLU MATH FEVER Biographies
Metrics: accuracy anomaly detection rate false discovery rate API calls runtime
Baselines: LLM Debate DyLAN SelfCheckGPT Challenger Inspector
Builds on: LLM Debate DyLAN GCN Information Bottleneck SelfCheckGPT
Closest papers by meaning
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.89)
- SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication (AAAI 2026 · similarity 0.88)
- When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems (ICML 2026 · similarity 0.88)
- Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.87)
- Securing Multi-Agent Systems Against Corruptions via Node Contribution Backpropagation (ICML 2026 · similarity 0.87)
- Latent Collaboration in Multi-Agent Systems (ICML 2026 · similarity 0.87)
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.