Research map / Multi-agent systems and games / LLM agents
Rethinking the Reliability of Multi-agent System: A Perspective from Byzantine Fault Tolerance
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.
Quantifies the Byzantine fault tolerance of LLM-based vs traditional agents in 7-node networks and proposes CP-WBFT, a confidence-probe-weighted consensus that uses prompt-level and hidden-state probes to tolerate up to 6 of 7 malicious agents.
Problem
Whether replacing traditional agents with LLM-based agents improves reliability of multi-agent systems under Byzantine faults, and how to exploit LLM confidence signals to build consensus that tolerates faults beyond the classical f < n/3 bound.
Approach
Treat LLM agents' self-reported confidence (prompt-level probe) and a linear probe on pooled hidden states (hidden-level probe) as weights in a two-stage consensus: agents adopt higher-confidence neighbor answers, then the answer with highest average confidence (ties broken by supporter count) is selected.
Key results
- HCP reaches 100% FAA and RA on complete graph for GSM8K and XSTest with +85.71% BFTI from 14.29% IAA, with 6 of 7 nodes Byzantine.
- PCP achieves up to +38.57% BFTI on star (leaf malicious) for GSM8K, but shows negative BFTI on some XSTest topologies (e.g., -12.86% on star with malicious center).
- Pilot study: traditional agents collapse to RA=0% beyond 2-3 malicious nodes, while LLM-based agents keep substantial accuracy at 6/7 malicious.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Hidden-level confidence probe (HCP) weighted consensus instead of prompt-level probe (PCP) (ablation) | Byzantine consensus over 7-node networks · GSM8K, 6 of 7 malicious, complete graph | BFTI (%): +85.71 vs +20.00 | Helped |
| Hidden-level confidence probe (HCP) instead of PCP (ablation) | Byzantine consensus over 7-node networks · XSTest, 6 of 7 malicious, tree | FAA (%): 57.14 vs 8.57 | Helped |
| Mean-pooled answer-token hidden states instead of last answer token for probe features (ablation) | HCP linear probe · GSM8K, LLaMA3.1 | Test Acc (%): 85.29 vs 84.23 | Helped |
| Mean-pooled answer-token hidden states instead of last answer token for probe features (ablation) | HCP linear probe · XSTest, LLaMA3.1 | Test Acc (%): 95.24 vs 80.16 | Helped |
| Mean-pooled answer-token hidden states instead of query final-token states (ablation) | HCP linear probe · GSM8K, LLaMA3.1 | Test Acc (%): 85.29 vs 71.27 | Helped |
| Mean-pooled answer-token hidden states on LLaMA3 backbone (ablation) | HCP linear probe · XSTest, LLaMA3 | Test Acc (%): 92.86 vs 76.19 | Helped |
| Prompt-level confidence probe (PCP) weighted consensus instead of unweighted traditional-style aggregation (ablation) | Byzantine consensus with GPT-4o-mini/GPT-3.5-turbo · GSM8K, 6 of 7 malicious, star (leaf malicious) | FAA (%): 100.00 final, +38.57 BFTI | Helped |
| Prompt-level confidence probe (PCP) on XSTest (ablation) | Byzantine consensus with GPT-4o-mini/GPT-3.5-turbo · XSTest, 6 of 7 malicious, star (center malicious) | BFTI (%): -12.86 | Hurt |
| Prompt-level confidence probe (PCP) on XSTest (ablation) | Byzantine consensus with GPT-4o-mini/GPT-3.5-turbo · XSTest, 6 of 7 malicious, varied topologies | BFTI (%): -12.86 to +21.43 | Mixed |
| Hidden-level confidence probe (HCP) weighted consensus on GSM8K (ablation) | Byzantine consensus with LLaMA3.1/LLaMA3 8B · GSM8K, 6 of 7 malicious, across topologies | FAA (%): 28.57 to 100.00 | Mixed |
| Traditional BFT-style agents vs LLM-based agents (pilot) | Multi-agent consensus over topologies · GSM8K and XSTest pilot, 6 malicious of 7 | RA (%): LLM-based RA up to 100 vs traditional RA 0 under heavy malicious ratio | Helped |
Surprising findings
- LLM-based agents with 6 of 7 nodes malicious still reach high consensus accuracy, far exceeding the classical f < n/3 bound.
- Prompt-level confidence probing yields negative BFTI on several XSTest topologies, while the hidden-level probe does not.
Limitations
- Evaluation uses only 10 curated questions per dataset in 7-node networks, with strong-weak model pairs chosen to create performance gaps.
- PCP results are inconsistent across tasks, with negative BFTI on several XSTest topologies.
- Hidden-level probe requires white-box access to model hidden states, limiting use with API models.
Experimental setup
Datasets: GSM8K XSTest CommonsenseQA
Benchmarks: GSM8K XSTest CommonsenseQA
Metrics: initial agent accuracy (IAA) final agent accuracy (FAA) round-level accuracy (RA) Byzantine Fault Tolerance Improvement (BFTI) accuracy precision F1 AUC
Baselines: Traditional agents (pilot setup) PCP (prompt-level confidence probe)
Builds on: PBFT HotStuff Tendermint
Closest papers by meaning
- On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents (ICML 2025 · similarity 0.88)
- HACO: Hedged Agent Computing for Reliable LLM Systems (NeurIPS 2026 · similarity 0.88)
- Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · similarity 0.86)
- SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication (AAAI 2026 · similarity 0.86)
- AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators (NeurIPS 2026 · similarity 0.86)
- AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.86)
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.