H-CSC: Byzantine-Resilient LLM-Agent Consensus with Semantic Finality
95% commit accuracy under Byzantine attack with explicit abort reasons for malicious proposals.
Current multi-agent LLM systems lack a reliable finality primitive when agents act maliciously (Byzantine faults). Standard Byzantine fault tolerance (BFT) assumes byte-level identity, but LLM proposals are stochastic natural language with no fixed byte representation. A new paper from Haoran Xu et al. introduces Hierarchical Certified Semantic Commitment (H-CSC), a BFT-inspired protocol that converts embedding-derived similarity signals into three typed outcomes: a semantic_commit (backed by a 2f+1 semantic core), a verdict_commit (strong verdict margin but dispersed semantics), or an explicit typed abort. This replaces naive aggregation or single-verdict approaches with a structured finality mechanism that understands meaning, not just byte patterns.
Evaluated on a controlled semantic-poisoning diagnostic (BCS_v1, 120 episodes), H-CSC commits with angular deviation as low as 0.31° on BFT-feasible rounds and aborts 100% of beyond-BFT rounds (n < 3f+1). On the MVR-50 claim-verification benchmark under paired static and rushing Byzantine attacks, H-CSC achieves commit rates of 0.90/0.92 with honest-reference-invalid rates of 0.02/0.00, matching a strong basline while additionally emitting an embedding-backed semantic_commit digest on 74%/72% of rounds. A strict-semantic ablation drops to 0.54/0.48, proving the verdict-level fallback is essential for coverage (+0.36/+0.44) without sacrificing safety. Cross-model tests across four LLMs keep invalid_hmaj within 0.00–0.03. The protocol offers a practical path to trusted, resilient multi-agent AI systems.
- H-CSC provides three typed finality outcomes: semantic_commit, verdict_commit, or explicit typed abort.
- Achieves 0.90/0.92 commit accuracy under static and rushing Byzantine attacks on MVR-50 benchmark.
- Emits embedding-backed semantic digest on 74% of rounds, enabling typed provenance for accepted proposals.
Why It Matters
Enables reliable multi-agent LLM collaboration despite malicious agents, critical for autonomous AI systems.