ALAS: Multi-Agent LLM Planning System via Validator Isolation and Localized Cascading Repair Protocol

ICLR 2026 Conference Submission20691 Authors

Keywords: planning and scheduling, optimization, LLM, system

Abstract: Large language models enable flexible multi-agent planning but struggle with reliability: verification is often circular, state changes are not tracked for repair, and small faults trigger costly global replanning. We present ALAS, a multi-agent LLM Planning framework that separates planning from non-circular validation and performs localized repair guided by versioned execution logs. The validator operates independently of the planning LLM with fresh, bounded context, avoiding self-check loops and mid-context attrition. The repair protocol edits only the minimal affected region of the plan while preserving work-in-progress. On urban ride-sharing and job-shop scheduling across five classical benchmarks, the system matches or exceeds state-of-art Single-LLM and Multi-Agent System baselines, which achieves 83.7\% success rate, reduces 60\% token usage, and 1.82× Faster. A minimal reliability study shows that the validator detects injected structural faults with low overhead, and localized repair contains runtime perturbations with bounded edit radius and reduced makespan degradation versus global recompute. Code and seeds will be released. Results indicate that ALAS the combination of validator isolation and localized repair with execution logs provides measurable efficiency, optimality, and scalability for multi-agent LLM planning.

Primary Area: infrastructure, software libraries, hardware, systems, etc.

Submission Number: 20691