Research map / Multi-agent systems and games / LLM agents

Learning Latency-Aware Orchestration for Multi-Agent Systems

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 1 citations (OpenAlex)

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 LAMaS, a latency-aware orchestrator for LLM multi-agent systems that combines Lagrangian-constrained training with critical-path credit assignment and a learned inference-time controller that eliminates redundant agent interactions, cutting latency by over 50% on four benchmarks.

Problem

Multi-agent LLM systems incur high end-to-end latency governed by the critical execution path, but existing automated orchestration methods optimize accuracy and cost and leave latency largely unaddressed.

Approach

Treat latency as a constrained objective (minimize latency under an accuracy floor via adaptive Lagrangian dual), assign latency penalty to operators in proportion to their critical-path criticality, and add a lightweight MLP controller that predicts whether to stop launching further operators based on partial-execution features.

Key results

• On GSM8K, LAMaS reaches 93.65% accuracy with 11.73 s latency versus 48.11 s for MaAS (93.36%).
• On HumanEval, LAMaS reaches 95.42% pass@1 at 21.00 s versus 47.32 s for MaAS (93.38%).
• Plugging the LAMaS pipeline into GDesigner and AgentDropout reduces latency by 27–39% with accuracy preserved within noise.

Surprising findings

• Naive latency-only reward (no accuracy floor, no CP credit, no controller) drops MMLU-Pro accuracy by about 5 pp relative to LAMaS.
• Removing the latency objective (lambda_t=0) leaves accuracy unchanged but raises latency by 24-70%, indicating that the latency term reshapes graphs without accuracy loss.

Limitations

• Accuracy floor is set relative to the reference architecture's validation accuracy, so gains depend on that reference choice.
• Transfer gains to GDesigner and AgentDropout (27-39%) are smaller than in the native setting.
• Evaluated on small training splits (e.g., 33 HumanEval training queries) and a single closed-source backbone family as main setting.

Experimental setup

Datasets: GSM8K HumanEval MATH MMLU-Pro

Benchmarks: GSM8K HumanEval MATH MMLU-Pro

Metrics: accuracy pass@1 cost (USD per query) end-to-end latency (s) CP tokens

Baselines: Generate CoT SC GDesigner AgentDropout MaAS

Builds on: MaAS GDesigner AgentDropout Chain-of-thought prompting Self-consistency

Closest papers by meaning

• MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks (ICML 2026 · similarity 0.91)
• Reward Modeling for Multi-Agent Orchestration (NeurIPS 2026 · similarity 0.91)
• Learning to Share: Selective Memory for Efficient Parallel Agentic Systems (ICML 2026 · similarity 0.88)
• Agent Primitives: Reuseable Latent Building Blocks for Multi-Agent Systems (ICML 2026 · similarity 0.88)
• Recognize Your Orchestrator: An Entropy Dynamics Perspective for LLM Multi-Agent Systems (ICML 2026 · similarity 0.88)
• A Two-Tier Perspective on Inference-Time Parallelism in Multi-Agent LLM Systems (ICML 2026 · similarity 0.88)