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

MonoScale: Scaling Multi-Agent System with Monotonic Improvement

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent collaboration and safety · 0 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 MonoScale, an expansion-aware router update framework that onboards new agents via agent-conditioned warm-up tasks and trust-region natural-language memory updates, yielding monotonic non-decreasing performance across agent-pool expansions.

Problem

Adding new heterogeneous, unreliable agents to an LLM multi-agent system's pool causes router cold-start mis-routing and performance collapse during scale-up.

Approach

Treat sequential agent augmentation as a contextual bandit; proactively synthesize agent-conditioned familiarization tasks, distill successes and failures into auditable natural-language routing memory, and select memory updates under a TRPO-style KL trust region with a conservative fallback that disables the new agent, giving a conditional monotonic guarantee.

Key results

• On GAIA, Qwen3-30B-A3B MonoScale with memory improves overall accuracy from 44.84% (3 agents) to 55.15% (10 agents), while naive scale-up drops to 42.42% at 10 agents.
• On HLE MCQ, MonoScale rises from 11.70% (3 agents) to 19.88% (10 agents), surpassing GPT-5 fixed-pool at 19.63%.
• Under the malfunctioning agent pool, Gemini-3-Pro without memory shows performance collapse during scaling while MonoScale remains stable and slightly improves.

Surprising findings

• Naive scale-up with a strong router (Gemini-3-Pro) can collapse under noisy agent pools, even though it is the strongest model tested.
• A 30B open-weight router with memory updates at 10 agents outperforms a fixed-pool GPT-5 orchestration on HLE.

Limitations

• Monotonic guarantee is conditional on a non-interfering expansion assumption that may fail under environment drift.
• Experiments only up to 10 agents (20-agent preliminary probe); memory saturation and long horizons not established.
• Main baselines compared at fixed 3-agent pool rather than best-scaled strong-router configuration.

Experimental setup

Datasets: GAIA validation set Humanity's Last Exam MCQ subset

Benchmarks: GAIA validation Humanity's Last Exam MCQ

Metrics: accuracy

Baselines: Naive scale-up (Qwen3-30B-A3B-Instruct w/o memory) GPT-5 Gemini-3-Pro DeepSeek-V3.2 Qwen3-235B-Instruct GLM-4.7 Gemini-2.5-Pro GPT-5.2

Builds on: TRPO Training-Free GRPO camel-ai OWL

Closest papers by meaning

• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.88)
• Learning to Share: Selective Memory for Efficient Parallel Agentic Systems (ICML 2026 · similarity 0.88)
• ScaleSim: Serving Large-Scale Multi-Agent Simulation with Invocation Distance-Based Memory Management (ICML 2026 · similarity 0.87)
• The Scaling Laws of Skills in LLM Agent Systems (NeurIPS 2026 · similarity 0.87)
• G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
• Learning Latency-Aware Orchestration for Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)