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

Agent Primitives: Reuseable Latent Building Blocks for Multi-Agent Systems

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making

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 Agent Primitives (Review, Voting and Selection, Planning and Execution) as reusable latent building blocks for LLM multi-agent systems, communicating via KV cache and composed by an LLM Organizer with a knowledge pool.

Problem

Existing LLM multi-agent systems are task-specific, rely on handcrafted roles and prompts, and communicate in natural language, which degrades in long-context and noisy multi-stage interactions.

Approach

Decompose MAS into reusable latent computation patterns (primitives) that exchange information through KV cache rather than text, with an LLM Organizer that composes primitives per query guided by a knowledge pool of past successful MAS configurations.

Key results

• Primitives-based MAS reaches 75.3% average accuracy on Qwen3-8B vs 58.8% single-agent (+16.5 pp).
• On DeepSeek-R1-Distill-Llama-70B, primitives-based MAS achieves 76.4% average vs 70.1% single-agent and 60.0% for LatentMAS.
• Primitives-based MAS reaches 53.2% on GPQA vs 33.6-40.2% for prior MAS methods under Llama-3-70B.

Surprising findings

• LatentMAS drops below single-agent accuracy by 10.1 pp on DeepSeek-R1-Distill-Llama-70B, showing latent communication can be unstable across backbones.
• Removing RoPE re-encoding for KV communication causes severe drops on the Llama-based model (e.g., HumanEval+ 85.3% to 31.1%).

Limitations

• Restricted to LLM agents sharing the same model parameters and positional encoding, so cross-model configurations are not considered.
• Primitive set is limited to three primitives; Organizer depends on a knowledge pool of 45 MAS structures.

Experimental setup

Datasets: InfiniteBench En.QA GSM8K

Benchmarks: AIME24 AIME25 MATH GSM8K HumanEval+ MBPP+ MedQA GPQA-Diamond

Metrics: accuracy average accuracy token usage inference latency

Baselines: Single agent TextMAS LatentMAS LLM-Debate Self-Refine Quality-Diversity SPP AgentVerse GPTSwarm DyLAN MAS-GPT Chain-of-Thought

Builds on: LatentMAS MAS-GPT AFlow DyLAN Multi-Agent Debate

Closest papers by meaning

• Latent Collaboration in Multi-Agent Systems (ICML 2026 · similarity 0.91)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.90)
• Multi-agent Architecture Search via Agentic Supernet (ICML 2025 · similarity 0.89)
• AgentRefine: Enhancing Agent Generalization through Refinement Tuning (ICLR 2025 · similarity 0.89)
• AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.89)
• MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems (ICML 2025 · similarity 0.89)