Paper 2511.20639
Latent Collaboration in Multi-Agent Systems
- Published
- Nov 2025
- Research lab
- Independent
- Citations
- 35
- GitHub
- 1.1K stars
01 In brief
Summary
LatentMAS is a training-free framework enabling multi-agent systems (MAS) to collaborate entirely in the continuous latent space, bypassing text-based communication.
Each agent generates latent thoughts via auto-regressive last-layer hidden states, and a shared latent working memory (KV caches) transfers these representations losslessly to subsequent agents.
A linear alignment operator (W_a) maps hidden states back to valid input embeddings, preventing distribution drift.
Theoretical analyses show latent thoughts are more expressive (requiring Ω(d_h m / log|V|) fewer tokens) and that latent working memory preserves information exactly.
Empirically, across 9 benchmarks (math, science, commonsense, code) and 5 LLM backbones (Qwen3, Llama3), LatentMAS outperforms single models and text-based MAS, improving accuracy by up to 14.6%, reducing token usage by 70.8%–83.7%, and achieving 4×–4.3× faster inference.
The framework is agnostic to MAS architectures (sequential or hierarchical) and includes a debug mode for interpreting latent thoughts.
LatentMAS demonstrates that latent collaboration enhances reasoning quality and efficiency without additional training.
02 From the paper
Abstract
Multi-agent systems (MAS) extend large language models (LLMs) from independent single-model reasoning to coordinative system-level intelligence. While existing LLM agents depend on text-based mediation for reasoning and communication, we take a step forward by enabling models to collaborate directly within the continuous latent space. We introduce LatentMAS, an end-to-end training-free framework that enables pure latent collaboration among LLM agents. In LatentMAS, each agent first performs auto-regressive latent thoughts generation through last-layer hidden embeddings instead of text. Then, a shared latent working memory preserves and transfers each agent's internal representations and latent thoughts, ensuring lossless information exchange without re-encoding. We provide detailed theoretical analyses showing that LatentMAS achieves higher expressiveness and lossless information preservation with lower overall complexity than standard text-based MAS. In addition, empirical evaluations across 9 comprehensive benchmarks spanning math and science reasoning, commonsense understanding, and code generation show that LatentMAS outperforms advanced single agents and text-based MAS baselines, achieving up to 14.6% higher accuracy, reducing output token usage by 70.8%-83.7%, and providing 4$\times$-4.3$\times$ faster end-to-end inference. Code and data are fully open-sourced at https://github.com/Gen-Verse/LatentMAS.