# Gen-Verse/LatentMAS

[ICML 2026 Spotlight] Latent Collaboration in Multi-Agent Systems

Repository: https://github.com/Gen-Verse/LatentMAS
Canonical: https://ross.abutalabs.com/products/latentmas
Homepage: https://arxiv.org/abs/2511.20639
Language: Python
License: Apache-2.0
License Family: permissive
Topics: latent-reasoning, model-collaboration, multi-agent-systems, continuous-reasoning, large-language-models, latent-space-model
Last push: 2026-06-18T06:56:18+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 88, release rhythm 35, longevity 20
- inputs: {"age_days": 282, "days_push": 76, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1099, forks 168 (observed 2026-08-28T04:03:34.975934+00:00)

## What it is
LatentMAS is a training-free multi-agent reasoning framework that lets LLM agents collaborate in continuous latent space by passing hidden-state 'latent thoughts' through a shared latent working memory instead of text. It works with any Hugging Face model (optionally vLLM) and delivers higher accuracy, 70-84% fewer output tokens, and 4x faster inference than text-based multi-agent baselines.

## Use cases
- run multi-agent LLM collaboration with fewer tokens
- speed up multi-agent reasoning pipelines
- make LLM agents communicate in latent space instead of text
- improve math and science reasoning accuracy with multiple agents
- reduce inference cost of multi-agent systems
- experiment with latent-space agent communication on Hugging Face models
- apply latent collaboration to scientific discovery workflows

## When to choose
- you want multi-agent LLM reasoning with lower token usage and faster wall-clock time
- you need a training-free technique compatible with any Hugging Face model
- you're researching latent-space reasoning or agent communication
- you run math, science, commonsense, or code generation benchmarks with LLM agents

## When to avoid
- you need a production multi-agent orchestration platform with tool use and APIs
- your models or backend aren't supported by Hugging Face or vLLM
- you require text-auditable agent communication traces
- you need a battle-tested framework with long-term community support

## Facets
- artifact type: library
- maturity: active
- function: agent-framework, llm-inference, machine-learning
- domain: large-language-models, machine-learning, artificial-intelligence
- platform: python
- tags: latent-reasoning, multi-agent-systems, training-free, latent-space, vllm, huggingface, research-code, icml-2026, ai-agents, gpu

## Member repositories
- Gen-Verse/LatentMAS (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:34.975934+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:45:57.165263+00:00, confidence not recorded.
  - readme: https://github.com/Gen-Verse/LatentMAS (fetched 2026-08-28T04:03:34.975934+00:00, sha c039b5c48f3d)
  - homepage: https://arxiv.org/abs/2511.20639 (fetched 2026-08-29T12:49:30.795481+00:00, sha 22d66d0bc4b8)
  - site_page: https://info.arxiv.org/about/donate.html (fetched 2026-08-29T12:49:30.798198+00:00, sha cca9c3a11c56)
  - site_page: https://info.arxiv.org/about/ourmembers.html (fetched 2026-08-29T12:49:30.801526+00:00, sha 47cbc55ff1de)
  - site_page: https://info.arxiv.org/about (fetched 2026-08-29T12:49:30.803344+00:00, sha a1f16f915a9a)
  - site_page: https://info.arxiv.org/labs/index.html (fetched 2026-08-29T12:49:30.799872+00:00, sha b14a8d05a0ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.