#138138th of 142 in Agent frameworks and SDKsFor science, health and research, research and papers
latentmas
Multi-agent LLM reasoning that passes latent states between agents instead of text
- Stars
- 1.2k
- License
- Apache-2.0
- Last commit
- Jun 2026
- Language
- Python
870 on 21 Apr 20261,185 on 10 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
LatentMAS is the reference code for an ICML 2026 paper on multi-agent systems in which agents hand off latent thoughts and KV-cache working memory rather than text. It works with Hugging Face models, with an optional hybrid vLLM backend, and ships single-agent and text-based multi-agent baselines. The repo reproduces experiments on GSM8K, AIME24/25, GPQA, ARC, MBPP+, HumanEval+ and MedQA.
Who it is for: Researchers reproducing or extending latent-space multi-agent reasoning
Strengths
- Baseline, TextMAS and LatentMAS methods side by side for comparison on nine benchmarks
- Paper reports about 50-80% fewer tokens and 3x-7x faster wall-clock than text multi-agent baselines
- Training-free; works with any Hugging Face model, with optional vLLM backend
- Example logs include full agent interaction traces
Weaknesses
- Experiment runner (run.py), not an installable library or agent-building SDK
- vLLM path relies on patched vLLM internals and can differ numerically from HF results
- Recommended vLLM setup uses two GPUs
- No release published; example commands only cover Qwen3 models
What it needs
- GPU required
- Needs PyTorch, Hugging Face Transformers, vLLM (optional), conda
- Models: Hugging Face models, Qwen3-4B, Qwen3-8B, Qwen3-14B, vLLM
Also in Agent frameworks and SDKs
See all 142| Rank | Project | Score |
|---|---|---|
| 1 | ppt-masterAgent skill that turns documents into natively editable PowerPoint decks | 96 out of 100 |
| 2 | openai-agents-pythonPython SDK for multi-agent workflows with handoffs, guardrails and tracing | 92 out of 100 |
| 3 | mastraTypeScript framework for building AI agents, workflows and MCP servers | 92 out of 100 |
| #4 | browser-usePython library that lets an LLM agent control a web browser | 90 out of 100 |
| #5 | copilotkitReact and Angular SDK for adding agent chat and generative UI to apps | 90 out of 100 |