AI Mindmeld connectors
AI Mindmeld connects scientific specialist models to a large language model through the Latent Interconnect Protocol (LIP). Its connectors are learned mappings that bring a specialist's representation of an object directly into the LLM's hidden states at a slot in the prompt, while both models stay frozen; independently trained connectors can be combined in one prompt without joint training.
This repository holds the 29 trained release connectors, plus two DEMO connectors for software examples. They connect three protein specialist families (ESM-C 600M and 6B, PGLM-1B and MIMIC) and three molecule specialist families (ChemFM-1B and 3B, MoLFormer-c3 and Uni-Mol2 1.1B) to Gemma 3 4B (8 connectors), Gemma 3 12B (10) and Kimi K2.6 (11). Each connector is specific to its specialist, extraction recipe and LLM; specialist and LLM weights are separate downloads.
Four of them (*-all-tasks) are trained on all tasks, so they have no zero-shot evaluation.
Download
Install the release and set LIP_DATA as shown in the Quickstart. These commands use the checkout’s pinned release revisions by default.
python -m latent_interconnect.release_download \
--connector gemma-3-4b-it__esmc-600m --connector gemma-3-4b-it__chemfm-1b \
--data protein-esmc-600m-mini --data molecule-chemfm-1b-mini \
--dest $LIP_DATA
Use python -m latent_interconnect.release_download --list --kinds to see bundle names, sizes, kinds, and DEMO
markers. The data inventory lists the data each connector needs.
Links
- Code and documentation: https://github.com/PolymathicAI/AI-Mindmeld
- Paper: https://arxiv.org/abs/2610.30215
- LIP library (connector descriptions and scores) and ranking details
- Training data: sgolkar/ai-mindmeld-data
License
The AI Mindmeld authors license the connector weights, including the DEMO connectors, under CC BY 4.0. Please credit:
Siavash Golkar, Jake Kovalic, and Shirley Ho. AI Mindmeld: A Latent Interconnect Protocol. arXiv:2610.30215, 2026.
Software: Siavash Golkar, Dennis Bollweg, and the AI Mindmeld contributors, AI Mindmeld, https://github.com/PolymathicAI/AI-Mindmeld.
The grant covers only rights the authors can license. Upstream terms of the receiver LLMs and specialist models apply on top. For ChemFM-1B and ChemFM-3B, this release follows the ChemFM project license, CC BY-NC 4.0; their Hugging Face metadata instead lists MIT. The PGLM-1B model card specifies CC BY-NC 4.0. See the list of non-commercial items and the inventory's license notes. The code is Apache-2.0.
Citation
Use the credit line above: cite the paper first, then the repository, with the software revision you used (CITATION.cff).