Membrane Labs · Research
new capability without retraining the base model. the gate decides what crosses — and when it closes, the base model answers alone.
Thesis
Most AI systems are monoliths: one model trained on everything, updated as a whole, opaque about which capability shaped an answer. Fine-tuning every new task into the same weights costs compute, risks forgetting, and makes audits hard.
Membrane studies a different contract: train focused specialists independently, attach them to a frozen base through learned interfaces, and activate them only when relevant. New capability, new specialist interface, unchanged base model.
Compositional intelligence is not only about how models communicate. It is also about when communication is licensed.
First result
In our first controlled experiment, we attach a biomedical negation specialist to a frozen Qwen2.5-0.5B base model using a residual cross-attention bridge and a prompt-level relevance gate. The base model, specialist, and LM head remain frozen.
Specialist hidden states improve narrow-task F1 when routed through the bridge.
Relevance gating preserves nearly all in-domain performance vs the unattached specialist.
Off-task prompts no longer collapse into the specialist label space.
Base model, specialist model, and LM head remain fully frozen throughout.
The gate makes off-task preservation part of the architecture. Without it, a useful narrow specialist can still contaminate general answers.
What we're testing next
Multi-specialist composition
Can several focused specialists activate without interfering with one another?
Focused generative specialists
SQL, unit-test, arithmetic — do they improve generation without hijacking the base model?
Router reliability
Can relevance gates stay calibrated under ambiguous, out-of-domain, and adversarial prompts?
Sparse execution
Can the system avoid running every specialist on every prompt?
Baselines
How does representation-level composition compare against LoRA, adapters, RAG, tool routing, and output-level routing?
Papers
Technical Report · 2026 · Rohith V. Putha
A first controlled result showing that a frozen specialist can improve a frozen base model through a learned residual interface, while relevance gating suppresses off-task contamination.
Technical Essay · 2026 · Membrane Labs
Why frontier models became the default substrate for agents, and why all of that reliability cannot live inside one model.