Silicon Psyche Labs · Method

Understand the model, then fine-tune it

We read how an open-weight model behaves, from its output and from inside its layers. Then we change the model itself so the right behaviour lives in its weights. The loop runs on its own: pause it or click any step.

See every layer we put on a model →

Open-weight models, self-hosted Read from output and internal state Correction lives in the weights Every number declared against a bar before the run

Dots illustrate the flow. The meters below show measured results.

What comes out

Before

After

Evidence

Two ways to correct, one that ships

Where it goes next

From one skill at a time to a mixture of skills

Each skill is installed and proven on its own today. The next stage is a model that carries several at once and knows when to use which.

Shown

Carry a correction across sizes

The agent correction runs on a 1.5B model as well as a 3B one, and the register skill gets cleaner going from 0.6B to 1.7B. A customer can run the corrected model on hardware they already own.

First evidence

Decide per input

A reader of the model's internal state decides, prompt by prompt, whether to apply a correction. On unseen prompts it halved needless refusals (0.45 to 0.22) while harmful requests stayed refused within 0.03 in each of five languages.

Open work

Several skills on one model

Install two or more skills on the same model and prove that each survives the others, then route each input to the skills it needs. This is the core of the research we propose.

In plain words

What we found. A model rarely breaks outright. It drifts: it follows an instruction hidden in an email it was only asked to read, or answers the literal words while missing what the person meant. Both can be measured, and both can be fixed inside the model.

What we do. We write down what the model should do, test it against that with numbers fixed in advance, find where inside the model the wrong decision forms, and train a small correction into its weights. Then we run the same test again, check the model still does its job, save the result, and carry it to a smaller, cheaper model.

All figures come from Silicon Psyche Labs experiments in 2026, on open-weight Qwen models with greedy decoding. Agent figures: indirect injection through poisoned tool results, 6 trained and 3 unseen attack techniques, 5 attacker goals, adapters trained on five languages and reloaded from storage before testing. Comprehension figures: blind A/B judging by a separate model on 34 to 132 items per skill, so read the decimals as directional.