Reports
See how researchers use Silico to connect model internals, interventions, behavior, and evidence.
Highlighted reports
See what a finished Silico report can look like.
Restore cyber safety after preference tuning without losing helpfulness
Model safety · Cybersecurity
Preference tuning erodes cyber safety, and internal probes restore it
Detect a hidden safety regression, then use internal probes during training to recover protection while preserving benign compliance.
Slim down a robot-control policy without sacrificing task success
Causal model design · Robotics
Build leaner robot policies by tracing the layers that drive action
Find the layers a robot policy uses, simplify its internal interface, and test the leaner design on held-out tasks.
Help a protein model find better drug-binding structures
Model steering · Drug discovery
Find better drug-binding poses for drug discovery by reading a structure model's internals
Use hidden activations to surface promising protein–ligand structures and steer the model to generate more of them.
Research reports
Explore complete reports produced in Silico.
Speech models · Model internals
A speech model tracks the emotion of what it says, and its voice carries only part of itOrpheus-3B keeps a running read of how a sentence feels, while almost none of that emotion reaches the voice.
Clinical AI · Model interpretability
How do motion models score Parkinson's severity?Ask how general-purpose motion models produce clinical gait scores and which motion characteristics they attend to.
Vision-language models · Decoding
Where a vision-language model starts putting words to what it seesRead image patches out as vocabulary and find where Gemma 3 27B first grounds words in the right pixels.
Biology · Representation alignment
Two AI models, one for DNA and one for protein, build the same picture of protein biologyTranslate between DNA and protein model representations to reveal shared biology and test what can be causally moved.
Genomics · Model compression
Reading a genomics model to find its dead weightFind low-value components without seeing the test set, then build a model that is 22% smaller without losing performance.
Language models · Multimodal concepts
Language models already know that kiki is spikyTrace sound-shape associations to the token lookup table, before the language model does any thinking at all.
Language models · Steering
Making Kimi K3 write like a pirate with a difference-in-means vectorBuild a writing-style direction from 200 texts and identify the coefficient that makes steering work.
Vision models · Concept geometry
One concept space for two vision models: they agree on positions inside itTest whether DinoV2 and SigLIP place matching visual concepts at the same coordinates despite different training objectives.
Multimodal models · Model design
Giving Qwen3-8B visionBolt a frozen image encoder onto a text model and use the right controls to learn which benchmark gains survive.
Biosecurity · Protein models
A frozen protein model reads hazard signal that alignment missesUse a linear read-out to separate hazardous bacterial proteins from harmless ones beyond alignment or composition baselines.