Skip to content

Capabilities

Silico uses researcher-written capabilities to bring specialized methods into an investigation. Each capability gives the agent focused guidance for one type of work. Silico uses progressive disclosure: it first exposes a short catalog, then loads the full instructions and supporting material only when a task needs that capability.

The catalog below describes the current capability set. One capability can use several related methods during execution.

Interpretability and model analysis

  • Activation harvesting. Collect model activations at scale for feature training, probing, and automated interpretation. This capability covers harvest design, token budgets, data mix, provenance, and backend selection for different model sizes and architectures.
  • Featurizer training. Train and evaluate sparse autoencoders, block-sparse featurizers, and other sparse dictionaries over activations. This capability covers hyperparameter selection, sweep evaluation, capacity tradeoffs, and interpretation.
  • Probing. Train supervised probes on activations to test whether a concept is decodable at a given layer. This capability covers architecture choice, hyperparameter selection, and probe-quality evaluation.
  • Activation monitoring. Build an input or output monitor from structured agent rollouts. This capability covers exchange-level labels and activation capture, localized span probes, calibration, multi-layer fusion, adversarial and distribution-shift tests, and cost-constrained routing to an LLM judge.
  • Automated interpretation. Use agentic workflows to label SAE features or neurons from their activating examples. This capability also covers label-quality measurement and iterative improvement.
  • Steering. Build and validate interventions from SAE features, contrastive directions, learned probe directions, or learned geometric structure. Use them to study causal effects on model behavior.
  • Causal analysis. Apply causal abstraction and mediation analysis to language-model internals. Methods include interchange interventions, distributed alignment search, ablation, and counterfactual dataset design.
  • Parameter decomposition. Decompose model weights into rank-one components with learned causal-importance functions.
  • J-lens. Use Jacobian-based readouts to inspect model computation.
  • Dimension reduction. Reduce and visualize high-dimensional representations with methods such as PCA, SVD, UMAP, ICA, CCA, manifold analysis, and intrinsic-dimension estimation. This capability helps select a suitable method.
  • Model-difference amplification. Surface rare behavioral differences between checkpoints or fine-tuned models. It supports black-box prompt search and white-box amplification methods that expose uncommon post-training behavior on ordinary prompts.
  • Post-training debugging. Explain what fine-tuning changed, or is likely to change, inside a model using named and interpretable features.
  • Architecture diagrams. Trace a live forward pass through a PyTorch model and create an interactive architecture page.

Interactive dashboards and live monitoring

  • Feature explorers. Explore features or neurons in an interactive viewer.
  • Guardrail monitor. Serve a model with per-token activation streaming and monitor a trained probe's scores on live traffic.
  • Interactive data viewer. Browse and search tabular datasets or inspect image galleries within the workspace.

Data

  • Dataset inspection. Inspect a dataset quickly before deeper analysis or training.
  • Data attribution. Attribute an output, behavior, or internal mechanism to the training examples most responsible for it. Methods include influence functions, TracIn, and concept-level influence.
  • Synthetic data generation. Generate labeled synthetic datasets for training and evaluation from a declarative specification.

Training and fine-tuning

  • Training pipeline. Plan and run pretraining, continued pretraining, supervised fine-tuning, LoRA, direct preference optimization, and reinforcement learning on a suitable backend.
  • Interpretability in training. Shape what a model learns with methods such as concept-score reward shaping and inoculation prompting during preference training.

Serving and evaluation

  • Serving. Start high-throughput inference endpoints with a topology suited to the model and hardware. Endpoints can support activation harvesting, steering, and other interpretability methods.
  • Evaluations. Select useful metrics and build well-validated LLM judges.

Life sciences

  • Biological foundation models. Apply domain-correct interpretability to DNA, RNA, protein, and structure models. This capability maps model units to biological meaning while controlling for confounds.
  • Bioinformatics. Handle biological data correctly, including sequence and variant formats, coordinate systems, reference builds, annotation sources, and expression data.
  • Biological database selection. Choose authoritative molecular, cellular, clinical, and population data sources for a research question.
  • Biological model comparison. Test whether two biological foundation models share internal representations and locate differences across modalities and scales.
  • Causal behaviors for biological models. Define observable behaviors against which internal interventions on biological models can be tested.

Research planning and design

  • Research design. Plan an interpretability study with a clear question, suitable methods, controls, and evidence.
  • Access verification. Check access to every required dependency before execution begins.
  • Paper replication. Reproduce a published result with a traceable and thorough workflow.

Reporting and synthesis

  • Interactive HTML. Turn a thread into an interactive HTML report with claim-level provenance.
  • LaTeX. Produce a LaTeX paper with claim-level provenance.
  • Results publication. Finish each worker run with a published results summary.
  • Figures. Create clear figures with consistent visual conventions.

Cluster work

  • Job orchestration. Prepare, launch, and monitor jobs across supported GPU backends.
  • GPU job checklist. Run focused smoke tests before spending substantial GPU time.
  • Environment builds. Create reusable compute environments with Goodfire's core libraries.