Building an Interpretability Playground with Activation Steering

ramp · Tech · 2026 · source Agentic: no · promotional bias: med · summary-only (no fetched source) · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: LLM interpretability and controllability

Ramp wanted an interpretability playground that could steer model behavior at inference time without retraining, while exposing the practical engineering problems of layer choice, multiplier calibration, model instability, and GPU startup latency.

Constraints:

• The fetched social source is short, so most technical detail comes from the ZenML summary.
• Activation steering had to modify behavior during the forward pass without fine-tuning.
• Qwen 2.5 7B was unstable and capacity-limited for the intended demonstrations.
• Serverless GPU cold starts needed to be short enough for interactive exploration.

Architecture

Ramp built Steer, an activation-steering playground. It computes steering vectors for target concepts, injects those vectors into selected model layers during inference, migrated from Qwen 2.5 7B to Gemma 3 27B-IT, tested 1,280 layer/multiplier configurations, and deployed on Modal serverless GPUs with snapshots to reduce cold starts.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  user[User] --> steer-playground[Steer Playground]
  steer-playground --> steering-vector-engine[Steering Vector Engine]
  steering-vector-engine --> gemma-model[Gemma Model]
  steer-playground --> modal-gpu-runtime[Modal GPU Runtime]
  modal-gpu-runtime --> gemma-model[Gemma Model]
  experiment-grid[Experiment Grid] --> steering-vector-engine[Steering Vector Engine]

Components

Data flow

A user selects a concept and steering intensity. Steer applies a steering vector to targeted activations during the model forward pass, runs the selected model on Modal GPU infrastructure, and returns generated behavior that reflects the steering experiment.

Why this architecture

Ramp used activation steering because it let the system demonstrate controllable model behavior without retraining. The move to Gemma 3 27B-IT and calibration over many configurations addressed the model stability and layer-selection problems that made the initial approach unreliable.

Models used: Qwen 2.5 7B · Gemma 3 27B-IT

Scale

• Ramp tested 1,280 steering configurations.
  “After initially struggling with Qwen 2.5 7B's instability and limited capacity, they migrated to Gemma 3 27B-IT and conducted extensive experiments across 1,280 configurations to determine optimal layer selection and multiplier calibration.”summary, stated
• Modal snapshots reduced cold starts from 60-120 seconds to 5-12 seconds.
  “The final implementation targets sparse global attention layers with carefully calibrated multipliers, deployed on Modal's serverless GPU infrastructure with snapshot capabilities to reduce cold starts from 60-120 seconds to 5-12 seconds.”summary, stated
• The final implementation used Gemma 3 27B-IT.
  “After initially struggling with Qwen 2.5 7B's instability and limited capacity, they migrated to Gemma 3 27B-IT and conducted extensive experiments across 1,280 configurations to determine optimal layer selection and multiplier calibration.”summary, stated

Alternatives considered (roads not taken)

Fine-tune or retrain the model for each behavior.

Rejected because: Steer modifies behavior at inference time without retraining.

Continue with Qwen 2.5 7B.

Rejected because: The summary reports instability and limited capacity with that model.

Use uncalibrated layer injection.

Rejected because: Layer selection and multiplier calibration required broad experimentation.

Patterns

latency-optimization (inference)

Modal snapshots reduce serverless GPU cold-start latency for interactive steering.

eval-driven-development (experimentation)

Ramp used a 1,280-configuration experiment grid to calibrate layers and multipliers.

model-routing (model selection)

The team changed model family after Qwen 2.5 7B proved unstable.

Tradeoffs

Control technique

Chose Activation steering over Fine-tuning or retraining.

Rationale: The playground needed interactive behavior changes at inference time.

Model choice

Chose Gemma 3 27B-IT over Qwen 2.5 7B.

Rationale: Gemma provided better stability and capacity for the demonstrations.

Hosting

Chose Modal serverless GPU snapshots over Unoptimized cold-start GPU serving.

Rationale: Interactive exploration required shorter startup latency.

Gotchas (what bit them)

Activation steering is model-architecture sensitive.

Root cause: Layer choice and sparse global attention behavior affect steering quality.

Resolution: Target sparse global attention layers with calibrated multipliers.

A smaller model may not be stable enough for interpretability demos.

Root cause: Qwen 2.5 7B had instability and limited capacity.

Resolution: Migrate to Gemma 3 27B-IT.

Evaluation

Ramp evaluated Steer by sweeping 1,280 configurations, comparing model stability, choosing layers and multipliers, and measuring cold-start reduction on Modal.

Metrics: Configuration count · Cold-start latency · Layer selection quality · Multiplier calibration · Model stability

Outcomes

• Steer demonstrates concept steering without retraining.
• Ramp migrated from Qwen 2.5 7B to Gemma 3 27B-IT.
• Modal snapshots reduced cold starts from 60-120 seconds to 5-12 seconds.

The summary reports 1,280 configurations, Gemma 3 27B-IT, and cold-start reduction from 60-120 seconds to 5-12 seconds.