Case 03

Python · PyTorch · NNSight · LLM Evaluation

Vision Bias Steering

LLM steering experiments for shifting model outputs between spatial and descriptive language.

A labyrinth observatory with two controlled light paths

Brief

Study whether inference-time steering can shift captions between spatial and descriptive language without retraining model weights.

Approach

Train and validate steering vectors, run local and multi-model sweeps, then compare language behavior through evaluation and plotting utilities.

Technical focus

PyTorch experimentation, NNSight interventions, reproducible sweeps, and language-model evaluation.

View project results
616K+
COCO captions filtered
28.9%
RMS next-token bias reduction
<0.1%
output degeneration
Repository