Unified access to Large Language Model modules using NNsight
-
Updated
Jul 2, 2026 - Python
Unified access to Large Language Model modules using NNsight
Token-time interpretability instrument for language models — measures how generation trajectories move, branch, and respond to perturbation. Deterministic intervention comparisons via SeedCache branchpoints, hysteresis protocol, and SAE-feature steering.
Code for the paper "Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models." Activation patching, knockout ablation, and mechanistic analysis of how VLMs resolve perception-knowledge conflicts, across Qwen 2.5 VL, LLaVA-NeXT, and PaliGemma 2.
ADI — AWS Deep Inference: ephemeral remote access to model internals (nnsight) on right-sized EC2 GPUs, per session, in your own AWS account.
A small, extensible mechanistic-interpretability lab — logit lens & activation patching on GPT-2 and Qwen3 behind a unified backend adapter. Config-driven, tested, laptop-friendly.
Training-free cross-model capability transfer via linear activation alignment | Master Key Hypothesis (arXiv:2604.06377)
Add a description, image, and links to the nnsight topic page so that developers can more easily learn about it.
To associate your repository with the nnsight topic, visit your repo's landing page and select "manage topics."