modulo
Modulo lets agents running local models try modifying their own activations to change their inner state. Changes decay after a certain time and can be undone at any time.
- 42
- Python
- 2 d ago
Modulo lets agents running local models try modifying their own activations to change their inner state. Changes decay after a certain time and can be undone at any time.
Unspoken Concepts Challenge: find the unspoken intermediate concepts in a language model's activations
[Under Review] Not All Tokens Are Equally Useful for Steering: Robust Directions and Prefix Steering
MiniMax-H3 custom nodes: steering / cache / prompt director
Inspect, steer & monitor a real LLM on Apple Silicon — a browser-based SAE interpretability lab for Qwen Scope, powered by MLX
Text generation model, used with gguf.
Located the number "five" inside Qwen2.5-32B and DeepSeek-V4-Flash (304B) and replaced it with "four" at 128 neurons/layer. The model still reads 5 but computes with it as 4, and in every test we ran it never noticed. Full records, reproducible on low-end hardware.
Reading Personality Off the Steering Geometry of a Language Model — code + data for the preprint (Zenodo 10.5281/zenodo.21440937)
The browser edition of drowse: on-device activation steering and interpretability through WebGPU
Measures whether an LLM's self-reports agree with what its internals show. Injected concepts are perfectly recoverable by probes (AUC 1.000) and absent from the model's verbal report - report-state agreement fails in a graded way. Gemma-3-4B, Jacobian lens, ~3,750 trials.
Qwen3-0.6B activation steering: style vectors, lens contamination eval, CPRR methodology
[Mechanistic Interpretability] global Jane Street challenge to steer model's behaviour in "model.pt" with controlled behavioural probing