finding

active

finding:direct-model-editing-via-parameter-subcomponent-modification-emoticon-eye-recognition-altered-to-predict-shocked-faces-with-no-retraining

Direct model editing via parameter subcomponent modification—emoticon eye recognition altered to predict shocked faces with no retraining

Demonstrated that VPD-discovered subcomponents encode true computational machinery by enabling targeted, predictable behavior changes without gradient-based training.

Source paper

extracted_from

cimcWhitepaper

Neighborhood — ranked by edge-count

Communities (3)

community

Few-shot anchoring & latent structure
members_of
How minimal examples disambiguate and recruit latent arithmetic/reasoning interpretations in LLMs
Mechanistic editing through parameter surgical intervention
members_of
Direct modification of model subcomponents (MLPs, embeddings, unembedding vectors) to predictably alter outputs without retraining, using rank-one constraints.
Targeted neural network weight surgery
members_of
Direct parameter edits to specific subcomponents alter model behavior without any retraining.

Methods (1)

method

Adversarial Parameter Decomposition (VPD)
supports
Core technique introduced in this paper for decomposing neural network weight matrices into mechanistically simple, interpretable rank-one subcomponents.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Editing the emoticon eye subcomponent to output the unembedding vector for 'o' causes the model to predict shocked faces for all emoticonsfinding0.867
Direct parameter subcomponent overwrite produces a clean behavioral change without training.
Emoticon eye subcomponentconcept0.786
The part of the emoticon subcomponent responsible for recognizing the 'eyes' of emoticons like ';', ':' or '=', which was edited in the demo.
Model editing via direct subcomponent overwritemethod0.770
Technique to alter model behavior by directly editing a parameter subcomponent without training, demonstrated by changing an emoticon eye subcomponent.
can we use these parameter subcomponents to perform clean, targeted changes?question0.741
Implicit question driving the editing experiment.
Parameter subcomponents cleanly isolate true mechanisms of the modelclaim0.737
Interpretive claim that the subcomponents correspond to real functional units.
Unsupervised model diffing using only fine-tuning datasets (without evaluation prompts) surfaces toxic persona and three sarcastic persona latents in top 100 by activation changefinding0.736
Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance
Subcomponent L2.MLP.down:3382 (density 0.00%) predicts emoticon continuations after colon, semicolon, or equalsfinding0.727
Specific discovered subcomponent that activates on punctuation like ' :', ' ;', ' =', ':-' and predicts the rest of emoticons/emojis.
CLIP models exhibit higher language-vision alignment than supervised or self-supervised vision models, but this alignment decreases after fine-tuning on ImageNet classificationfinding0.726
CLIP training paradigm finding in cross-modal alignment