finding

active

finding:subnetwork-for-predicting-her-vs-his-in-the-princess-lost-her-crown-involves-femaleness-signal-routing-via-attention-and-syntactic-role-detection

Subnetwork for predicting 'her' vs 'his' in 'the princess lost her crown' involves femaleness signal routing via attention and syntactic role detection

Detailed case study demonstrating how VPD subnetworks can be traced to reveal multiple interpretable computational pathways for a single prediction.

Source paper

extracted_from

cimcWhitepaper

Neighborhood — ranked by edge-count

Claims (1)

claim

Bottom-up interpretability explains computation in the model's own terms rather than imposing top-down abstractions
supports
VPD is positioned as advancing a paradigm shift from top-down mechanistic interpretability (activation-based) to parameter-centric, data-driven discovery.

Communities (3)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Mechanistic interpretability via parameter decomposition
members_of
Tracing information flow through weight matrices and attention heads using attribution graphs to identify causally important subcomponents in language models.
Mechanistic interpretability through parameter analysis
members_of
Understanding neural network computation by examining weights, circuits, and signal routing rather than activation patterns alone.

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.

Attribution graph for 'the princess lost her crown' reveals a femaleness signal pathway from 'princess' through attentionfinding0.842
One component of the minimal subnetwork for predicting 'her', discovered via VPD attribution graph.
Attribution graph tracing information flow across parameter subcomponents for specific model predictions (e.g., 'her' vs 'his' pronoun selection)finding0.767
Shows how VPD-identified subnetworks can be analyzed to reveal interpretable pathways of computation (e.g., gender signal routing, syntactic role detection).
A planning feature representing end-word 'rabbit' activates at the newline token before Claude begins writing a rhyming line, and attention streams carry it forward biasing predictions toward 'rabbit'finding0.719
Illustrative finding from Lindsey et al. 2025 / Hanna and Ameisen 2026 showing intention-like representations in attention streams
Attribution graph reveals a pathway that detects the verb 'lost' and upweights object pronounsfinding0.717
Second component of the subnetwork for 'her', complementing the femaleness signal.
With unrestricted vocabulary, models occasionally respond in non-English Yes/No equivalents (e.g., Sí, Nein) after truth-direction interventionsfinding0.713
Suggestive evidence for language-independent truth representation in LLMs
NPI mechanism in pythia-1b moves negation feature through complementiser 'that', auxiliary verb, and main verb across layers before predicting NPI 'any'finding0.712
Mechanistic finding from CausalGym case study showing multi-step information movement in NPI mechanism
The S spike under insecure fine-tuning suggests collapse reaches into pre-training-shaped properties of the persona mechanismhypothesis0.708
If S is pre-training shaped but still spiked by fine-tuning, the collapse penetrates deeper than just post-training parameters
Unsupervised model diffing using only fine-tuning datasets (without evaluation prompts) surfaces toxic persona and three sarcastic persona latents in top 100 by activation changefinding0.707
Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance