finding
active
finding:no-activation-tool-ao-nla-sae-improves-on-transcript-only-baseline-for-gemma-3-27b-it-targetNo activation tool (AO, NLA, SAE) improves on transcript-only baseline for Gemma-3-27B-IT target
Main negative result; baseline AUROC ~81 with tools within ~1 point.
Source paper
extracted_from(2026) · Adam Karvonen · Euan Ong · Subhash Kantamneni · Samuel Marks
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Unlike fine-tuned quirks, natural behaviors do not produce anomalous activation tool outputssupportsExplains null result versus auditing games.
- Mechanism for null result.
Hypotheses (2)
hypothesis
- Activations may be inherently incomplete for causal claims, requiring weight- or circuit-level accessassociated_withOne explanation for tool failure.
- If the relationship is present in activations, better tools (e.g. trained on these datasets) could surface itassociated_withAlternative explanation.
Findings (5)
finding
- Replication across predictor families.
- Null result persists with reasoning.
- Replication on second target model.
- Filters do not drive the null.
- Robustness to hyperparameters on Gemma.
Questions (1)
question
- Unclear per Section 3.1.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Table 1.
- Vulnerability profile for Gemma-3-27B showing SP dominance
- Fig. 2a,e.
- Qualitatively different defense profile compared to Llama-3.1-8B
- Interpretation of probe vs judge gap.
- Paper describing Gemma 2 model family used in this study
- Tool-induced answer flips mostly toward 'no effect' (NLA 82% positive, 84% negative), cancelling in accuracyfinding0.741Direction of tool-driven prediction changes.
- The paper positions NLAs as combining unsupervised learning with direct readability.