finding
active
finding:in-llama-qwen-5-dfc-qwen-exclusives-90-5-not-dead-78-3-interpretable-llama-79-0-not-dead-63-1-interpretable-claude-flagged-251-qwen-and-351-llama-featuresIn Llama-Qwen 5% DFC, Qwen exclusives: 90.5% not dead, 78.3% interpretable; Llama: 79.0% not dead, 63.1% interpretable; Claude flagged 251 Qwen and 351 Llama features.
C.5.1.
Source paper
extracted_from(2026) · Thomas Jiralerspong · Trenton Bricken
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Crosscoder model diffing is best used as a high-recall pre-screening tool followed by verification.supportsRecommended workflow.
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.
- Figure 5.
- Appendix C.4.
- Validation step 1.
- Model-specific baseline personality difference revealed by revealed preferences experiment
- Moderation statistics.
- Model-specific difference in persona susceptibility
- Case study demonstrating mechanism behind flat harness-updating: smaller models reach same procedural content
- Transferring the Llama sycophantic vector to Qwen via DFC induced strikingly similar sycophantic behavior.finding0.779Figure 3/7.