finding
active
finding:activation-alignment-algorithm-succeeded-on-99-2-of-llama-qwen-and-99-1-of-gpt-oss-deepseek-sequencesActivation alignment algorithm succeeded on 99.2% of Llama/Qwen and 99.1% of GPT-OSS/Deepseek sequences.
Alignment failure analysis.
Source paper
extracted_from(2026) · Thomas Jiralerspong · Trenton Bricken
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Transfer of independently derived persona vectors shows the DFC learns a meaningfully aligned representation space.associated_withValidity of cross-architecture diffing.
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.
- Demonstrates that high IIA can be obtained even when model cannot solve the task
- Only model showing marginal benefit from increased reflection, at substantial token cost
- Key empirical result showing that optimizing for behavioral outputs and fitting representation geometry produce the same path in activation space.
- Second central claim of the paper.
- Replication in second Chinese model.
- Harder questions lower faithfulness.
- Greedy algorithm aligning activations across different tokenizers.
- Unexpected positive finding suggesting capping may sometimes help capabilities