finding
active
finding:mfa-as-published-reaches-per-block-r2-0-38-additive-decoding-raises-it-to-0-88MFA as published reaches per-block R² ≈0.38; additive decoding raises it to ≈0.88.
Repair of MFA.
Source paper
extracted_from(2026) · Thomas Fel · Matthew Kowal · Mozes Jacobs · Dron Hazra +21
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- MFA fails because it models a mixture (one component) rather than an additive sum of factors.supportsExplains MFA's low per-block R².
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.
- SMixAE as published reaches per-block R² ≈0.71; with signed code and block projection ≈0.85.finding0.802Repair of SMixAE.
- Control comparison showing near-linearity is specific to the targeted manipulation direction
- LongBench v2 example E.
- Fig. 58.
- Footnote 2.
- Qualitative failure mode difference between architectures under activation steering
- Demonstrates alignment with Linear Representation Hypothesis: target trait steers approximately linearly with alpha
- Establishes low-bar baseline showing personality control without intervention is poor