claim
active
claim:token-filtering-works-well-when-behavior-is-localizable-to-short-spans-but-underperforms-for-broad-diffuse-traitsToken filtering works well when behavior is localizable to short spans but underperforms for broad diffuse traits
Poisoning results.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Fig. 3 on Llama-3.1-8B.
- Poisoning induces trait.
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.
- Fig. 8.
- Averaging acts as a low-pass filter that smooths away per-token detail and retains broad conceptsclaim0.772Intuition.
- Design choice validated empirically.
- Fig. 7.
- With raw data, wider token strides give more accurate approximations at same token budget.finding0.765Cost tradeoff.
- A better basis (e.g., factoring tokens into languages or parts of speech) may yield a sparse interpretable modelhypothesis0.763Same model can be dense in one basis and sparse in another.
- Prediction from coverage results.
- Choosing token indices to sample adaptively, as in optimal experiment design, could be more efficient than post-hoc smoothing.hypothesis0.759Future work.