hypothesis
active
hypothesis:larger-models-learn-rare-tasks-via-reduced-interference-and-better-retentionLarger models learn rare tasks via reduced interference and better retention
Hypothesis box in §3.2.
Source paper
extracted_from(2026) · Jing Huang · Daniel Wurgaft · Rachit Bansal · Laura Ruis +5
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (8)
finding
- 1B model non-task gradient nearly orthogonal to task reference (7.58e-5±0.02)associated_withsupportsFig. 9.
- In matched-frequency injection, rare-task signal decays with gap G at all widths, much more steeply for narrow models, while wide models retain and accumulate signalassociated_withsupportsFig. 4.
- Fig. 6b.
- Fig. 7.
- Fig. 9.
- Fig. 9.
- Fig. 22.
- Fig. 4 caption.
Claims (1)
claim
- Example 7.
Questions (1)
question
- Motivating question.
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.
- Core framing.
- Fig. 23.
- Selective pressure toward convergence via task generality
- Larger OLMo models achieve both low training loss and high test accuracy, i.e. learn generalizable structurefinding0.823Fig. 5.
- Inverse scaling interpretation.
- Data-induced competition over resources.
- Motivating question.
- Practical implication.