finding
active
finding:1b-model-batch-gradient-similarity-to-task-reference-is-0-08-0-02-at-injection-steps-300m-is-0-04-0-041B model batch-gradient similarity to task reference is 0.08±0.02 at injection steps; 300M is 0.04±0.04
Fig. 9.
Source paper
extracted_from(2026) · Jing Huang · Daniel Wurgaft · Rachit Bansal · Laura Ruis +5
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Hypothesis box in §3.2.
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. 9.
- Fig. 9.
- Fig. 23.
- Task accuracy on CausalGym increases consistently with model scale from 0.62 (14M) to 0.89 (6.9B)finding0.764Scaling result showing larger pythia models perform better on CausalGym linguistic tasks
- Establishes generalizability of the core difficulty-boundary finding across model families.
- Demonstrates the failure mode of surface-level similarity for instruction discovery.
- Strength comparison accuracy reaches 73% at layer 3 for injection pair (2,6) vs. 50% chancefinding0.754Secondary positive result for strength comparison showing graded sensitivity to perturbation magnitude
- When task gradient norms differ greatly, large-norm tasks have not converged while small-norm tasks have nearly convergedhypothesis0.752Motivates setting αk = max norm to enable further learning on under-converged tasks