benchmarks evals
The Evidence Was There. Averaging Hid It.
A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.
Summary
A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.
Language-model data attribution asks which training records caused a particular output, but ordinary evaluations often change several properties at once. CATT instead fine-tunes models on fictitious facts with known provenance, then varies redundancy and lexical distractors separately. Both conditions degraded attribution. The researchers found that causal evidence within a document clustered in a few tokens, while whole-document averaging diluted it; negative-log-odds gradients plus top-token aggregation improved ground-truth recall by 35 to 72 percentage points across tested baselines. On FTRACE-TREX, the gradient change improved counterfactual-retraining performance by as much as 4.9 times. Those gains diagnose the tested methods and datasets, not a guarantee that any production model's training history can be reconstructed.
Why it matters
A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.
Limits and context
- Those gains diagnose the tested methods and datasets, not a guarantee that any production model's training history can be reconstructed.
Key claims
A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.
Qualification: Those gains diagnose the tested methods and datasets, not a guarantee that any production model's training history can be reconstructed.
Evidence: source-2026-11-01-001
Sources
- arXiv preprint 2610.40655, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.