Table 1; Extended Data 5c; SI: editing
Matched fixed fact edits
On 30 matched GPT-2 edits, the geometric-mean Euclidean/Fisher other-token KL ratio is 10.57, with 95% interval 8.52–13.37. Both attain the fixed +5 log-odds target.
Loading recorded measurements…
Editing comparisons under distinct protocols
Loading recorded comparison…
Deciding when an edit should apply
Loading recorded comparison…
The operator used to apply an edit
Loading recorded comparison…
Methods and interpretation
The same prompt-selection rule routes both operators. Secondary GPT-2-XL and method comparisons each use their own evaluation protocol.
Procedure in the paper: Acquire the frozen CounterFact records, solve each operator once at the matched calibration target, apply the common routing rule and evaluate the declared other-token KL and generalisation metrics.
Matched fixed edits
Thirty routed GPT-2 edits, each matched to a +5 target log-odds change. The controlled comparison uses other-token KL. Secondary editing benchmarks use distinct protocols.
Code and data
v0.1.0 · 67 KB · View source on GitHub ↗
The primary matched comparison on 30 held-out GPT-2 fact edits.
v0.1.0 · 87 KB · View source on GitHub ↗
Routed and learned-scope editing, descriptive editor comparisons, 300-edit probability margins, and apply-operator/fluency controls.
Each standalone package includes code, shared helpers, required small inputs and reference results, with setup and commands in its README. Model weights and public datasets are obtained separately where needed.
Download complete source (v0.1.0) for all experiments and the companion website.