CORPUS EVIDENCE AND FACTUAL RECALL CAUSAL TRACING WITH AUDITED LABELS
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Factual recall and corpus evidence are distinct observations: a model may answer a question even when the supporting subject–object pair is absent from a reference corpus. We compare activation-patching profiles under these operational labels on Pythia-2.8B and Pythia-6.9B. Across 2,216 traced fact–model pairs, recovery concentrates early in both networks. Among recallable facts, the groups have equal median peak depths, with no detected rank-distribution difference (p = 0.568/0.390). These reference-corpus labels do not establish training membership. We then compare independently restored single-layer rank-one interventions at relation-selected, fixed-middle, and control layers under equal relative update norms. On 100 held-out facts per model, selected and fixed-middle layers both achieve near-ceiling rewrite success. Their side effects differ: selected layers preserve other factual answers more often, by 3.3 and 7.0 percentage points. A post-hoc strength analysis reveals a target-success advantage at one-quarter of the calibrated norm: selected layers outperform the fixed-middle layer by 18 and 20 percentage points. Layer choice therefore interacts with intervention strength, distinguishing recovery location from a uniquely effective storage address. Repeating all dose cells with two additional full-pipeline split seeds preserves the quarter-strength advantage in both models, while showing that half-strength effects are small and model-specific. Transferring the selected layers to the non-deduplicated checkpoint twins yields quarter-strength advantages of 16 and 24 percentage points. A predeclared projected-gradient intervention saturates across sites and does not reproduce the selected–middle top-1 gap, bounding the layer-selection result to the rank-one update mechanism.
- Primary area:
- interpretability and explainable AI