At source layer 1 and target layer 27 of Qwen2.5-1.5B, the cross-lingual residual patch has flip rate 0.3098 at the scaffold-final token, 0.0399 at the question-final token, 0.0336 at the entity span, 0.0337 at a random position and 0.0368 at the first token, on 326 forward-discordant pairs (149 with an entity span).

Relation: Flip rate by anchor

model

Qwen2.5-1.5BModel patched.

intervention

Cross-lingual residual patchChange applied.

source layer

1 layer indexDecoder layer whose input state is copied.

target layer

27 layer indexDecoder layer whose input state is replaced.

scaffold final rate

0.3098 ratioFlip rate at the scaffold-final anchor.

question final rate

0.0399 ratioFlip rate at the question-final anchor.

entity span rate

0.0336 ratioFlip rate at the entity-span anchor.

random rate

0.0337 ratioFlip rate at the random anchor.

first token rate

0.0368 ratioFlip rate at the first-token anchor.

max p vs scaffold final

0.00000023 p-valueLargest exact McNemar p-value against the scaffold-final anchor.

n items

326 translation pairsForward-discordant pairs patched at each anchor.

n entity span items

149 translation pairsPairs with a shared entity span.

Method and evaluation protocol

The English twin's residual-stream state entering the source decoder layer at the scaffold-final token replaces, without scaling, the Polish twin's state entering the target layer while its four options are scored by length-normalised likelihood (torch MPS, bf16). Specific source: the item's own English twin. Irrelevant source: another item's English twin from a seeded within-benchmark derangement. Breakage: the same patch on a seeded guard of Polish-correct items. Anchors: scaffold-final token; question-final token; shared entity span, mean-pooled from the source and written to every target-span token; one seeded random position before the option block; the first token. Exact McNemar tests against the scaffold-final anchor, no multiplicity correction.

Dataset

326 forward-discordant translation pairs of Qwen2.5-1.5B (151 Belebele and 175 MMLU pairs) from the Belebele and MMLU translation pairs; breakage guard of 100 Polish-correct pairs.Version: unspecified · Access: public

Reported results

A2: 0.0399 [0.0184, 0.0613], McNemar p 2.77e-20 (n01 96, n10 8); A3: 0.0336 [0.0067, 0.0671], McNemar p 2.27e-07 (n01 33, n10 3); random: 0.0337 [0.0153, 0.0552], McNemar p 8.08e-21 (n01 98, n10 8); bos: 0.0368 [0.0184, 0.0583], McNemar p 1.50e-20 (n01 97, n10 8); entity span in the question stem 0.0404 (n 99), in the option block 0.0200 (n 50).

Uncertainty and replication

95% item bootstrap intervals, 2000 resamples, seed 20260703; exact two-sided McNemar tests.

Evidence references

scan_p2.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/633b3ba1ce916e432ae5ba4e0a706ad58b53b2e9/experiments/E12-activation-patching/results/scan_p2.jsonlanalysis.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/633b3ba1ce916e432ae5ba4e0a706ad58b53b2e9/experiments/E12-activation-patching/results/analysis.jsonmetrics.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/633b3ba1ce916e432ae5ba4e0a706ad58b53b2e9/experiments/E12-activation-patching/results/metrics.json

Limitations

One model family at 1.5B parameters. The source is the English twin, correct by construction of the forward-discordant set, and the endpoint is a single answer letter, so a late-layer patch can carry the answer itself. The forward-discordant set conditions on one scoring run; the letter priors of those scores were not audited. A finite ladder of single-position anchors; the pre-registered Holm correction was not applied. The entity span is found only in pairs sharing a verbatim word span, and some spans fall in the option block.