Cross-lingual residual-stream patching from English into Polish twins of translation-paired multiple-choice items on Qwen2.5-1.5B, with behavioural controls and two continued-pretraining arms

https://github.com/stw2/tokenizer-science-tax @ 6819d364ee1992d0734a3cea7f5e54de0cb8e74e

Reference checked 2026-09-14 13:18 UTC. No code was executed or scientific result verified.

• experiments/E12-activation-patching/DESIGN.md

Prediction

Patching the English twin's state into the Polish pass makes at least half of the forward-discordant pairs correct at the best confirmed layer pair with the validity gate passing; effective pairs target layers near 13; the entity span carries at least the scaffold-final effect; continued pretraining leaves the flip rate within 10 percentage points; the reverse direction reaches at most half the forward rate; swapping the scaffold language moves accuracy by less than 3 percentage points.

Protocol

01 derives from the committed per-item scores of Qwen2.5-1.5B, with structurally malformed pairs excluded, the pairs answered correctly in English and incorrectly in Polish, gated to 151 Belebele and 175 MMLU pairs, and the reverse set (33 and 49); it freezes with seed 20260703 a scan subsample of 100 items, a within-benchmark derangement for irrelevant sources, a breakage guard of 100 Polish-correct items and shared entity spans. P0 (02, no hooks) scores the three models with the scaffold language swapped and with the English twin's prompt prepended to the Polish prompt. 03 builds the patch: the English twin's residual-stream state entering a source decoder layer at an anchor replaces, without scaling, the Polish twin's state entering a target layer while its options are scored; its smoke gates are a bitwise self-identity check, a byte-identical 10-item repeat and a calibration of every benchmark and anchor. P1 (04) scans the subsample at the scaffold-final anchor over every second layer as source and target, plus the pairs 23 into 13 and 3 into 13. P2 confirms the top P1 pairs (at most 6, all within 2 percentage points of the best) on all forward-discordant pairs with an irrelevant-source control, the breakage guard, and the anchor ladder: scaffold-final token, question-final token, shared entity span (mean-pooled), a random position and the first token. P3 runs the reverse direction, the two continued-pretraining arms at the P2 pairs, and a STEM slice of P2. 05 computes item bootstrap intervals (2000 resamples, seed 20260703) and exact McNemar tests.

Dataset

Translation pairs of 600 Belebele test items (English and Polish) and 600 MMLU test items with Polish machine translations from openGPT-X mmlux, with committed per-item scores of Qwen2.5-1.5B and its two continued-pretraining arms; the patch sets exclude structurally malformed pairs.

Split

None: confirmation patches every forward-discordant pair; the coarse scan uses a seeded subsample of 100 (50 per benchmark).

Access needs

The two continued-pretraining checkpoints are restricted materials: ask the Room owner. The APT4 tokenizer reference used for the APT4 arm is gated on Hugging Face.

Configurations

Anchors: scaffold-final token (primary), question-final token, shared entity span, random position, first token. Layers: every second decoder layer as source and target, plus 23 into 13 and 3 into 13. Sources: the item's own English twin (specific) and another item's (irrelevant). Models: Qwen2.5-1.5B; continued-pretraining arms A and B at 500M tokens. Behavioural conditions: baseline, scaffold swap, translation assist.

Metric

Flip rate: share of forward-discordant pairs whose patched Polish prediction is correct; irrelevant-source flip rate; breakage on Polish-correct pairs; translation-assist recovery of the English-minus-Polish accuracy gap; accuracy change under a scaffold-language swap.

Seeds

20260703 (PCG64: scan subsample, derangement, breakage guard, bootstrap)

Interpretation rule

Validity gate first: specific flip rate at least 3 times the irrelevant-source flip rate, and breakage at most 5 percentage points on the Polish-correct guard; if it fails, patching claims are void and the experiment is reported as an instrument failure. Then the flip rate R at the best confirmed pair: R at least 0.50 is routing-dominant, R at most 0.15 is representation-bound, which also requires the entity-span anchor to fail, otherwise mixed, reported by benchmark. Translation-assist recovery of the gap on the base model, per benchmark: at least 70% comprehension-dominant, at most 30% deeper than comprehension, otherwise mixed. Scaffold swap: absolute accuracy change below 3 percentage points means the gap tracks question language. Arm A's flip rate within 10 percentage points of the base model's with an interval including zero means continued pretraining left routing unchanged. Reverse flip rate at most half the forward rate means an English-sided store. Fixed pairs 23 into 13 and 3 into 13, read only if routing is dominant or mixed: at least 50% of the best R. STEM against non-STEM flip rate is exploratory, without a bin.

Resources

Apple silicon, 128 GB unified memory; bf16 inference with forward pre-hooks on the MPS backend, about 5 to 7 hours; no training.

Prior work

arXiv:2607.08393 patches residual states between layers of the same failing prompt (self-patching), reports effective pairs from about 0.10L and 0.82L into 0.45L, and uses an irrelevant-patching control.