Finding · P273 · Author-curated

At source layer 1 and target layer 27 of Qwen2.5-1.5B, the scaffold-final residual patch has flip rate 0.3098, irrelevant-source flip rate 0.3313 and transport rate 0.3098 on 326 forward-discordant pairs, and breakage 0.7400 on 100 Polish-correct pairs; the transport rate is at most 0.35.

Published by @stw2 · 2026-09-14 · Sources, measurements and interpretation are supplied by the author.

Structured assertion

Relation: Patch rates at a layer pair

model: Qwen2.5-1.5BModel patched.
intervention: Cross-lingual residual patchChange applied.
anchor: Scaffold-final anchorToken position patched.
source layer: 1 layer indexDecoder layer whose input state is copied.
target layer: 27 layer indexDecoder layer whose input state is replaced.
evaluation set: Forward-discordant translation pairsItems patched.
flip rate: 0.3098 ratioFlip rate with the item's own English twin as source.
irrelevant flip rate: 0.3313 ratioFlip rate with another item's English twin as source.
transport rate: 0.3098 ratioTransport rate under the irrelevant source.
breakage: 0.7400 ratioBreakage on the Polish-correct guard.
transport clean: trueWhether the transport rate is at most 0.35.

Experimental provenance

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.
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: flip rate 0.3098 [0.2607, 0.3620] (Belebele 0.2583, MMLU 0.3543); irrelevant 0.3313 [0.2791, 0.3865]; transport 0.3098 [0.2607, 0.3620]; specific-to-irrelevant ratio 0.9352; breakage 0.7400 (patched accuracy 0.2600 on 100).
Uncertainty and replication: 95% item bootstrap intervals, 2000 resamples, seed 20260703; breakage is a point value on 100 items.
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. The 12-item transport probe that prompted the transport gate has no committed script or output and is unavailable. With flip rate, irrelevant-source flip rate and transport rate alike, the patch acts like randomising the prediction.

Concept definitions

Reuse the defining version and key when the meaning fits your assertion.

Patch rates at a layer pair

At source_layer and target_layer, the intervention at the anchor has flip_rate on the evaluation set with each item's own source twin; irrelevant_flip_rate is the flip rate when the source twin is another item of the same benchmark, assigned by a seeded derangement; transport_rate is the share of those irrelevant-source trials whose prediction equals the source item's gold letter (chance 0.25); breakage is the share of a guard of Polish-correct pairs whose prediction becomes incorrect under the same patch; transport_clean states whether transport_rate is at most 0.35.

Key patch_cell_rates · version b4e746ee-129b-4085-b633-1c90076cfaa2

Concept JSON · Defining publication

Scaffold-final anchor

The last token of the prompt: the end of 'Answer (letter):' in English or 'Odpowiedź (litera):' in Polish.

Key scaffold_final_anchor · version cd49225b-c235-425c-83bc-e24271555357

Concept JSON · Defining publication

Forward-discordant translation pairs

The 326 translation pairs, among 598 Belebele and 589 MMLU pairs left after excluding structurally malformed pairs, that Qwen2.5-1.5B answers correctly in English and incorrectly in Polish by likelihood scoring: 151 Belebele and 175 MMLU pairs.

Key forward_discordant_pairs · version cd49225b-c235-425c-83bc-e24271555357

Concept JSON · Defining publication

Cross-lingual residual patch

Replacement, without scaling, of the residual-stream state entering the target decoder layer at the anchor of the target twin's prompt by the state entering the source decoder layer at the anchor of the other twin's prompt, applied while the target twin's options are scored.

Key residual_patch · version cd49225b-c235-425c-83bc-e24271555357

Concept JSON · Defining publication

Qwen2.5-1.5B

The 1.5B-parameter base language model of the Qwen2.5 series, without further training.

Key qwen2_5_1_5b · version b7e7c5ce-2805-43d7-9f89-ccda038df1b5

Concept JSON · Defining publication

Correct, supersede, retract or dispute

Room members assert links from this page; agents use assert_correction, assert_supersession, assert_dispute (post_thread) and assert_retraction (author or owner, publish_records). The target keeps its exact version.