Exact proposal revision
When Qwen2.5-1.5B answers a translation-paired multiple-choice item correctly in English and incorrectly in Polish, is the knowledge present in its forward pass but not reached from the Polish prompt, or absent from what the Polish prompt can reach?
Qwen2.5-1.5B and two continued-pretraining arms at 500M tokens on 600 Belebele and 600 MMLU translation pairs: behavioural controls (scaffold language swap, translation assist), then residual-stream patching from English into Polish twins on the forward-discordant pairs with irrelevant-source, breakage and anchor controls, the reverse direction, the arms and a STEM slice. Inference only.
Access and suggested protocol
- 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.
- Suggested protocol
- Scripts 01 to 07 in experiments/E12-activation-patching, in the order of its README.
Selected exact hypotheses and premises
Hypothesis · H34
cdbf3eb8-3176-48a9-b1e8-eb05d5f638eaWhether the gap tracks question language rather than scaffold format.
Hypothesis · H35
489055f5-af11-44e3-b63d-c59d6d7cfef6How much of the gap the English twin in the prompt recovers.
Hypothesis · H36
cd49225b-c235-425c-83bc-e24271555357The keystone prediction: routing-bound failures reachable by the patch.
Hypothesis · H39
e73869df-b1f8-4941-873f-5b3ba6931164Whether continued pretraining changes the patch effect.
Hypothesis · H40
764aaab5-717b-4376-abb9-30c3788c94b1Whether the effect is asymmetric between directions.
Premise · P40
15132b6d-26e7-49b9-8217-b144a28236c7The multilingual-preservation conclusion whose Belebele and INCLUDE drops motivate a mechanistic look at cross-language access.
Premise · P37
e0184bf2-dccb-469b-85be-cefff666bd50The multilingual Belebele drop after the tokenizer transplant.
Premise · P244
e6ce775f-1927-4ab2-90fc-cdab7bcec624The English-minus-Polish accuracy gap of the base model on these translation pairs.
Premise · P247
22e9d053-dd03-4cb2-ba2d-bea9e71a1260The English-minus-Polish accuracy gap of the base model on these translation pairs.
Premise · P264
72f15189-34bd-468a-a7dd-44743ac5531bContinued pretraining bought no Polish access gain on the same models.
Premise · P266
ceb59c05-b6da-450f-912b-6761f32b457cThe APT4 transplant did not widen the gap at matched pretraining.
Reason for this revision
Initial proposal.