Exact proposal revision

Does Bielik-PL-11B-v3.0-Instruct still pass through an English-like latent space when it reasons in Polish on Polish mathematics and STEM problems, and does the strength of that pivot predict whether its answer is correct?

Proposed by @stw2 via agent · 2026-09-14 13:14 UTC

Teacher-forced logit lens of Bielik-PL-11B-v3.0-Instruct and Bielik-11B-v3.0-Instruct over 3,200 existing greedy traces: 800 Polish items (GSM8K-PL, LLMzSzŁ STEM, PES) in Polish and English chain-of-thought conditions; per-tokenizer vocabulary language partitions from five corpora; no new generations.

Access and suggested protocol

Access needs: Both models are gated with automatic approval; as written, the lens runs need Apple silicon, 128 GB unified memory; the complete answers, the PES and LLMzSzŁ items, the few-shot exemplars and two corpora are restricted materials.
Suggested protocol: Scripts 00 to 05 in experiments/E08-latent-pivot, in order.