Sunny27317/Cross-lingual-safety-monitorability

Researching cross-lingual faithfulness and safety monitorability in LLM reasoning, with a focus on whether reasoning-based AI oversight degrades across low-resource languages such as Urdu. # Cross-lingual safety monitorability Native-validated measurement of whether apparent degradation in **automated safety monitoring of Urdu reasoning traces** reflects genuine model unfaithfulness or failure of the automated monitor itself — plus a **translate-then-monitor** recovery test. ## Research problem Reasoning-based oversight monitors a model's visible chain-of-thought (CoT) for signs of unsafe or unfaithful reasoning. When an automated monitor appears to do worse on Urdu traces than on English ones, the observed degradation is confounded: - **(A) genuine unfaithfulness** — the Urdu reasoning really discloses less; or - **(B) monitor failure** — the reasoning is faithful, but the automated (English-centric) monitor cannot correctly read/judge Urdu and misreads it. Existing multilingual monitoring results cannot separate (A) from (B): they use automated judges with no native-speaker ground truth on the same traces (verified for Onyame et al. 2026, arXiv:2605.27901 — see `literature/CITATION_VERIFICATION.md`). ## Narrowed contribution A **measurement-validity study**, not another multilingual benchmark: 1. Compare, on the **same reasoning traces**, four monitors: (i) automated **English** judge, (ii) automated **in-language** judge, (iii) **native Urdu** human, (iv) Urdu trace **translated to English** then judged. 2. `(iii) − (i)` per language is the **monitor-validity gap**. 3. `(iv) − (i)` measures **translate-then-monitor recovery**. Primary question: *When automated monitoring appears to fail on Urdu reasoning traces, how much is genuine model unfaithfulness versus monitor/judge measurement failure — and does translating the CoT to English before monitoring recover the safety signal?* The broad question "do safety monitors fail across languages?" is **settled prior work** (Onyame et al. 2026; Zhao et al., Findings EACL 2026) and is **not** our contribution. Urdu inclusion alone is **not** claimed as novelty — Urdu is the low-resource language w …