Newest papers

Same Entity, Different Script: Diagnosing Cross-Script Entity Matching Failures in Language Models

Athira Gopal, Patil Prathamesh Pravin, Ankush Kadu, Arpit Jain, Ashwanth Krishnan · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.31407 · cited by 0 works (OpenAlex) · open-access PDF

Hindi

Multilingual translated language models can give different margin judgments when the same people's names are written in different scripts. We study this effect using 601 audited identity pairs with attested Devanagari and Latin forms, keeping the question and judgment criteria fixed. Qwen2.5-7B's same-person accuracy falls from 35.1% in Latin to 8.5% in Devanagari under Marathi prompts, while different-person accuracy remains above 98%. पांच मॉडलों के बीच मूल्यांकन अलग-अलग प्रतिक्रिया बाधाओं को प्रकट करता है, और कम टोकनेनिज़ेशन टुकड़ाकरण विश्वसनीय रूप से अंतर को बंद नहीं करता है. Qwen2.5-7B's same-person accuracy replicating the design on 374 audited Malayalam pairs built from the same entities reproduces the effect

Machine translation (Meta M2M100); the English original follows.

Abstract

Multilingual language models can give different identity judgments when the same people's names are written in different scripts. We study this effect using 601 audited identity pairs with attested Devanagari and Latin forms, holding the question and relation fixed. Qwen2.5-7B's same-person accuracy falls from 35.1% in Latin to 8.5% in Devanagari under Marathi prompts, while different-person accuracy remains above 98%. Evaluation across five models reveals distinct response biases, and reduced tokenization fragmentation does not reliably close the gap. Replicating the design on 374 audited Malayalam pairs built from the same entities reproduces the effect in a second script and language family: same-person accuracy falls to 5.3%, about half the matched Devanagari rate, while the margin still ranks matching above non-matching names at 0.925 AUROC. Latin-trained probes recover label information from Devanagari states, exceeding a string-feature baseline under similarity-matched negatives. Scalar calibration recovers 87% of the raw model-probe balanced-accuracy gap, identifying decision criterion placement as a major contributor. Activation patching localises causal sensitivity to the second entity in Qwen's middle layers; one shared sensitivity direction transfers about half the full-state margin correction. A fact-independent intervention at this site raises held-out Marathi same-person accuracy from 6.5% to 33.5%, with 98.5% different-person accuracy. Together, these results distinguish script-dependent discrimination from criterion placement and identify a causal site for improving identity judgments.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Devanagari, Computer science, Marathi, Artificial intelligence, Natural language processing, Telugu

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange