Hemant Palivela

PACT: Evaluating Tool Calling in Quantized Small Language Models on Parallel English, Arabic and Synthetic Code-Switched Requests

Small language models are increasingly deployed as tool-calling agents in quantized form, yet their reliability has been evaluated almost only in English, although many users in the Arab world write requests in Arabic or mix Arabic and English. The paper (i) audits a public Arabic tool-calling dataset and shows that it...