LIT @ UMich

Multilingual Deception Detection of GPT-generated Hotel Reviews

🎯 Accepted to COLM 2026 — One flipped example is all it takes: how one-shot GRPO breaks LLM alignment and induces systematic, generalizing bias. Code, data & models for 'It Takes One to Bias Them All.'

Wait, am I Being Fair? Characterizing deductive stereotyping in LLM reasoning and mitigating it with Fair-GCG (reasoning-time fairness steering).

The Language–Energy Divide: measuring energy costs of multilingual LLM inference across 122 languages. Analysis code + per-language energy/accuracy results.

Code and the VETO benchmark for 'The Wrong Kind of Right: Quantifying and Localizing Misfired Alignment in LLMs'.

AtlasNLP is a resource that maps dataset representation by country and shows that coverage, production, and task diversity are deeply uneven across the global NLP ecosystem.

Repository for SIFT: Benchmarking and Evaluating Implicit Biases in Single- and Multi-Agent Language Models