MSc Master's @mila-quebec.bsky.social @mcgill-nlp.bsky.social
Research Fellow @ RBC Borealis
Model analysis, interpretability, reasoning and hallucination
Studying model behaviours to make them better :))
Looking for Fall '26 PhD
Ziling Cheng
Loading...
π§ TL;DR: These irrelevant context hallucinations show that LLMs go beyond mere parroting π¦ β they do generalize, based on contextual cues and abstract classes. But not reliably. They're more like chameleons π¦ β blending with the context, even when they shouldnβt. 6/n
What do systematic hallucinations in LLMs tell us about their generalization abilities?
Come to our poster at #ACL2025 on July 29th at 4 PM in Level 0, Halls X4/X5. Would love to chat about interpretability, hallucinations, and reasoning :)
@mcgill-nlp.bsky.social @mila-quebec.bsky.social
These examples show answers β even to the same query β can shift under different irrelevant contexts. Can we predict these shifts? 2/n
π Whatβs going on inside?
With mechanistic interpretability, we found:
- LLMs first compute abstract classes (like βlanguageβ) before narrowing to specific answers
- Competing circuits inside the model: one based on context, one based on query. Whichever is stronger wins. 5/n
Sometimes this yields the right answer for the wrong reasoning (βPortugueseβ from βBrazilβ), other times, it produces confident errors (βJapaneseβ from βHondaβ). 4/n
Turns out, we can. They follow a systematic failure mode we call class-based (mis)generalization: the model abstracts the class from the query (e.g., languages) and generalizes based on features from the irrelevant context (e.g., Honda β Japan). 3/n
Do LLMs hallucinate randomly? Not quite.
Our #ACL2025 (Main) paper shows that hallucinations under irrelevant contexts follow a systematic failure mode β revealing how LLMs generalize using abstract classes + context cues, albeit unreliably.
π Paper: arxiv.org/abs/2505.22630 1/n
π Huge thanks to my collaborators @mengcao.bsky.social, Marc-Antoine Rondeau, and my advisor Jackie Cheung for their invaluable guidance and support throughout this work, and to friends at @mila-quebec.bsky.social and @mcgill-nlp.bsky.social π 7/n
How can we use models of cognition to help LLMs interpret figurative language (irony, hyperbole) in a more human-like manner? Come to our #ACL2025NLP poster on Wednesday at 11AM (exhibit hall - exact location TBA) to find out! @mcgill-nlp.bsky.social @mila-quebec.bsky.social @aclmeeting.bsky.social
