Houjun Liu

ACL2025 Tuesday Morning Posters

ACL2025 Katz: segment based attention masking

Key insight: allow by directional attention

ACL2025 Monodorf: exploring modular sturctures transformer based language models

Key insight: learn circuit compositions by learning a binary mask for both faithfulness and scarcity

ACL2025 Li: some more samples of next token prediction

Key insight: when there’s a high difference between generation probability and ground truth, those samples when intervene will cause a more dramatic effect

ACL2025 Kim: counterfactual consistency prompting

Key insight: prompt with counter factual for temporal order to be able to be more consistent temporally

Key insight: breakdown original query into sub parts and answers, using Monte Carlo tree search, with gold traces, then fine tune reasoning model

ACL2025 Liu: Judge as a Judge: Improving the Evaluation of Retrieval-Augmented Generation

Key insight: is a panel of different LLMs to be synthesized together to improve retrieve augmented generation reward signal

Good baseline; also GLM-4

ACL2025 Huang: low entropy watermark detection

Key insight: current red green detection underestimate low entry positions because it can’t be altered as much, fixed with Bayes rule