Mulang' Onando, PhD.’s Post

💡ACL 2025: A Retrospection of Computational Linguists💡 “Imagine you were a mechanic. You’ve spent your life building cars and fixing them. Suddenly, somebody brings a new car. This car is everything you ever dreamt of. Except, occasionally, it just drives off the road and kills people. But you can’t fix it.” Prof. Eduard Hovy, who moderated the panel debate, observes that many papers today feel like LLM popcorn—flashy demos without underpining theory. We need more than benchmarks: reproducibility, reliability, and predictive science. 🤔 💎Reclaiming Control Prof. Luke Zettlemoyer’s keynote argued that data fully determines model behavior. Model Architectures should respect data quirks, not just scale blindly. He proposes tokenizer-free, hierarchical, modular models that are steerable during post-training and incorporate privacy. 💎3 Perspectives on Generalization for NLP The panel's message: keep the generative core, but scaffold it with adaptation, causal discipline, and verifiable tools. ⭐ Prof. Mirella Lapata (the adapter): Shift from abstract generalization to adaptation. Models should learn on the fly, adapt to users, and leverage tools. ⭐Prof. Yue Zhang (the optimist): Scaling has given us zero-shot generalization, but we need mechanisms for causal reasoning. ⭐ Prof. Dan Roth (the pragmatist): LLMs are great at similarity, not structure. External tools assist in logic. 💎A Call for Pluralism The ACL Presidential Prof. Chengqing Zong's address warned against homogenization: one paradigm dominating, academia lagging behind industry, and rising environmental costs. There are many paths up the mountain (exploring brain-inspired computing, lightweight models, & greener methods); why climb only one? 💎Awards & Recognition Best papers - indicate introspection 🌟A Theory of Response Sampling in LLMs: Part Descriptive and Part Prescriptive discusses why and how models generate text. 🌟Fairness through Difference Awareness: Measuring Desired Group Discrimination in LLMs champions explicit modeling of differences for fairness. 🌟Language Models Resist Alignment: Evidence From Data Comprehension explores the friction between pretraining & alignment. 🌟Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention focuses on improving model efficiency. The Test of Time Award, a nostalgia 🔥 25-year ToT Gildea & Jurafsky's - Automatic Labeling of Semantic Roles, and to be frank, SRLs are not yet fully explored to date. 🔥 10-year ToT Thang Luong "Effective Approaches to Attention-based Neural Machine Translation," the starting point for understanding the fundamentals of LLMs. Remember Dzmitry Bahdanau Neural Machine Translation by Jointly Learning to Align and Translate. 💎Way forward for Linguists LLMs may infer linguistic structure, but linguistics is essential for diagnosis, control, fairness, and bias detection. ✨ ACL’25 was one of triumph and reckoning. LLMs dazzle, but we must build fresh foundations.