https://fdalvi.github.io/tags/emnlp/ Recent content in Emnlp on Fahim Dalvi Hugo en Sat, 15 Nov 2025 13:00:00 +0300 https://fdalvi.github.io/blog/2025-11-15-two-papers-accepted-at-emnlp-2025/ Sat, 15 Nov 2025 13:00:00 +0300 https://fdalvi.github.io/blog/2025-11-15-two-papers-accepted-at-emnlp-2025/ <p>Thrilled to announce that we have two papers have been accepted at <a href="https://2025.emnlp.org">EMNLP 2025</a>, and one in the co-located <a href="https://arabicnlp2025.sigarab.org">ArabicNLP 2025</a> conference.</p> <p><strong>Beyond the Leaderboard: Understanding Performance Disparities in Large Language Models via Model Diffing</strong></p> <p>Fine-tuning LLMs is the go-to way to improve them, but most evaluations only tell you <em>that</em> a model got better, not <em>why</em>. Our paper takes a different approach using <strong>model diffing</strong> in order to compare the internal representations of two models. We specifically choose SimPO as a case study, and use crosscoders to find and categorize the latent concepts that differentiate the original Gemma-2-9b-it from its fine-tuned version. By looking at the mechanistic changes, we can attribute performance gains to concrete capabilities. This gives us a much richer, more actionable picture of what fine-tuning actually does.</p> https://fdalvi.github.io/blog/2024-11-12-paper-accepted-at-emnlp-2024/ Tue, 12 Nov 2024 13:00:00 +0300 https://fdalvi.github.io/blog/2024-11-12-paper-accepted-at-emnlp-2024/ <p>Pleased to share that our paper, <a href="https://aclanthology.org/2024.emnlp-main.692">Latent Concept-based Explanation of NLP Models</a>, has been accepted at EMNLP 2024!</p> <p>This paper continues our series of work on interpretability. We introduce a method called <strong>LACOAT</strong> (Latent Concept Attribution) that connects predictions with latent concepts present in a model&rsquo;s representation. Hence, we move beyond attribution to individual tokens in the input to a more holistic concept.</p> <p>The code is available on <a href="https://github.com/xuemin-yu/eraser_movie_latentConcept">GitHub</a>. Congratulations to Xuemin, Nadir, Marzia, and Hassan on this work!</p>