Program

09:00 Opening Remarks

09:05–09:50 Invited Talk 1: Anya Belz (Dublin City University)

09:50–10:30 Oral Session 1: Semantic Representation

• Geometric Separation Without Functional Specialization: Auditing Subspace Disentanglement in Dual-Encoder LLM Paraphrase Detection
  Govind Arun, Ashish Abraham and Gadha Lekshmi P

• Exploring Dowker Homology for Sentence Similarity
  Marius Huber and Juri Opitz

• Does Pragmatic Transfer Help Euphemism Detection? Evidence from Target-Span Marking
  Whitney Poh, Noriko Takahashi, Patrick Lee, Julia Sammartino, Libby Barak, Jing Peng and Anna Feldman

10:30–11:00 Coffee Break

11:00–11:45 Invited Talk 2: Ryan Boyd (Vanderbilt University)

What Is a Result a Result About?

11:45–12:25 Oral Session 2: Tuning Recipes that Fail

• Risk-Averse Failure Modes of Jensen-Shannon Distance in RLVR Distributional Alignment
  Sebastian Loftus, Xinpeng Wang, Marie-Catherine De Marneffe and Barbara Plank

• Controlled Fine-Tuning for Two-Tower Recommenders: Why Gradual Unfreezing Can Fail and Cosine Regularization Suffices
  Gabriel Dershowitz and Djallel Bouneffouf

• Do Soft Prompts Have Useful Wavelet Structure? Negative Evidence In A Controlled Setup
  Soumyadeep Dey, Srikanth Tenneti, Bamdev Mishra, Shobhit Niranjan and Mayur Datar

12:25–14:05 Lunch

14:05–14:50 Invited Talk 3: Dirk Hovy (Bocconi University)

14:50–15:30 Oral Session 3: Test-Time Behavior and Evaluation

• Decomposing LLM-Judge–Human Agreement: How Much Is Inherited Surface Preference?
  Gaurav Kumar

• The Flip-Test: A Behavioral Diagnostic for Multi-Agent Debate Failures
  Julia Hu, Alfred Shen and Kumar Lakshmipathi

• Activation-Based Active Learning for In-Context Learning: Challenges and Insights
  Yaseen Osman, Geoff Merrett and Stuart Middleton

15:30–16:05 Coffee Break

16:05–17:30 Poster Session

• The Architecture Was Doing the Work: Why NLI Debiasing Results Do Not Generalise
  Prabhjot Singh

• When Learned Context Planning Fails to Beat Strong Retrieval: A Controlled Study of Planning, Routing, and Reranking for Long-Context QA
  Yingrui Li and Han Chen

• No Optimal Language Set Exists for Multilingual Instruction Tuning: Insights from a Linguistically-Informed Study
  Gürkan Soykan and Gözde Şahin

• Short-Context Gains, Long-Context Drift: Negative Results for Token-Conditioned RoPE
  Allan Kazakov

• When Gradient Importance Lies: Adaptive LoRA Rank Allocation Fails Under GRPO
  Yash Sawant

• Do Not Score the Edited Answer: A Negative Result on Loose Evaluation for Verifiable Instruction Following
  José Luis Delgado

• Perturbation, Not Provenance: Why Activation Steering Does Not Make a Robust Text Watermark
  Sergey Pletenev

17:30 Closing Remarks