Workshop format

EARL 2026 will be a half-day workshop held in person at CIKM 2026.

Programme highlights

• Invited talk: Bum Jun Kim (The University of Tokyo) will give an invited talk on gradient vulnerability and its implications for LLM-based recommender systems (details below).
• Oral presentations: All accepted papers will be presented orally, as long talks (15 min + 5 min Q&A) or short talks (8 min + 2 min Q&A).
• Best Paper Award: We will continue the award introduced in 2025 to recognize originality, technical excellence, and impact.
• Best Paper Runner-Up recognitions: Two runner-up recognitions will acknowledge additional outstanding contributions.

Invited Talk

Gradient Vulnerability and Its Implications for LLM-Based Recommender Systems

Speaker: Bum Jun Kim (The University of Tokyo)
Time: 8 November 2026, 09:10–09:40 (local Rome time, CET) · 25 min talk + 5 min Q&A

Abstract. Gradients are widely used not only for training AI models but also for measuring importance and guiding decisions. In this talk, I will introduce a vulnerability in gradient measurement: even when two models implement the same function, their gradients can differ. The gradient is not uniquely determined within a specific model, and rather, multiple gradients can be measured in a Transformer. The problem will be demonstrated with empirical and theoretical evidence on LLMs, especially how gradient-based decisions can become unreliable in Transformers. I will deliver how to ensure reliable gradients to develop more trustworthy gradient-based decision-making methods. I will then discuss its implication and challenges in LLM-based recommender systems with several examples.

Programme

The full schedule for 8 November 2026 is available on the home page. See also the list of accepted papers.