Safety-aligned LLMs frequently generate refusal responses to harmless queries due to superficial lexical similarity with malicious ones — a phenomenon known as over-refusal. Existing approaches either reduce over-refusals or preserve safety, but rarely achieve both simultaneously. We propose AdaCD, a training-free and model-agnostic adaptive contrastive decoding method that dynamically adjusts the refusal token distribution to mitigate over-refusal while maintaining or even enhancing model safety.

We study the problem of user preference drift in LLM-based recommendation and propose RAIE, a region-aware incremental editing framework. Instead of global updates or instance-level edits, RAIE introduces preference regions as structured units for localized adaptation. This design enables efficient continual learning while preserving stable preferences.

Large Language Models (LLMs) need to be both helpful and safe, but achieving both is a major challenge. We propose MidPO, a Mixture of Experts (MoE) framework that fine-tunes two specialized experts for safety and helpfulness and uses a dynamic routing mechanism to adaptively balance them, outperforming existing methods.

Dual-perspective, dual-granularity fairness evaluation for LLM-based job recommendation: we assess bias from both the user and the recruiter sides, at individual and group levels.

AI-generated images are more visually stunning than ever, but how do they really stack up against natural, real-world photos? We introduce D-Judge, a large-scale benchmark designed to systematically investigate and quantify the discrepancies that remain.