chenlong zhangchenlong zhang张晨龙

MMR-Life introduces a benchmark for evaluating multimodal reasoning in real-life scenarios with complex visual-textual contexts, aiming to better measure practical reasoning capabilities of modern MLLMs.

We propose RULE, an on-policy RL-based unlearning framework that performs refusal boundary optimization using only 12% forget data and 8% synthetic queries. We introduce the concept of "naturalness" as a novel evaluation dimension. Experimental results demonstrate that RULE achieves Pareto-optimal trade-offs between forgetting and utility.

We introduce DTELS, a task for timeline summarization with controllable granularity levels. Alongside, we release the DTELS-Bench dataset containing 543 topics and 55k articles annotated at three granularity levels. We further propose event-centric evaluation metrics and reveal that existing LLMs struggle with maintaining temporal consistency across granularities.

We introduce Continual Few-shot Event Detection (CFED), a challenging task that requires continual learning of event detection with limited exemplars. To address this, we present HANet, a hierarchical augmentation network for continual few-shot event detection. By leveraging one-exemplar learning and contrastive augmentation, HANet achieves up to +8.4% micro-F1 improvement over prior methods, outperforming both full model retraining and GPT-3.5-based baselines.