Despite progress in LLM summarization, factual hallucinations persist, motivating Attributed Summary Generation (ASG), which requires sentence-level citations. However, existing prompt-based approaches face severe challenges such as positional preference, poor citation quality and sensitivity to uninformative documents. In view of these limitations, we propose RAAC, a framework of 𝐑eflective 𝐀gents with 𝐀daptive 𝐂ollaboration for attributed summarization. RAAC performs iterative summarization via reflective agents’ collaboration, where a post reflection module evaluates the consistency between the summary and the input documents, based on which it critiques the summary and uses the resulting feedback to recalibrate the inputs to the next adaptive iteration. The agents’ collaboration involves two components: TextAgent and CitationAgent. Experimental results on the ALCE benchmark demonstrate that our framework outperforms existing baselines in both factual correctness and citation quality.

Discovering new intents is crucial for many systems to adapt to evolving users’ needs, making it inherently a continual process. However, most related studies have not explored such continual learning scenarios for new intent discovery. To bridge this research gap, we formalize the Continual Open-Stream Intent Discovery (COSID) task, which requires simultaneous detection of historical intents and discovery of new ones from hybrid streaming data. Our central contribution is a configurable open-stream dataset construction protocol that transforms a static intent dataset into a stream of temporally partitioned datasets, as well as a set of evaluation metrics for the COSID task. We further propose the Continual Intent Discovery Agent (CIDA) for the task, a memory-augmented agent that screens the incoming stream, clusters new-intents’ samples, and employs an LLM for intent abstraction and memory expansion. Experiments on the ClincCL and BankCL data streams constructed according to our protocol validate both the protocol expressiveness and agent superiority over existing baselines.

New Intent Discovery (NID) aims to uncover unseen intent categories from unlabeled user queries, with the core objective of learning an intent representation space with a clear clustering structure. However, in the absence of explicit supervision, how to reliably form a global clustering structure remains challenging. To address this issue, we start from the constituent units of the global structure, namely local relations, and introduce more accessible and more stable forms of local supervision. Based on this idea, we propose LR2GS, which directly trains the intent encoder by leveraging multiple complementary local supervision signals, allowing local constraints to continuously act on the training process and thereby progressively shaping a well-separated global intent structure. Extensive experimental results demonstrate that LR2GS achieves state-of-the-art performance in both unsupervised and semi-supervised settings.

Few-shot intent detection (FSID) targets the classification of user queries into in-scope intent categories or detecting them as out-of-scope, with only a few or even zero labeled examples per class. Existing PLM-based methods struggle in low-resource situations; while LLM-based methods face high inference cost and label interference. To harness their complementary strengths, we propose the FCSLM, a framework that collaborates a small prediction model with a large language model for the FSID task. During training, we leverage LLMs for data augmentation in self-supervised pretraining and supervised fine-tuning a task-specific prediction model. During inference, a multi-round reasoning process first applies the small prediction model to output candidate intents with uncertainty estimations, then invokes an LLM with enriched intent descriptions for refined prediction and OOS detection. Extensive experiments on three benchmark datasets demonstrate that our FCSLM outperforms strong competitors, achieving the new state-of-the-art performance in both intent classification and OOS detection. Our code is available at: https://github.com/hustchenpeng/FCSLM