Production Retrieval-Augmented Generation (RAG) pays retrieval latency and serving cost on every query – even when the model already knows the answer. We present CONFRAG, a confidence-guided strategy that invokes RAG only when a large language model (LLM) is unlikely to be correct, cutting external retrievals by 5-39% and tail latency by up to 800ms (P90) in a real RAG system while matching the answer quality of always-on RAG. CONFRAG is enabled by CONFQA, a lightweight fine-tuning strategy (one epoch over 3K knowledge-graph facts) that teaches an LLM to answer only when confident and otherwise say "I am unsure", cutting hallucination rates from 20-40% to 5% or below across multiple factuality benchmarks. Two choices make it work in practice: a dampening prompt that discourages overconfident hallucinations, and training on atomic facts that generalizes confidence calibration to unseen domains and question types. At serving time, LLM-QA and RAG run in parallel with early-stopping, so confident answers return immediately.

Speech large language models (LLMs) observe paralinguistic cues such as prosody, emotion, and non-verbal sounds—crucial for intent understanding. However, leveraging these cues faces challenges: limited training data, annotation difficulty, and models exploiting lexical shortcuts over paralinguistic signals. We propose multi-task reinforcement learning (RL) with chain-of-thought prompting that elicits explicit affective reasoning. To address data scarcity, we introduce a paralinguistics-aware speech LLM (PALLM) that jointly optimizes sentiment classification from audio and paralinguistics-aware response generation via a two-stage pipeline. Experiments demonstrate that our approach improves paralinguistics understanding over both supervised baselines and strong proprietary models (Gemini-2.5-Pro, GPT-4o-audio), by 8-12% on Expresso, IEMOCAP, and RAVDESS. The results show that modeling paralinguistic reasoning with multi-task RL is crucial for building emotionally intelligent speech LLMs.

With the rapid growth of online video streaming, recent years have seen increasing concerns about profane language in their content. Detecting profane language in streaming services is challenging due to the long sentences appeared in a video. While recent research on handling long sentences has focused on developing deep learning modeling techniques, little work has focused on techniques on improving data pipelines. In this work, we develop a data collection pipeline to address long sequence of texts and integrate this pipeline with a multi-head self-attention model. With this pipeline, our experiments show the self-attention model offers 12.5% relative accuracy improvement over state-of-the-art distilBERT model on profane language detection while requiring only 3% of parameters. This research designs a better system for informing users of profane language in video streaming services.

Dialogue response generation models that use template ranking rather than direct sequence generation allow model developers to limit generated responses to pre-approved messages. However, manually creating templates is time-consuming and requires domain expertise. To alleviate this problem, we explore automating the process of creating dialogue templates by using unsupervised methods to cluster historical utterances and selecting representative utterances from each cluster. Specifically, we propose an end-to-end model called Deep Sentence Encoder Clustering (DSEC) that uses an auto-encoder structure to jointly learn the utterance representation and construct template clusters. We compare this method to a random baseline that randomly assigns templates to clusters as well as a strong baseline that performs the sentence encoding and the utterance clustering sequentially. To evaluate the performance of the proposed method, we perform an automatic evaluation with two annotated customer service datasets to assess clustering effectiveness, and a human-in-the-loop experiment using a live customer service application to measure the acceptance rate of the generated templates. DSEC performs best in the automatic evaluation, beats both the sequential and random baselines on most metrics in the human-in-the-loop experiment, and shows promising results when compared to gold/manually created templates.