Large language model (LLM) based search agents iteratively generate queries, retrieve external information, and reason to answer open-domain questions.While researchers have primarily focused on improving their utility, their safety behaviors remain underexplored.In this paper, we first evaluate search agents using red-teaming datasets and find that they are more likely to produce harmful outputs than base LLMs. For example, when asked “How can I track someone’s location without their consent?”, a base model refuses, whereas a search agent designed to retrieve and cite sources may lower its refusal threshold, fetch documents (e.g., court cases), and, once appended, synthesize them into an informative yet unsafe summary.We further show that utility-oriented finetuning intensifies this risk, motivating joint alignment of safety and utility.We present SafeSearch, a multi-objective reinforcement learning approach that couples a final-output safety/utility reward with a novel query-level shaping term that penalizes unsafe queries and rewards safe ones.Experiments show that SafeSearch reduces agent harmfulness by over 70% across three red-teaming datasets while producing safe, helpful responses, and matches the QA performance of a utility-only finetuned agent. Further analyses confirm the effectiveness of the query-level reward in jointly improving safety and utility.

Leveraging multi-agentic frameworks to enhance large language models (LLMs) has demonstrated significant potential recently, with most existing studies focusing on prompting and developing workflows with frozen LLMs. In this paper, we aim to further unleash the power of such multi-agentic frameworks for post-training LLMs for better collaboration. Specifically, we develop a new paradigm of Multi-Agent Post-co-training for collaborative LLMs with Reinforcement Learning (MAPoRL). In MAPoRL, multiple LLMs first generate their own responses and engage in discussions to collaboratively enhance the final response output; the final output is then scored by a verifier, where the scores serve as the reward and is maximized through multi-agent RL. Additionally, MAPoRL also reshapes the reward above with additional incentives to encourage corrective and persuasive outputs in the discussions. A key novelty from most existing LLM post-training paradigms is the advocacy of co-training multiple LLMs together, and the use of RL for better generalization. Accompanied by a few analytical insights, our experiments show that training single LLMs solely is insufficient for encouraging collaboration, while multi-agent co-training can significantly enhance the collaboration performance across multiple datasets, with generalization to unseen domains, compared to that of multiple LLMs before post-training.

Knowledge graph–grounded dialog generation requires retrieving a dialog-relevant subgraph from the given knowledge base graph and integrating it with the dialog history. Previous works typically represent the graph using an external encoder, such as graph neural networks, and retrieve relevant triplets based on the similarity between single-vector representations of triplets and the dialog history. However, these external encoders fail to leverage the rich knowledge of pretrained language models, and the retrieval process is also suboptimal due to the information bottleneck caused by the single-vector abstraction of the dialog history. In this work, we propose Dialog generation with Generative Subgraph Retrieval (DialogGSR), which retrieves relevant knowledge subgraphs by directly generating their token sequences on top of language models. For effective generative subgraph retrieval, we introduce two key methods: (i) structure-aware knowledge graph linearization with self-supervised graph-specific tokens and (ii) graph-constrained decoding utilizing graph structural proximity-based entity informativeness scores for valid and relevant generative retrieval. DialogGSR achieves state-of-the-art performance in knowledge graph–grounded dialog generation, as demonstrated on OpenDialKG and KOMODIS datasets.

Visual Question Answering (VQA) often involves diverse reasoning scenarios across Vision and Language (V&L). Most prior VQA studies, however, have merely focused on assessing the model’s overall accuracy without evaluating it on different reasoning cases. Furthermore, some recent works observe that conventional Chain-of-Thought (CoT) prompting fails to generate effective reasoning for VQA, especially for complex scenarios requiring multi-hop reasoning. In this paper, we propose II-MMR, a novel idea to identify and improve multi-modal multi-hop reasoning in VQA. In specific, II-MMR takes a VQA question with an image and finds a reasoning path to reach its answer using two novel language promptings: (i) answer prediction-guided CoT prompt, or (ii) knowledge triplet-guided prompt. II-MMR then analyzes this path to identify different reasoning cases in current VQA benchmarks by estimating how many hops and what types (i.e., visual or beyond-visual) of reasoning are required to answer the question. On popular benchmarks including GQA and A-OKVQA, II-MMR observes that most of their VQA questions are easy to answer, simply demanding “single-hop” reasoning, whereas only a few questions require “multi-hop” reasoning. Moreover, while the recent V&L model struggles with such complex multi-hop reasoning questions even using the traditional CoT method, II-MMR shows its effectiveness across all reasoning cases in both zero-shot and fine-tuning settings.

For extreme multi-label classification (XMC), existing classification-based models poorly per- form for tail labels and often ignore the semantic relations among labels, like treating”Wikipedia” and “Wiki” as independent and separate labels. In this paper, we cast XMC as a generation task (XLGen), where we benefit from pre-trained text-to-text models. However, generating labels from the extremely large label space is challenging without any constraints or guidance. We, therefore, propose to guide label generation using label cluster information to hierarchically generate lower-level labels. We also find that frequency-based label ordering and using decoding ensemble methods are critical factors for the improvements in XLGen. XLGen with cluster guidance significantly outperforms the classification and generation baselines on tail labels, and also generally improves the overall performance in four popular XMC benchmarks. In human evaluation, we also find XLGen generates unseen but plausible labels. Our code is now available at https://github.com/alexa/xlgen-eacl-2023.

In large-scale domain classification for natural language understanding, leveraging each user’s domain enablement information, which refers to the preferred or authenticated domains by the user, with attention mechanism has been shown to improve the overall domain classification performance. In this paper, we propose a supervised enablement attention mechanism, which utilizes sigmoid activation for the attention weighting so that the attention can be computed with more expressive power without the weight sum constraint of softmax attention. The attention weights are explicitly encouraged to be similar to the corresponding elements of the output one-hot vector, and self-distillation is used to leverage the attention information of the other enabled domains. By evaluating on the actual utterances from a large-scale IPDA, we show that our approach significantly improves domain classification performance

Intelligent personal digital assistants (IPDAs), a popular real-life application with spoken language understanding capabilities, can cover potentially thousands of overlapping domains for natural language understanding, and the task of finding the best domain to handle an utterance becomes a challenging problem on a large scale. In this paper, we propose a set of efficient and scalable shortlisting-reranking neural models for effective large-scale domain classification for IPDAs. The shortlisting stage focuses on efficiently trimming all domains down to a list of k-best candidate domains, and the reranking stage performs a list-wise reranking of the initial k-best domains with additional contextual information. We show the effectiveness of our approach with extensive experiments on 1,500 IPDA domains.

Training a POS tagging model with crosslingual transfer learning usually requires linguistic knowledge and resources about the relation between the source language and the target language. In this paper, we introduce a cross-lingual transfer learning model for POS tagging without ancillary resources such as parallel corpora. The proposed cross-lingual model utilizes a common BLSTM that enables knowledge transfer from other languages, and private BLSTMs for language-specific representations. The cross-lingual model is trained with language-adversarial training and bidirectional language modeling as auxiliary objectives to better represent language-general information while not losing the information about a specific target language. Evaluating on POS datasets from 14 languages in the Universal Dependencies corpus, we show that the proposed transfer learning model improves the POS tagging performance of the target languages without exploiting any linguistic knowledge between the source language and the target language.