Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and generating human-like text, yet they largely operate as reactive agents, responding only when directly prompted. This passivity creates an "awareness gap," limiting their potential as truly collaborative partners in dynamic human discussions. We introduce DiscussLLM, a framework designed to bridge this gap by training models to proactively decide not just what to say, but critically, when to speak. Our primary contribution is a scalable two-stage data generation pipeline that synthesizes a large-scale dataset of realistic multi-turn human discussions. Each discussion is annotated with one of five intervention types (e.g., Factual Correction, Concept Definition) and contains an explicit conversational trigger where an AI intervention adds value. By training models to predict a special silent token when no intervention is needed, they learn to remain quiet until a helpful contribution can be made. We explore two architectural baselines: an integrated end-to-end model and a decoupled classifier-generator system optimized for low-latency inference. We evaluate these models on their ability to accurately time interventions and generate helpful responses, paving the way for more situationally aware and proactive conversational AI.

Question Answering (QA) accounts for a significant portion of LLM usage "in the wild". However, LLMs sometimes produce false or misleading responses, also known as "hallucinations". Therefore, grounding the generated answers in contextually provided information -- i.e., providing evidence for the generated text -- is paramount for LLMs' trustworthiness. Providing this information is the task of context attribution. In this paper, we systematically study LLM-based approaches for this task, namely we investigate (i) zero-shot inference, (ii) LLM ensembling, and (iii) fine-tuning of small LMs on synthetic data generated by larger LLMs. Our key contribution is SynQA: a novel generative strategy for synthesizing context attribution data. Given selected context sentences, an LLM generates QA pairs that are supported by these sentences. This leverages LLMs' natural strengths in text generation while ensuring clear attribution paths in the synthetic training data. We show that the attribution data synthesized via SynQA is highly effective for fine-tuning small LMs for context attribution in different QA tasks and domains. Finally, with a user study, we validate the usefulness of small LMs (fine-tuned on synthetic data from SynQA) in context attribution for QA.

Multimodal Large Language Models (MLLMs) have shown impressive performance in vision and text tasks. However, hallucination remains a major challenge, especially in fields like healthcare where details are critical. In this work, we show how MLLMs may be enhanced to support Visual RAG (V-RAG), a retrieval-augmented generation framework that incorporates both text and visual data from retrieved images. On the MIMIC-CXR chest X-ray report generation and Multicare medical image caption generation datasets, we show that Visual RAG improves the accuracy of entity probing, which asks whether a medical entities is grounded by an image. We show that the improvements extend both to frequent and rare entities, the latter of which may have less positive training data. Downstream, we apply V-RAG with entity probing to correct hallucinations and generate more clinically accurate X-ray reports, obtaining a higher RadGraph-F1 score.

Radiology report generation requires advanced medical image analysis, effective temporal reasoning, and accurate text generation. Although recent innovations, particularly multimodal large language models, have shown improved performance, their supervised fine-tuning (SFT) objective is not explicitly aligned with clinical efficacy. In this work, we introduce EditGRPO, a mixed-policy reinforcement learning algorithm designed specifically to optimize the generation through clinically motivated rewards. EditGRPO integrates on-policy exploration with off-policy guidance by injecting sentence-level detailed corrections during training rollouts. This mixed-policy approach addresses the exploration dilemma and sampling efficiency issues typically encountered in RL. Applied to a Qwen2.5-VL-3B, EditGRPO outperforms both SFT and vanilla GRPO baselines, achieving an average improvement of 3.4% in clinical metrics across four major datasets. Notably, EditGRPO also demonstrates superior out-of-domain generalization, with an average performance gain of 5.9% on unseen datasets.

Self-consistency has emerged as a powerful method for improving the accuracy of short answers generated by large language models. As previously defined, it only concerns the accuracy of a final answer parsed from generated text. In this work, we extend the idea to open response generation, by integrating voting into the decoding method. Each output sentence is selected from among multiple samples, conditioning on the previous selections, based on a simple token overlap score. We compare this "Sample & Select" method to greedy decoding, beam search, nucleus sampling, and the recently introduced hallucination avoiding decoders of DoLA, P-CRR, and S-CRR. We show that Sample & Select improves factuality by a 30% relative margin against these decoders in NLI-based evaluation on the subsets of CNN/DM and XSum used in the FRANK benchmark, while maintaining comparable ROUGE-1 F1 scores against reference summaries. We collect human verifications of the generated summaries, confirming the factual superiority of our method.

When performing complex multi-step reasoning tasks, the ability of Large Language Models (LLMs) to derive structured intermediate proof steps is important for ensuring that the models truly perform the desired reasoning and for improving models' explainability. This paper is centred around a focused study: whether the current state-of-the-art generalist LLMs can leverage the structures in a few examples to better construct the proof structures with in-context learning. Our study specifically focuses on structure-aware demonstration and structure-aware pruning. We demonstrate that they both help improve performance. A detailed analysis is provided to help understand the results.

Separating disinformation from fact on the web has long challenged both the search and the reasoning powers of humans. We show that the reasoning power of large language models (LLMs) and the retrieval power of modern search engines can be combined to automate this process and explainably verify claims. We integrate LLMs and search under a multi-hop evidence pursuit strategy. This strategy generates an initial question based on an input claim using a sequence to sequence model, searches and formulates an answer to the question, and iteratively generates follow-up questions to pursue the evidence that is missing using an LLM. We demonstrate our system on the FEVER 2024 (AVeriTeC) shared task. Compared to a strategy of generating all the questions at once, our method obtains .045 higher label accuracy and .155 higher AVeriTeC score (evaluating the adequacy of the evidence). Through ablations, we show the importance of various design choices, such as the question generation method, medium-sized context, reasoning with one document at a time, adding metadata, paraphrasing, reducing the problem to two classes, and reconsidering the final verdict. Our submitted system achieves .510 AVeriTeC score on the dev set and .477 AVeriTeC score on the test set.

When faced with a large number of product reviews, it is not clear that a human can remember all of them and weight opinions representatively to write a good reference summary. We propose an automatic metric to test the prevalence of the opinions that a summary expresses, based on counting the number of reviews that are consistent with each statement in the summary, while discrediting trivial or redundant statements. To formulate this opinion prevalence metric, we consider several existing methods to score the factual consistency of a summary statement with respect to each individual source review. On a corpus of Amazon product reviews, we gather multiple human judgments of the opinion consistency, to determine which automatic metric best expresses consistency in product reviews. Using the resulting opinion prevalence metric, we show that a human authored summary has only slightly better opinion prevalence than randomly selected extracts from the source reviews, and previous extractive and abstractive unsupervised opinion summarization methods perform worse than humans. We demonstrate room for improvement with a greedy construction of extractive summaries with twice the opinion prevalence achieved by humans. Finally, we show that preprocessing source reviews by simplification can raise the opinion prevalence achieved by existing abstractive opinion summarization systems to the level of human performance.

We study few-shot debugging of transformer based natural language understanding models, using recently popularized test suites to not just diagnose but correct a problem. Given a few debugging examples of a certain phenomenon, and a held-out test set of the same phenomenon, we aim to maximize accuracy on the phenomenon at a minimal cost of accuracy on the original test set. We examine several methods that are faster than full epoch retraining. We introduce a new fast method, which samples a few in-danger examples from the original training set. Compared to fast methods using parameter distance constraints or Kullback-Leibler divergence, we achieve superior original accuracy for comparable debugging accuracy.

Knowledge Graphs (KGs) store information in the form of (head, predicate, tail)-triples. To augment KGs with new knowledge, researchers proposed models for KG Completion (KGC) tasks such as link prediction; i.e., answering (h; p; ?) or (?; p; t) queries. Such models are usually evaluated with averaged metrics on a held-out test set. While useful for tracking progress, averaged single-score metrics cannot reveal what exactly a model has learned -- or failed to learn. To address this issue, we propose KGxBoard: an interactive framework for performing fine-grained evaluation on meaningful subsets of the data, each of which tests individual and interpretable capabilities of a KGC model. In our experiments, we highlight the findings that we discovered with the use of KGxBoard, which would have been impossible to detect with standard averaged single-score metrics.

One major source of vulnerability of neural nets in classification tasks is from overparameterized fully connected layers near the end of the network. In this paper, we propose a new neighborhood preserving layer which can replace these fully connected layers to improve the network robustness. Networks including these neighborhood preserving layers can be trained efficiently. We theoretically prove that our proposed layers are more robust against distortion because they effectively control the magnitude of gradients. Finally, we empirically show that networks with our proposed layers are more robust against state-of-the-art gradient descent based attacks, such as a PGD attack on the benchmark image classification datasets MNIST and CIFAR10.

Modern natural language understanding models depend on pretrained subword embeddings, but applications may need to reason about words that were never or rarely seen during pretraining. We show that examples that depend critically on a rarer word are more challenging for natural language inference models. Then we explore how a model could learn to use definitions, provided in natural text, to overcome this handicap. Our model's understanding of a definition is usually weaker than a well-modeled word embedding, but it recovers most of the performance gap from using a completely untrained word.

We develop a system for the FEVEROUS fact extraction and verification task that ranks an initial set of potential evidence and then pursues missing evidence in subsequent hops by trying to generate it, with a “next hop prediction module” whose output is matched against page elements in a predicted article. Seeking evidence with the next hop prediction module continues to improve FEVEROUS score for up to seven hops. Label classification is trained on possibly incomplete extracted evidence chains, utilizing hints that facilitate numerical comparison. The system achieves .281 FEVEROUS score and .658 label accuracy on the development set, and finishes in second place with .259 FEVEROUS score and .576 label accuracy on the test set.

Image captioning systems are expected to have the ability to combine individual concepts when describing scenes with concept combinations that are not observed during training. In spite of significant progress in image captioning with the help of the autoregressive generation framework, current approaches fail to generalize well to novel concept combinations. We propose a new framework that revolves around probing several similar image caption training instances (retrieval), performing analogical reasoning over relevant entities in retrieved prototypes (analogy), and enhancing the generation process with reasoning outcomes (composition). Our method augments the generation model by referring to the neighboring instances in the training set to produce novel concept combinations in generated captions. We perform experiments on the widely used image captioning benchmarks. The proposed models achieve substantial improvement over the compared baselines on both composition-related evaluation metrics and conventional image captioning metrics.

We propose a framework for answering open domain multi-hop questions in which partial information is read and used to generate followup questions, to finally be answered by a pretrained single-hop answer extractor. This framework makes each hop interpretable, and makes the retrieval associated with later hops as flexible and specific as for the first hop. As a first instantiation of this framework, we train a pointer-generator network to predict followup questions based on the question and partial information. This provides a novel application of a neural question generation network, which is applied to give weak ground truth single-hop followup questions based on the final answers and their supporting facts. Learning to generate followup questions that select the relevant answer spans against downstream supporting facts, while avoiding distracting premises, poses an exciting semantic challenge for text generation. We present an evaluation using the two-hop bridge questions of HotpotQA.

Learning disentangled representations of natural language is essential for many NLP tasks, e.g., conditional text generation, style transfer, personalized dialogue systems, etc. Similar problems have been studied extensively for other forms of data, such as images and videos. However, the discrete nature of natural language makes the disentangling of textual representations more challenging (e.g., the manipulation over the data space cannot be easily achieved). Inspired by information theory, we propose a novel method that effectively manifests disentangled representations of text, without any supervision on semantics. A new mutual information upper bound is derived and leveraged to measure dependence between style and content. By minimizing this upper bound, the proposed method induces style and content embeddings into two independent low-dimensional spaces. Experiments on both conditional text generation and text-style transfer demonstrate the high quality of our disentangled representation in terms of content and style preservation.

We develop a system for the FEVER fact extraction and verification challenge that uses a high precision entailment classifier based on transformer networks pretrained with language modeling, to classify a broad set of potential evidence. The precision of the entailment classifier allows us to enhance recall by considering every statement from several articles to decide upon each claim. We include not only the articles best matching the claim text by TFIDF score, but read additional articles whose titles match named entities and capitalized expressions occurring in the claim text. The entailment module evaluates potential evidence one statement at a time, together with the title of the page the evidence came from (providing a hint about possible pronoun antecedents). In preliminary evaluation, the system achieves .5736 FEVER score, .6108 label accuracy, and .6485 evidence F1 on the FEVER shared task test set.

Existing entailment datasets mainly pose problems which can be answered without attention to grammar or word order. Learning syntax requires comparing examples where different grammar and word order change the desired classification. We introduce several datasets based on synthetic transformations of natural entailment examples in SNLI or FEVER, to teach aspects of grammar and word order. We show that without retraining, popular entailment models are unaware that these syntactic differences change meaning. With retraining, some but not all popular entailment models can learn to compare the syntax properly.