Large language models (LLMs) are increasingly used for long-document question answering, where reliable attribution to sources is critical for trust. Existing post-hoc attribution methods work well for extractive QA but struggle in multi-hop, abstractive, and semi-extractive settings, where answers synthesize information across passages. To address these challenges, we argue that post-hoc attribution can be reframed as a reasoning problem, where answers are decomposed into constituent units, each tied to specific context. We first show that prompting models to generate such decompositions alongside attributions improves performance. Building on this, we introduce DecompTune, a post-training method that teaches models to produce answer decompositions as intermediate reasoning steps. We curate a diverse dataset of complex QA tasks, annotated with decompositions by a strong LLM, and post-train Qwen-2.5 (7B and 14B) using a two-stage SFT + GRPO pipeline with task-specific curated rewards. Across extensive experiments and ablations, DecompTune substantially improves attribution quality, outperforming prior methods and matching or exceeding state-of-the-art frontier models.

People intuitively ask natural language dialogue systems for advice on nutrition and dietary guidelines, but systems based on prompted text generation are susceptible to fabricating details, which could be hazardous to non-specialist users. The FoodBench-QA shared task grounds answers in knowledge bases with linked ontologies, in order to evaluate and mitigate fabrication of nutrition information. Our system treats nutrient estimation and entity linking not as a generative problem (predicting numbers from scratch), but as a retrieval problem. We operate on the hypothesis that for structured data like food composition, finding a “real” recipe that is 95% similar is more likely to approximate the correct values than letting the language model fabricate values from sparse context. Our system performed well on food safety labeling from recipe ingredients alone, and it did not benefit from the additional information of recipe titles. In the NER and NEL tasks, our system handled the recipe-focused FCD corpus well, but suffered from poor recall on scientific abstracts and the artificial dataset. These results show the importance of basing information retrieval and question answering in data that is well-matched to the target data.

Lexica – words and associated scores – are widely used as simple, interpretable, generalizable language features to predict sentiment, emotions, mental health, and personality. They also provide insight into the psychological features behind those moods and traits. Such lexica, historically created by human experts, are valuable to linguists, psychologists, and social scientists, but they take years of refinement and have limited coverage. In this paper, we investigate how the lexica that provide psycholinguistic insights could be computationally induced and how they should be assessed. We identify generalizability and interpretability as two essential properties of such lexica. We induce lexica using both context-oblivious and context-aware approaches, compare their predictive performance both within the training corpus and across various corpora, and evaluate their quality using crowd-worker assessment. We find that lexica induced from context-oblivious models are more generalizable and interpretable than those from more accurate context-aware transformer models. In addition, lexicon scores can identify explanatory words more reliably than a high performing transformer with feature-importance measures like SHAP.

In this paper, we present an iterative graph-based approach for the detection of symptoms of COVID-19, the pathology of which seems to be evolving. More generally, the method can be applied to finding context-specific words and texts (e.g. symptom mentions) in large imbalanced corpora (e.g. all tweets mentioning #COVID-19). Given the novelty of COVID-19, we also test if the proposed approach generalizes to the problem of detecting Adverse Drug Reaction (ADR). We find that the approach applied to Twitter data can detect symptom mentions substantially before to their being reported by the Centers for Disease Control (CDC).