Uncertainty estimates in language models are typically computed locally, from the predictive distribution at an individual decoding step. In multi-step reasoning, however, the outcome of interest is often trajectory-level: whether the final answer produced after many intermediate steps will be correct. We show that these two notions of uncertainty can diverge sharply, and study what an uncertainty signal should look like when the prediction target is failure of the complete reasoning trajectory. We first quantify the limitation of local uncertainty for predicting final-answer failure. On GSM8K with a Qwen-1.5B draft model, token-level entropy predicts final incorrectness at AUC 0.531, negative log-likelihood at 0.528, and logit margin at 0.474. Short-window changes in these quantities remain between 0.498 and 0.503. A model can therefore remain locally confident while following an incorrect derivation, or become locally uncertain at a step from which it subsequently recovers. Next-token uncertainty does not directly measure the risk accumulated over a reasoning trajectory. The shape of the confidence trajectory is substantially more informative. We represent each observed prefix using approximately 85 statistics describing the evolution of entropy, likelihood, logit margin, and probability concentration, including means, extrema, percentiles, slopes, variances, temporal changes, and position-dependent behavior. A lightweight logistic predictor trained only on draft trajectories labeled by final correctness reaches AUC 0.864, while the strongest individual trajectory feature reaches 0.803. Its learned weights associate failure with declining logit-margin separation, increasing temporal volatility, and weaker top-k probability concentration. A complementary wavelet analysis finds elevated mid-frequency oscillation in incorrect trajectories, suggesting that persistent instability and unresolved competition between reasoning paths carry more information than isolated uncertainty spikes.

For an uncertainty estimate used during generation, full-trajectory discrimination can be misleading because the estimate must become reliable before the trajectory is complete. We therefore evaluate prefix-conditioned uncertainty. On GSM8K, the Qwen-1.5B predictor rises from AUC 0.60 after 16 tokens to 0.82 after 128 tokens and 0.86 on the complete trajectory. Across GSM8K drafts, AUC at 128 tokens ranges from 0.69 to 0.82. On MATH-500, full-trajectory AUC reaches 0.73-0.78, yet AUC at 128 tokens remains only 0.57-0.59. Harder symbolic problems accumulate evidence more slowly. These results motivate prefix-conditioned AUC as an operational measure of uncertainty quality for systems that must act during decoding.

We test whether this distinction matters in a downstream decision problem. In sparse collaborative decoding, a small draft model generates most of a reasoning trajectory, while a stronger target model is invoked only for short textual corrections when predicted failure risk crosses a threshold. Perturbation experiments reveal a fundamental timing tradeoff: early interventions have greater influence on the eventual trajectory, precisely when uncertainty estimates are least reliable, while later estimates are more accurate but have less ability to redirect generation.

Across five GSM8K draft-target configurations, trajectory-triggered intervention improves over the best fixed insertion schedule by 5.3-20.0 accuracy points and over draft-only decoding by as much as 43.0 points. On StrategyQA it improves over the best open-loop schedule by 5.0-8.3 points, demonstrating transfer beyond arithmetic reasoning. On MATH-500 it gains 7.0-13.5 points in successful configurations but loses when early trajectory prediction is weak. The failure cases expose an important practical distinction between retrospective and actionable uncertainty: a predictor can discriminate failures well after a trajectory is complete while remaining too unreliable at the prefix where a decision must be made. Target behavior introduces a second source of uncertainty. Corrective interventions that degrade an otherwise correct draft exhibit higher entropy and negative log-likelihood and lower maximum probability and logit margin than successful interventions. Consequently, stronger standalone target accuracy does not guarantee better corrective behavior. Reliable collaborative reasoning requires both an uncertainty estimate that identifies failing trajectories early enough to act and a target continuation that is itself sufficiently reliable.

Overall, our results suggest that uncertainty for multi-step language-model reasoning should be treated as a property of an evolving trajectory rather than an isolated token. Aggregate confidence dynamics reveal failure signals that local uncertainty largely misses, while prefix-conditioned evaluation distinguishes uncertainty that predicts eventual failure from uncertainty that is available early enough.

Large language models still struggle with faithfulness and hallucinations despite their remarkable reasoning abilities. In Knowledge Graph Question Answering (KGQA), semantic parsing-based approaches address the limitations by understanding constraints in a user’s question and converting them into a logical form to execute on a knowledge graph. However, existing KGQA benchmarks and methods are biased toward positive and calculation constraints. Negative constraints are neglected, although they frequently appear in real-world questions. In this paper, we introduce a new task, NEgative-conSTrained (NEST) KGQA, where each question contains at least one negative constraint, and a corresponding dataset, NestKGQA. We also design PyLF, a Python-formatted logical form, since existing logical forms are hardly suitable to express negation clearly while maintaining readability. Furthermore, NEST questions naturally contain multiple constraints. To mitigate their semantic complexity, we present a novel framework named CUCKOO, specialized to multiple-constrained questions and ensuring semantic executability. CUCKOO first generates a constraint-aware logical form draft and performs schema-guided semantic matching. It then selectively applies self-directed refinement only when executing improper logical forms yields an empty result, reducing cost while improving robustness. Experimental results demonstrate that CUCKOO consistently outperforms baselines on both conventional KGQA and NEST-KGQA benchmarks under few-shot settings.

Large language model (LLM)-based conversational recommender systems (CRSs) have demonstrated strong capabilities in capturing user preferences and generating contextually relevant recommendations. Nevertheless, the recommendation quality of the models frozen after training inevitably degrades under contextual shifts, such as changes in language and social trends. While periodic model updates are essential to maintain alignment with real-world preferences, training on large-scale data incurs substantial costs. This motivates data-efficient adaptation. However, existing data selection methods struggle to distinguish learnable samples under contextual shifts. To address this, we propose Contextual Shift-Adaptive Data Pruning and Training (CAPT), a framework agnostic to underlying LLM-based CRSs. Specifically, we conceptualize a three-class data taxonomy comprising familiar, valuable, and outlier samples to formalize data behavior under contextual shifts. Based on this taxonomy, we design an importance score estimation scheme that quantifies a sample’s relative learnability for shift adaptation. Leveraging these importance scores, CAPT prioritizes highly learnable samples and further guides shift-adaptive training to actively steer the model toward evolving preferences. Experiments on three CRS benchmarks with real-world temporal splits demonstrate that CAPT outperforms baselines, matching or surpassing full-data fine-tuning performance using only 10-50% of the training data.

Knowledge graphs (KGs) provide a structured representation of real-world facts as triples consisting of entities and their relationships. With the rapid progress of large language models (LLMs), recent studies increasingly explore LLMs for end-to-end KG construction from text. In particular, generative knowledge extraction (GKE) builds KGs by directly generating structured triples from documents. However, generation errors are inevitable, and the resulting KGs often contain triples that do not align with the facts expressed in the source text. To address these issues, we propose GraphRefine, a framework that performs triple-level refinement on KGs constructed via GKE. We first analyze factual inconsistencies that arise in GKE and categorize their types based on a human evaluation. We then construct training data reflecting these types and fine-tune an LLM as a KG refiner. Given a draft KG, the fine-tuned refiner selects a refinement operation for each triple and, if needed, deletes, edits, or rewrites it to reduce factual inconsistencies. Extensive experiments demonstrate that GraphRefine goes beyond deletion-only approaches and improves KG quality from diverse perspectives.

Generating diverse and consistent responses is the ultimate goal of a persona-based dialogue. Although many studies have been conducted, the generated responses tend to be generic and bland due to the personas’ limited descriptiveness. Therefore, it is necessary to expand the given personas for more attractive responses. However, indiscriminate expansion of personas threaten the consistency of responses and therefore reduce the interlocutor’s interest in conversation. To alleviate this issue, we propose a consistent persona expansion framework that improves not only the diversity but also the consistency of persona-based responses. To do so, we define consistency criteria to avoid possible contradictions among personas as follows: 1) Intra-Consistency and 2) Inter-Consistency. Then, we construct a silver profile dataset to deliver the ability to conform with the consistency criteria to the expansion model. Finally, we propose a persona expansion model with an encoder-decoder structure, which considers the relatedness and consistency among personas. Our experiments on the Persona-Chat dataset demonstrate the superiority of the proposed framework.

A persona-grounded dialogue model aims to improve the quality of responses to promote user engagement. However, because the given personas are mostly short and limited to only a few informative words, it is challenging to utilize them to generate diverse responses. To tackle this problem, we propose a novel persona expansion framework, Concept-based Persona eXpansion (CPX). CPX takes the original persona as input and generates expanded personas that contain conceptually rich content. We constitute CPX with two task modules: 1) Concept Extractor and 2) Sentence Generator. To train these modules, we exploit the duality of two tasks with a commonsense dataset consisting of a concept set and the corresponding sentences which contain the given concepts. Extensive experiments on persona expansion and response generation show that our work sufficiently contributes to improving the quality of responses in diversity and richness.

Conversational recommender systems (CRSs) capture a user preference through a conversation. However, the existing CRSs lack capturing comprehensive user preferences. This is because the items mentioned in a conversation are mainly regarded as a user preference. Thus, they have limitations in identifying a user preference from a dialogue context expressed without preferred items. Inspired by the characteristic of an online recommendation community where participants identify a context of a recommendation request and then comment with appropriate items, we exploit the Reddit data. Specifically, we propose a Contrastive Learning approach for Injecting Contextual Knowledge (CLICK) from the Reddit data to the CRS task, which facilitates the capture of a context-level user preference from a dialogue context, regardless of the existence of preferred item-entities. Moreover, we devise a relevance-enhanced contrastive learning loss to consider the fine-grained reflection of multiple recommendable items. We further develop a response generation module to generate a persuasive rationale for a recommendation. Extensive experiments on the benchmark CRS dataset show the effectiveness of CLICK, achieving significant improvements over state-of-the-art methods.

Each utterance in multi-turn empathetic dialogues has features such as emotion, keywords, and utterance-level meaning. Feature transitions between utterances occur naturally. However, existing approaches fail to perceive the transitions because they extract features for the context at the coarse-grained level. To solve the above issue, we propose a novel approach of recognizing feature transitions between utterances, which helps understand the dialogue flow and better grasp the features of utterance that needs attention. Also, we introduce a response generation strategy to help focus on emotion and keywords related to appropriate features when generating responses. Experimental results show that our approach outperforms baselines and especially, achieves significant improvements on multi-turn dialogues.

We propose a novel topic-guided coherence modeling (TGCM) for sentence ordering. Our attention based pointer decoder directly utilize sentence vectors in a permutation-invariant manner, without being compressed into a single fixed-length vector as the paragraph representation. Thus, TGCM can improve global dependencies among sentences and preserve relatively informative paragraph-level semantics. Moreover, to predict the next sentence, we capture topic-enhanced sentence-pair interactions between the current predicted sentence and each next-sentence candidate. With the coherent topical context matching, we promote local dependencies that help identify the tight semantic connections for sentence ordering. The experimental results show that TGCM outperforms state-of-the-art models from various perspectives.