Prior works have shown that fine-tuning on new knowledge can induce factual hallucinations in large language models (LLMs), leading to incorrect outputs when evaluated on previously known information. However, the specific manifestations of such hallucination and its underlying mechanisms remain insufficiently understood. Our work addresses this gap by designing a controlled dataset Biography-Reasoning, and conducting a fine-grained analysis across multiple knowledge types and two task types, including knowledge question answering (QA) and knowledge reasoning tasks. We find that hallucinations not only severely affect tasks involving newly introduced knowledge, but also propagate to other evaluation tasks. Moreover, when fine-tuning on a dataset in which a specific knowledge type consists entirely of new knowledge, LLMs exhibit elevated hallucination tendencies. This suggests that the degree of unfamiliarity within a particular knowledge type, rather than the overall proportion of new knowledge, is a stronger driver of hallucinations. Through interpretability analysis, we show that learning new knowledge weakens the model’s attention to key entities in the input question, leading to an over-reliance on surrounding context and a higher risk of hallucination. Conversely, reintroducing a small amount of known knowledge during the later stages of training restores attention to key entities and substantially mitigates hallucination behavior. Finally, we demonstrate that disrupted attention patterns can propagate across lexically similar contexts, facilitating the spread of hallucinations beyond the original task.

Idiom translation remains a formidable challenge for Large Language Models (LLMs), as the constraints of static parametric memory and the noise in sentence-level retrieval often lead to literal misinterpretations. To address this, we propose DeReA, a detect-retrieve-arbitrate framework. The system employs a preference-aligned detector to identify idiomatic spans by reasoning over semantic conflicts between literal and contextual meanings. Subsequently, an idiom-centric translator invokes a fine-tuned embedding model to efficiently retrieve canonical definitions from an external knowledge base. The translator then utilizes a dual-path arbitration mechanism to select the optimal rendering by weighing the retrieval-augmented translations against direct translation. To evaluate our framework, we introduce LoMI, a high-difficulty benchmark with low data contamination. Experimental results demonstrate that DeReA significantly enhances performance across various model scales, improving GPT-5-mini by over 5.2 points in both idiomatic quality and consistency according to LLM-based metrics. Furthermore, evaluations on an emerging slang dataset from Urban Dictionary validate the potential of our approach in handling novel and evolving linguistic data. Our code is available at https://github.com/jrongqing/DeReA.

Document-level context is crucial for handling discourse challenges in text-to-text document-level machine translation (MT). Despite the increased discourse challenges introduced by noise from automatic speech recognition (ASR), the integration of document-level context in speech translation (ST) remains insufficiently explored. In this paper, we develop DoCIA, an online framework that enhances ST performance by incorporating document-level context. DoCIA decomposes the ST pipeline into four stages. Document-level context is integrated into the ASR refinement, MT, and MT refinement stages through auxiliary LLM (large language model)-based modules. Furthermore, DoCIA leverages document-level information in a multi-level manner while minimizing computational overhead. Additionally, a simple yet effective determination mechanism is introduced to prevent hallucinations from excessive refinement, ensuring the reliability of the final results. Experimental results show that DoCIA significantly outperforms traditional ST baselines in both sentence and discourse metrics across four LLMs, demonstrating its effectiveness in improving ST performance.

Connecting audio encoders with large language models (LLMs) allows the LLM to perform various audio understanding tasks, such as automatic speech recognition (ASR) and audio captioning (AC). Most research focuses on training an adapter layer to generate a unified audio feature for the LLM. However, different tasks may require distinct features that emphasize either semantic or acoustic aspects, making task-specific audio features more desirable. In this paper, we propose Prompt-aware Mixture (PaM) to enhance the Speech LLM that uses multiple audio encoders. Our approach involves using different experts to extract different features based on the prompt that indicates different tasks. Experiments demonstrate that with PaM, only one Speech LLM surpasses the best performances achieved by all single-encoder Speech LLMs on ASR, speaker number verification, and AC tasks. PaM also outperforms other feature fusion baselines, such as concatenation and averaging.

Recent research has shown that large language models (LLMs) can enhance translation quality through self-refinement. In this paper, we build on this idea by extending the refinement from sentence-level to document-level translation, specifically focusing on document-to-document (Doc2Doc) translation refinement. Since sentence-to-sentence (Sent2Sent) and Doc2Doc translation address different aspects of the translation process, we propose fine-tuning LLMs for translation refinement using two intermediate translations, combining the strengths of both Sent2Sent and Doc2Doc. Additionally, recognizing that the quality of intermediate translations varies, we introduce an enhanced fine-tuning method with quality awareness that assigns lower weights to easier translations and higher weights to more difficult ones, enabling the model to focus on challenging translation cases. Experimental results across ten translation tasks with LLaMA-3-8B-Instruct and Mistral-Nemo-Instruct demonstrate the effectiveness of our approach. We will release our code on GitHub.

In this paper, we present an effective method for TextGraphs-17 Shared Task. This task requires selecting an entity from the candidate entities that is relevant to the given question and answer. The selection process is aided by utilizing the shortest path graph in the knowledge graph, connecting entities in the query to the candidate entity. This task aims to explore how to enhance LLMs output with KGs, although current LLMs have certain logical reasoning capabilities, they may not be certain about their own outputs, and the answers they produce may be correct by chance through incorrect paths. In this case, we have introduced a LLM prompt design strategy based on self-ranking and emotion. Specifically, we let the large model score its own answer choices to reflect its confidence in the answer. Additionally, we add emotional incentives to the prompts to encourage the model to carefully examine the questions. Our submissions was conducted under zero-resource setting, and we achieved the second place in the task with an F1-score of 0.8321.

Audio deepfake detection (ADD) is essential for preventing the misuse of synthetic voices that may infringe on personal rights and privacy. Recent zero-shot text-to-speech (TTS) models pose higher risks as they can clone voices with a single utterance. However, the existing ADD datasets are outdated, leading to suboptimal generalization of detection models. In this paper, we construct a new cross-domain ADD dataset comprising over 300 hours of speech data that is generated by five advanced zero-shot TTS models. To simulate real-world scenarios, we employ diverse attack methods and audio prompts from different datasets. Experiments show that, through novel attack-augmented training, the Wav2Vec2-large and Whisper-medium models achieve equal error rates of 4.1% and 6.5% respectively. Additionally, we demonstrate our models’ outstanding few-shot ADD ability by fine-tuning with just one minute of target-domain data. Nonetheless, neural codec compressors greatly affect the detection accuracy, necessitating further research. Our dataset is publicly available (https://github.com/leolya/CD-ADD).

In the paper, we describe a unified system for task 3 of SemEval-2022. The task aims to recognize the semantic structures of sentences by providing two nominal arguments and to evaluate the degree of taxonomic relations. We utilise the strategy that adding language prefix tag in the training set, which is effective for the model. We split the training set to avoid the translation information to be learnt by the model. For the task, we propose a unified model fine-tuned on the multilingual pretrained model, XLM-RoBERTa. The model performs well in subtask 1 (the binary classification subtask). In order to verify whether our model could also perform better in subtask 2 (the regression subtask), the ranking score is transformed into classification labels by an up-sampling strategy. With the ensemble strategy, the performance of our model can be also improved. As a result, the model obtained the second place for subtask 1 and subtask 2 in the competition evaluation.

This paper describes the system for the identifying Plausible Clarifications of Implicit and Underspecified Phrases. This task was set up as an English cloze task, in which clarifications are presented as possible fillers and systems have to score how well each filler plausibly fits in a given context. For this shared task, we propose our own solutions, including supervised proaches, unsupervised approaches with pretrained models, and then we use these models to build an ensemble model. Finally we get the 2nd best result in the subtask1 which is a classification task, and the 3rd best result in the subtask2 which is a regression task.

Natural Language Inference (NLI) datasets contain examples with highly ambiguous labels due to its subjectivity. Several recent efforts have been made to acknowledge and embrace the existence of ambiguity, and explore how to capture the human disagreement distribution. In contrast with directly learning from gold ambiguity labels, relying on special resource, we argue that the model has naturally captured the human ambiguity distribution as long as it’s calibrated, i.e. the predictive probability can reflect the true correctness likelihood. Our experiments show that when model is well-calibrated, either by label smoothing or temperature scaling, it can obtain competitive performance as prior work, on both divergence scores between predictive probability and the true human opinion distribution, and the accuracy. This reveals the overhead of collecting gold ambiguity labels can be cut, by broadly solving how to calibrate the NLI network.