Shudong Liu, Hongwei Liu, Junnan Liu, Linchen Xiao, Songyang Gao, Chengqi Lyu, Yuzhe Gu, Wenwei Zhang, Derek F. Wong, Songyang Zhang, Kai Chen. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large language models (LLMs) exhibit hallucinations (i.e., unfaithful or nonsensical information) when serving as AI assistants in various domains. Since hallucinations always come with truthful content in the LLM responses, previous factuality alignment methods that conduct response-level preference learning inevitably introduced noises during training. Therefore, this paper proposes a fine-grained factuality alignment method based on Direct Preference Optimization (DPO), called Mask-DPO. Incorporating sentence-level factuality as mask signals, Mask-DPO only learns from factually correct sentences in the preferred samples and prevents the penalty on factual contents in the not preferred samples, which resolves the ambiguity in the preference learning. Extensive experimental results demonstrate that Mask-DPO can significantly improve the factuality of LLMs responses to questions from both in-domain and out-of-domain datasets, although these questions and their corresponding topics are unseen during training. Only trained on the ANAH train set, the score of Llama3.1-8B-Instruct on the ANAH test set is improved from 49.19% to 77.53%, even surpassing the score of Llama3.1-70B-Instruct (53.44%), while its FactScore on the out-of-domain Biography dataset is also improved from 30.29% to 39.39%. We further study the generalization property of Mask-DPO using different training sample scaling strategies and find that scaling the number of topics in the dataset is more effective than the number of questions. We provide a hypothesis of what factual alignment is doing with LLMs, on the implication of this phenomenon, and conduct proof-of-concept experiments to verify it. We hope the method and the findings pave the way for future research on scaling factuality alignment.

Enhancing large vision-language models (LVLMs) with visual slow-thinking reasoning is crucial for solving complex multimodal tasks. However, since LVLMs are mainly trained with vision-language alignment, it is difficult to adopt on-policy reinforcement learning (RL) to develop the slow thinking ability because the rollout space is restricted by its initial abilities. Off-policy RL offers a way to go beyond the current policy, but directly distilling trajectories from external models may cause visual hallucinations due to mismatched visual perception abilities across models. To address these issues, this paper proposes **SOPHIA**, a simple and scalable **S**emi-**O**ff-**P**olicy RL for vision-language slow-t**HI**nking re**A**soning. SOPHIA builds a semi-off-policy behavior model by combining on-policy visual understanding from a trainable LVLM with off-policy slow-thinking reasoning from a language model, assigns outcome-based rewards to reasoning, and propagates visual rewards backward. Then LVLM learns slow-thinking reasoning ability from the obtained reasoning trajectories using propagated rewards via off-policy RL algorithms. Extensive experiments with InternVL2.5 and InternVL3.0 with 8B and 38B sizes show the effectiveness of SOPHIA. Notably, SOPHIA improves InternVL3.0-38B by 8.50\% in average, reaching state-of-the-art performance among open-source LVLMs on multiple multimodal reasoning benchmarks, and even outperforms some closed-source models (e.g., GPT-4.1) on the challenging MathVision and OlympiadBench, achieving 49.08\% and 49.95\% pass@1 accuracy, respectively. Analysis shows SOPHIA outperforms supervised fine-tuning and direct on-policy RL methods, offering a better policy initialization for further on-policy training.

Reducing the 'hallucination' problem of Large Language Models (LLMs) is crucial for their wide applications.A comprehensive and finegrained measurement of the hallucination is the first key step for the governance of this issue but is under-explored in the community.Thus, we present ANAH, a bilingual dataset that offers ANalytical Annotation of Hallucinations in LLMs within Generative Question Answering.Each answer sentence in our dataset undergoes rigorous annotation, involving the retrieval of a reference fragment, the judgment of the hallucination type, and the correction of hallucinated content.ANAH consists of ∼12k sentence-level annotations for ∼4.3kLLM responses covering over 700 topics, constructed by a human-in-the-loop pipeline.Thanks to the fine granularity of the hallucination annotations, we can quantitatively confirm that the hallucinations of LLMs progressively accumulate in the answer and use ANAH to train and evaluate hallucination annotators.We conduct extensive experiments on studying generative and discriminative annotators and show that, although current open-source LLMs have difficulties in fine-grained hallucination annotation, the generative annotator trained with ANAH can surpass all open-source LLMs and GPT-3.5, obtain performance competitive with GPT-4, and exhibits better generalization ability on unseen questions. 1

Neural speech codecs aim to compress input signals into minimal bits while maintaining content quality in a low-latency manner.However, existing neural codecs often trade model complexity for reconstruction performance.These codecs primarily use convolutional blocks for feature transformation, which are not inherently suited for capturing the local redundancies in speech signals.To compensate, they require either adversarial discriminators or a large number of model parameters to enhance audio quality.In response to these challenges, we introduce the Efficient Speech Codec (ESC) 1 , a lightweight, parameter-efficient speech codec based on a cross-scale residual vector quantization scheme and transformers.Our model employs mirrored hierarchical window transformer blocks and performs step-wise decoding from coarse-tofine feature representations.To enhance bitrate efficiency, we propose a novel combination of vector quantization techniques along with a pre-training paradigm.Extensive experiments demonstrate that ESC can achieve high-fidelity speech reconstruction with significantly lower model complexity, making it a promising alternative to existing convolutional audio codecs.

COVID-19 has become the most devastating infectious disease of the 21st century. Governments worldwide devised a range of policies to control the pandemic and reduce losses in multiple aspects. To make timely and wise decisions, a thorough analysis of policies with reliable statistical evidence is crucial. Obtaining these statistical references to policies would help decision-makers optimize intervention policies to the maximum. In this work, we designed a policy-aware time series forecasting model based on the transformer architecture and successfully estimated COVID-19 epidemic trends. Through incorporating temporal information from 16 policy indicators, we developed a policy-aware time series model that demonstrated high forecasting performance. We further quantify the causal effect of indicators by employing a counterfactual approach and propose two static metrics lag period and average effect. The empirical results demonstrate that our model causally verifies the effectiveness of all 16 policy indicators in controlling COVID-19 virus transmission in the US. From qualitative analysis, we conclude that frequent adjustments to the extent of policy actions may intensify epidemic spread. Our study provides insight into modeling government policies from a statistical angle, and it has practical applications when confronting potential influenza-like diseases in the future.

Yilun Hua, Nicholas Chernogor, Yuzhe Gu, Seoyeon Jeong, Miranda Luo, Cristian Danescu-Niculescu-Mizil. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Large language models (LLMs) exhibit hallucinations in long-form question-answering tasks across various domains and wide applications. Current hallucination detection and mitigation datasets are limited in domain and size, which struggle to scale due to prohibitive labor costs and insufficient reliability of existing hallucination annotators. To facilitate the scalable oversight of LLM hallucinations, this paper introduces an iterative self-training framework that simultaneously and progressively scales up the annotation dataset and improves the accuracy of the annotator. Based on the Expectation Maximization algorithm, in each iteration, the framework first applies an automatic hallucination annotation pipeline for a scaled dataset and then trains a more accurate annotator on the dataset. This new annotator is adopted in the annotation pipeline for the next iteration. Extensive experimental results demonstrate that the finally obtained hallucination annotator with only 7B parameters surpasses GPT-4 and obtains new state-of-the-art hallucination detection results on HaluEval and HalluQA by zero-shot inference. Such an annotator can not only evaluate the hallucination levels of various LLMs on the large-scale dataset but also help to mitigate the hallucination of LLMs generations, with the Natural Language Inference metric increasing from 25% to 37% on HaluEval.

Recent works have revealed the vulnerability of deep neural networks to backdoor attacks, where a backdoored model orchestrates targeted or untargeted misclassification when activated by a trigger. A line of purification methods (e.g., fine-pruning, neural attention transfer, MCR [69]) have been proposed to remove the backdoor in a model. However, they either fail to reduce the attack success rate of more advanced backdoor attacks or largely degrade the prediction capacity of the model for clean samples. In this paper, we put forward a new purification defense framework, dubbed SAGE, which utilizes self-attention distillation to purge models of backdoors. Unlike traditional attention transfer mechanisms that require a teacher model to supervise the distillation process, SAGE can realize self-purification with a small number of clean samples. To enhance the defense performance, we further propose a dynamic learning rate adjustment strategy that carefully tracks the prediction accuracy of clean samples to guide the learning rate adjustment. We compare the defense performance of SAGE with 6 state-of-the-art defense approaches against 8 backdoor attacks on 4 datasets. It is shown that SAGE can reduce the attack success rate by as much as 90% with less than 3% decrease in prediction accuracy for clean samples. We will open-source our codes upon publication.