Wei Ye

Yidong Wang, Hao Chen, Yue Fan et al. · cmu, pku

Semi-supervised learning (SSL) improves model generalization by leveraging massive unlabeled data to augment limited labeled samples. However, currently, popular SSL evaluation protocols are often constrained to computer vision (CV) tasks. In addition, previous work typically trains deep neural networks from scratch, which is time-consuming and environmentally unfriendly. To address the above issues, we construct a Unified SSL Benchmark (USB) for classification by selecting 15 diverse, challenging, and comprehensive tasks from CV, natural language processing (NLP), and audio processing (Audio), on which we systematically evaluate the dominant SSL methods, and also open-source a modular and extensible codebase for fair evaluation of these SSL methods. We further provide the pre-trained versions of the state-of-the-art neural models for CV tasks to make the cost affordable for further tuning. USB enables the evaluation of a single SSL algorithm on more tasks from multiple domains but with less cost. Specifically, on a single NVIDIA V100, only 39 GPU days are required to evaluate FixMatch on 15 tasks in USB while 335 GPU days (279 GPU days on 4 CV datasets except for ImageNet) are needed on 5 CV tasks with TorchSSL.

Yupeng Chang, Xu Wang, Jindong Wang et al. · cmu, pku

Large language models (LLMs) are gaining increasing popularity in both academia and industry, owing to their unprecedented performance in various applications. As LLMs continue to play a vital role in both research and daily use, their evaluation becomes increasingly critical, not only at the task level, but also at the society level for better understanding of their potential risks. Over the past years, significant efforts have been made to examine LLMs from various perspectives. This paper presents a comprehensive review of these evaluation methods for LLMs, focusing on three key dimensions: what to evaluate, where to evaluate, and how to evaluate. Firstly, we provide an overview from the perspective of evaluation tasks, encompassing general natural language processing tasks, reasoning, medical usage, ethics, educations, natural and social sciences, agent applications, and other areas. Secondly, we answer the `where' and `how' questions by diving into the evaluation methods and benchmarks, which serve as crucial components in assessing performance of LLMs. Then, we summarize the success and failure cases of LLMs in different tasks. Finally, we shed light on several future challenges that lie ahead in LLMs evaluation. Our aim is to offer invaluable insights to researchers in the realm of LLMs evaluation, thereby aiding the development of more proficient LLMs. Our key point is that evaluation should be treated as an essential discipline to better assist the development of LLMs. We consistently maintain the related open-source materials at: https://github.com/MLGroupJLU/LLM-eval-survey.

Yidong Wang, Zhuohao Yu, Zhengran Zeng et al. · cmu, pku

Instruction tuning large language models (LLMs) remains a challenging task, owing to the complexity of hyperparameter selection and the difficulty involved in evaluating the tuned models. To determine the optimal hyperparameters, an automatic, robust, and reliable evaluation benchmark is essential. However, establishing such a benchmark is not a trivial task due to the challenges associated with evaluation accuracy and privacy protection. In response to these challenges, we introduce a judge large language model, named PandaLM, which is trained to distinguish the superior model given several LLMs. PandaLM's focus extends beyond just the objective correctness of responses, which is the main focus of traditional evaluation datasets. It addresses vital subjective factors such as relative conciseness, clarity, adherence to instructions, comprehensiveness, and formality. To ensure the reliability of PandaLM, we collect a diverse human-annotated test dataset, where all contexts are generated by humans and labels are aligned with human preferences. Our results indicate that PandaLM-7B achieves 93.75% of GPT-3.5's evaluation ability and 88.28% of GPT-4's in terms of F1-score on our test dataset. PandaLM enables the evaluation of LLM to be fairer but with less cost, evidenced by significant improvements achieved by models tuned through PandaLM compared to their counterparts trained with default Alpaca's hyperparameters. In addition, PandaLM does not depend on API-based evaluations, thus avoiding potential data leakage. All resources of PandaLM are released at https://github.com/WeOpenML/PandaLM.

Yidong Wang, Zhuohao Yu, Jindong Wang et al. · pku

Vision-Language models (VLMs) that use contrastive language-image pre-training have shown promising zero-shot classification performance. However, their performance on imbalanced dataset is relatively poor, where the distribution of classes in the training dataset is skewed, leading to poor performance in predicting minority classes. For instance, CLIP achieved only 5% accuracy on the iNaturalist18 dataset. We propose to add a lightweight decoder to VLMs to avoid OOM (out of memory) problem caused by large number of classes and capture nuanced features for tail classes. Then, we explore improvements of VLMs using prompt tuning, fine-tuning, and incorporating imbalanced algorithms such as Focal Loss, Balanced SoftMax and Distribution Alignment. Experiments demonstrate that the performance of VLMs can be further boosted when used with decoder and imbalanced methods. Specifically, our improved VLMs significantly outperforms zero-shot classification by an average accuracy of 6.58%, 69.82%, and 6.17%, on ImageNet-LT, iNaturalist18, and Places-LT, respectively. We further analyze the influence of pre-training data size, backbones, and training cost. Our study highlights the significance of imbalanced learning algorithms in face of VLMs pre-trained by huge data. We release our code at https://github.com/Imbalance-VLM/Imbalance-VLM.

Botao Yu, Peiling Lu, Rui Wang et al. · microsoft-research

Symbolic music generation aims to generate music scores automatically. A recent trend is to use Transformer or its variants in music generation, which is, however, suboptimal, because the full attention cannot efficiently model the typically long music sequences (e.g., over 10,000 tokens), and the existing models have shortcomings in generating musical repetition structures. In this paper, we propose Museformer, a Transformer with a novel fine- and coarse-grained attention for music generation. Specifically, with the fine-grained attention, a token of a specific bar directly attends to all the tokens of the bars that are most relevant to music structures (e.g., the previous 1st, 2nd, 4th and 8th bars, selected via similarity statistics); with the coarse-grained attention, a token only attends to the summarization of the other bars rather than each token of them so as to reduce the computational cost. The advantages are two-fold. First, it can capture both music structure-related correlations via the fine-grained attention, and other contextual information via the coarse-grained attention. Second, it is efficient and can model over 3X longer music sequences compared to its full-attention counterpart. Both objective and subjective experimental results demonstrate its ability to generate long music sequences with high quality and better structures.

Kaijie Zhu, Jindong Wang, Jiaheng Zhou et al. · cmu, pku

The increasing reliance on Large Language Models (LLMs) across academia and industry necessitates a comprehensive understanding of their robustness to prompts. In response to this vital need, we introduce PromptRobust, a robustness benchmark designed to measure LLMs' resilience to adversarial prompts. This study uses a plethora of adversarial textual attacks targeting prompts across multiple levels: character, word, sentence, and semantic. The adversarial prompts, crafted to mimic plausible user errors like typos or synonyms, aim to evaluate how slight deviations can affect LLM outcomes while maintaining semantic integrity. These prompts are then employed in diverse tasks including sentiment analysis, natural language inference, reading comprehension, machine translation, and math problem-solving. Our study generates 4,788 adversarial prompts, meticulously evaluated over 8 tasks and 13 datasets. Our findings demonstrate that contemporary LLMs are not robust to adversarial prompts. Furthermore, we present a comprehensive analysis to understand the mystery behind prompt robustness and its transferability. We then offer insightful robustness analysis and pragmatic recommendations for prompt composition, beneficial to both researchers and everyday users.

Jindong Wang, Xixu Hu, Wenxin Hou et al. · cmu, pku

ChatGPT is a recent chatbot service released by OpenAI and is receiving increasing attention over the past few months. While evaluations of various aspects of ChatGPT have been done, its robustness, i.e., the performance to unexpected inputs, is still unclear to the public. Robustness is of particular concern in responsible AI, especially for safety-critical applications. In this paper, we conduct a thorough evaluation of the robustness of ChatGPT from the adversarial and out-of-distribution (OOD) perspective. To do so, we employ the AdvGLUE and ANLI benchmarks to assess adversarial robustness and the Flipkart review and DDXPlus medical diagnosis datasets for OOD evaluation. We select several popular foundation models as baselines. Results show that ChatGPT shows consistent advantages on most adversarial and OOD classification and translation tasks. However, the absolute performance is far from perfection, which suggests that adversarial and OOD robustness remains a significant threat to foundation models. Moreover, ChatGPT shows astounding performance in understanding dialogue-related texts and we find that it tends to provide informal suggestions for medical tasks instead of definitive answers. Finally, we present in-depth discussions of possible research directions.

Hao Chen, Ran Tao, Han Zhang et al. · cmu, pku

While parameter efficient tuning (PET) methods have shown great potential with transformer architecture on Natural Language Processing (NLP) tasks, their effectiveness with large-scale ConvNets is still under-studied on Computer Vision (CV) tasks. This paper proposes Conv-Adapter, a PET module designed for ConvNets. Conv-Adapter is light-weight, domain-transferable, and architecture-agnostic with generalized performance on different tasks. When transferring on downstream tasks, Conv-Adapter learns tasks-specific feature modulation to the intermediate representations of backbones while keeping the pre-trained parameters frozen. By introducing only a tiny amount of learnable parameters, e.g., only 3.5% full fine-tuning parameters of ResNet50. It can also be applied for transformer-based backbones. Conv-Adapter outperforms previous PET baseline methods and achieves comparable or surpasses the performance of full fine-tuning on 23 classification tasks of various domains. It also presents superior performance on the few-shot classification with an average margin of 3.39%. Beyond classification, Conv-Adapter can generalize to detection and segmentation tasks with more than 50% reduction of parameters but comparable performance to the traditional full fine-tuning.

Yeying Jin, Beibei Lin, Wending Yan et al.

Visibility in hazy nighttime scenes is frequently reduced by multiple factors, including low light, intense glow, light scattering, and the presence of multicolored light sources. Existing nighttime dehazing methods often struggle with handling glow or low-light conditions, resulting in either excessively dark visuals or unsuppressed glow outputs. In this paper, we enhance the visibility from a single nighttime haze image by suppressing glow and enhancing low-light regions. To handle glow effects, our framework learns from the rendered glow pairs. Specifically, a light source aware network is proposed to detect light sources of night images, followed by the APSF (Atmospheric Point Spread Function)-guided glow rendering. Our framework is then trained on the rendered images, resulting in glow suppression. Moreover, we utilize gradient-adaptive convolution, to capture edges and textures in hazy scenes. By leveraging extracted edges and textures, we enhance the contrast of the scene without losing important structural details. To boost low-light intensity, our network learns an attention map, then adjusted by gamma correction. This attention has high values on low-light regions and low values on haze and glow regions. Extensive evaluation on real nighttime haze images, demonstrates the effectiveness of our method. Our experiments demonstrate that our method achieves a PSNR of 30.38dB, outperforming state-of-the-art methods by 13% on GTA5 nighttime haze dataset. Our data and code is available at https://github.com/jinyeying/nighttime_dehaze.

Yeying Jin, Wei Ye, Wenhan Yang et al.

Removing soft and self shadows that lack clear boundaries from a single image is still challenging. Self shadows are shadows that are cast on the object itself. Most existing methods rely on binary shadow masks, without considering the ambiguous boundaries of soft and self shadows. In this paper, we present DeS3, a method that removes hard, soft and self shadows based on adaptive attention and ViT similarity. Our novel ViT similarity loss utilizes features extracted from a pre-trained Vision Transformer. This loss helps guide the reverse sampling towards recovering scene structures. Our adaptive attention is able to differentiate shadow regions from the underlying objects, as well as shadow regions from the object casting the shadow. This capability enables DeS3 to better recover the structures of objects even when they are partially occluded by shadows. Different from existing methods that rely on constraints during the training phase, we incorporate the ViT similarity during the sampling stage. Our method outperforms state-of-the-art methods on the SRD, AISTD, LRSS, USR and UIUC datasets, removing hard, soft, and self shadows robustly. Specifically, our method outperforms the SOTA method by 16\% of the RMSE of the whole image on the LRSS dataset. Our data and code is available at: \url{https://github.com/jinyeying/DeS3_Deshadow}

Bo Li, Gexiang Fang, Yang Yang et al.

The capability of Large Language Models (LLMs) like ChatGPT to comprehend user intent and provide reasonable responses has made them extremely popular lately. In this paper, we focus on assessing the overall ability of ChatGPT using 7 fine-grained information extraction (IE) tasks. Specially, we present the systematically analysis by measuring ChatGPT's performance, explainability, calibration, and faithfulness, and resulting in 15 keys from either the ChatGPT or domain experts. Our findings reveal that ChatGPT's performance in Standard-IE setting is poor, but it surprisingly exhibits excellent performance in the OpenIE setting, as evidenced by human evaluation. In addition, our research indicates that ChatGPT provides high-quality and trustworthy explanations for its decisions. However, there is an issue of ChatGPT being overconfident in its predictions, which resulting in low calibration. Furthermore, ChatGPT demonstrates a high level of faithfulness to the original text in the majority of cases. We manually annotate and release the test sets of 7 fine-grained IE tasks contains 14 datasets to further promote the research. The datasets and code are available at https://github.com/pkuserc/ChatGPT_for_IE.

Xuanwang Zhang, Yunze Song, Yidong Wang et al. · pku

Large Language Models (LLMs) demonstrate human-level capabilities in dialogue, reasoning, and knowledge retention. However, even the most advanced LLMs face challenges such as hallucinations and real-time updating of their knowledge. Current research addresses this bottleneck by equipping LLMs with external knowledge, a technique known as Retrieval Augmented Generation (RAG). However, two key issues constrained the development of RAG. First, there is a growing lack of comprehensive and fair comparisons between novel RAG algorithms. Second, open-source tools such as LlamaIndex and LangChain employ high-level abstractions, which results in a lack of transparency and limits the ability to develop novel algorithms and evaluation metrics. To close this gap, we introduce RAGLAB, a modular and research-oriented open-source library. RAGLAB reproduces 6 existing algorithms and provides a comprehensive ecosystem for investigating RAG algorithms. Leveraging RAGLAB, we conduct a fair comparison of 6 RAG algorithms across 10 benchmarks. With RAGLAB, researchers can efficiently compare the performance of various algorithms and develop novel algorithms.

Chenfei Kang, Peiling Lu, Botao Yu et al.

Music is used to convey emotions, and thus generating emotional music is important in automatic music generation. Previous work on emotional music generation directly uses annotated emotion labels as control signals, which suffers from subjective bias: different people may annotate different emotions on the same music, and one person may feel different emotions under different situations. Therefore, directly mapping emotion labels to music sequences in an end-to-end way would confuse the learning process and hinder the model from generating music with general emotions. In this paper, we propose EmoGen, an emotional music generation system that leverages a set of emotion-related music attributes as the bridge between emotion and music, and divides the generation into two stages: emotion-to-attribute mapping with supervised clustering, and attribute-to-music generation with self-supervised learning. Both stages are beneficial: in the first stage, the attribute values around the clustering center represent the general emotions of these samples, which help eliminate the impacts of the subjective bias of emotion labels; in the second stage, the generation is completely disentangled from emotion labels and thus free from the subjective bias. Both subjective and objective evaluations show that EmoGen outperforms previous methods on emotion control accuracy and music quality respectively, which demonstrate our superiority in generating emotional music. Music samples generated by EmoGen are available via this link:https://ai-muzic.github.io/emogen/, and the code is available at this link:https://github.com/microsoft/muzic/.

Yongrui Heng, Chaoya Jiang, Han Yang et al.

Self-evolution of multimodal large language models (MLLMs) remains a critical challenge: pseudo-label-based methods suffer from progressive quality degradation as model predictions drift, while template-based methods are confined to a static set of transformations that cannot adapt in difficulty or diversity. We contend that robust, continuous self-improvement requires not only deterministic external feedback independent of the model's internal certainty, but also a mechanism to perpetually diversify the training distribution. To this end, we introduce EVE (Executable Visual transformation-based self-Evolution), a novel framework that entirely bypasses pseudo-labels by harnessing executable visual transformations continuously enriched in both variety and complexity. EVE adopts a Challenger-Solver dual-policy architecture. The Challenger maintains and progressively expands a queue of visual transformation code examples, from which it synthesizes novel Python scripts to perform dynamic visual transformations. Executing these scripts yields VQA problems with absolute, execution-verified ground-truth answers, eliminating any reliance on model-generated supervision. A multi-dimensional reward system integrating semantic diversity and dynamic difficulty calibration steers the Challenger to enrich its code example queue while posing progressively more challenging tasks, preventing mode collapse and fostering reciprocal co-evolution between the two policies. Extensive experiments demonstrate that EVE consistently surpasses existing self-evolution methods, establishing a robust and scalable paradigm for verifiable MLLM self-evolution. The code is available at https://github.com/0001Henry/EVE .

Zhanghao Sun, Wei Ye, Jinhui Xiong et al.

Direct time-of-flight (dToF) sensors are promising for next-generation on-device 3D sensing. However, limited by manufacturing capabilities in a compact module, the dToF data has a low spatial resolution (e.g., $\sim 20\times30$ for iPhone dToF), and it requires a super-resolution step before being passed to downstream tasks. In this paper, we solve this super-resolution problem by fusing the low-resolution dToF data with the corresponding high-resolution RGB guidance. Unlike the conventional RGB-guided depth enhancement approaches, which perform the fusion in a per-frame manner, we propose the first multi-frame fusion scheme to mitigate the spatial ambiguity resulting from the low-resolution dToF imaging. In addition, dToF sensors provide unique depth histogram information for each local patch, and we incorporate this dToF-specific feature in our network design to further alleviate spatial ambiguity. To evaluate our models on complex dynamic indoor environments and to provide a large-scale dToF sensor dataset, we introduce DyDToF, the first synthetic RGB-dToF video dataset that features dynamic objects and a realistic dToF simulator following the physical imaging process. We believe the methods and dataset are beneficial to a broad community as dToF depth sensing is becoming mainstream on mobile devices. Our code and data are publicly available: https://github.com/facebookresearch/DVSR/

Xiangyu Xi, Jianwei Lv, Shuaipeng Liu et al.

Event detection (ED) identifies and classifies event triggers from unstructured texts, serving as a fundamental task for information extraction. Despite the remarkable progress achieved in the past several years, most research efforts focus on detecting events from formal texts (e.g., news articles, Wikipedia documents, financial announcements). Moreover, the texts in each dataset are either from a single source or multiple yet relatively homogeneous sources. With massive amounts of user-generated text accumulating on the Web and inside enterprises, identifying meaningful events in these informal texts, usually from multiple heterogeneous sources, has become a problem of significant practical value. As a pioneering exploration that expands event detection to the scenarios involving informal and heterogeneous texts, we propose a new large-scale Chinese event detection dataset based on user reviews, text conversations, and phone conversations in a leading e-commerce platform for food service. We carefully investigate the proposed dataset's textual informality and multi-source heterogeneity characteristics by inspecting data samples quantitatively and qualitatively. Extensive experiments with state-of-the-art event detection methods verify the unique challenges posed by these characteristics, indicating that multi-source informal event detection remains an open problem and requires further efforts. Our benchmark and code are released at \url{https://github.com/myeclipse/MUSIED}.

Marcos V. Conde, Florin-Alexandru Vasluianu, Jinhui Xiong et al.

The increasing demand for augmented reality (AR) and virtual reality (VR) applications highlights the need for efficient depth information processing. Depth maps, essential for rendering realistic scenes and supporting advanced functionalities, are typically large and challenging to stream efficiently due to their size. This challenge introduces a focus on developing innovative depth upsampling techniques to reconstruct high-quality depth maps from compressed data. These techniques are crucial for overcoming the limitations posed by depth compression, which often degrades quality, loses scene details and introduces artifacts. By enhancing depth upsampling methods, this challenge aims to improve the efficiency and quality of depth map reconstruction. Our goal is to advance the state-of-the-art in depth processing technologies, thereby enhancing the overall user experience in AR and VR applications.

Zile Qiao, Wei Ye, Tong Zhang et al. · pku

Answering natural language questions on knowledge graphs (KGQA) remains a great challenge in terms of understanding complex questions via multi-hop reasoning. Previous efforts usually exploit large-scale entity-related text corpora or knowledge graph (KG) embeddings as auxiliary information to facilitate answer selection. However, the rich semantics implied in off-the-shelf relation paths between entities is far from well explored. This paper proposes improving multi-hop KGQA by exploiting relation paths' hybrid semantics. Specifically, we integrate explicit textual information and implicit KG structural features of relation paths based on a novel rotate-and-scale entity link prediction framework. Extensive experiments on three existing KGQA datasets demonstrate the superiority of our method, especially in multi-hop scenarios. Further investigation confirms our method's systematical coordination between questions and relation paths to identify answer entities.

Hongrui Jia, Chaoya Jiang, Shikun Zhang et al.

As Large Multimodal Models (LMMs) scale up and reinforcement learning (RL) methods mature, LMMs have made notable progress in complex reasoning and decision making. Yet training still relies on static data and fixed recipes, making it difficult to diagnose capability blind spots or provide dynamic, targeted reinforcement. Motivated by findings that test driven error exposure and feedback based correction outperform repetitive practice, we propose Diagnostic-driven Progressive Evolution (DPE), a spiral loop where diagnosis steers data generation and reinforcement, and each iteration re-diagnoses the updated model to drive the next round of targeted improvement. DPE has two key components. First, multiple agents annotate and quality control massive unlabeled multimodal data, using tools such as web search and image editing to produce diverse, realistic samples. Second, DPE attributes failures to specific weaknesses, dynamically adjusts the data mixture, and guides agents to generate weakness focused data for targeted reinforcement. Experiments on Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B-Instruct show stable, continual gains across eleven benchmarks, indicating DPE as a scalable paradigm for continual LMM training under open task distributions. Our code, models, and data are publicly available at https://github.com/hongruijia/DPE.

Chaoya Jiang, Haiyang Xu, Wei Ye et al.

Vision Transformer (ViT) based Vision-Language Pre-training (VLP) models have demonstrated impressive performance in various tasks. However, the lengthy visual token sequences fed into ViT can lead to training inefficiency and ineffectiveness. Existing efforts address the challenge by either bottom-level patch extraction in the ViT backbone or top-level patch abstraction outside, not balancing training efficiency and effectiveness well. Inspired by text summarization in natural language processing, we propose a Bottom-Up Patch Summarization approach named BUS, coordinating bottom-level extraction and top-level abstraction to learn a concise summary of lengthy visual token sequences efficiently. Specifically, We incorporate a Text-Semantics-Aware Patch Selector (TSPS) into the ViT backbone to perform a coarse-grained visual token extraction and then attach a flexible Transformer-based Patch Abstraction Decoder (PAD) upon the backbone for top-level visual abstraction. This bottom-up collaboration enables our BUS to yield high training efficiency while maintaining or even improving effectiveness. We evaluate our approach on various visual-language understanding and generation tasks and show competitive downstream task performance while boosting the training efficiency by 50\%. Additionally, our model achieves state-of-the-art performance on many downstream tasks by increasing input image resolution without increasing computational costs over baselines.

Weizheng Gu, Chengze Li, Zhuohao Yu et al.

Large language models are increasingly evaluated as interactive agents, yet standard agent benchmarks conflate two qualitatively distinct sources of success: semantic tool-use and interface-specific interaction pattern memorization. Because both mechanisms can yield identical task success on the original interface, benchmark scores alone are not identifiable evidence of environment-invariant capability. We propose PIPE, a protocol-level evaluation augmentation for diagnosing interface reliance by minimally rewriting environment interfaces while preserving task semantics and execution behavior. Across 16 environments from AgentBench and AgentGym and a range of open-source and API-based agents, PIPE reveals that trajectory-SFT substantially amplifies interface shortcutting: trained agents degrade sharply under minimal interface rewrites, while non-trajectory-trained models remain largely stable. We further introduce Interface Reliance (IR), a counterbalanced alias-based metric that quantifies preference for training-time interfaces, and show that interface shortcutting exhibits environment-dependent, non-monotonic training dynamics that remain invisible under standard evaluation. Our code is available at https://anonymous.4open.science/r/What-Do-Agents-Learn-from-Trajectory-SFT-Semantics-or-Interfaces--0831/.

Bo Li, Dingyao Yu, Wei Ye et al.

Sequence generation demonstrates promising performance in recent information extraction efforts, by incorporating large-scale pre-trained Seq2Seq models. This paper investigates the merits of employing sequence generation in relation extraction, finding that with relation names or synonyms as generation targets, their textual semantics and the correlation (in terms of word sequence pattern) among them affect model performance. We then propose Relation Extraction with Label Augmentation (RELA), a Seq2Seq model with automatic label augmentation for RE. By saying label augmentation, we mean prod semantically synonyms for each relation name as the generation target. Besides, we present an in-depth analysis of the Seq2Seq model's behavior when dealing with RE. Experimental results show that RELA achieves competitive results compared with previous methods on four RE datasets.

Dingyao Yu, Kaitao Song, Peiling Lu et al.

AI-empowered music processing is a diverse field that encompasses dozens of tasks, ranging from generation tasks (e.g., timbre synthesis) to comprehension tasks (e.g., music classification). For developers and amateurs, it is very difficult to grasp all of these task to satisfy their requirements in music processing, especially considering the huge differences in the representations of music data and the model applicability across platforms among various tasks. Consequently, it is necessary to build a system to organize and integrate these tasks, and thus help practitioners to automatically analyze their demand and call suitable tools as solutions to fulfill their requirements. Inspired by the recent success of large language models (LLMs) in task automation, we develop a system, named MusicAgent, which integrates numerous music-related tools and an autonomous workflow to address user requirements. More specifically, we build 1) toolset that collects tools from diverse sources, including Hugging Face, GitHub, and Web API, etc. 2) an autonomous workflow empowered by LLMs (e.g., ChatGPT) to organize these tools and automatically decompose user requests into multiple sub-tasks and invoke corresponding music tools. The primary goal of this system is to free users from the intricacies of AI-music tools, enabling them to concentrate on the creative aspect. By granting users the freedom to effortlessly combine tools, the system offers a seamless and enriching music experience.

Wei Ye, Yixin Su, Yueguo Chen et al.

Visual Question Answering (VQA) is the task of answering questions based on image content. Building upon this, Knowledge-Based VQA (KB-VQA) requires models to answer questions that depend on external knowledge beyond the visual content of an image. In such settings, effective knowledge filtering is essential for achieving high question answering accuracy. Typical filtering methods suffer from two issues: they fail to focus on parts relevant to the question during candidate section encoding, and they use similarity metrics to locate a section from a single article, resulting in information limitation. To address these issues, this paper proposes a question-focused, cross-article filtering method. Specifically, we design a trainable Question-Focused Filter (QFF) and a Chunk-based Dynamic Cross-Article Selection module (CDA). This approach maintains inference time comparable to the optimal method with the shorter context length, efficiently obtaining high-quality filtered knowledge. The accuracy outperforms current state-of-the-art methods by 3.2 and 2.2 percentage points on Encyclopedic-VQA and InfoSeek, respectively. The code is publicly available at: https://github.com/leaffeall/QKVQA.

Bo Li, Mingda Wang, Gexiang Fang et al.

We revisit retrieval-augmented generation (RAG) by embedding retrieval control directly into generation. Instead of treating retrieval as an external intervention, we express retrieval decisions within token-level decoding, enabling end-to-end coordination without additional controllers or classifiers. Under the paradigm of Retrieval as Generation, we propose \textbf{GRIP} (\textbf{G}eneration-guided \textbf{R}etrieval with \textbf{I}nformation \textbf{P}lanning), a unified framework in which the model regulates retrieval behavior through control-token emission. Central to GRIP is \textit{Self-Triggered Information Planning}, which allows the model to decide when to retrieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples retrieval and reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, partially answerable, and multi-hop queries, each aligned with specific token patterns. Experiments on five QA benchmarks show that GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters.

Peiyang Liu, Ziqiang Cui, Xi Wang et al.

Iterative Retrieval-Augmented Generation (iRAG) has emerged as a powerful paradigm for answering complex multi-hop questions by progressively retrieving and reasoning over external documents. However, current systems predominantly operate on parsed text, which creates two critical bottlenecks: (1) \textit{Coarse-grained attribution}, where users are burdened with manually locating evidence within lengthy documents based on vague text-level citations; and (2) \textit{Visual semantic loss}, where the conversion of visually rich documents (e.g., slides, PDFs with charts) into text discards spatial logic and layout cues essential for reasoning. To bridge this gap, we present \textbf{Chain of Evidence (CoE)}, a retriever-agnostic visual attribution framework that leverages Vision-Language Models to reason directly over screenshots of retrieved document candidates. CoE eliminates format-specific parsing and outputs precise bounding boxes, visualizing the complete reasoning chain within the retrieved candidate set. We evaluate CoE on two distinct benchmarks: \textbf{Wiki-CoE}, a large-scale dataset of structured web pages derived from 2WikiMultiHopQA, and \textbf{SlideVQA}, a challenging dataset of presentation slides featuring complex diagrams and free-form layouts. Experiments demonstrate that fine-tuned Qwen3-VL-8B-Instruct achieves robust performance, significantly outperforming text-based baselines in scenarios requiring visual layout understanding, while establishing a retriever-agnostic solution for pixel-level interpretable iRAG. Our code is available at https://github.com/PeiYangLiu/CoE.git.

Peiyang Liu, Qiang Yan, Ziqiang Cui et al.

Standard Retrieval-Augmented Generation (RAG) systems predominantly rely on semantic relevance as a proxy for utility. However, this assumption collapses in realistic decision-making scenarios where user queries are laden with cognitive biases, such as false premises or confirmation bias. In such cases, maximizing relevance paradoxically promotes the retrieval of sycophantic evidence that reinforces hallucinations, a critical failure we term the ``Relevance-Robustness Gap''. To bridge this gap, we propose CoRM-RAG (Counterfactual Risk Minimization for RAG), a framework that aligns retrieval with decision safety rather than mere similarity. Grounded in causal intervention, we introduce a Cognitive Perturbation Protocol to simulate user biases during training, which is then distilled into a lightweight Evidence Critic. This scoring module learns to identify documents that possess sufficient evidential strength to steer the model toward correctness despite adversarial query perturbations. Extensive experiments on decision-making benchmarks demonstrate that CoRM-RAG significantly outperforms strong dense retrievers and LLM-based rerankers in adversarial settings, while enabling effective risk-aware abstention through reliable robustness scoring. Our code is available at https://github.com/PeiYangLiu/CoRM-RAG.git.

Jiayi Yang, Sourav Medya, Wei Ye

Graph Neural Networks (GNNs) often assume strong homophily for graph classification, seldom considering heterophily, which means connected nodes tend to have different class labels and dissimilar features. In real-world scenarios, graphs may have nodes that exhibit both homophily and heterophily. Failing to generalize to this setting makes many GNNs underperform in graph classification. In this paper, we address this limitation by identifying three effective designs and develop a novel GNN architecture called IHGNN (short for Incorporating Heterophily into Graph Neural Networks). These designs include the combination of integration and separation of the ego- and neighbor-embeddings of nodes, adaptive aggregation of node embeddings from different layers, and differentiation between different node embeddings for constructing the graph-level readout function. We empirically validate IHGNN on various graph datasets and demonstrate that it outperforms the state-of-the-art GNNs for graph classification.

Yinyi Wei, Shuaipeng Liu, Hailei Yan et al. · pku

With the extensive accumulation of conversational data on the Internet, emotion recognition in conversations (ERC) has received increasing attention. Previous efforts of this task mainly focus on leveraging contextual and speaker-specific features, or integrating heterogeneous external commonsense knowledge. Among them, some heavily rely on future contexts, which, however, are not always available in real-life scenarios. This fact inspires us to generate pseudo future contexts to improve ERC. Specifically, for an utterance, we generate its future context with pre-trained language models, potentially containing extra beneficial knowledge in a conversational form homogeneous with the historical ones. These characteristics make pseudo future contexts easily fused with historical contexts and historical speaker-specific contexts, yielding a conceptually simple framework systematically integrating multi-contexts. Experimental results on four ERC datasets demonstrate our method's superiority. Further in-depth analyses reveal that pseudo future contexts can rival real ones to some extent, especially in relatively context-independent conversations.

Bo Li, Mingda Wang, Shikun Zhang et al.

Instruction tuning relies on large instruction-response corpora whose quality and composition strongly affect downstream performance. We propose Answer Divergence-Guided Selection (ADG), which selects instruction data based on the geometric structure of multi-sample outputs. ADG draws several high-temperature generations per instruction, maps responses into an embedding space, and computes an output divergence score that jointly encodes dispersion magnitude and shape anisotropy. High scores correspond to instructions whose answers are both far apart and multi-modal, rather than clustered paraphrases along a single direction. Across two backbones and three public instruction pools, fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors on six benchmarks spanning reasoning, knowledge, and coding. Analyses further show that both dispersion magnitude and shape anisotropy are necessary, supporting answer divergence as a practical signal for instruction data selection. Code and appendix are included in the supplementary materials.

Bo Li, Wei Ye, Jinglei Zhang et al.

The typical way for relation extraction is fine-tuning large pre-trained language models on task-specific datasets, then selecting the label with the highest probability of the output distribution as the final prediction. However, the usage of the Top-k prediction set for a given sample is commonly overlooked. In this paper, we first reveal that the Top-k prediction set of a given sample contains useful information for predicting the correct label. To effectively utilizes the Top-k prediction set, we propose Label Graph Network with Top-k Prediction Set, termed as KLG. Specifically, for a given sample, we build a label graph to review candidate labels in the Top-k prediction set and learn the connections between them. We also design a dynamic $k$-selection mechanism to learn more powerful and discriminative relation representation. Our experiments show that KLG achieves the best performances on three relation extraction datasets. Moreover, we observe that KLG is more effective in dealing with long-tailed classes.

Zhiyin Yu, Yuchen Mou, Juncheng Yan et al. · pku

Reinforcement learning (RL) has emerged as a powerful post-training paradigm for enhancing the reasoning capabilities of large language models (LLMs). However, reinforcement learning for LLMs faces substantial data scarcity challenges, including the limited availability of high-quality external supervision and the constrained volume of model-generated experience. These limitations make data-efficient reinforcement learning a critical research direction. In this survey, we present the first systematic review of reinforcement learning for LLMs under data scarcity. We propose a bottom-up hierarchical framework built around three complementary perspectives: the data-centric perspective, the training-centric perspective, and the framework-centric perspective. We develop a taxonomy of existing methods, summarize representative approaches in each category, and analyze their strengths and limitations. Our taxonomy aims to provide a clear conceptual foundation for understanding the design space of data-efficient RL for LLMs and to guide researchers working in this emerging area. We hope this survey offers a comprehensive roadmap for future research and inspires new directions toward more efficient and scalable reinforcement learning post-training for LLMs.

Chaoya Jiang, Jia Hongrui, Haiyang Xu et al.

This paper presents MaVEn, an innovative Multi-granularity Visual Encoding framework designed to enhance the capabilities of Multimodal Large Language Models (MLLMs) in multi-image reasoning. Current MLLMs primarily focus on single-image visual understanding, limiting their ability to interpret and integrate information across multiple images. MaVEn addresses this limitation by combining discrete visual symbol sequences, which abstract coarse-grained semantic concepts, with traditional continuous representation sequences that model fine-grained features. This dual approach bridges the semantic gap between visual and textual data, thereby improving the model's ability to process and interpret information from multiple images effectively. Additionally, we design a dynamic reduction mechanism by for long-sequence continuous features to enhance multi-image processing efficiency. Experimental results demonstrate that MaVEn significantly enhances MLLMs' understanding in complex multi-image scenarios, while also improving performance in single-image contexts.

Wei Ye, Hao Tian, Qijun Chen

Graph kernels are conventional methods for computing graph similarities. However, the existing R-convolution graph kernels cannot resolve both of the two challenges: 1) Comparing graphs at multiple different scales, and 2) Considering the distributions of substructures when computing the kernel matrix. These two challenges limit their performances. To mitigate both of the two challenges, we propose a novel graph kernel called the Multi-scale Wasserstein Shortest-Path graph kernel (MWSP), at the heart of which is the multi-scale shortest-path node feature map, of which each element denotes the number of occurrences of the shortest path around a node. The shortest path is represented by the concatenation of all the labels of nodes in it. Since the shortest-path node feature map can only compare graphs at local scales, we incorporate into it the multiple different scales of the graph structure, which are captured by the truncated BFS trees of different depths rooted at each node in a graph. We use the Wasserstein distance to compute the similarity between the multi-scale shortest-path node feature maps of two graphs, considering the distributions of shortest paths. We empirically validate MWSP on various benchmark graph datasets and demonstrate that it achieves state-of-the-art performance on most datasets.

Bo Li, Shikun Zhang, Wei Ye

Instruction-tuned language models increasingly rely on large multi-turn dialogue corpora, but these datasets are often noisy and structurally inconsistent, with topic drift, repetitive chitchat, and mismatched answer formats across turns. We address this from a data selection perspective and propose \textbf{MDS} (Multi-turn Dialogue Selection), a dialogue-level framework that scores whole conversations rather than isolated turns. MDS combines a global coverage stage that performs bin-wise selection in the user-query trajectory space to retain representative yet non-redundant dialogues, with a local structural stage that evaluates within-dialogue reliability through entity-grounded topic grounding and information progress, together with query-answer form consistency for functional alignment. MDS outperforms strong single-turn selectors, dialogue-level LLM scorers, and heuristic baselines on three multi-turn benchmarks and an in-domain Banking test set, achieving the best overall rank across reference-free and reference-based metrics, and is more robust on long conversations under the same training budget. Code and resources are included in the supplementary materials.

Yi Lu, Deyang Kong, Jianing Wang et al.

Recent advances in block diffusion language models have demonstrated competitive performance and strong scalability on reasoning tasks. However, existing BDLMs have limited exploration under the test-time scaling setting and face more severe decoding challenges in long Chain-of-Thought reasoning, particularly in balancing the decoding speed and effectiveness. In this work, we propose a unified framework for test-time scaling in BDLMs that introduces adaptivity in both decoding and block-wise generation. At the decoding level, we propose Bounded Adaptive Confidence Decoding (BACD), a difficulty-aware sampling strategy that dynamically adjusts denoising based on model confidence, accelerating inference while controlling error accumulation. Beyond step-wise adaptivity, we introduce Think Coarse, Critic Fine (TCCF), a test-time scaling paradigm that allocates large block sizes to exploratory reasoning and smaller block sizes to refinement, achieving an effective efficiency-effectiveness balance. To enable efficient and effective decoding with a large block size, we adopt Progressive Block Size Extension, which mitigates performance degradation when scaling block sizes. Extensive experiments show that applying BACD and TCCF to TDAR-8B yields significant improvements over strong baselines such as TraDo-8B (2.26x speedup, +11.2 points on AIME24). These results mark an important step toward unlocking the potential of BDLMs for test-time scaling in complex reasoning tasks.

Chaoya Jiang, Haiyang Xu, Mengfan Dong et al.

Multi-modal large language models (MLLMs) have been shown to efficiently integrate natural language with visual information to handle multi-modal tasks. However, MLLMs still face a fundamental limitation of hallucinations, where they tend to generate erroneous or fabricated information. In this paper, we address hallucinations in MLLMs from a novel perspective of representation learning. We first analyzed the representation distribution of textual and visual tokens in MLLM, revealing two important findings: 1) there is a significant gap between textual and visual representations, indicating unsatisfactory cross-modal representation alignment; 2) representations of texts that contain and do not contain hallucinations are entangled, making it challenging to distinguish them. These two observations inspire us with a simple yet effective method to mitigate hallucinations. Specifically, we introduce contrastive learning into MLLMs and use text with hallucination as hard negative examples, naturally bringing representations of non-hallucinative text and visual samples closer while pushing way representations of non-hallucinating and hallucinative text. We evaluate our method quantitatively and qualitatively, showing its effectiveness in reducing hallucination occurrences and improving performance across multiple benchmarks. On the MMhal-Bench benchmark, our method obtains a 34.66% /29.5% improvement over the baseline MiniGPT-4/LLaVA. Our code is available on https://github.com/X-PLUG/mPLUG-HalOwl/tree/main/hacl.

Chengli Xing, Zhengran Zeng, Gexiang Fang et al. · pku

Recent advancements in code large language models (Code-LLMs) have demonstrated remarkable capabilities in resolving programming related tasks. Meanwhile, researchers have recognized that the quality of pre-training data is crucial for improving LLM performance. However, most of the existing research on pre-training data filtering has focused on general datasets, and little attention for programming datasets. In this paper, we aim to address this gap by exploring the effectiveness of a widely used general data filtering technique, i.e., data-influence-score filtering, within the context of programming-related datasets. To this end, we first introduce a method for calculating data-influence-score for generative programming tasks which involves transforming a variety of downstream coding tasks into validation sets and using the models loss on these sets as a performance metric. Next, we pre-train a Code-LLMs with 1 billion parameters from scratch on a dataset of 100 billion code tokens. Based on it, we conduct an extensive empirical study to evaluate the effectiveness of data-influence-score filtering methods. Specifically, we examine how well this technique improves model performance, investigate how the characteristics of beneficial training data vary across different training stages and programming tasks, and assess the feasibility of prediction-based data-influence-score filtering method. Our findings show that data-influence-score filtering based on validation-set-loss can enhance models programming performance. Moreover, we observe that the criteria of beneficial training data differ significantly across various downstream programming tasks.

Dingyao Yu, Yang An, Wei Ye et al.

Chinese Spelling Correction (CSC) commonly lacks large-scale high-quality corpora, due to the labor-intensive labeling of spelling errors in real-life human writing or typing scenarios. Two data augmentation methods are widely adopted: (1) \textit{Random Replacement} with the guidance of confusion sets and (2) \textit{OCR/ASR-based Generation} that simulates character misusing. However, both methods inevitably introduce noisy data (e.g., false spelling errors), potentially leading to over-correction. By carefully analyzing the two types of corpora, we find that though the latter achieves more robust generalization performance, the former yields better-calibrated CSC models. We then provide a theoretical analysis of this empirical observation, based on which a corpus refining strategy is proposed. Specifically, OCR/ASR-based data samples are fed into a well-calibrated CSC model trained on random replacement-based corpora and then filtered based on prediction confidence. By learning a simple BERT-based model on the refined OCR/ASR-based corpus, we set up impressive state-of-the-art performance on three widely-used benchmarks, while significantly alleviating over-correction (e.g., lowering false positive predictions).

Qiming Xia, Hongwei Lin, Wei Ye et al.

LiDAR-based outdoor 3D object detection has received widespread attention. However, training 3D detectors from the LiDAR point cloud typically relies on expensive bounding box annotations. This paper presents SC3D, an innovative label-efficient method requiring only a single coarse click on the bird's eye view of the 3D point cloud for each frame. A key challenge here is the absence of complete geometric descriptions of the target objects from such simple click annotations. To address this issue, our proposed SC3D adopts a progressive pipeline. Initially, we design a mixed pseudo-label generation module that expands limited click annotations into a mixture of bounding box and semantic mask supervision. Next, we propose a mix-supervised teacher model, enabling the detector to learn mixed supervision information. Finally, we introduce a mixed-supervised student network that leverages the teacher model's generalization ability to learn unclicked instances.Experimental results on the widely used nuScenes and KITTI datasets demonstrate that our SC3D with only coarse clicks, which requires only 0.2% annotation cost, achieves state-of-the-art performance compared to weakly-supervised 3D detection methods.The code will be made publicly available.

Yuncheng Hua, Xiangyu Xi, Zheng Jiang et al.

End-to-end generation-based approaches have been investigated and applied in task-oriented dialogue systems. However, in industrial scenarios, existing methods face the bottlenecks of controllability (e.g., domain-inconsistent responses, repetition problem, etc) and efficiency (e.g., long computation time, etc). In this paper, we propose a task-oriented dialogue system via action-level generation. Specifically, we first construct dialogue actions from large-scale dialogues and represent each natural language (NL) response as a sequence of dialogue actions. Further, we train a Sequence-to-Sequence model which takes the dialogue history as input and outputs sequence of dialogue actions. The generated dialogue actions are transformed into verbal responses. Experimental results show that our light-weighted method achieves competitive performance, and has the advantage of controllability and efficiency.

Xiangyu Xi, Chenxu Lv, Yuncheng Hua et al.

Though widely used in industry, traditional task-oriented dialogue systems suffer from three bottlenecks: (i) difficult ontology construction (e.g., intents and slots); (ii) poor controllability and interpretability; (iii) annotation-hungry. In this paper, we propose to represent utterance with a simpler concept named Dialogue Action, upon which we construct a tree-structured TaskFlow and further build task-oriented chatbot with TaskFlow as core component. A framework is presented to automatically construct TaskFlow from large-scale dialogues and deploy online. Our experiments on real-world after-sale customer services show TaskFlow can satisfy the major needs, as well as reduce the developer burden effectively.

Jiaheng Liu, Dawei Zhu, Zhiqi Bai et al. · pku

Efficient processing of long contexts has been a persistent pursuit in Natural Language Processing. With the growing number of long documents, dialogues, and other textual data, it is important to develop Long Context Language Models (LCLMs) that can process and analyze extensive inputs in an effective and efficient way. In this paper, we present a comprehensive survey on recent advances in long-context modeling for large language models. Our survey is structured around three key aspects: how to obtain effective and efficient LCLMs, how to train and deploy LCLMs efficiently, and how to evaluate and analyze LCLMs comprehensively. For the first aspect, we discuss data strategies, architectural designs, and workflow approaches oriented with long context processing. For the second aspect, we provide a detailed examination of the infrastructure required for LCLM training and inference. For the third aspect, we present evaluation paradigms for long-context comprehension and long-form generation, as well as behavioral analysis and mechanism interpretability of LCLMs. Beyond these three key aspects, we thoroughly explore the diverse application scenarios where existing LCLMs have been deployed and outline promising future development directions. This survey provides an up-to-date review of the literature on long-context LLMs, which we wish to serve as a valuable resource for both researchers and engineers. An associated GitHub repository collecting the latest papers and repos is available at: \href{https://github.com/LCLM-Horizon/A-Comprehensive-Survey-For-Long-Context-Language-Modeling}{\color[RGB]{175,36,67}{LCLM-Horizon}}.

Qi Guo, Jianing Wang, Deyang Kong et al.

Parallel thinking has emerged as a new paradigm for large reasoning models (LRMs) in tackling complex problems. Recent methods leverage Reinforcement Learning (RL) to enhance parallel thinking, aiming to address the limitations in computational resources and effectiveness encountered with supervised fine-tuning. However, most existing studies primarily focus on optimizing the aggregation phase, with limited attention to the path exploration stage. In this paper, we theoretically analyze the optimization of parallel thinking under the Reinforcement Learning with Verifiable Rewards (RLVR) setting, and identify that the mutual information bottleneck among exploration paths fundamentally restricts overall performance. To address this, we propose Outline-Guided Path Exploration (OPE), which explicitly partitions the solution space by generating diverse reasoning outlines prior to parallel path reasoning, thereby reducing information redundancy and improving the diversity of information captured across exploration paths. We implement OPE with an iterative RL strategy that optimizes outline planning and outline-guided reasoning independently. Extensive experiments across multiple challenging mathematical benchmarks demonstrate that OPE effectively improves reasoning performance in different aggregation strategies, enabling LRMs to more reliably discover correct solutions.

Yutao Mou, Zhangchi Xue, Lijun Li et al.

While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains under-explored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action-attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, we introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65 percent on average and improves benign task completion by approximately 10 percent under prompt injection attacks.

Yutao Mou, Shikun Zhang, Wei Ye

Ensuring the safety of large language model (LLM) applications is essential for developing trustworthy artificial intelligence. Current LLM safety benchmarks have two limitations. First, they focus solely on either discriminative or generative evaluation paradigms while ignoring their interconnection. Second, they rely on standardized inputs, overlooking the effects of widespread prompting techniques, such as system prompts, few-shot demonstrations, and chain-of-thought prompting. To overcome these issues, we developed SG-Bench, a novel benchmark to assess the generalization of LLM safety across various tasks and prompt types. This benchmark integrates both generative and discriminative evaluation tasks and includes extended data to examine the impact of prompt engineering and jailbreak on LLM safety. Our assessment of 3 advanced proprietary LLMs and 10 open-source LLMs with the benchmark reveals that most LLMs perform worse on discriminative tasks than generative ones, and are highly susceptible to prompts, indicating poor generalization in safety alignment. We also explain these findings quantitatively and qualitatively to provide insights for future research.

Chaoya Jiang, Hongrui Jia, Wei Ye et al.

Large Vision Language Models exhibit remarkable capabilities but struggle with hallucinations inconsistencies between images and their descriptions. Previous hallucination evaluation studies on LVLMs have identified hallucinations in terms of objects, attributes, and relations but overlooked complex hallucinations that create an entire narrative around a fictional entity. In this paper, we introduce a refined taxonomy of hallucinations, featuring a new category: Event Hallucination. We then utilize advanced LLMs to generate and filter fine grained hallucinatory data consisting of various types of hallucinations, with a particular focus on event hallucinations, laying the groundwork for integrating discriminative and generative evaluation methods within our universal evaluation framework. The proposed benchmark distinctively assesses LVLMs ability to tackle a broad spectrum of hallucinations, making it a reliable and comprehensive tool for gauging LVLMs efficacy in handling hallucinations. We will release our code and data.

Linyi Yang, Shuibai Zhang, Zhuohao Yu et al. · pku

Large Language Models (LLMs) exhibit emerging in-context learning abilities through prompt engineering. The recent progress in large-scale generative models has further expanded their use in real-world language applications. However, the critical challenge of improving the generalizability and factuality of LLMs in natural language understanding and question answering remains under-explored. While previous in-context learning research has focused on enhancing models to adhere to users' specific instructions and quality expectations, and to avoid undesired outputs, little to no work has explored the use of task-Specific fine-tuned Language Models (SLMs) to improve LLMs' in-context learning during the inference stage. Our primary contribution is the establishment of a simple yet effective framework that enhances the reliability of LLMs as it: 1) generalizes out-of-distribution data, 2) elucidates how LLMs benefit from discriminative models, and 3) minimizes hallucinations in generative tasks. Using our proposed plug-in method, enhanced versions of Llama 2 and ChatGPT surpass their original versions regarding generalizability and factuality. We offer a comprehensive suite of resources, including 16 curated datasets, prompts, model checkpoints, and LLM outputs across 9 distinct tasks. The code and data are released at: https://github.com/YangLinyi/Supervised-Knowledge-Makes-Large-Language-Models-Better-In-context-Learners. Our empirical analysis sheds light on the advantages of incorporating discriminative models into LLMs and highlights the potential of our methodology in fostering more reliable LLMs.

Peiyang Liu, Zhirui Chen, Xi Wang et al.

Monte Carlo Tree Search (MCTS) has been widely used for automated reasoning data exploration, but current supervision extraction methods remain inefficient. Standard approaches retain only the single highest-reward trajectory, discarding the comparative signals present in the many explored paths. Here we introduce \textbf{Contrastive Reasoning Path Synthesis (CRPS)}, a framework that transforms supervision extraction from a filtering process into a synthesis procedure. CRPS uses a structured reflective process to analyze the differences between high- and low-quality search trajectories, extracting explicit information about strategic pivots and local failure modes. These insights guide the synthesis of reasoning chains that incorporate success patterns while avoiding identified pitfalls. We show empirically that models fine-tuned on just 60K CRPS-synthesized examples match or exceed the performance of baselines trained on 590K examples derived from standard rejection sampling, a 20$\times$ reduction in dataset size. Furthermore, CRPS improves generalization on out-of-domain benchmarks, demonstrating that learning from the contrast between success and failure produces more transferable reasoning capabilities than learning from success alone.

Bo Li, Tian Tian, Zhenghua Xu et al.

Dynamic retrieval-augmented generation (RAG) allows large language models (LLMs) to fetch external knowledge on demand, offering greater adaptability than static RAG. A central challenge in this setting lies in determining the optimal timing for retrieval. Existing methods often trigger retrieval based on low token-level confidence, which may lead to delayed intervention after errors have already propagated. We introduce Entropy-Trend Constraint (ETC), a training-free method that determines optimal retrieval timing by modeling the dynamics of token-level uncertainty. Specifically, ETC utilizes first- and second-order differences of the entropy sequence to detect emerging uncertainty trends, enabling earlier and more precise retrieval. Experiments on six QA benchmarks with three LLM backbones demonstrate that ETC consistently outperforms strong baselines while reducing retrieval frequency. ETC is particularly effective in domain-specific scenarios, exhibiting robust generalization capabilities. Ablation studies and qualitative analyses further confirm that trend-aware uncertainty modeling yields more effective retrieval timing. The method is plug-and-play, model-agnostic, and readily integrable into existing decoding pipelines. Implementation code is included in the supplementary materials.