Zhaowei Wang

Changjian Chen, Yukai Guo, Fengyuan Tian et al. · tsinghua

Existing model evaluation tools mainly focus on evaluating classification models, leaving a gap in evaluating more complex models, such as object detection. In this paper, we develop an open-source visual analysis tool, Uni-Evaluator, to support a unified model evaluation for classification, object detection, and instance segmentation in computer vision. The key idea behind our method is to formulate both discrete and continuous predictions in different tasks as unified probability distributions. Based on these distributions, we develop 1) a matrix-based visualization to provide an overview of model performance; 2) a table visualization to identify the problematic data subsets where the model performs poorly; 3) a grid visualization to display the samples of interest. These visualizations work together to facilitate the model evaluation from a global overview to individual samples. Two case studies demonstrate the effectiveness of Uni-Evaluator in evaluating model performance and making informed improvements.

Kimi Team, Tongtong Bai, Yifan Bai et al.

We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

Zhaowei Wang, Hongming Zhang, Tianqing Fang et al. · tencent-ai

In this paper, we propose a new task of sub-event generation for an unseen process to evaluate the understanding of the coherence of sub-event actions and objects. To solve the problem, we design SubeventWriter, a sub-event sequence generation framework with a coherence controller. Given an unseen process, the framework can iteratively construct the sub-event sequence by generating one sub-event at each iteration. We also design a very effective coherence controller to decode more coherent sub-events. As our extensive experiments and analysis indicate, SubeventWriter can generate more reliable and meaningful sub-event sequences for unseen processes.

Tianqing Fang, Quyet V. Do, Zihao Zheng et al. · tencent-ai

Commonsense Knowledge Bases (CSKB) Population, which aims at automatically expanding knowledge in CSKBs with external resources, is an important yet hard task in NLP. Fang et al. (2021a) proposed a CSKB Population (CKBP) framework with an evaluation set CKBP v1. However, CKBP v1 relies on crowdsourced annotations that suffer from a considerable number of mislabeled answers, and the evaluationset lacks alignment with the external knowledge source due to random sampling. In this paper, we introduce CKBP v2, a new high-quality CSKB Population evaluation set that addresses the two aforementioned issues by employing domain experts as annotators and incorporating diversified adversarial samples to make the evaluation data more representative. We show that CKBP v2 serves as a challenging and representative evaluation dataset for the CSKB Population task, while its development set aids in selecting a population model that leads to improved knowledge acquisition for downstream commonsense reasoning. A better population model can also help acquire more informative commonsense knowledge as additional supervision signals for both generative commonsense inference and zero-shot commonsense question answering. Specifically, the question-answering model based on DeBERTa-v3-large (He et al., 2023b) even outperforms powerful large language models in a zero-shot setting, including ChatGPT and GPT-3.5.

Sehyun Choi, Tianqing Fang, Zhaowei Wang et al. · tencent-ai

Large Language Models (LLMs) have demonstrated remarkable human-level natural language generation capabilities. However, their potential to generate misinformation, often called the hallucination problem, poses a significant risk to their deployment. A common approach to address this issue is to retrieve relevant knowledge and fine-tune the LLM with the knowledge in its input. Unfortunately, this method incurs high training costs and may cause catastrophic forgetting for multi-tasking models. To overcome these limitations, we propose a knowledge-constrained decoding method called KCTS (Knowledge-Constrained Tree Search), which guides a frozen LM to generate text aligned with the reference knowledge at each decoding step using a knowledge classifier score and MCTS (Monte-Carlo Tree Search). To adapt the sequence-level knowledge classifier to token-level guidance, we also propose a novel token-level hallucination detection method called RIPA (Reward Inflection Point Approximation). Our empirical results on knowledge-grounded dialogue and abstractive summarization demonstrate the strength of KCTS as a plug-and-play, model-agnostic decoding method that can effectively reduce hallucinations in natural language generation.

Haobo Li, Zhaowei Wang, Jiachen Wang et al.

Forecasting weather and climate events is crucial for making appropriate measures to mitigate environmental hazards and minimize losses. However, existing environmental forecasting research focuses narrowly on predicting numerical meteorological variables (e.g., temperature), neglecting the translation of these variables into actionable textual narratives of events and their consequences. To bridge this gap, we proposed Weather and Climate Event Forecasting (WCEF), a new task that leverages numerical meteorological raster data and textual event data to predict weather and climate events. This task is challenging to accomplish due to difficulties in aligning multimodal data and the lack of supervised datasets. To address these challenges, we present CLLMate, the first multimodal dataset for WCEF, using 26,156 environmental news articles aligned with ERA5 reanalysis data. We systematically benchmark 23 existing MLLMs on CLLMate, including closed-source, open-source, and our fine-tuned models. Our experiments reveal the advantages and limitations of existing MLLMs and the value of CLLMate for the training and benchmarking of the WCEF task.

Zhaowei Wang, Haochen Shi, Weiqi Wang et al. · tencent-ai

Cognitive research indicates that abstraction ability is essential in human intelligence, which remains under-explored in language models. In this paper, we present AbsPyramid, a unified entailment graph of 221K textual descriptions of abstraction knowledge. While existing resources only touch nouns or verbs within simplified events or specific domains, AbsPyramid collects abstract knowledge for three components of diverse events to comprehensively evaluate the abstraction ability of language models in the open domain. Experimental results demonstrate that current LLMs face challenges comprehending abstraction knowledge in zero-shot and few-shot settings. By training on our rich abstraction knowledge, we find LLMs can acquire basic abstraction abilities and generalize to unseen events. In the meantime, we empirically show that our benchmark is comprehensive to enhance LLMs across two previous abstraction tasks.

Yiming Du, Baojun Wang, Yifan Xiang et al.

Temporal reasoning over long, multi-session dialogues is a critical capability for conversational agents. However, existing works and our pilot study have shown that as dialogue histories grow in length and accumulate noise, current long-context models struggle to accurately identify temporally pertinent information, significantly impairing reasoning performance. To address this, we introduce Memory-T1, a framework that learns a time-aware memory selection policy using reinforcement learning (RL). It employs a coarse-to-fine strategy, first pruning the dialogue history into a candidate set using temporal and relevance filters, followed by an RL agent that selects the precise evidence sessions. The RL training is guided by a multi-level reward function optimizing (i) answer accuracy, (ii) evidence grounding, and (iii) temporal consistency. In particular, the temporal consistency reward provides a dense signal by evaluating alignment with the query time scope at both the session-level (chronological proximity) and the utterance-level (chronological fidelity), enabling the agent to resolve subtle chronological ambiguities. On the Time-Dialog benchmark, Memory-T1 boosts a 7B model to an overall score of 67.0\%, establishing a new state-of-the-art performance for open-source models and outperforming a 14B baseline by 10.2\%. Ablation studies show temporal consistency and evidence grounding rewards jointly contribute to a 15.0\% performance gain. Moreover, Memory-T1 maintains robustness up to 128k tokens, where baseline models collapse, proving effectiveness against noise in extensive dialogue histories. The code and datasets are publicly available at https://github.com/Elvin-Yiming-Du/Memory-T1/

Zhaowei Wang, Zhisheng Yin, Xiucheng Wang et al.

In Space-air-ground integrated networks (SAGIN), the inherent openness and extensive broadcast coverage expose these networks to significant eavesdropping threats. Considering the inherent co-channel interference due to spectrum sharing among multi-tier access networks in SAGIN, it can be leveraged to assist the physical layer security among heterogeneous transmissions. However, it is challenging to conduct a secrecy-oriented access strategy due to both heterogeneous resources and different eavesdropping models. In this paper, we explore secure access selection for a scenario involving multi-mode users capable of accessing satellites, unmanned aerial vehicles, or base stations in the presence of eavesdroppers. Particularly, we propose a Q-network approximation based deep learning approach for selecting the optimal access strategy for maximizing the sum secrecy rate. Meanwhile, the power optimization is also carried out by an unsupervised learning approach to improve the secrecy performance. Remarkably, two neural networks are trained by unsupervised learning and Q-network approximation which are both label-free methods without knowing the optimal solution as labels. Numerical results verify the efficiency of our proposed power optimization approach and access strategy, leading to enhanced secure transmission performance.

Aaron Branson Cigres Li, Zhaowei Wang, Yu Zhao et al.

Large vision-language models increasingly rely on long-context modeling to reason over documents, hour-level videos, and long-horizon agent trajectories, requiring them to locate relevant evidence across interleaved text and images. Prior work has studied this behavior using retrieval heads in large language models, but its copy-based criterion does not directly apply when evidence appears in images. We introduce a multimodal retrieval head detection method that scores attention from question tokens to textual or visual evidence. With this method, we show that multimodal retrieval heads are sparse, intrinsic, and causally important: only 4.4-10.2% of attention heads account for 50% of the positive retrieval-score mass, and masking the top-5% selected heads drops MMLongBench-Doc from 48.2% to 5.7% and SlideVQA from 71.2% to 8.9%, while random-head masking is far less damaging. Further analysis shows that these heads are partly shared across modalities yet remain dynamic within each modality, with image retrieval heads changing more than text retrieval heads as context length and haystack modality change. Without further training, we find that these heads can also be used directly to rank visually rich documents: on MMDocIR, Qwen3-VL-8B selected-head scoring improves Recall@1 by 7.7/7.4 macro/micro points for page retrieval and 6.3/6.8 points for layout retrieval over the strongest reported baseline.

Xiyu Ren, Zhaowei Wang, Yiming Du et al.

Memory is essential for large vision-language models (LVLMs) to handle long, multimodal interactions, with two method directions providing this capability: long-context LVLMs and memory-augmented agents. However, no existing benchmark conducts a systematic comparison of the two on questions that genuinely require multimodal evidence. To close this gap, we introduce MEMLENS, a comprehensive benchmark for memory in multimodal multi-session conversations, comprising 789 questions across five memory abilities (information extraction, multi-session reasoning, temporal reasoning, knowledge update, and answer refusal) at four standard context lengths (32K-256K tokens) under a cross-modal token-counting scheme. An image-ablation study confirms that solving MEMLENS requires visual evidence: removing evidence images drops two frontier LVLMs below 2% accuracy on the 80.4% of questions whose evidence includes images. Evaluating 27 LVLMs and 7 memory-augmented agents, we find that long-context LVLMs achieve high short-context accuracy through direct visual grounding but degrade as conversations grow, whereas memory agents are length-stable but lose visual fidelity under storage-time compression. Multi-session reasoning caps most systems below 30%, and neither approach alone solves the task. These results motivate hybrid architectures that combine long-context attention with structured multimodal retrieval. Our code is available at https://github.com/xrenaf/MEMLENS.

Zhaowei Wang, Ying Hao, Hao Wei et al.

Recent advancements in text-to-image diffusion models have significantly transformed visual content generation, yet their application in specialized fields such as interior design remains underexplored. In this paper, we present RoomDiffusion, a pioneering diffusion model meticulously tailored for the interior design industry. To begin with, we build from scratch a whole data pipeline to update and evaluate data for iterative model optimization. Subsequently, techniques such as multiaspect training, multi-stage fine-tune and model fusion are applied to enhance both the visual appeal and precision of the generated results. Lastly, leveraging the latent consistency Distillation method, we distill and expedite the model for optimal efficiency. Unlike existing models optimized for general scenarios, RoomDiffusion addresses specific challenges in interior design, such as lack of fashion, high furniture duplication rate, and inaccurate style. Through our holistic human evaluation protocol with more than 20 professional human evaluators, RoomDiffusion demonstrates industry-leading performance in terms of aesthetics, accuracy, and efficiency, surpassing all existing open source models such as stable diffusion and SDXL.

Junde Wu, Huihui Fang, Fangxin Shang et al.

Clinically, the accurate annotation of lesions/tissues can significantly facilitate the disease diagnosis. For example, the segmentation of optic disc/cup (OD/OC) on fundus image would facilitate the glaucoma diagnosis, the segmentation of skin lesions on dermoscopic images is helpful to the melanoma diagnosis, etc. With the advancement of deep learning techniques, a wide range of methods proved the lesions/tissues segmentation can also facilitate the automated disease diagnosis models. However, existing methods are limited in the sense that they can only capture static regional correlations in the images. Inspired by the global and dynamic nature of Vision Transformer, in this paper, we propose Segmentation-Assisted diagnosis Transformer (SeATrans) to transfer the segmentation knowledge to the disease diagnosis network. Specifically, we first propose an asymmetric multi-scale interaction strategy to correlate each single low-level diagnosis feature with multi-scale segmentation features. Then, an effective strategy called SeA-block is adopted to vitalize diagnosis feature via correlated segmentation features. To model the segmentation-diagnosis interaction, SeA-block first embeds the diagnosis feature based on the segmentation information via the encoder, and then transfers the embedding back to the diagnosis feature space by a decoder. Experimental results demonstrate that SeATrans surpasses a wide range of state-of-the-art (SOTA) segmentation-assisted diagnosis methods on several disease diagnosis tasks.

Junde Wu, Huihui Fang, Fangxin Shang et al.

The segmentation of optic disc(OD) and optic cup(OC) from fundus images is an important fundamental task for glaucoma diagnosis. In the clinical practice, it is often necessary to collect opinions from multiple experts to obtain the final OD/OC annotation. This clinical routine helps to mitigate the individual bias. But when data is multiply annotated, standard deep learning models will be inapplicable. In this paper, we propose a novel neural network framework to learn OD/OC segmentation from multi-rater annotations. The segmentation results are self-calibrated through the iterative optimization of multi-rater expertness estimation and calibrated OD/OC segmentation. In this way, the proposed method can realize a mutual improvement of both tasks and finally obtain a refined segmentation result. Specifically, we propose Diverging Model(DivM) and Converging Model(ConM) to process the two tasks respectively. ConM segments the raw image based on the multi-rater expertness map provided by DivM. DivM generates multi-rater expertness map from the segmentation mask provided by ConM. The experiment results show that by recurrently running ConM and DivM, the results can be self-calibrated so as to outperform a range of state-of-the-art(SOTA) multi-rater segmentation methods.

Zheye Deng, Weiqi Wang, Zhaowei Wang et al.

Commonsense Knowledge Graphs (CSKGs) are crucial for commonsense reasoning, yet constructing them through human annotations can be costly. As a result, various automatic methods have been proposed to construct CSKG with larger semantic coverage. However, these unsupervised approaches introduce spurious noise that can lower the quality of the resulting CSKG, which cannot be tackled easily by existing denoising algorithms due to the unique characteristics of nodes and structures in CSKGs. To address this issue, we propose Gold (Global and Local-aware Denoising), a denoising framework for CSKGs that incorporates entity semantic information, global rules, and local structural information from the CSKG. Experiment results demonstrate that Gold outperforms all baseline methods in noise detection tasks on synthetic noisy CSKG benchmarks. Furthermore, we show that denoising a real-world CSKG is effective and even benefits the downstream zero-shot commonsense question-answering task.

Qing Zong, Zhaowei Wang, Baixuan Xu et al.

A main goal of Argument Mining (AM) is to analyze an author's stance. Unlike previous AM datasets focusing only on text, the shared task at the 10th Workshop on Argument Mining introduces a dataset including both text and images. Importantly, these images contain both visual elements and optical characters. Our new framework, TILFA (A Unified Framework for Text, Image, and Layout Fusion in Argument Mining), is designed to handle this mixed data. It excels at not only understanding text but also detecting optical characters and recognizing layout details in images. Our model significantly outperforms existing baselines, earning our team, KnowComp, the 1st place in the leaderboard of Argumentative Stance Classification subtask in this shared task.

Qiaolong Cai, Zhaowei Wang, Shizhe Diao et al.

With the increasing popularity of large language models (LLMs), reasoning on basic graph algorithm problems is an essential intermediate step in assessing their abilities to process and infer complex graph reasoning tasks. Existing methods usually convert graph-structured data to textual descriptions and then use LLMs for reasoning and computation. However, LLMs often produce computation errors on arithmetic parts in basic graph algorithm problems, such as counting number of edges. In addition, they struggle to control or understand the output of the reasoning process, raising concerns about whether LLMs are simply guessing. In this paper, we introduce CodeGraph, a method that encodes graph problem solutions as code. The methods solve new graph problems by learning from exemplars, generating programs, and executing them via a program interpreter. Using the few-shot setting, we evaluate CodeGraph with the base LLM being GPT-3.5 Turbo, Llama3-70B Instruct, Mixtral-8x22B Instruct, and Mixtral-8x7B Instruct. Experimental results on six tasks with six graph encoding methods in the GraphQA dataset demonstrate that CodeGraph can boost performance on graph reasoning tasks inside LLMs by 1.3% to 58.6%, depending on the task. Compared to the existing methods, CodeGraph demonstrates strong performance on arithmetic problems in graph tasks and offers a more controllable and interpretable approach to the reasoning process.

Weiqi Wang, Tianqing Fang, Chunyang Li et al. · tencent-ai

The sequential process of conceptualization and instantiation is essential to generalizable commonsense reasoning as it allows the application of existing knowledge to unfamiliar scenarios. However, existing works tend to undervalue the step of instantiation and heavily rely on pre-built concept taxonomies and human annotations to collect both types of knowledge, resulting in a lack of instantiated knowledge to complete reasoning, high cost, and limited scalability. To tackle these challenges, we introduce CANDLE, a distillation framework that iteratively performs contextualized conceptualization and instantiation over commonsense knowledge bases by instructing large language models to generate both types of knowledge with critic filtering. By applying CANDLE to ATOMIC, we construct a comprehensive knowledge base comprising six million conceptualizations and instantiated commonsense knowledge triples. Both types of knowledge are firmly rooted in the original ATOMIC dataset, and intrinsic evaluations demonstrate their exceptional quality and diversity. Empirical results indicate that distilling CANDLE on student models provides benefits across four downstream tasks. Our code, data, and models are publicly available at https://github.com/HKUST-KnowComp/CANDLE.

Zhaowei Wang

Legal case retrieval, which aims to retrieve relevant cases given a query case, plays an essential role in the legal system. While recent research efforts improve the performance of traditional ad-hoc retrieval models, legal case retrieval is still challenging since queries are legal cases, which contain hundreds of tokens. Legal cases are much longer and more complicated than keywords queries. Apart from that, the definition of legal relevance is beyond the general definition. In addition to general topical relevance, the relevant cases also involve similar situations and legal elements, which can support the judgment of the current case. In this paper, we propose an interaction-focused network for legal case retrieval with a multi-view contrastive learning objective. The contrastive learning views, including case-view and element-view, aim to overcome the above challenges. The case-view contrastive learning minimizes the hidden space distance between relevant legal case representations produced by a pre-trained language model (PLM) encoder. The element-view builds positive and negative instances by changing legal elements of cases to help the network better compute legal relevance. To achieve this, we employ a legal element knowledge-aware indicator to detect legal elements of cases. We conduct extensive experiments on the benchmark of relevant case retrieval. Evaluation results indicate our proposed method obtains significant improvement over the existing methods.

Yiming Du, Wenyu Huang, Danna Zheng et al.

Memory is a fundamental component of AI systems, underpinning large language models (LLMs)-based agents. While prior surveys have focused on memory applications with LLMs (e.g., enabling personalized memory in conversational agents), they often overlook the atomic operations that underlie memory dynamics. In this survey, we first categorize memory representations into parametric and contextual forms, and then introduce six fundamental memory operations: Consolidation, Updating, Indexing, Forgetting, Retrieval, and Compression. We map these operations to the most relevant research topics across long-term, long-context, parametric modification, and multi-source memory. By reframing memory systems through the lens of atomic operations and representation types, this survey provides a structured and dynamic perspective on research, benchmark datasets, and tools related to memory in AI, clarifying the functional interplay in LLMs based agents while outlining promising directions for future research\footnote{The paper list, datasets, methods and tools are available at \href{https://github.com/Elvin-Yiming-Du/Survey_Memory_in_AI}{https://github.com/Elvin-Yiming-Du/Survey\_Memory\_in\_AI}.}.

Zihao Zheng, Zhaowei Wang, Qing Zong et al.

Dialogical Argument Mining(DialAM) is an important branch of Argument Mining(AM). DialAM-2024 is a shared task focusing on dialogical argument mining, which requires us to identify argumentative relations and illocutionary relations among proposition nodes and locution nodes. To accomplish this, we propose a two-stage pipeline, which includes the Two-Step S-Node Prediction Model in Stage 1 and the YA-Node Prediction Model in Stage 2. We also augment the training data in both stages and introduce context in Stage 2. We successfully completed the task and achieved good results. Our team Pokemon ranked 1st in the ARI Focused score and 4th in the Global Focused score.

Zhaowei Wang, Wenhao Yu, Xiyu Ren et al.

The rapid extension of context windows in large vision-language models has given rise to long-context vision-language models (LCVLMs), which are capable of handling hundreds of images with interleaved text tokens in a single forward pass. In this work, we introduce MMLongBench, the first benchmark covering a diverse set of long-context vision-language tasks, to evaluate LCVLMs effectively and thoroughly. MMLongBench is composed of 13,331 examples spanning five different categories of downstream tasks, such as Visual RAG and Many-Shot ICL. It also provides broad coverage of image types, including various natural and synthetic images. To assess the robustness of the models to different input lengths, all examples are delivered at five standardized input lengths (8K-128K tokens) via a cross-modal tokenization scheme that combines vision patches and text tokens. Through a thorough benchmarking of 46 closed-source and open-source LCVLMs, we provide a comprehensive analysis of the current models' vision-language long-context ability. Our results show that: i) performance on a single task is a weak proxy for overall long-context capability; ii) both closed-source and open-source models face challenges in long-context vision-language tasks, indicating substantial room for future improvement; iii) models with stronger reasoning ability tend to exhibit better long-context performance. By offering wide task coverage, various image types, and rigorous length control, MMLongBench provides the missing foundation for diagnosing and advancing the next generation of LCVLMs.

Zhaowei Wang, Lishu Luo, Haodong Duan et al.

Long-context modeling is becoming a core capability of modern large vision-language models (LVLMs), enabling sustained context management across long-document understanding, video analysis, and multi-turn tool use in agentic workflows. Yet practical training recipes remain insufficiently explored, particularly for designing and balancing long-context data mixtures. In this work, we present a systematic study of long-context continued pre-training for LVLMs, extending a 7B model from 32K to 128K context with extensive ablations on long-document data. We first show that long-document VQA is substantially more effective than OCR transcription. Building on this observation, our ablations further yield three key findings: i) for sequence-length distribution, balanced data outperforms target-length-focused data (e.g., 128K), suggesting that long-context ability requires generalizable key-information retrieval across various lengths and positions; ii) retrieval remains the primary bottleneck, favoring retrieval-heavy mixtures with modest reasoning data for task diversity; and iii) pure long-document VQA largely preserves short-context capabilities, suggesting that instruction-formatted long data reduces the need for short-data mixing. Based on these findings, we introduce MMProLong, obtained by long-context continued pre-training from Qwen2.5-VL-7B with only a 5B-token budget. MMProLong improves long-document VQA scores by 7.1% and maintains strong performance at 256K and 512K contexts beyond its 128K training window, without additional training. It further generalizes to webpage-based multimodal needle retrieval, long-context vision-text compression, and long-video understanding without task-specific supervision. Overall, our study establishes a practical LongPT recipe and an empirical foundation for advancing long-context vision-language models.

Xingyue Huang, Rishabh, Gregor Franke et al.

Recent advances in Large Language Models (LLMs) have shown that their reasoning capabilities can be significantly improved through Reinforcement Learning with Verifiable Reward (RLVR), particularly in domains like mathematics and programming, where ground-truth correctness can be automatically evaluated. However, extending this success to other reasoning-intensive domains remains challenging due to the scarcity of high-quality, verifiable datasets and the high cost of human supervision. In this work, we introduce the Loong Project: an open-source framework for scalable synthetic data generation and verification across a diverse range of reasoning-intensive domains. The framework consists of two key components: (1) LoongBench, a curated seed dataset containing 8,729 human-vetted examples across 12 domains (e.g., Advanced Mathematics, Chemistry, Logic), each paired with executable code and rich metadata; and (2) LoongEnv, a modular synthetic data generation environment that supports multiple prompting strategies to produce new question-answer-code triples. Together, these components form an agent-environment loop that enables reinforcement learning, where an LLM-based agent is rewarded for generating Chain-of-Thought (CoT) solutions that align with code-executed answers. Empirically, we benchmark LoongBench on a broad suite of both open-source and proprietary LLMs to evaluate domain coverage and reveal performance bottlenecks. In addition, we conduct a comprehensive analysis of synthetic data generated by LoongEnv, examining correctness, difficulty, and diversity. Code and documentation are available at https://github.com/camel-ai/loong.

Haobo Li, Eunseo Jung, Zixin Chen et al.

Multimodal time series forecasting is foundational in various fields, such as utilizing satellite imagery and numerical data for predicting typhoons in climate science. However, existing multimodal approaches primarily focus on utilizing text data to help time series forecasting, leaving the visual data in existing time series datasets untouched. Furthermore, it is challenging for models to effectively capture the physical information embedded in visual data, such as satellite imagery's temporal and geospatial context, which extends beyond images themselves. To address this gap, we propose physics-informed positional encoding (PIPE), a lightweight method that embeds physical information into vision language models (VLMs). PIPE introduces two key innovations: (1) a physics-informed positional indexing scheme for mapping physics to positional IDs, and (2) a variant-frequency positional encoding mechanism for encoding frequency information of physical variables and sequential order of tokens within the embedding space. By preserving both the physical information and sequential order information, PIPE significantly improves multimodal alignment and forecasting accuracy. Through the experiments on the most representative and the largest open-sourced satellite image dataset, PIPE achieves state-of-the-art performance in both deep learning forecasting and climate domain methods, demonstrating superiority across benchmarks, including a 12% improvement in typhoon intensity forecasting over prior works. Our code is provided in the supplementary material.

Quyet V. Do, Tianqing Fang, Shizhe Diao et al.

Reasoning over Commonsense Knowledge Bases (CSKB), i.e. CSKB reasoning, has been explored as a way to acquire new commonsense knowledge based on reference knowledge in the original CSKBs and external prior knowledge. Despite the advancement of Large Language Models (LLM) and prompt engineering techniques in various reasoning tasks, they still struggle to deal with CSKB reasoning. One of the problems is that it is hard for them to acquire explicit relational constraints in CSKBs from only in-context exemplars, due to a lack of symbolic reasoning capabilities (Bengio et al., 2021). To this end, we proposed **ConstraintChecker**, a plugin over prompting techniques to provide and check explicit constraints. When considering a new knowledge instance, ConstraintChecker employs a rule-based module to produce a list of constraints, then it uses a zero-shot learning module to check whether this knowledge instance satisfies all constraints. The acquired constraint-checking result is then aggregated with the output of the main prompting technique to produce the final output. Experimental results on CSKB Reasoning benchmarks demonstrate the effectiveness of our method by bringing consistent improvements over all prompting methods. Codes and data are available at \url{https://github.com/HKUST-KnowComp/ConstraintChecker}.

Zhaowei Wang, Wei Fan, Qing Zong et al. · tencent-ai

Abstraction ability is crucial in human intelligence, which can also benefit various tasks in NLP study. Existing work shows that LLMs are deficient in abstract ability, and how to improve it remains unexplored. In this work, we design the framework AbsInstruct to enhance LLMs' abstraction ability through instruction tuning. The framework builds instructions with in-depth explanations to assist LLMs in capturing the underlying rationale of abstraction. Meanwhile, we introduce a plausibility estimator to select instructions that are more consistent with the abstraction knowledge of LLMs to be aligned. Then, our framework combines abstraction instructions with general-purpose ones to build a hybrid dataset. Extensive experiments and analyses demonstrate that our framework can considerably enhance LLMs' abstraction ability with strong generalization performance while maintaining their general instruction-following abilities.

Qing Zong, Zhaowei Wang, Tianshi Zheng et al.

The rapid development of LLMs has sparked extensive research into their factual knowledge. Current works find that LLMs fall short on questions around low-frequency entities. However, such proofs are unreliable since the questions can differ not only in entity frequency but also in difficulty themselves. So we introduce ComparisonQA benchmark, containing 283K abstract questions, each instantiated by a pair of high-frequency and low-frequency entities. It ensures a controllable comparison to study the role of knowledge frequency in the performance of LLMs. Because the difference between such a pair is only the entity with different frequencies. In addition, we use both correctness and uncertainty to develop a two-round method to evaluate LLMs' knowledge robustness. It aims to avoid possible semantic shortcuts which is a serious problem of current QA study. Experiments reveal that LLMs, including GPT-4o, exhibit particularly low robustness regarding low-frequency knowledge. Besides, we find that uncertainty can be used to effectively identify high-quality and shortcut-free questions while maintaining the data size. Based on this, we propose an automatic method to select such questions to form a subset called ComparisonQA-Hard, containing only hard low-frequency questions.

Kaiqiao Han, Tianqing Fang, Zhaowei Wang et al. · tencent-ai

While Large Language Models (LLMs) have showcased remarkable proficiency in reasoning, there is still a concern about hallucinations and unreliable reasoning issues due to semantic associations and superficial logical chains. To evaluate the extent to which LLMs perform robust reasoning instead of relying on superficial logical chains, we propose a new evaluation dataset, the Concept-Reversed Winograd Schema Challenge (CR-WSC), based on the famous Winograd Schema Challenge (WSC) dataset. By simply reversing the concepts to those that are more associated with the wrong answer, we find that the performance of LLMs drops significantly despite the rationale of reasoning remaining the same. Furthermore, we propose Abstraction-of-Thought (AoT), a novel prompt method for recovering adversarial cases to normal cases using conceptual abstraction to improve LLMs' robustness and consistency in reasoning, as demonstrated by experiments on CR-WSC.

Liang Cheng, Tianyi Li, Zhaowei Wang et al.

LLMs are often claimed to be capable of Natural Language Inference (NLI), which is widely regarded as a cornerstone of more complex forms of reasoning. However, recent works show that LLMs still suffer from hallucinations in NLI due to attestation bias, where LLMs overly rely on propositional memory to build shortcuts. To solve the issue, we design an unsupervised framework to construct counterfactual reasoning data and fine-tune LLMs to reduce attestation bias. To measure bias reduction, we build bias-adversarial variants of NLI datasets with randomly replaced predicates in premises while keeping hypotheses unchanged. Extensive evaluations show that our framework can significantly reduce hallucinations from attestation bias. Then, we further evaluate LLMs fine-tuned with our framework on original NLI datasets and their bias-neutralized versions, where original entities are replaced with randomly sampled ones. Extensive results show that our framework consistently improves inferential performance on both original and bias-neutralized NLI datasets.

Quyet V. Do, Junze Li, Tung-Duong Vuong et al.

Commonsense datasets have been well developed in Natural Language Processing, mainly through crowdsource human annotation. However, there are debates on the genuineness of commonsense reasoning benchmarks. In specific, a significant portion of instances in some commonsense benchmarks do not concern commonsense knowledge. That problem would undermine the measurement of the true commonsense reasoning ability of evaluated models. It is also suggested that the problem originated from a blurry concept of commonsense knowledge, as distinguished from other types of knowledge. To demystify all of the above claims, in this study, we survey existing definitions of commonsense knowledge, ground into the three frameworks for defining concepts, and consolidate them into a multi-framework unified definition of commonsense knowledge (so-called consolidated definition). We then use the consolidated definition for annotations and experiments on the CommonsenseQA and CommonsenseQA 2.0 datasets to examine the above claims. Our study shows that there exists a large portion of non-commonsense-knowledge instances in the two datasets, and a large performance gap on these two subsets where Large Language Models (LLMs) perform worse on commonsense-knowledge instances.

Qing Zong, Jiayu Liu, Tianshi Zheng et al.

Accurate confidence calibration in Large Language Models (LLMs) is critical for safe use in high-stakes domains, where clear verbalized confidence enhances user trust. Traditional methods that mimic reference confidence expressions often fail to capture the reasoning needed for accurate confidence assessment. We propose natural language critiques as a solution, ideally suited for confidence calibration, as precise gold confidence labels are hard to obtain and often require multiple generations. This paper studies how natural language critiques can enhance verbalized confidence, addressing: (1) What to critique: uncertainty (question-focused) or confidence (answer-specific)? Analysis shows confidence suits multiple-choice tasks, while uncertainty excels in open-ended scenarios. (2) How to critique: self-critique or critique calibration training? We propose Self-Critique, enabling LLMs to critique and optimize their confidence beyond mere accuracy, and CritiCal, a novel Critique Calibration training method that leverages natural language critiques to improve confidence calibration, moving beyond direct numerical optimization. Experiments show that CritiCal significantly outperforms Self-Critique and other competitive baselines, even surpassing its teacher model, GPT-4o, in complex reasoning tasks. CritiCal also shows robust generalization in out-of-distribution settings, advancing LLM's reliability.

Tianshi Zheng, Kelvin Kiu-Wai Tam, Newt Hue-Nam K. Nguyen et al. · tencent-ai

Large language models are emerging as powerful tools for scientific law discovery, a foundational challenge in AI-driven science. However, existing benchmarks for this task suffer from a fundamental methodological trilemma, forcing a trade-off between scientific relevance, scalability, and resistance to memorization. Furthermore, they oversimplify discovery as static function fitting, failing to capture the authentic scientific process of uncovering embedded laws through the interactive exploration of complex model systems. To address these critical gaps, we introduce NewtonBench, a benchmark comprising 324 scientific law discovery tasks across 12 physics domains. Our design mitigates the evaluation trilemma by using metaphysical shifts - systematic alterations of canonical laws - to generate a vast suite of problems that are scalable, scientifically relevant, and memorization-resistant. Moreover, we elevate the evaluation from static function fitting to interactive model discovery, requiring agents to experimentally probe simulated complex systems to uncover hidden principles. Our extensive experiment reveals a clear but fragile capability for discovery in frontier LLMs: this ability degrades precipitously with increasing system complexity and exhibits extreme sensitivity to observational noise. Notably, we uncover a paradoxical effect of tool assistance: providing a code interpreter can hinder more capable models by inducing a premature shift from exploration to exploitation, causing them to satisfice on suboptimal solutions. These results demonstrate that robust, generalizable discovery in complex, interactive environments remains the core challenge. By providing a scalable, robust, and scientifically authentic testbed, NewtonBench offers a crucial tool for measuring true progress and guiding the development of next-generation AI agents capable of genuine scientific discovery.

Baixuan Xu, Tianshi Zheng, Zhaowei Wang et al. · tencent-ai

Enabling LLMs to effectively operate long-horizon task which requires long-term planning and multiple interactions is essential for open-world autonomy. Conventional methods adopt planning with actions where a executable action list would be provided as reference. However, this action representation choice would be impractical when the environment action space is combinatorial exploded (e.g., open-ended real world). This naturally leads to a question: As environmental action space scales, what is the optimal action representation for long-horizon agents? In this paper, we systematically study the effectiveness of two different action representations. The first one is conventional planning with actions (PwA) which is predominantly adopted for its effectiveness on existing benchmarks. The other one is planning with schemas (PwS) which instantiate an action schema into action lists (e.g., "move [OBJ] to [OBJ]" -> "move apple to desk") to ensure concise action space and reliable scalability. This alternative is motivated by its alignment with human cognition and its compliance with environment-imposed action format restriction. We propose cognitive bandwidth perspective as a conceptual framework to qualitatively understand the differences between these two action representations and empirically observe a representation-choice inflection point between ALFWorld (~35 actions) and SciWorld (~500 actions), which serve as evidence of the need for scalable representations. We further conduct controlled experiments to study how the location of this inflection point interacts with different model capacities: stronger planning proficiency shifts the inflection rightward, whereas better schema instantiation shifts it leftward. Finally, noting the suboptimal performance of PwS agents, we provide an actionable guide for building more capable PwS agents for better scalable autonomy.

Jipeng Zhang, Kehao Miao, Renjie Pi et al.

Reinforcement Fine-Tuning (RFT) with verifiable rewards has advanced large language models but remains underexplored for Vision-Language (VL) models. The Vision-Language Reward Model (VL-RM) is key to aligning VL models by providing structured feedback, yet training effective VL-RMs faces two major challenges. First, the bootstrapping dilemma arises as high-quality training data depends on already strong VL models, creating a cycle where self-generated supervision reinforces existing biases. Second, modality bias and negative example amplification occur when VL models hallucinate incorrect visual attributes, leading to flawed preference data that further misguides training. To address these issues, we propose an iterative training framework leveraging vision experts, Chain-of-Thought (CoT) rationales, and Margin-based Rejection Sampling. Our approach refines preference datasets, enhances structured critiques, and iteratively improves reasoning. Experiments across VL-RM benchmarks demonstrate superior performance in hallucination detection and multimodal reasoning, advancing VL model alignment with reinforcement learning.

Zhaowei Wang, Yunsong Huang, Weicheng Liu et al.

The utilization of radio frequency (RF) signals for wireless sensing has garnered increasing attention. However, the radio environment is unpredictable and often unfavorable, the sensing accuracy of traditional RF sensing methods is often affected by adverse propagation channels from the transmitter to the receiver, such as fading and noise. In this paper, we propose employing distributed Reconfigurable Intelligent Metasurface Antennas (RIMSA) to detect the presence and location of objects where multiple RIMSA receivers (RIMSA Rxs) are deployed on different places. By programming their beamforming patterns, RIMSA Rxs can enhance the quality of received signals. The RF sensing problem is modeled as a joint optimization problem of beamforming pattern and mapping of received signals to sensing outcomes. To address this challenge, we introduce a deep reinforcement learning (DRL) algorithm aimed at calculating the optimal beamforming patterns and a neural network aimed at converting received signals into sensing outcomes. In addition, the malicious attacker may potentially launch jamming attack to disrupt sensing process. To enable effective sensing in interferenceprone environment, we devise a combined loss function that takes into account the Signal to Interference plus Noise Ratio (SINR) of the received signals. The simulation results show that the proposed distributed RIMSA system can achieve more efficient sensing performance and better overcome environmental influences than centralized implementation. Furthermore, the introduced method ensures high-accuracy sensing performance even under jamming attack.

Liang Cheng, Zhaowei Wang, Mark Steedman

While fine-tuning LLMs on NLI corpora improves their inferential performance, the underlying mechanisms driving this improvement remain largely opaque. In this work, we conduct a series of experiments to investigate what LLMs actually learn during fine-tuning. We begin by analyzing predicate frequencies in premises and hypotheses across NLI datasets and identify a consistent frequency bias, where predicates in hypotheses occur more frequently than those in premises for positive instances. To assess the impact of this bias, we evaluate both standard and NLI fine-tuned LLMs on bias-consistent and bias-adversarial cases. We find that LLMs exploit frequency bias for inference and perform poorly on adversarial instances. Furthermore, fine-tuned LLMs exhibit significantly increased reliance on this bias, suggesting that they are learning these frequency patterns from datasets. Finally, we compute the frequencies of hyponyms and their corresponding hypernyms from WordNet, revealing a correlation between frequency bias and textual entailment. These findings help explain why learning frequency patterns can enhance model performance on inference tasks.

Jiangshu Du, Yibo Wang, Wenting Zhao et al.

This work is motivated by two key trends. On one hand, large language models (LLMs) have shown remarkable versatility in various generative tasks such as writing, drawing, and question answering, significantly reducing the time required for many routine tasks. On the other hand, researchers, whose work is not only time-consuming but also highly expertise-demanding, face increasing challenges as they have to spend more time reading, writing, and reviewing papers. This raises the question: how can LLMs potentially assist researchers in alleviating their heavy workload? This study focuses on the topic of LLMs assist NLP Researchers, particularly examining the effectiveness of LLM in assisting paper (meta-)reviewing and its recognizability. To address this, we constructed the ReviewCritique dataset, which includes two types of information: (i) NLP papers (initial submissions rather than camera-ready) with both human-written and LLM-generated reviews, and (ii) each review comes with "deficiency" labels and corresponding explanations for individual segments, annotated by experts. Using ReviewCritique, this study explores two threads of research questions: (i) "LLMs as Reviewers", how do reviews generated by LLMs compare with those written by humans in terms of quality and distinguishability? (ii) "LLMs as Metareviewers", how effectively can LLMs identify potential issues, such as Deficient or unprofessional review segments, within individual paper reviews? To our knowledge, this is the first work to provide such a comprehensive analysis.

Tianqing Fang, Zhaowei Wang, Wenxuan Zhou et al.

Event temporal reasoning aims at identifying the temporal relations between two or more events from narratives. However, knowledge conflicts arise when there is a mismatch between the actual temporal relations of events in the context and the prior knowledge or biases learned by the model. In this paper, we propose to detect knowledge-conflict examples in event temporal reasoning using bias indicators, which include event relation prior bias, tense bias, narrative bias, and dependency bias. We define conflict examples as those where event relations are opposite to biased or prior relations. To mitigate event-related knowledge conflicts, we introduce a Counterfactual Data Augmentation (CDA) based method that can be applied to both Pre-trained Language Models (PLMs) and Large Language Models (LLMs) either as additional training data or demonstrations for In-Context Learning. Experiments suggest both PLMs and LLMs suffer from knowledge conflicts in event temporal reasoning, and CDA has the potential for reducing hallucination and improving model performance.

Zhaowei Wang, Quyet V. Do, Hongming Zhang et al.

Detecting commonsense causal relations (causation) between events has long been an essential yet challenging task. Given that events are complicated, an event may have different causes under various contexts. Thus, exploiting context plays an essential role in detecting causal relations. Meanwhile, previous works about commonsense causation only consider two events and ignore their context, simplifying the task formulation. This paper proposes a new task to detect commonsense causation between two events in an event sequence (i.e., context), called contextualized commonsense causal reasoning. We also design a zero-shot framework: COLA (Contextualized Commonsense Causality Reasoner) to solve the task from the causal inference perspective. This framework obtains rich incidental supervision from temporality and balances covariates from multiple timestamps to remove confounding effects. Our extensive experiments show that COLA can detect commonsense causality more accurately than baselines.

Junde Wu, Huihui Fang, Dalu Yang et al.

With the advancement of deep learning techniques, an increasing number of methods have been proposed for optic disc and cup (OD/OC) segmentation from the fundus images. Clinically, OD/OC segmentation is often annotated by multiple clinical experts to mitigate the personal bias. However, it is hard to train the automated deep learning models on multiple labels. A common practice to tackle the issue is majority vote, e.g., taking the average of multiple labels. However such a strategy ignores the different expertness of medical experts. Motivated by the observation that OD/OC segmentation is often used for the glaucoma diagnosis clinically, in this paper, we propose a novel strategy to fuse the multi-rater OD/OC segmentation labels via the glaucoma diagnosis performance. Specifically, we assess the expertness of each rater through an attentive glaucoma diagnosis network. For each rater, its contribution for the diagnosis will be reflected as an expertness map. To ensure the expertness maps are general for different glaucoma diagnosis models, we further propose an Expertness Generator (ExpG) to eliminate the high-frequency components in the optimization process. Based on the obtained expertness maps, the multi-rater labels can be fused as a single ground-truth which we dubbed as Diagnosis First Ground-truth (DiagFirstGT). Experimental results show that by using DiagFirstGT as ground-truth, OD/OC segmentation networks will predict the masks with superior glaucoma diagnosis performance.

Kelong Mao, Jieming Zhu, Xi Xiao et al.

With the recent success of graph convolutional networks (GCNs), they have been widely applied for recommendation, and achieved impressive performance gains. The core of GCNs lies in its message passing mechanism to aggregate neighborhood information. However, we observed that message passing largely slows down the convergence of GCNs during training, especially for large-scale recommender systems, which hinders their wide adoption. LightGCN makes an early attempt to simplify GCNs for collaborative filtering by omitting feature transformations and nonlinear activations. In this paper, we take one step further to propose an ultra-simplified formulation of GCNs (dubbed UltraGCN), which skips infinite layers of message passing for efficient recommendation. Instead of explicit message passing, UltraGCN resorts to directly approximate the limit of infinite-layer graph convolutions via a constraint loss. Meanwhile, UltraGCN allows for more appropriate edge weight assignments and flexible adjustment of the relative importances among different types of relationships. This finally yields a simple yet effective UltraGCN model, which is easy to implement and efficient to train. Experimental results on four benchmark datasets show that UltraGCN not only outperforms the state-of-the-art GCN models but also achieves more than 10x speedup over LightGCN. Our source code will be available at https://reczoo.github.io/UltraGCN.

Yang Bai, Xiaoguang Li, Gang Wang et al.

Term-based sparse representations dominate the first-stage text retrieval in industrial applications, due to its advantage in efficiency, interpretability, and exact term matching. In this paper, we study the problem of transferring the deep knowledge of the pre-trained language model (PLM) to Term-based Sparse representations, aiming to improve the representation capacity of bag-of-words(BoW) method for semantic-level matching, while still keeping its advantages. Specifically, we propose a novel framework SparTerm to directly learn sparse text representations in the full vocabulary space. The proposed SparTerm comprises an importance predictor to predict the importance for each term in the vocabulary, and a gating controller to control the term activation. These two modules cooperatively ensure the sparsity and flexibility of the final text representation, which unifies the term-weighting and expansion in the same framework. Evaluated on MSMARCO dataset, SparTerm significantly outperforms traditional sparse methods and achieves state of the art ranking performance among all the PLM-based sparse models.