Jiajun Song

Pengfei Zhou, Weiqing Min, Yang Zhang et al.

Food detection is becoming a fundamental task in food computing that supports various multimedia applications, including food recommendation and dietary monitoring. To deal with real-world scenarios, food detection needs to localize and recognize novel food objects that are not seen during training, demanding Zero-Shot Detection (ZSD). However, the complexity of semantic attributes and intra-class feature diversity poses challenges for ZSD methods in distinguishing fine-grained food classes. To tackle this, we propose the Semantic Separable Diffusion Synthesizer (SeeDS) framework for Zero-Shot Food Detection (ZSFD). SeeDS consists of two modules: a Semantic Separable Synthesizing Module (S$^3$M) and a Region Feature Denoising Diffusion Model (RFDDM). The S$^3$M learns the disentangled semantic representation for complex food attributes from ingredients and cuisines, and synthesizes discriminative food features via enhanced semantic information. The RFDDM utilizes a novel diffusion model to generate diversified region features and enhances ZSFD via fine-grained synthesized features. Extensive experiments show the state-of-the-art ZSFD performance of our proposed method on two food datasets, ZSFooD and UECFOOD-256. Moreover, SeeDS also maintains effectiveness on general ZSD datasets, PASCAL VOC and MS COCO. The code and dataset can be found at https://github.com/LanceZPF/SeeDS.

Zhang Li, Zhibo Lin, Qiang Liu et al.

We introduce Multilingual Document Parsing Benchmark, the first benchmark for multilingual digital and photographed document parsing. Document parsing has made remarkable strides, yet almost exclusively on clean, digital, well-formatted pages in a handful of dominant languages. No systematic benchmark exists to evaluate how models perform on digital and photographed documents across diverse scripts and low-resource languages. MDPBench comprises 3,400 document images spanning 17 languages, diverse scripts, and varied photographic conditions, with high-quality annotations produced through a rigorous pipeline of expert model labeling, manual correction, and human verification. To ensure fair comparison and prevent data leakage, we maintain separate public and private evaluation splits. Our comprehensive evaluation of both open-source and closed-source models uncovers a striking finding: while closed-source models (notably Gemini3-Pro) prove relatively robust, open-source alternatives suffer dramatic performance collapse, particularly on non-Latin scripts and real-world photographed documents, with an average drop of 17.8% on photographed documents and 14.0% on non-Latin scripts. These results reveal significant performance imbalances across languages and conditions, and point to concrete directions for building more inclusive, deployment-ready parsing systems. Source available at https://github.com/Yuliang-Liu/MultimodalOCR.

Sijia Cui, Pengyu Cheng, Jiajun Song et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capacity of Large Language Models (LLMs). However, RLVR solely relies on final answers as outcome rewards, neglecting the correctness of intermediate reasoning steps. Training on these process-wrong but outcome-correct rollouts can lead to hallucination and answer-copying, severely undermining the model's generalization and robustness. To address this, we incorporate a Contrastive Learning mechanism into the Policy Optimization (CLIPO) to generalize the RLVR process. By optimizing a contrastive loss over successful rollouts, CLIPO steers the LLM to capture the invariant structure shared across correct reasoning paths. This provides a more robust cross-trajectory regularization than the original single-path supervision in RLVR, effectively mitigating step-level reasoning inconsistencies and suppressing hallucinatory artifacts. In experiments, CLIPO consistently improves multiple RLVR baselines across diverse reasoning benchmarks, demonstrating uniform improvements in generalization and robustness for policy optimization of LLMs. Our code and training recipes are available at https://github.com/Qwen-Applications/CLIPO.

Zhuo Li, Yupeng Zhang, Pengyu Cheng et al.

Hallucination remains a critical bottleneck for large language models (LLMs), undermining their reliability in real-world applications, especially in Retrieval-Augmented Generation (RAG) systems. While existing hallucination detection methods employ LLM-as-a-judge to verify LLM outputs against retrieved evidence, they suffer from inherent confirmation bias, where the verifier inadvertently reproduces the errors of the original generation. To address this, we introduce Multi-Agent Reinforced Self-Check for Hallucination (MARCH), a framework that enforces rigorous factual alignment by leveraging deliberate information asymmetry. MARCH orchestrates a collaborative pipeline of three specialized agents: a Solver, a Proposer, and a Checker. The Solver generates an initial RAG response, which the Proposer decomposes into claim-level verifiable atomic propositions. Crucially, the Checker validates these propositions against retrieved evidence in isolation, deprived of the Solver's original output. This well-crafted information asymmetry scheme breaks the cycle of self-confirmation bias. By training this pipeline with multi-agent reinforcement learning (MARL), we enable the agents to co-evolve and optimize factual adherence. Extensive experiments across hallucination benchmarks demonstrate that MARCH substantially reduces hallucination rates. Notably, an 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models. MARCH paves a scalable path for factual self-improvement of LLMs through co-evolution. The code is at https://github.com/Qwen-Applications/MARCH.

Jiajun Song, Zhuoyan Xu, Yiqiao Zhong

Large language models (LLMs) such as GPT-4 sometimes appear to be creative, solving novel tasks often with a few demonstrations in the prompt. These tasks require the models to generalize on distributions different from those from training data -- which is known as out-of-distribution (OOD) generalization. Despite the tremendous success of LLMs, how they approach OOD generalization remains an open and underexplored question. We examine OOD generalization in settings where instances are generated according to hidden rules, including in-context learning with symbolic reasoning. Models are required to infer the hidden rules behind input prompts without any fine-tuning. We empirically examined the training dynamics of Transformers on a synthetic example and conducted extensive experiments on a variety of pretrained LLMs, focusing on a type of components known as induction heads. We found that OOD generalization and composition are tied together -- models can learn rules by composing two self-attention layers, thereby achieving OOD generalization. Furthermore, a shared latent subspace in the embedding (or feature) space acts as a bridge for composition by aligning early layers and later layers, which we refer to as the common bridge representation hypothesis.

Pengfei Zhou, Weiqing Min, Jiajun Song et al.

Food computing brings various perspectives to computer vision like vision-based food analysis for nutrition and health. As a fundamental task in food computing, food detection needs Zero-Shot Detection (ZSD) on novel unseen food objects to support real-world scenarios, such as intelligent kitchens and smart restaurants. Therefore, we first benchmark the task of Zero-Shot Food Detection (ZSFD) by introducing FOWA dataset with rich attribute annotations. Unlike ZSD, fine-grained problems in ZSFD like inter-class similarity make synthesized features inseparable. The complexity of food semantic attributes further makes it more difficult for current ZSD methods to distinguish various food categories. To address these problems, we propose a novel framework ZSFDet to tackle fine-grained problems by exploiting the interaction between complex attributes. Specifically, we model the correlation between food categories and attributes in ZSFDet by multi-source graphs to provide prior knowledge for distinguishing fine-grained features. Within ZSFDet, Knowledge-Enhanced Feature Synthesizer (KEFS) learns knowledge representation from multiple sources (e.g., ingredients correlation from knowledge graph) via the multi-source graph fusion. Conditioned on the fusion of semantic knowledge representation, the region feature diffusion model in KEFS can generate fine-grained features for training the effective zero-shot detector. Extensive evaluations demonstrate the superior performance of our method ZSFDet on FOWA and the widely-used food dataset UECFOOD-256, with significant improvements by 1.8% and 3.7% ZSD mAP compared with the strong baseline RRFS. Further experiments on PASCAL VOC and MS COCO prove that enhancement of the semantic knowledge can also improve the performance on general ZSD. Code and dataset are available at https://github.com/LanceZPF/KEFS.

Jiajun Song, Jiajun Luo, Rongwei Lu et al.

Asynchronous Federated Learning (AFL) confronts inherent challenges arising from the heterogeneity of devices (e.g., their computation capacities) and low-bandwidth environments, both potentially causing stale model updates (e.g., local gradients) for global aggregation. Traditional approaches mitigating the staleness of updates typically focus on either adjusting the local updating or gradient compression, but not both. Recognizing this gap, we introduce a novel approach that synergizes local updating with gradient compression. Our research begins by examining the interplay between local updating frequency and gradient compression rate, and their collective impact on convergence speed. The theoretical upper bound shows that the local updating frequency and gradient compression rate of each device are jointly determined by its computing power, communication capabilities and other factors. Building on this foundation, we propose an AFL framework called FedLuck that adaptively optimizes both local update frequency and gradient compression rates. Experiments on image classification and speech recognization show that FedLuck reduces communication consumption by 56% and training time by 55% on average, achieving competitive performance in heterogeneous and low-bandwidth scenarios compared to the baselines.

Jiajun Song, Yiqiao Zhong

Transformers are widely used to extract semantic meanings from input tokens, yet they usually operate as black-box models. In this paper, we present a simple yet informative decomposition of hidden states (or embeddings) of trained transformers into interpretable components. For any layer, embedding vectors of input sequence samples are represented by a tensor $\boldsymbol{h} \in \mathbb{R}^{C \times T \times d}$. Given embedding vector $\boldsymbol{h}_{c,t} \in \mathbb{R}^d$ at sequence position $t \le T$ in a sequence (or context) $c \le C$, extracting the mean effects yields the decomposition \[ \boldsymbol{h}_{c,t} = \boldsymbolμ + \mathbf{pos}_t + \mathbf{ctx}_c + \mathbf{resid}_{c,t} \] where $\boldsymbolμ$ is the global mean vector, $\mathbf{pos}_t$ and $\mathbf{ctx}_c$ are the mean vectors across contexts and across positions respectively, and $\mathbf{resid}_{c,t}$ is the residual vector. For popular transformer architectures and diverse text datasets, empirically we find pervasive mathematical structure: (1) $(\mathbf{pos}_t)_{t}$ forms a low-dimensional, continuous, and often spiral shape across layers, (2) $(\mathbf{ctx}_c)_c$ shows clear cluster structure that falls into context topics, and (3) $(\mathbf{pos}_t)_{t}$ and $(\mathbf{ctx}_c)_c$ are mutually nearly orthogonal. We argue that smoothness is pervasive and beneficial to transformers trained on languages, and our decomposition leads to improved model interpretability.

Jiajun Song, Xiaoou Liu

Food recognition has gained significant attention, but the rapid emergence of new dishes requires methods for recognizing unseen food categories, motivating Zero-Shot Food Learning (ZSFL). We propose the task of Compositional Zero-Shot Food Recognition (CZSFR), where cuisines and ingredients naturally align with attributes and objects in Compositional Zero-Shot learning (CZSL). However, CZSFR faces three challenges: (1) Redundant background information distracts models from learning meaningful food features, (2) Role confusion between staple and side dishes leads to misclassification, and (3) Semantic bias in a single attribute can lead to confusion of understanding. Therefore, we propose SalientFusion, a context-aware CZSFR method with two components: SalientFormer, which removes background redundancy and uses depth features to resolve role confusion; DebiasAT, which reduces the semantic bias by aligning prompts with visual features. Using our proposed benchmarks, CZSFood-90 and CZSFood-164, we show that SalientFusion achieves state-of-the-art results on these benchmarks and the most popular general datasets for the general CZSL. The code is avaliable at https://github.com/Jiajun-RUC/SalientFusion.

Hiba Maryam, Ling Fu, Jiajun Song et al.

The development of Urdu scene text detection, recognition, and Visual Question Answering (VQA) technologies is crucial for advancing accessibility, information retrieval, and linguistic diversity in digital content, facilitating better understanding and interaction with Urdu-language visual data. This initiative seeks to bridge the gap between textual and visual comprehension. We propose a new multi-task Urdu scene text dataset comprising over 1000 natural scene images, which can be used for text detection, recognition, and VQA tasks. We provide fine-grained annotations for text instances, addressing the limitations of previous datasets for facing arbitrary-shaped texts. By incorporating additional annotation points, this dataset facilitates the development and assessment of methods that can handle diverse text layouts, intricate shapes, and non-standard orientations commonly encountered in real-world scenarios. Besides, the VQA annotations make it the first benchmark for the Urdu Text VQA method, which can prompt the development of Urdu scene text understanding. The proposed dataset is available at: https://github.com/Hiba-MeiRuan/Urdu-VQA-Dataset-/tree/main

Zhuo Li, Weiqing Min, Jiajun Song et al.

Optimizing the approximation of Average Precision (AP) has been widely studied for image retrieval. Limited by the definition of AP, such methods consider both negative and positive instances ranking before each positive instance. However, we claim that only penalizing negative instances before positive ones is enough, because the loss only comes from these negative instances. To this end, we propose a novel loss, namely Penalizing Negative instances before Positive ones (PNP), which can directly minimize the number of negative instances before each positive one. In addition, AP-based methods adopt a fixed and sub-optimal gradient assignment strategy. Therefore, we systematically investigate different gradient assignment solutions via constructing derivative functions of the loss, resulting in PNP-I with increasing derivative functions and PNP-D with decreasing ones. PNP-I focuses more on the hard positive instances by assigning larger gradients to them and tries to make all relevant instances closer. In contrast, PNP-D pays less attention to such instances and slowly corrects them. For most real-world data, one class usually contains several local clusters. PNP-I blindly gathers these clusters while PNP-D keeps them as they were. Therefore, PNP-D is more superior. Experiments on three standard retrieval datasets show consistent results with the above analysis. Extensive evaluations demonstrate that PNP-D achieves the state-of-the-art performance. Code is available at https://github.com/interestingzhuo/PNPloss

Fanqing Meng, Lingxiao Du, Zijian Wu et al.

Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce \bench{}, a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0\%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.

Ling Fu, Zhebin Kuang, Jiajun Song et al.

Scoring the Optical Character Recognition (OCR) capabilities of Large Multimodal Models (LMMs) has witnessed growing interest. Existing benchmarks have highlighted the impressive performance of LMMs in text recognition; however, their abilities in certain challenging tasks, such as text localization, handwritten content extraction, and logical reasoning, remain underexplored. To bridge this gap, we introduce OCRBench v2, a large-scale bilingual text-centric benchmark with currently the most comprehensive set of tasks (4x more tasks than the previous multi-scene benchmark OCRBench), the widest coverage of scenarios (31 diverse scenarios), and thorough evaluation metrics, with 10,000 human-verified question-answering pairs and a high proportion of difficult samples. Moreover, we construct a private test set with 1,500 manually annotated images. The consistent evaluation trends observed across both public and private test sets validate the OCRBench v2's reliability. After carefully benchmarking state-of-the-art LMMs, we find that most LMMs score below 50 (100 in total) and suffer from five-type limitations, including less frequently encountered text recognition, fine-grained perception, layout perception, complex element parsing, and logical reasoning. The project website is at: https://99franklin.github.io/ocrbench_v2/

Yi-Long Lu, Jiajun Song, Wei Wang

A central architectural question for both biological and artificial intelligence is whether judgment relies on specialized modules or a unified, domain-general resource. While the discovery of decodable neural representations for distinct concepts in Large Language Models (LLMs) has suggested a modular architecture, whether these representations are truly independent systems remains an open question. Here we provide evidence for a convergent architecture for evaluative judgment. Across a range of LLMs, we find that diverse evaluative judgments are computed along a dominant dimension, which we term the Valence-Assent Axis (VAA). This axis jointly encodes subjective valence ("what is good") and the model's assent to factual claims ("what is true"). Through direct interventions, we demonstrate this axis drives a critical mechanism, which is identified as the subordination of reasoning: the VAA functions as a control signal that steers the generative process to construct a rationale consistent with its evaluative state, even at the cost of factual accuracy. Our discovery offers a mechanistic account for response bias and hallucination, revealing how an architecture that promotes coherent judgment can systematically undermine faithful reasoning.

Pengfei Zhou, Xiaopeng Peng, Jiajun Song et al.

Multimodal Large Language Models (MLLMs) have made significant strides in visual understanding and generation tasks. However, generating interleaved image-text content remains a challenge, which requires integrated multimodal understanding and generation abilities. While the progress in unified models offers new solutions, existing benchmarks are insufficient for evaluating these methods due to limitations in data size and diversity. To bridge this gap, we introduce OpenING, a comprehensive benchmark comprising 5,400 high-quality human-annotated instances across 56 real-world tasks. OpenING covers diverse daily scenarios such as travel guide, design, and brainstorming, offering a robust platform for challenging interleaved generation methods. In addition, we present IntJudge, a judge model for evaluating open-ended multimodal generation methods. Trained with a novel data pipeline, our IntJudge achieves an agreement rate of 82.42% with human judgments, outperforming GPT-based evaluators by 11.34%. Extensive experiments on OpenING reveal that current interleaved generation methods still have substantial room for improvement. Key findings on interleaved image-text generation are further presented to guide the development of next-generation models.

Chi Zhang, Jiajun Song, Siyu Li et al.

Mathematics olympiads are prestigious competitions, with problem proposing and solving highly honored. Building artificial intelligence that proposes and solves olympiads presents an unresolved challenge in automated theorem discovery and proving, especially in geometry for its combination of numerical and spatial elements. We introduce TongGeometry, a Euclidean geometry system supporting tree-search-based guided problem proposing and solving. The efficient geometry system establishes the most extensive repository of geometry theorems to date: within the same computational budget as the existing state-of-the-art, TongGeometry discovers 6.7 billion geometry theorems requiring auxiliary constructions, including 4.1 billion exhibiting geometric symmetry. Among them, 10 theorems were proposed to regional mathematical olympiads with 3 of TongGeometry's proposals selected in real competitions, earning spots in a national team qualifying exam or a top civil olympiad in China and the US. Guided by fine-tuned large language models, TongGeometry solved all International Mathematical Olympiad geometry in IMO-AG-30, outperforming gold medalists for the first time. It also surpasses the existing state-of-the-art across a broader spectrum of olympiad-level problems. The full capabilities of the system can be utilized on a consumer-grade machine, making the model more accessible and fostering widespread democratization of its use. By analogy, unlike existing systems that merely solve problems like students, TongGeometry acts like a geometry coach, discovering, presenting, and proving theorems.

Yi-Long Lu, Chunhui Zhang, Jiajun Song et al.

Theory of Mind (ToM), the ability to attribute mental states to others, is fundamental for human social intelligence and a critical capability for advanced Artificial Intelligence. Recent advancements in Large Language Models (LLMs) have shown promising performance on ToM benchmarks, raising the question: Do these benchmarks necessitate explicit human-like reasoning processes, or can models succeed through alternative strategies? We investigate this question empirically by applying Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT) to LLMs of varying scales (0.5B to 7B parameters) and evaluating them across multiple ToM datasets. Our results reveal a scale-dependent impact of RL: while RL significantly improves accuracy and fosters high-quality, interpretable, and transferable belief-tracking reasoning in larger models (7B), it leads to "reasoning collapse" in smaller models ($\leq$3B), where high accuracy and generalization ability are achieved via drastically shortened, less meaningful responses. Surprisingly, further SFT achieves competitive and generalizable performance across these benchmarks, often matching or exceeding RL models in accuracy, despite not being explicitly trained to produce structured reasoning traces. These findings highlight a critical discrepancy between benchmark accuracy and the nature of learned reasoning. Our work suggests that current ToM benchmarks may be solvable without requiring the explicit, human-like simulation of mental states they were designed to probe. LLMs, particularly when scale is limited or training signals focus solely on output correctness, may leverage alternative rules effective for benchmark data structures.

Jiajun Song, Xiaoou Liu

Transformer-based models have significantly advanced time series forecasting. Recent work, like the Cross-Attention-only Time Series transformer (CATS), shows that removing self-attention can make the model more accurate and efficient. However, these streamlined architectures may overlook the fine-grained, local temporal dependencies effectively captured by classical statistical models like Vector AutoRegressive Moving Average model (VARMA). To address this gap, we propose VARMAformer, a novel architecture that synergizes the efficiency of a cross-attention-only framework with the principles of classical time series analysis. Our model introduces two key innovations: (1) a dedicated VARMA-inspired Feature Extractor (VFE) that explicitly models autoregressive (AR) and moving-average (MA) patterns at the patch level, and (2) a VARMA-Enhanced Attention (VE-atten) mechanism that employs a temporal gate to make queries more context-aware. By fusing these classical insights into a modern backbone, VARMAformer captures both global, long-range dependencies and local, statistical structures. Through extensive experiments on widely-used benchmark datasets, we demonstrate that our model consistently outperforms existing state-of-the-art methods. Our work validates the significant benefit of integrating classical statistical insights into modern deep learning frameworks for time series forecasting.

Yi-Long Lu, Jiajun Song, Chunhui Zhang et al.

Humans constantly generate a diverse range of tasks guided by internal motivations. While generative agents powered by large language models (LLMs) aim to simulate this complex behavior, it remains uncertain whether they operate on similar cognitive principles. To address this, we conducted a task-generation experiment comparing human responses with those of an LLM agent (GPT-4o). We find that human task generation is consistently influenced by psychological drivers, including personal values (e.g., Openness to Change) and cognitive style. Even when these psychological drivers are explicitly provided to the LLM, it fails to reflect the corresponding behavioral patterns. They produce tasks that are markedly less social, less physical, and thematically biased toward abstraction. Interestingly, while the LLM's tasks were perceived as more fun and novel, this highlights a disconnect between its linguistic proficiency and its capacity to generate human-like, embodied goals. We conclude that there is a core gap between the value-driven, embodied nature of human cognition and the statistical patterns of LLMs, highlighting the necessity of incorporating intrinsic motivation and physical grounding into the design of more human-aligned agents.