Multimodal large language models (MLLMs) remain prone to visual hallucinations when correct responses depend on subtle visual details. One line of post-training methods constructs visual contrasts through image retrieval, generation, or editing to improve fine-grained perception. However, these contrasts are not built around the model's own hallucinations, making it difficult to ensure that supervision targets its actual perceptual confusions. Another line suppresses hallucinated outputs through response correction or preference learning under fixed image conditions, but leaves the design of supervision signals on the image-input side underexplored. Inspired by how humans correct misconceptions, we introduce DreamRealm, a post-training framework that explores diverse latent hallucinated responses, minimally edits original images to make them valid, and trains the model to assign each response to its supporting image. Our data construction pipeline, RealmForgePipe, mines observed and latent hallucinations from a target model and materializes them through image editing. Using hallucinations mined from Qwen3.5-9B, we construct the Realized Hallucination Dataset (RealmSet). Our training objective, RealmMatch, constructs a likelihood score matrix across image--response combinations and combines supervised learning with bidirectional contrastive losses to reinforce correct matches and suppress mismatches. Compared with the original Qwen3.5-9B model, DreamRealm improves performance across six visual perception and hallucination benchmarks, including gains of 7.26 percentage points in HallusionBench accuracy and 11.34 percentage points in MMVP pair accuracy. Together with ablation studies on data construction and training methods, these results demonstrate that hallucination assignment is a promising paradigm for improving visual perception and mitigating hallucinations.
Coding agents must deliver correct software while respecting users' expertise, preferences, and working conditions. Existing benchmarks increasingly incorporate interaction, yet personalized requirement fulfillment remains underexplored. We introduce PersonaCodeBench, which turns user characteristics into verifiable obligations in persona-driven multi-turn coding sessions. Our persona pool spans roles, technical abilities, interaction preferences, and practical constraints, yielding task-grounded evolved requirements for implementation, permissions, testing, explanation, and delivery. At runtime, a controller governs user simulator responses and requirement disclosure, using online requirement evaluation to provide feedback and advance the interaction. Requirements emerge progressively, with their content, disclosure conditions, and acceptance criteria fixed before evaluation. A completion verifier and a personalization verifier separately assess native correctness and persona fulfillment. Joint pass requires both adherence to the declared budget and compliance with validity conditions. We evaluate all major frontier agents, establish a comprehensive leaderboard, and conduct detailed experimental analyses to validate effectiveness. PersonaCodeBench supports reproducible evaluation of coding agents' ability to meet explicit, personalized user expectations alongside coding objectives.
Explicit reasoning in large language models (LLMs) is often associated with reliability, yet it can still pose risks when a model infers harmful goals under a benign context. We identify a distinct challenge: ensuring the safety of the \emph{multi-step reasoning process} itself, where locally coherent steps may still lead to globally harmful outcomes. To explore this vulnerability, we propose PDRA (Paradox-driven Reasoning-based Attack), a novel attack method that programs a harmful reasoning trajectory. PDRA first constructs a paradoxical narrative core by pairing a malicious actor (e.g., a terrorist) with a benign task (e.g., writing a safety bulletin). It then embeds this core into a structured analytical framework that forces the model through a mandatory three-stage analysis -- surface purpose, latent intent, and operational reconstruction -- guiding it to produce detailed harmful content while maintaining narrative coherence. Mechanistically, we show PDRA steers the model's internal representations away from refusal-related patterns, suppressing safety-triggering lexicons while activating planning-oriented terms. Empirically, PDRA achieves state-of-the-art attack success rates across diverse LLMs with high efficiency. \textcolor{red}{\textbf{Warning:} This paper may contain sensitive content.}
Parameter-Efficient Fine-Tuning (PEFT) has emerged as a prevalent strategy for adapting large pre-trained models. Among PEFT techniques, reparameterization-based methods, particularly those that employ low-rank or sparsity structures, have attracted significant attention. However, these approaches often underperform full fine-tuning. Specifically, low-rank methods constrain the update rank, which is misaligned with the inherently high-rank nature of full fine-tuning. In contrast, sparsity-based techniques limit the flexibility of the parameter space and compromise connectivity across weights. To overcome these limitations, we explore the use of Toeplitz matrices, whose entries are constant along each diagonal, therefore providing a compact parameterization without imposing explicit rank or sparsity constraints. A straightforward approach is to update weights via a product of Toeplitz matrices. However, similar to RNNs, long multiplicative chains often lead to gradient instability. To address this issue, we propose \textbf{ToPA}, a more refined version of this idea that replaces standard Toeplitz matrices with block-wise Toeplitz, greatly increasing each factor's capacity and shortening the chain, thereby stabilizing training and improving expressiveness. We theoretically show that products of block-wise Toeplitz matrices can approximate arbitrary matrices under shorter chain length, justifying the structural design of ToPA. Furthermore, we demonstrate that ToPA offers greater expressivity than existing low-rank and sparse parameterizations. Empirical evaluations across 21 NLP and CV datasets, spanning 5 model architectures, consistently validate the effectiveness of ToPA.
Yuancheng Wei, Haojie Zhang, Linli Yao, Lei Li, Jiali Chen, Tao Huang, Yiting Lu, Duojun Huang 等 10 位作者
Image Difference Captioning (IDC) generates natural language descriptions that precisely identify differences between two images, serving as a key benchmark for fine-grained change perception, cross-modal reasoning, and image editing data construction. However, existing benchmarks lack diversity and compositional complexity, and standard lexical-overlap metrics (e.g., BLEU, METEOR) fail to capture semantic consistency or penalize hallucinations, which together prevent a comprehensive and robust evaluation of multimodal large language models (MLLMs) on IDC. To address these gaps, we introduce DiffCap-Bench, a comprehensive IDC benchmark covering ten distinct difference categories to ensure diversity and compositional complexity. Furthermore, we propose an LLM-as-a-Judge evaluation protocol grounded in human-validated Difference Lists, enabling a robust assessment of models' ability to both capture and describe visual changes. Through extensive evaluation of state-of-the-art MLLMs, we reveal significant performance gaps between proprietary and open-source models, highlight the critical importance of reasoning capability, and identify clear limitations in model scaling. Our framework also demonstrates strong alignment with human expert judgments and strong correlation with downstream image editing data construction quality. These findings establish DiffCap-Bench as both a reliable IDC evaluation framework and a practical predictor of downstream utility. The benchmark and code will be made publicly available to support further research.
Li Kang, Yutao Fan, Rui Li, Heng Zhou, Yiran Qin, Zhemeng Zhang, Songtao Huang, Xiufeng Song 等 14 位作者
General-purpose coding agents can now write, debug, and refine entire robot programs, but only if they have somewhere to practice. Physical hardware is a poor practice ground: exploratory rollouts are slow to reset and, when several arms share a workspace, unsafe. Generic simulation benchmarks are safe but do not match the geometry, assets, or viewpoints of the workspace where the robot must actually work. We present CoEx (Code-driven Exploration), which lets the agent practice in an executable digital twin of its own deployment workspace. The twin is built by CoEnv, our real-to-sim framework, which reconstructs the physical scene from multi-view RGBD into a physics simulation. Inside it, a centralized code agent explores at the program level, writing a complete multi-arm program, running it in the twin, and refining it against grounded verification (privileged state checks, checkpoint predicates on risky steps, and adaptive viewpoints against inter-arm occlusion). Verified programs transfer to hardware through collision-checked deployment. On five real-world multi-arm tasks with up to three heterogeneous robots, CoEx reaches 70% average success, against 50% for the same system without the code agent (where a VLM must choose every primitive step by step), outperforms two prior VLM-planning frameworks, and produces verified episodes at 3.8-11.7x the ablated rate. Because every verified rollout is also a demonstration, exploration leaves reusable experience behind: a $π_{0.5}$ policy fine-tuned purely on 200 agent-generated rollouts reaches 100% and 70% real-robot success on two tasks, with no human demonstrations.
Siyang Wu, Weizhi Zhang, Chen Yang, Esther Xu Fei, Jorin Overwiening, Xu Pan, Jing Tang, Yixuan He 等 37 位作者
Scientific agents collaborate with researchers whose expertise and preferences differ, but benchmarks usually hold the collaborator fixed. We introduce PersonaScienceBench: 200 tasks from five benchmarks across four scientific domains. Simulated researchers provide context and feedback according to persona profiles, while task objectives and grading remain fixed. Agents encounter researcher preferences only through conversation. Across eighteen modelconfigurations, task performance, fact recovery, and self-repetition reveal different strengths. Detailed GPT 6 Astra analyses show distinct agent responses across researcher profiles; interaction traces illustrate targeted clar ification and workflow revision. These behaviors can differ even when final scores match. PersonaScienceBench therefore evaluates how agents collaborate with diverse researchers alongside whether they complete the task.
World action models (WAMs) increasingly couple internal visual or world representations with action generation, making their world-to-action pathways a critical yet poorly understood source of robustness. We investigate when perturbations to these internal representations actually become behaviorally consequential, and uncover a pronounced mismatch between representation perturbability and closed-loop vulnerability: attack effectiveness varies sharply across visual K/V routes, while larger representation displacement does not reliably imply stronger control failure. Our analysis attributes this mismatch to the action-conditioned readout, where keys determine where action queries read and values determine what content they receive. This motivates an action-facing view of vulnerability, in which the effect of a perturbation depends on how it is routed into action generation rather than on representation-space magnitude alone. We instantiate this view with Tail-AV, a short-probe signal for identifying vulnerable routes, and operationalize it in PASA-KV, which selects action-facing K/V targets before training a fresh task-shared attack. Experiments across multiple WAM systems and a physical robot validate readout-informed target selection, demonstrate strong closed-loop attacks, and reveal substantial system-level differences in resistance.
One-step distillation turns pretrained diffusion models into fast generators, but strong results often require substantial training machinery and compute. Pure adversarial distillation offers a simpler route, yet has struggled at modern text-to-image scale. To address this limitation, we propose ROAD (Representation-grounded One-step Adversarial Distillation), a one-stage framework that distills text-to-image models with one adversarial objective and a lightweight representation-space critic. First, we replace the classical convolutional head with a Lipschitz-controlled Transformer head with token-level text conditioning, which matches the semantic tokens of modern visual encoders. Second, we observed a fixed random feature mixing leads to long-horizon degradation. We attribute it to generator-critic co-adaptation around one realization-specific view, and periodically re-draw the mixing to keep the randomization effective. Third, for our dense and vector-valued logits, we introduce Jacobian R1 which penalizes the full input-logit Jacobian, overcoming pooled R1's limitation of neglecting most of logit sensitivities. A single adversarial objective then suffices: in one step, ROAD reaches 11.03 FID on SDXL and 13.73 on SD3.5-M on COCO-10K, surpassing the multi-step teachers at a fraction of distribution matching's training compute.
Ptychography is a computational imaging method that reconstructs a sample from overlapping diffraction measurements. This technique relies on these overlapping measurements to provide the redundancy information needed for robust reconstruction, but dense scanning increases acquisition time and data volume. Reducing scan overlaps can improve measurement efficiency but weakens the physical constraints available for reconstruction. Diffusion models provide strong learned prior information, offering a promising approach to compensate for the loss of measurement redundancy. However, effectively utilizing this prior information under weak measurement constraints remains a challenge. We propose a progressive coupling posterior sampling (PCPS) method, which allows for a broader exploration of the solution space under high uncertainty while maintaining stronger diffusion-guided consistency as the reconstruction becomes more refined. Experiments on real X-ray ptychographic data at various scan overlaps demonstrate that PCPS consistently improves reconstruction accuracy compared to traditional iterative methods and existing diffusion posterior samplers, especially at low overlaps. These results show that diffusion priors can reduce the measurement redundancy required for accurate ptychography reconstruction.
As recommendation scenarios become increasingly diverse and multimodal information grows more prevalent, effectively modeling cross-scenario user preferences and multimodal semantics has become a critical challenge in recommender systems. When incorporating multimodal information, existing multi-scenario recommendation (MSR) methods typically treat textual and visual representations as generic input features and directly feed them into their original parameter-sharing architectures. However, this extension may not fully accommodate the different modeling requirements of collaborative and multimodal information and may not be the best modeling paradigm. Our controlled experiments show that directly adopting these architectures is not the optimal modeling approach. To this end, we propose Dual-Branch Modeling with Triadic Factorization (DuTri), a framework for multimodal multi-scenario recommendation comprising a collaborative branch and a multimodal branch. Specifically, the collaborative branch learns structured collaborative preference representations through CANDECOMP/PARAFAC (CP) decomposition of user-scenario-item interactions and residual quantization, thereby capturing cross-scenario commonalities and scenario-specific differences. The multimodal branch leverages the learned user and scenario preferences to guide adaptive modality fusion; additionally, item-level contrastive learning aligns collaborative item preference representations with multimodal representations to facilitate effective collaboration between the two branches. Extensive experiments on two real-world datasets, AntM$^2$C and Amazon, demonstrate that DuTri outperforms existing multimodal and multi-scenario recommendation baselines in overall recommendation performance, validating the effectiveness of the proposed framework.
Wentao Huang, Qiming Huang, Yunhong Ji, Mian Lu, Yuqiang Chen, Bingsheng He, Kian-Lee Tan
Sparse general matrix-matrix multiplication (SpGEMM) is a key kernel in graph analytics, relational queries, and machine learning. Yet, its performance remains a significant challenge, due to the inherent sparsity and, more critically, the prevalent skew patterns often found in real-world workloads. This paper investigates this problem and demonstrates that an unbalanced workload distribution is the main culprit, harming performance through two co-existing root causes: first, multi-dimensional skew, where imbalance across computation cost, output density, and data locality cannot be captured by any single balancing metric; and second, cache contention, where dense substructures induce severe on-chip interference. To that end, we propose adaptive morsel scheduling (AMS), a dynamic scheduling framework that decouples skew dimensions via fixed-result-size morsels and integrates contention-aware load balancing (CALB) and data-conscious simultaneous multithreading activator (SMTact) to alleviate cache-level contention, alongside a hybrid accumulation mechanism (HYB) that adaptively selects the best accumulator per morsel. Evaluations across matrix squaring, relational queries, graph analytics, and other sparse matrix kernels show that AMS significantly outperforms state-of-the-art methods on skewed workloads while maintaining comparable performance on non-skewed ones. These gains, along with improved bandwidth utilization, remain valid across a wide range of workloads and hardware systems.
Runze Jiang, Yan Wang, Boran Jin, Bingyu Zhu, Longtao Huang
Knowledge improvement in large language models is shifting from predefined knowledge incorporation pipelines toward evaluation-guided updating. Existing methods use evaluation to localize model-specific knowledge weaknesses and guide targeted post-training. However, identifying what to improve does not necessarily provide sufficient information for deciding how to improve it. The same localized weakness can be compatible with distinct operational knowledge states: the relevant knowledge may be unavailable, available but difficult to access reliably, or accessible but ineffective for downstream use. We call this ambiguity knowledge-state aliasing. When compatible states favor different interventions, greater diagnostic resolution alone is insufficient for intervention selection. Moreover, model updates can change the underlying knowledge state before the observed weakness disappears, such that an intervention appropriate in one round may no longer address the bottleneck in the next. To address these challenges, we propose Recursive Knowledge Improvement (RKI), which formulates knowledge improvement as a sequential, state-aware intervention process. RKI organizes knowledge into graph-structured units that provide a shared basis for diagnostic evaluation and targeted intervention across rounds. At each round, RKI localizes knowledge weaknesses, estimates the model's operational knowledge state from diagnostic evidence and intervention history, selects a corresponding intervention, and re-evaluates the updated model before determining the next intervention. On HotpotEval and SafeInternalEval, RKI outperforms existing knowledge incorporation methods by 18.41 points on average. Further analyses show that knowledge states provide additional guidance for intervention selection beyond weakness localization and can change across model updates even when the observed weakness persists.
Computer-use agents (CUAs) are commonly evaluated against fixed task specifications, but users may omit requirements, add constraints, or require approval before consequential actions. Completing the benchmark instruction does not guarantee serving the person who asked. We introduce PersonaCUABench, which holds 74 tasks across 11 domains fixed and varies only who asks. A persona agent, playing one of 44 personas in four groups, rewrites each instruction in that user's own words and answers the CUA's questions, while the CUA sees only the dialogue, never the profile or the original instruction. Across nine backbones (3,330 runs), mean task completion falls by 36% relative to the original requests. The losses have two sources: in 59.9% of cells that lose points, the agent engages with what the user asked for and still gets it wrong, and 11.2% are stops the user requested that the checker scores as failures. Even when a persona request is fully completed, the user gives the top satisfaction rating in only 62.7% of cells, against 84.1% for the original requests. CUAs should therefore be evaluated not only on task completion but also on whether they meet the requirements and constraints users express. Code: https://github.com/MatrAIx-ai/PersonaCUABench. Project page: https://matraix.ai/research/persona-cua.html
Haolei Xu Zhejiang University, Haiwen Hong Alibaba Group, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang Alibaba Group, Hui Xue, Yongliang Shen Zhejiang University 等 10 位作者
Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerBench, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Ziwen Xu Zhejiang University, Chenyan WU Zhejiang University of Technology, Hengyu Sun, Haiwen Hong Alibaba Group, Mengru Wang, Yunzhi Yao Zhejiang University, Longtao Huang Alibaba Group, Hui Xue 等 12 位作者
Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model’s valid-generation manifold. Finally, we introduce a new steering approach guided by this analysis that improves preference while better preserving utility.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yumeng Fu, Weitao Huang Harbin Institute of Technology, Junjie Wu Suzhou University, Hao Teng Harbin Institute of Technology, Meishan Zhang Harbin Institute of Technology (Shenzhen), China, Bingquan Liu Harbin Institute of Technology
Emotion Recognition in Conversation (ERC), the task of identifying the emotion of each utterance in a conversation, is crucial for human-machine interaction. Existing LLM-based ERC methods focus on standard prompting and slow thinking for emotion analysis. However, they suffer from the lack of human-like emotion reasoning and discrimination between similar emotions, thus limiting accurate emotion predictions. To this end, we present JoPR, jointing perception-curriculum learning and emotional reasoning for conversational emotion recognition. Specifically, we devise a multi-dimension curriculum with long CoT fine-tuning to clone human-like emotion reasoning. We further design an emotion-specific reward function in a novel reinforcement learning framework, thereby enhancing the discernment between similar emotions. Our proposal is extensively evaluated over three widely used benchmark datasets, and experimental results confirm the superiority of JoPR. Furthermore, we provide an in-depth analysis to confirm the emotion perception and reasoning capabilities of JoPR.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yumeng Fu, Weitao Huang Harbin Institute of Technology, Junjie Wu Suzhou University, Hao Teng Harbin Institute of Technology, Shouduo Shang, Meishan Zhang Harbin Institute of Technology (Shenzhen), China, Bingquan Liu Harbin Institute of Technology
Emotion Recognition in Conversation (ERC) aims to identify the emotional states of speakers in conversations. Existing ERC methods perform either fast thinking or slow thinking for emotion predictions. The former lacks interpretability of emotion predictions, and the latter focuses on emotion analysis at shallow semantics. Such insufficient reasoning chains fall short in capturing deep semantics within conversations. To address these limitations, we propose ERCThinker, a Fast-Slow thinking framework for the task of ERC. First, we design different thinking strategies with fine-grained emotion reasoning chains to capture deep semantics that contain topic, discourse structure, speaker characteristic, scene, and emotion shift. Second, we develop an adaptive thinking mechanism in both strategy-level and utterance-level, guiding the model to dynamically perform thinking switching across various scenarios. Furthermore, we utilize Agent-as-Judge to score reasoning chains as reward signals for more accurate emotion predictions. To support training, we construct EmotionCueCoT, the emotion reasoning dataset with supervision in both explanation and judgment. Extensive experiments on various ERC benchmark datasets demonstrate that ERCThinker achieves state-of-the-art performance in both explanation and judgment, making progress in the realm of ERC.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表