iFLYTEK（科大讯飞）

Vision-language agents (VLAs) are increasingly used to interpret complex driving scenes and support safety-critical reasoning. This report presents the CVPR 2026@AdvML Workshop Challenge on adversarial multimodal attacks against autonomous-driving VLAs. Built on DriveLM-style multi-view visual question answering, the challenge represents each scene with six synchronized camera images and a structured collection of driving-related question-answer pairs. Participants generate adversarial images and suffix-only textual perturbations that induce model responses to deviate from reference answers while preserving image fidelity and limiting textual cost. The competition comprises two phases, with Phase II adding a hidden black-box model to assess transferability. We describe the task design, submission rules, evaluation protocol, and leaderboard results, and then examine five leading submissions for which technical reports were available. Across these reports, several recurring patterns emerge: image-side attacks are favored by the suffix penalty; scene-level, multi-view optimization is more effective than treating views in isolation; QA types and graph structure provide useful priors for allocating attack budget; feature-space objectives can improve black-box transfer; and typographic content embedded in camera images exposes a persistent vulnerability in driving VLAs. These findings provide a practical reference for future robustness evaluation and defense design in multimodal autonomous-driving systems.

Retrieval-Augmented Generation (RAG) streamlines long-document understanding by leveraging retrieval mechanisms to restrict input images to a highly curated subset. However, existing multimodal RAG pipelines primarily face two critical challenges: first, standard semantic similarity retrievers frequently fetch topically overlapping yet answer-void distractor pages that mislead downstream generation; second, rigid single-pass pipelines heavily depend on initial retrieval success, where any omission of core evidence inevitably causes cascading errors. To address these challenges, we introduce HIEVI-RAG, a hierarchical, evidence-driven multimodal RAG framework for closed-domain document understanding. HIEVI-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (1) hierarchical question decomposition to break multi-hop root queries into atomic child questions; (2) coarse visual page retrieval leveraging a multimodal retriever to fetch candidate pages based on semantic similarity; (3) fine-grained page verification via EVIAGENT, a specialized multi-page verifier trained with GRPO to execute cross-page reasoning over multi-image blocks; and (4) memory-guided iterative generation that leverages accumulated sub-question context to execute multi-round, dynamic reasoning over the prioritized sequence. Extensive evaluations across four benchmarks demonstrate the robust efficacy and synergy of our framework, which significantly outperforms existing open-source baselines and exceeds the strongest reported baseline by an average of 8.05% in accuracy.

General-purpose embodied agents must understand multimodal instructions, anticipate how their environment will evolve, and produce precise control actions over extended horizons. Existing approaches typically specialize in visual-language reasoning, video-based world modeling, or action generation, while cascaded pipelines that first synthesize future observations and then infer actions can introduce interface bottlenecks and compound prediction errors. We present iFLYTEK-Embodied-Omni, a unified multimodal foundation model that jointly models vision(videos and images), language, and action within a single Omni framework. Its modality-specific visual-language, video-generation, and action-generation components communicate through shared multimodal self-attention. This design establishes brain-cerebellum collaboration: the vision-language modeland video generation model form a high-level brain for instruction understanding, task planning, progress tracking, and future visual-state prediction, whereas the action generation modelserves as a low-level cerebellum that directly converts planned subgoals and shared multimodal context into executable action chunks. To develop these capabilities, we combine action-annotated and action-free embodied videos from human demonstrations and robot interactions with embodied reasoning, embodied perception, and general-purpose image-text data to construct a comprehensive dataset. We further adopt a four-stage strategy that progressively trains the VLM, VGM, and AGM before jointly fine-tuning the complete model.

Multi-Behavior Recommendation (MBR) leverages multiple user interaction types (e.g., views, clicks, purchases) to enrich preference modeling and alleviate data sparsity issues in traditional single-behavior approaches. However, existing MBR methods face fundamental challenges: they lack principled frameworks to model complex confounding effects from user behavioral habits and item multi-behavior distributions, struggle with effective aggregation of heterogeneous auxiliary behaviors, and fail to align behavioral representations across semantic gaps while accounting for bias distortions. To address these limitations, we propose MCLMR, a novel model-agnostic causal learning framework that can be seamlessly integrated into various MBR architectures. MCLMR first constructs a causal graph to model confounding effects and performs interventions for unbiased preference estimation. Under this causal framework, it employs an Adaptive Aggregation module based on Mixture-of-Experts to dynamically fuse auxiliary behavior information and a Bias-aware Contrastive Learning module to align cross-behavior representations in a bias-aware manner. Extensive experiments on three real-world datasets demonstrate that MCLMR achieves significant performance improvements across various baseline models, validating its effectiveness and generality. All data and code will be made publicly available. For anonymous review, our code is available at the following the link: https://github.com/gitrxh/MCLMR.

Large language models (LLMs) have shown remarkable capabilities in natural language processing. However, in knowledge graph question answering tasks (KGQA), there remains the issue of answering questions that require multi-hop reasoning. Existing methods rely on entity vector matching, but the purpose of the question is abstract and difficult to match with specific entities. As a result, it is difficult to establish reasoning paths to the purpose, which leads to information loss and redundancy. To address this issue, inspired by human reverse thinking, we propose Ontology-Guided Reverse Thinking (ORT), a novel framework that constructs reasoning paths from purposes back to conditions. ORT operates in three key phases: (1) using LLM to extract purpose labels and condition labels, (2) constructing label reasoning paths based on the KG ontology, and (3) using the label reasoning paths to guide knowledge retrieval. Experiments on the WebQSP and CWQ datasets show that ORT achieves state-of-the-art performance and significantly enhances the capability of LLMs for KGQA.

Instruction-based controllable speech synthesis enables users to specify emotions through natural language. However, existing approaches often rely on coarse emotion labels and lack explicit modeling of fine-grained intensity. We propose EmoInstruct-TTS, a dual-path instruction-guided framework for emotional speech synthesis. We introduce Emotion2embed, a supervised semantic-acoustic emotion embedding covering 48 emotional states, including fine-grained categories and intensity levels. To infer embeddings from free-form instructions, we design an Instruction-Conditioned Emotion Flow Model (ICE-Flow) that generates acoustically grounded emotion representations. The inferred embeddings are integrated into an LLM-based synthesis pipeline to provide explicit emotional control while preserving semantic planning. Experiments show improved emotional controllability and speech naturalness over strong baselines.

Large Language Models (LLMs) excel at complex reasoning through search algorithms, yet current strategies often suffer from massive token consumption due to redundant exploration of semantically equivalent steps. Existing semantic similarity methods struggle to accurately identify such equivalence in domain-specific contexts like mathematical reasoning. To address this, we propose EquivPruner, a simple yet effective approach that identifies and prunes semantically equivalent actions during LLM reasoning search. We also introduce MathEquiv, the first dataset we created for mathematical statement equivalence, which enables the training of a lightweight equivalence detector. Extensive experiments across various models and tasks demonstrate that EquivPruner significantly reduces token consumption, improving searching efficiency and often bolstering reasoning accuracy. For instance, when applied to Qwen2.5-Math-7B-Instruct on GSM8K, EquivPruner reduced token consumption by 48.1\% while also improving accuracy. Our code is available at https://github.com/Lolo1222/EquivPruner.

Time series forecasting plays a vital role in supporting decision-making across a wide range of critical applications, including energy, healthcare, and finance. Despite recent advances, forecasting accuracy remains limited due to the challenge of integrating historical numerical sequences with contextual features, which often comprise unstructured textual data. To address this challenge, we propose TokenCast, a large language model (LLM) driven framework that leverages language-based symbolic representations as a unified intermediary for context-aware time series forecasting. Specifically, TokenCast employs a discrete tokenizer to transform continuous numerical sequences into temporal tokens, enabling structural alignment with language-based inputs. To effectively bridge the semantic gap between modalities, both temporal and contextual tokens are embedded into a shared representation space via a pre-trained LLM, further optimized with generative objectives. Building upon this unified semantic space, the aligned LLM is subsequently fine-tuned in a supervised manner to predict future temporal tokens, which are then decoded back into the original numerical space. Extensive experiments on real-world datasets demonstrate the effectiveness of our framework and highlight its potential as a generative framework for context-aware time series forecasting. The code is available at https://github.com/Xiaoyu-Tao/TokenCast.

Neural operator models trained on simulation data often lose accuracy when applied to experimental measurements due to the sim-to-real gap. Standard fine-tuning with limited real data can reduce this gap, but it may also damage the core physics-relevant representations learned during pretraining. Although knowledge-preserving adaptation has been widely investigated in vision or language tasks, it remains unclear whether these methods are suitable for neural operators whose architectures and protected knowledge are fundamentally different. Neural operators need to preserve core-scale physical structures rather than semantic or visual features. We propose PhysGuard, a physics-preserving framework for accurate sim-to-real adaptation of neural operators. Specifically, PhysGuard uses the empirical Fisher Information Matrix computed on simulation data to identify physics-critical parameter directions, then restricts fine-tuning updates to directions that do not interfere with them. A layer-wise Gram-matrix formulation makes this efficient for models with millions of parameters, while an adaptive threshold automatically determines the protected subspace size. A spectral probe experiment shows that the dominant Fisher directions are strongly associated with low-frequency output structures. Experiments on benchmark across four neural operator architectures and different physical systems show that PhysGuard performs strongly on most evaluation metrics compared to baselines. The benefits are most evident under severe domain shift, where it reduces low-frequency error by up to 32\% compared to standard fine-tuning while maintaining adaptability. Our code is available at https://github.com/ZhouChaunge/PhysGuard.

Personalized text-to-speech (TTS) aims to clone the target speaker in the synthesized speech, imitating both the voice and speaking style. Current large language model (LLM)-based TTS methods ignore the style-specific prosodic patterns in generated speech, resulting in deficient style learning and thus limiting speaker similarity in synthesized speech. To this end, we investigate the prosody learning conditioned on the synthesized speech, and propose to predict the prosody of the current syllable based on previously predicted speech. Experimental results obtained on three datasets demonstrated the efficacy of the proposed dynamic prosody prediction method in enhancing the prosody learning capability, thereby improving the speaker similarity of the generated speech. Audio samples are available at https://muzw.github.io/dynapros/.

Large-scale learner-task interaction data are crucial for intelligent educational systems but are costly to collect and constrained by privacy and learner engagement. Learner simulators play a critical role in simulating scalable learner behavior without the need for continuous involvement of real learners. However, existing methods are predominantly \textbf{individual-centric}, pairing a simulator with each learner to iteratively infer latent knowledge states from dense interaction histories, which is both data- and computation-intensive, and fragile in cold-start scenarios. We propose a \textbf{cohort-aware roll-call simulation paradigm} that first constructs cohort-level proficiency priors and refines individual learner states through a small number of targeted diagnostic queries. Based on this paradigm, we introduce \textbf{Edu-Theater}, an LLM-powered agent system that performs cohort-aware learner simulation via a teacher agent and retrospective roll-call probing over learner logs. Edu-Theater enables scalable future behavior simulation without the need for dense per-learner histories. Experiments on two real-world datasets demonstrate that Edu-Theater achieves higher simulation accuracy with significantly fewer LLM calls, producing synthetic data that enhances downstream applications such as adaptive testing.

Vision-Language-Action (VLA) models achieve strong benchmark performance but still struggle in real-world deployment with unseen objects, background shifts, and different robot embodiments. We argue that this stems from the lack of a unified geometry-aware manipulation representation, leaving existing VLAs vulnerable to low-level trajectory supervision, misaligned 3D features, and embodiment differences. To address this, we propose GEAR-VLA, a VLA framework for learning unified geometry-aware action representations for generalizable robotic manipulation. GEAR-VLA adopts coarse-to-fine action learning, where multi-source embodied pretraining equips the VLM with embodied reasoning and discrete action understanding before latent action tokens connect action semantics to a gradient-decoupled DiT continuous action expert. It further performs semantic-aligned 3D integration by aligning a trainable 3D spatial backbone with the VLA representation while freezing the original VLM-aligned visual pathway. To share this representation across robots, GEAR-VLA uses embodiment canonicalization, where embodiment-aware states and embodiment-invariant actions confine robot differences to the low-level interface. Extensive simulation and real-world experiments demonstrate strong generalization: GEAR-VLA achieves state-of-the-art performance on LIBERO, zero-shot LIBERO-Plus, and RoboTwin 2.0, reaches 85.9% success on AgileX and 81.0% on the pretraining-unseen LDT-01 embodiment, and obtains 90.1% success on a 6,360-trial universal grasping benchmark with 212 unseen objects. Code and models will be released at https://github.com/babynabeauty/GEAR-VLA.

Reinforcement learning (RL) for large language models usually supervises reasoning with scalar outcome rewards, such as binary correctness. Such rewards provide an optimization direction but rarely explain how a model should revise its mistaken reasoning, which can encourage shortcut learning and brittle policies. We propose \textbf{SocraticPO} (Socratic Policy Optimization), a policy-optimization framework that augments RL rollouts with Socratic-style natural-language guidance. During rollout, the student first answers independently; if the answer is incorrect, a teacher diagnoses the attempt and provides concise corrective guidance, after which the student continues under the expanded context. Crucially, this guidance is paired with reward decay: correct answers obtained after teacher intervention only receive decayed rewards, preventing the policy from treating teacher help as a free path to reward. Since SocraticPO only modifies the rollout process while leaving the standard expected-reward objective intact, it can be plugged into existing policy-gradient backends such as Reinforce++. Moreover, because the teacher provides only text-level guidance, SocraticPO can leverage stronger black-box teacher models without requiring access to logits or distribution matching. On undergraduate-level scientific reasoning benchmarks from SciKnowEval, SocraticPO improves over strong RL and self-distillation baselines. Ablations show that both targeted guidance and reward decay are necessary, with reward decay mitigating reliance on assisted correction.

Ambient clinical scribes increasingly combine Automatic Speech Recognition with Large Language Models to automate documentation. However, traditional metrics like Word Error Rate mask systemic safety degradation. We present a paired acoustic stress test to isolate the causal impact of noise on clinical reasoning. For the same dialogues, we inject diverse noise types while keeping the downstream model configuration frozen. Crucially, we uncover a dangerous disconnect between signal fidelity and clinical safety. Stationary ambient noise increased the Word Error Rate by a negligible 0.71 percentage points yet nearly doubled the rate of unsafe outputs. Our analysis reveals that minor acoustic perturbations can invert clinical meaning without substantially inflating error rates. Furthermore, we demonstrate a lightweight mitigation strategy that mitigates safety degradation under noisy conditions without requiring model fine tuning.

Large language models (LLMs) are prone to hallucinate unintended text due to false or outdated knowledge. Since retraining LLMs is resource intensive, there has been a growing interest in model editing. Despite the emergence of benchmarks and approaches, existing unidirectional editing and evaluation paradigms have failed to explore the reversal curse. In this paper, we study bidirectional language model editing, aiming to provide a rigorous evaluation to assess if edited LLMs can recall the editing knowledge bidirectionally. A metric of reverse generalization is introduced and a benchmark dubbed Bidirectional Assessment for Knowledge Editing (BAKE) is constructed to evaluate if post-edited models can recall the edited knowledge in the reverse direction of editing. We conduct extensive experiments using a variety of editing methods and LLMs. The results show that while most editing methods are able to accurately recall editing facts along the modification direction, they exhibit substantial systematic deficiencies when evaluating in the reverse direction. To further investigate the underlying causes of reversal curse and to explore potential strategies for mitigation, a detailed analysis is conducted from three perspectives. Our findings reveal that although In-Context Learning (ICL) can mitigate the reversal curse to a certain extent, it lacks continuity, is limited by the input length, and may introduce hallucinations. Therefore, combining the advantages of ICL and other editing methods is a promising direction for developing new editing paradigms.

Learning transferable and interpretable representations from medical volumetric scans remains challenging due to complex anatomical structures and weak, heterogeneous supervision provided by radiology reports. In this paper, we propose Anatomy-aware Semantically-Adaptive Pre-training (ASAP), a principled vision-language pre-training framework for fine-grained medical volumetric representation learning from large-scale chest CT scans and their corresponding radiology reports. ASAP integrates three key components: (1) an anatomy-aware knowledge injection module that incorporates organ-level structural priors via off-the-shelf segmentation tool to encourage anatomically coherent representations; (2) a semantically-adaptive selective alignment mechanism that dynamically associates sentence-level findings with localized volumetric regions; and (3) a semantically-adaptive fusion module for effective interaction between anatomically informed visual features and grounded textual cues under dual-modal masked modeling paradigm. Beyond methodological contributions, we establish a comprehensive benchmark for medical volumetric vision-language pre-training on chest CT, covering 15 datasets and 22 downstream tasks spanning abnormality classification, segmentation, disease prognosis prediction, report generation, vocabulary classification, cross-modal retrieval and visual question answering. This benchmark provides standardized evaluation protocols to systematically assess representation quality under diverse clinical settings and data regimes. Extensive experiments demonstrate that ASAP consistently achieves state-of-the-art performance across tasks and datasets, with particularly pronounced gains under limited supervision and distribution shift, validating its effectiveness in learning transferable and clinically meaningful volumetric representations.

Large Language Models (LLMs) have shown promise as educational tutors, yet effective tutoring requires more than solving problems: it must provide progressive Socratic guidance and balance multiple pedagogical objectives across multi-turn interactions. However, training such tutors remains challenging due to limited-fidelity and weakly controllable student simulation, under-specified pedagogical reward modeling, and unstable multi-objective optimization. To overcome these limitations, we propose PEARL, a pedagogically aligned reinforcement learning framework for training Socratic tutoring agents, consisting of three key components. First, we introduce a controllable student simulator that decouples latent cognitive states from response generation to model diverse abilities and misconceptions. Second, we develop a generative reward model that jointly evaluates pedagogical quality and objective correctness for policy optimization. Finally, we propose a stable multi-objective RL scheme that discretizes rewards within each dimension and aggregates normalized advantages across dimensions, preventing high-variance objectives from dominating updates. Experiments on multiple benchmarks show that PEARL achieves the best performance among open-source models and remains competitive with leading proprietary LLMs, despite using only a 30B policy model.

Prompt-based verification is widely used to mitigate hallucinations in large vision-language models (LVLMs), yet when it helps remains poorly understood. We systematically study verification prompting across two representative LVLM architectures and hallucination benchmarks, and find that it is a risk-bearing intervention: its corrections increase with input difficulty, while newly introduced errors persist across difficulty levels. As a result, always-on prompting helps on hard inputs but offers little benefit -- and can harm -- easier ones. Our analysis further shows that this behavior is associated with a conservative output shift. Verification prompts redistribute attention from visual tokens toward instruction tokens and induce a distinct middle-layer entropy pattern absent in a neutral-prompt control, suggesting instruction-conditioned attention redistribution rather than uniformly improved visual grounding. Motivated by this input-dependent risk, we propose Risk-aware Selective Prompting (RSP), a training-free approach that uses pre-generation uncertainty signals to trigger verification selectively. RSP mitigates the degradation of always-on prompting while preserving baseline performance, and reveals that effective selection signals vary across architectures.

Parsing chemical reaction diagrams from scientific literature is challenging due to heterogeneous layouts, intertwined visual elements, and the difficulty of integrating recognition and reasoning. Existing vision-language models advance multimodal understanding but still fail on complex diagrams, struggling to maintain spatial coherence and to integrate multidimensional information during reasoning. To address these issues, we propose MACReD, a hierarchical multi-agent framework that coordinates specialized agents for molecular perception, arrow understanding, text extraction, and reaction reconstruction within a unified VLM-guided architecture. The planning and perception layers use flexible, fine-grained detection to handle visual complexity, while the reasoning layer uses a multigraph fusion mechanism to integrate heterogeneous cues and enforce chemically consistent global reasoning. Experiments on the RxnScribe benchmark show that MACReD achieves state-of-the-art performance, with F1 scores of 75.2% and 84.6% under hard and soft match criteria, outperforming the RxnScribe baseline, which obtains 69.1% and 80.0%, respectively. These results demonstrate the robustness of MACReD across diverse diagram layouts, including multi-step and tree-structured reactions.

We study training-free image editing on Qwen-Image-Edit-2511, a 60-block multi-modal diffusion transformer (MMDiT) that concatenates noise and source-image tokens within a single attention stream. We make three contributions. (i) We introduce KVInject, a single-forward attention manipulation that alpha-blends source-half key/value projections into the noise-half within a localized layer/step band. KVInject is simpler than the classical two-pass MasaCtrl recipe and avoids the prompt-mismatch failure mode that disables MasaCtrl on MMDiT (composite score drops 31% versus baseline). (ii) We show that no single attention operation dominates across edit types, motivating AttnRouter, a per-category routing table that dispatches edits to the operation that best preserves source structure for that type. With ground-truth categories the router improves the CLIP-T+DINO-I composite by 6.4% over the editing baseline; an automatic CLIP zero-shot classifier closes 98% of this gap despite only 55% category accuracy. (iii) Through layer-, step-, and alpha-band ablations we localize the editing-effective attention sub-circuit: K/V injection in early denoising steps (S0-7) recovers nearly all of the gain of full-step injection, while injection in early (L0-15) or late (L45-60) layer bands fails to drive editing entirely; alpha in [0.3, 0.5] is a stable sweet spot. We also report negative results that highlight what does not transfer from the UNet folklore: simple K/V rescaling never beats baseline and aggressive variants collapse generation entirely (composite 0.084). We release code, pre-computed routing tables, and a 100-sample stratified subset of ImgEdit-Bench used in all ablations.

Image editing instructions are heterogeneous: a color swap, an object insertion, and a physical-action edit all demand different spatial coverage and different reasoning depth, yet existing reasoning-based editors apply a single fixed inference recipe to every instruction. We argue that adaptivity along both the spatial and temporal axes is the missing degree of freedom, and we present PhysEdit, an editing framework built around this principle. PhysEdit introduces two inference-time modules that compose without retraining the backbone. At its core, (1) Complexity-Adaptive Reasoning Depth (CARD) predicts edit complexity directly from the instruction and reference image and allocates the reasoning step count N_r and reasoning-token length r per sample -- turning a previously fixed inference schedule into a conditional-computation problem. CARD is supported by (2) a Spatial Reasoning Mask (SRM) that extracts an instruction-conditioned spatial prior from cross-attention to confine reasoning to regions that semantically require it. On the full 737-case ImgEdit Basic-Edit Suite, PhysEdit delivers a 1.18x wall-clock speedup (64.3s vs. 76.1s per sample) over a strong reasoning baseline while slightly improving instruction adherence (CLIP-T 0.2283 vs. 0.2266, +0.7%) and matching identity preservation within noise (CLIP-I 0.8246 vs. 0.8280). The speedup is category-dependent and reaches 1.52x on appearance-level edits, validating CARD's adaptive allocation as the principal source of efficiency gain. A 30-sample pilot with full ablations isolates the contribution of each module.

Although multi-step generative policies achieve strong performance in robotic manipulation by modeling multimodal action distributions, they require multi-step iterative denoising at inference time. Each action therefore needs tens to hundreds of network function evaluations (NFEs), making them costly for high-frequency closed-loop control and online reinforcement learning (RL). To address this limitation, we propose a two-stage framework for native one-step generative policies that shifts refinement from inference to training. First, we introduce the Drift-Based Policy (DBP), which leverages fixed-point drifting objectives to internalize iterative refinement into the model parameters, yielding a one-step generative backbone by design while preserving multimodal action modeling capacity. Second, we develop Drift-Based Policy Optimization (DBPO), an online RL framework that equips the pretrained backbone with a compatible stochastic interface, enabling stable on-policy updates without sacrificing the one-step deployment property. Extensive experiments demonstrate the effectiveness of the proposed framework across offline imitation learning, online fine-tuning, and real-world control scenarios. DBP matches or exceeds the performance of multi-step diffusion policies while achieving up to $100\times$ faster inference. It also consistently outperforms existing one-step baselines on challenging manipulation benchmarks. Moreover, DBPO enables effective and stable policy improvement in online settings. Experiments on a real-world dual-arm robot demonstrate reliable high-frequency control at 105.2 Hz.

Scene text editing (STE) has achieved remarkable progress in accurately rendering target text through diffusion-based methods. However, we identify a critical yet overlooked problem: edit spillover -- when editing a target text region, existing methods inadvertently modify non-target regions, particularly neighboring text. Through systematic evaluation on 50 real-world scenes across four categories, we reveal that state-of-the-art diffusion editing models exhibit a spillover rate of 94%, meaning nearly all non-target text regions are altered during editing. To address this, we propose the Edit Fidelity Field (EFF), a semantics-aware continuous field that controls per-pixel editing fidelity. Unlike binary masks, EFF leverages OCR-detected text regions to construct a four-zone field: Edit Core (fully editable), Transition Zone (smooth decay), Protected Zone (non-target text, explicitly locked), and Background (strictly preserved). EFF operates as a training-free, model-agnostic post-processing module applicable to any diffusion-based STE method. We further propose per-region spillover quantification, a novel evaluation protocol that measures edit leakage at each non-target text region individually. Experiments demonstrate that EFF reduces spillover rate from 94% to 25% while improving non-target region preservation by +91.4 dB PSNR.

Flow Matching models achieve state-of-the-art image generation quality but incur substantial inference cost due to iterative denoising through large Transformer networks. We observe that different layer groups within a Transformer exhibit markedly heterogeneous velocity dynamics: shallow layers are highly stable and amenable to aggressive caching, while deep layers undergo large velocity changes that demand full computation. Existing caching methods, however, treat the entire Transformer as a monolithic unit, applying a single caching decision per timestep and thus failing to exploit this heterogeneity. Based on this finding, we propose LayerCache, a layer-aware caching framework that partitions the Transformer into layer groups and makes independent, per-group caching decisions at each denoising step. LayerCache introduces an adaptive JVP span K selection mechanism that leverages per-group stability measurements to balance estimation accuracy and computational savings. We formulate a three-dimensional scheduling problem over timesteps, layer groups, and JVP span, and solve it with a greedy budget allocation algorithm. On Qwen-Image (1024x1024, 50 steps), LayerCache achieves PSNR 37.46 dB (+5.38 dB over MeanCache), SSIM 0.9834, and LPIPS 0.0178 (a 70% reduction over MeanCache) at 1.37x speedup, dominating all prior caching methods on the quality-speed Pareto frontier.

Powered by multimodal text-to-image priors, diffusion-based super-resolution excels at synthesizing intricate details; however, models trained on synthetic low-resolution (LR) and high-resolution (HR) image pairs often degrade when applied to real-world LR images due to significant distribution shifts. We propose Bird-SR, a bidirectional reward-guided diffusion framework that formulates super-resolution as trajectory-level preference optimization via reward feedback learning (ReFL), jointly leveraging synthetic LR-HR pairs and real-world LR images. For structural fidelity easily affected in ReFL, the model is directly optimized on synthetic pairs at early diffusion steps, which also facilitates structure preservation for real-world inputs under smaller distribution gap in structure levels. For perceptual enhancement, quality-guided rewards are applied to both synthetic and real LR images at the later trajectory phase. To mitigate reward hacking, the rewards for synthetic results are formulated in a relative advantage space bounded by their ground-truth counterparts, while real-world optimization is regularized via a semantic alignment constraint. Furthermore, to balance structural and perceptual learning, we introduce a dynamic fidelity-perception weighting strategy that emphasizes structure preservation at early stages and progressively shifts focus toward perceptual optimization at later diffusion steps. Extensive experiments on real-world SR benchmarks demonstrate that Bird-SR consistently outperforms state-of-the-art methods in perceptual quality while preserving structural consistency, validating its effectiveness for real-world super-resolution. Our code can be obtained at https://github.com/fanzh03/Bird-SR.

Audio-driven human video generation has achieved remarkable success in monologue scenarios, largely driven by advancements in powerful video generation foundation models. Moving beyond monologues, authentic human communication is inherently a full-duplex interactive process, requiring virtual agents not only to articulate their own speech but also to react naturally to incoming conversational audio. Most existing methods simply extend conventional audio-driven paradigms to listening scenarios. However, relying on strict frame-to-frame alignment renders the model's response to long-range conversational dynamics rigid, whereas directly introducing global attention catastrophically degrades lip synchronization. Recognizing the unique temporal Scale Discrepancy between talking and listening behaviors, we introduce a multi-head Gaussian kernel to explicitly inject this physical intuition into the model as a progressive temporal inductive bias. Building upon this, we construct a full-duplex interactive virtual agent capable of simultaneously processing dual-stream audio inputs for both talking and listening. Furthermore, we introduce a rigorously cleaned Talking-Listening dataset VoxHear featuring perfectly decoupled speech and background audio tracks. Extensive experiments demonstrate that our approach successfully fuses strong temporal alignment with deep contextual semantics, setting a new state-of-the-art for generating highly natural and responsive full-duplex interactive digital humans. The project page is available at https://warmcongee.github.io/beyond-monologue/ .

It is widely believed that tens of thousands of physical qubits are needed to build a practically useful quantum computer. Atom arrays formed by optical tweezers are among the most promising platforms for achieving this goal, owing to the excellent scalability and mobility of atomic qubits. However, assembling a defect-free atom array with ~ 10^4 qubits remains algorithmically challenging, alongside other hardware limitations. This is due to the computationally hard path-planning problems and the time-consuming generation of suffciently smooth trajectories for optical tweezer potentials by spatial light modulators (SLM). Here, we present a unified framework comprising two innovative components to fully address these algorithmic challenges: (1) a path-planning module that employs a supervised learning approach using a graph neural network combined with a modified auction decoder, and (2) a potential-generation module called the phase and profile-aware Weighted Gerchberg-Saxton algorithm. The inference time for the first module is nearly a size-independent constant overhead of ~ 5 ms, and the second module generates a potential frame with about 0.5 ms, a timescale shorter than the current commercial SLM refresh time. Altogether, our algorithm enables the assembly of an atom array with 10^4 qubits on a timescale much shorter than the typical vacuum lifetime of the trapped atoms.

With the growing demand for intelligent in-vehicle experiences, vehicle-based agents are evolving from simple assistants to long-term companions. This evolution requires agents to continuously model multi-user preferences and make reliable decisions in the face of inter-user preference conflicts and changing habits over time. However, existing benchmarks are largely limited to single-user, static question-answer settings, failing to capture the temporal evolution of preferences and the multi-user, tool-interactive nature of real vehicle environments. To address this gap, we introduce VehicleMemBench, a multi-user long-context memory benchmark built on an executable in-vehicle simulation environment. The benchmark evaluates tool use and memory by comparing the post-action environment state with a predefined target state, enabling objective and reproducible evaluation without LLM-based or human scoring. VehicleMemBench includes 23 tool modules, and each sample contains over 80 historical memory events. Experiments show that powerful models perform well on direct instruction tasks but struggle in scenarios involving memory evolution, particularly when user preferences change dynamically. Even advanced memory systems struggle to handle domain-specific memory requirements in this environment. These findings highlight the need for more robust and specialized memory management mechanisms to support long-term adaptive decision-making in real-world in-vehicle systems. To facilitate future research, we release the data and code.

Geometric Problem Solving (GPS) poses a unique challenge for Multimodal Large Language Models (MLLMs), requiring not only the joint interpretation of text and diagrams but also iterative visuospatial reasoning. While existing approaches process diagrams as static images, they lack the capacity for dynamic manipulation - a core aspect of human geometric reasoning involving auxiliary line construction and affine transformations. We present GeoSketch, a neural-symbolic framework that recasts geometric reasoning as an interactive perception-reasoning-action loop. GeoSketch integrates: (1) a Perception module that abstracts diagrams into structured logic forms, (2) a Symbolic Reasoning module that applies geometric theorems to decide the next deductive step, and (3) a Sketch Action module that executes operations such as drawing auxiliary lines or applying transformations, thereby updating the diagram in a closed loop. To train this agent, we develop a two-stage pipeline: supervised fine-tuning on 2,000 symbolic-curated trajectories followed by reinforcement learning with dense, symbolic rewards to enhance robustness and strategic exploration. To evaluate this paradigm, we introduce the GeoSketch Benchmark, a high-quality set of 390 geometry problems requiring auxiliary construction or affine transformations. Experiments on strong MLLM baselines demonstrate that GeoSketch significantly improves stepwise reasoning accuracy and problem-solving success over static perception methods. By unifying hierarchical decision-making, executable visual actions, and symbolic verification, GeoSketch advances multimodal reasoning from static interpretation to dynamic, verifiable interaction, establishing a new foundation for solving complex visuospatial problems.

Tables are pervasive in diverse documents, making table recognition (TR) a fundamental task in document analysis. Existing modular TR pipelines separately model table structure and content, leading to suboptimal integration and complex workflows. End-to-end approaches rely heavily on large-scale TR data and struggle in data-constrained scenarios. To address these issues, we propose TDATR (Table Detail-Aware Table Recognition) improves end-to-end TR through table detail-aware learning and cell-level visual alignment. TDATR adopts a ``perceive-then-fuse'' strategy. The model first performs table detail-aware learning to jointly perceive table structure and content through multiple structure understanding and content recognition tasks designed under a language modeling paradigm. These tasks can naturally leverage document data from diverse scenarios to enhance model robustness. The model then integrates implicit table details to generate structured HTML outputs, enabling more efficient TR modeling when trained with limited data. Furthermore, we design a structure-guided cell localization module integrated into the end-to-end TR framework, which efficiently locates cell and strengthens vision-language alignment. It enhances the interpretability and accuracy of TR. We achieve state-of-the-art or highly competitive performance on seven benchmarks without dataset-specific fine-tuning.