arXiv subjects

Chun Yuan

Publications and source records attributed to Chun Yuan.

At least 37 records · Page 2Linked to original sources

arXiv2026

AgentCVR: Active Multi-Agent Cross-Video Reasoning via Script-Simulated Reinforcement Learning

Cross-Video Reasoning (CVR) has emerged as a critical frontier in multimodal intelligence, requiring models to retrieve, align, and aggregate evidence distributed across multiple videos. Current Multimodal Large Language Models (MLLMs) often struggle with CVR, as simple single-pass strategies encode multiple videos into a shared compressed context, potentially obscuring rare but critical evidence. In this paper, we propose AgentCVR, a multi-agent framework that treats CVR as an active evidence-acquisition task. AgentCVR employs a Master Agent to iteratively coordinate specialized Visual and Audio Agents for targeted evidence extraction. To ensure efficient training, we introduce Script-Simulated RL, which optimizes the agent's policy with LLM-generated semantic scripts and a lightweight text-based simulator, bypassing costly multimodal inference during online exploration. Experimental results on a comprehensive CVR benchmark show that AgentCVR outperforms single-pass baselines and achieves comparable performance to state-of-the-art closed-source systems, particularly in complex cross-video alignment and localization. To ensure reproducibility, our code is available at https://github.com/wang-jh24/AgentCVR.

cs.CV↗

arXiv2026

VCap: Hypergeometric Rewards for Weak-to-Strong Visual Captioning

Visual captioning requires models to capture visual content faithfully while minimizing both omission and hallucination. As the dominant paradigm for captioning, MLLMs have achieved strong performance through scaling and high-quality data. Recently, RL has emerged as a key route to driving MLLMs toward higher precision and broader coverage, however, existing reward designs for captioning fail to provide fine-grained and reliable signals for factual verification, limiting their effectiveness. To address this, we propose VCap, a Witness-Adjudicator reward that pairs the reference caption (a witness) with the visual signal (an adjudicator). By explicitly verifying factual consistency between the reference and policy-generated captions grounded in the visual signal, VCap delivers a reward signal with hypergeometric-distribution-level precision for caption quality verification. This design enables effective learning even from imperfect references, facilitating weak-to-strong generalization in RL training. In our experiments, an 8B model trained with VCap outperforms open- and closed-source SOTA models on multiple image and video captioning benchmarks. Human evaluation further confirms its strong alignment with factual correctness. Additionally, VCap improves MLLM perceptual capability, generalizes across tasks, and surpasses best-of-N distillation, challenging prior assumptions about RLVR.

cs.CV↗

arXiv2026

Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation

Unified Multimodal Models (UMMs) integrate both visual understanding and generation within a single framework. Their ultimate aspiration is to create a cycle where understanding and generation mutually reinforce each other. While recent post-training methods have successfully leveraged understanding to enhance generation, the reverse direction of utilizing generation to improve understanding remains largely unexplored. In this work, we propose UniMRG (Unified Multi-Representation Generation), a simple yet effective architecture-agnostic post-training method. UniMRG enhances the understanding capabilities of UMMs by incorporating auxiliary generation tasks. Specifically, we train UMMs to generate multiple intrinsic representations of input images, namely pixel (reconstruction), depth (geometry), and segmentation (structure), alongside standard visual understanding objectives. By synthesizing these diverse representations, UMMs capture complementary information regarding appearance, spatial relations, and structural layout. Consequently, UMMs develop a deeper and more comprehensive understanding of visual inputs. Extensive experiments across diverse UMM architectures demonstrate that our method notably enhances fine-grained perception, reduces hallucinations, and improves spatial understanding, while simultaneously boosting generation capabilities.

cs.CV↗

arXiv2026

Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory

Scaling conditional memory offers a promising way to increase language-model capacity, but existing methods such as Engram learn large memory tables from scratch during pre-training, making memory scaling expensive and sometimes ineffective. We propose Memory Grafting, a conditional memory scaling method that utilizes frozen hidden states from a grafting model as conditional n-gram memory. Given frequent local n-grams, we run the grafting model offline, store final-token hidden representations as memory values, and let the recipient model retrieve them through exact longest-match suffix lookup. Retrieved memories are adapted by lightweight projections and gates, while a hash-based Engram fallback preserves coverage for unmatched contexts. Since the grafting model is only run offline and exact lookup has expected O(1) complexity with respect to memory-bank size, Memory Grafting expands external latent capacity with limited training and inference overhead. Experiments under matched recipient architectures and pre-training budgets show that Memory Grafting improves over both MoE and vanilla Engram baselines. In the 2.8B-scale setting, it improves the average benchmark score from 51.95 for MoE and 52.43 for vanilla Engram to 53.86. In the 0.92B-scale setting, all grafting-model variants improve over the baselines, with Qwen3.5-35B-A3B giving the strongest gains. These results suggest that pretrained models can serve as reusable constructors of external latent memory, providing a practical step toward scaling future language models beyond trainable parameters alone.

cs.CL↗

arXiv2026

AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents

Large language model (LLM) agents increasingly rely on external memory to support long-horizon interaction, personalized assistance, and multi-step reasoning. However, existing memory systems still face three core challenges: they often rely too heavily on semantic similarity, which can miss evidence crucial for user-centric understanding; they frequently store related experiences as isolated fragments, weakening temporal and causal coherence; and they typically use static memory granularities that do not adapt well to the requirements of different questions. We propose AdaMem, an adaptive user-centric memory framework for long-horizon dialogue agents. AdaMem organizes dialogue history into working, episodic, persona, and graph memories, enabling the system to preserve recent context, structured long-term experiences, stable user traits, and relation-aware connections within a unified framework. At inference time, AdaMem first resolves the target participant, then builds a question-conditioned retrieval route that combines semantic retrieval with relation-aware graph expansion only when needed, and finally produces the answer through a role-specialized pipeline for evidence synthesis and response generation. We evaluate AdaMem on the LoCoMo and PERSONAMEM benchmarks for long-horizon reasoning and user modeling. Experimental results show that AdaMem achieves state-of-the-art performance on both benchmarks. The code will be released upon acceptance.

cs.CL↗

arXiv2026

OmniAlpha: Aligning Transparency-Aware Generation via Multi-Task Unified Reinforcement Learning

Transparency-aware generation requires modeling not only RGB appearance but also alpha-based opacity and cross-layer composition, which are essential for tasks such as image matting, object removal, layer decomposition, and multi-layer content creation. However, existing RGBA-related methods remain largely fragmented, with separate pipelines designed for individual tasks. While a unified model is desirable, supervised fine-tuning alone is insufficient, as localized regression objectives cannot directly optimize the compositional fidelity, alpha-boundary precision, and structural consistency required for high-quality RGBA generation. To address this, we propose OmniAlpha, a unified multi-task reinforcement learning framework for transparency-aware generation and manipulation. OmniAlpha combines an end-to-end alpha-aware VAE and a sequence-to-sequence Diffusion Transformer, with a bi-directional layer axis in positional encoding to jointly model multiple RGBA inputs and outputs within a single forward pass. Built on a multi-task SFT cold start, it further performs GRPO-style post-training with layer-aware rewards defined on decoded RGBA outputs, enabling direct optimization of cross-layer coherence and fine transparency details. Experiments across five categories of transparency-aware tasks show that OmniAlpha consistently outperforms its unified SFT baseline and achieves strong performance against specialized expert models, including a 9.07% relative reduction in RGB L1 on layer decomposition and 74%/68% improvements over conventional matting tools on SAD/Grad for automatic matting.

cs.CV↗

arXiv2026

Task-Distributionally Robust Data-Free Meta-Learning

Data-Free Meta-Learning (DFML) aims to enable efficient learning of unseen few-shot tasks, by meta-learning from multiple pre-trained models without accessing their original training data. While existing DFML methods typically generate synthetic data from these models to perform meta-learning, a comprehensive analysis of DFML's robustness-particularly its failure modes and vulnerability to potential attacks-remains notably absent. Such an analysis is crucial as algorithms often operate in complex and uncertain real-world environments. This paper fills this significant gap by systematically investigating the robustness of DFML, identifying two critical but previously overlooked vulnerabilities: Task-Distribution Shift (TDS) and Task-Distribution Corruption (TDC). TDS refers to the sequential shifts in the evolving task distribution, leading to the catastrophic forgetting of previously learned meta-knowledge. TDC exposes a security flaw of DFML, revealing its susceptibility to attacks when the pre-trained model pool includes untrustworthy models that deceptively claim to be beneficial but are actually harmful. To mitigate these vulnerabilities, we propose a trustworthy DFML framework comprising three components: synthetic task reconstruction, meta-learning with task memory interpolation, and automatic model selection. Specifically, utilizing model inversion techniques, we reconstruct synthetic tasks from multiple pre-trained models to perform meta-learning. To prevent forgetting, we introduce a strategy to replay interpolated historical tasks to efficiently recall previous meta-knowledge. Furthermore, our framework seamlessly incorporates an automatic model selection mechanism to automatically filter out untrustworthy models during the meta-learning process. Code is available at https://github.com/Egg-Hu/Trustworthy-DFML.

cs.LG↗

arXiv2026

Love Me, Love My Label: Rethinking the Role of Labels in Prompt Retrieval for Visual In-Context Learning

Visual in-context learning (VICL) enables visual foundation models to handle multiple tasks by steering them with demonstrative prompts. The choice of such prompts largely influences VICL performance, standing out as a key challenge. Prior work has made substantial progress on prompt retrieval and reranking strategies, but mainly focuses on prompt images while overlooking labels. We reveal these approaches sometimes get visually similar but label-inconsistent prompts, which potentially degrade VICL performance. On the other hand, higher label consistency between query and prompts preferably indicates stronger VICL results. Motivated by these findings, we develop a framework named LaPR (Label-aware Prompt Retrieval), which highlights the role of labels in prompt selection. Our framework first designs an image-label joint representation for prompts to incorporate label cues explicitly. Besides, to handle unavailable query labels at test time, we introduce a mixture-of-expert mechanism to the dual encoders with query-adaptive routing. Each expert is expected to capture a specific label mode, while the router infers query-adaptive mixture weights and helps to learn label-aware representation. We carefully design alternative optimization for experts and router, with a VICL performance-guided contrastive loss and a label-guided contrastive loss, respectively. Extensive experiments show promising and consistent improvement of LaPR on in-context segmentation, detection, and colorization tasks. Moreover, LaPR generalizes well across feature extractors and cross-fold scenarios, suggesting the importance of label utilization in prompt retrieval for VICL. Code is available at https://github.com/luotc-why/CVPR26-LaPR.

cs.CV↗

arXiv2026

Learning Cross-View Object Correspondence via Cycle-Consistent Mask Prediction

We study the task of establishing object-level visual correspondence across different viewpoints in videos, focusing on the challenging egocentric-to-exocentric and exocentric-to-egocentric scenarios. We propose a simple yet effective framework based on conditional binary segmentation, where an object query mask is encoded into a latent representation to guide the localization of the corresponding object in a target video. To encourage robust, view-invariant representations, we introduce a cycle-consistency training objective: the predicted mask in the target view is projected back to the source view to reconstruct the original query mask. This bidirectional constraint provides a strong self-supervisory signal without requiring ground-truth annotations and enables test-time training (TTT) at inference. Experiments on the Ego-Exo4D and HANDAL-X benchmarks demonstrate the effectiveness of our optimization objective and TTT strategy, achieving state-of-the-art performance. The code is available at https://github.com/shannany0606/CCMP.

cs.CV↗

arXiv2026

PromptHub: Enhancing Multi-Prompt Visual In-Context Learning with Locality-Aware Fusion, Concentration and Alignment

Visual In-Context Learning (VICL) aims to complete vision tasks by imitating pixel demonstrations. Recent work pioneered prompt fusion that combines the advantages of various demonstrations, which shows a promising way to extend VICL. Unfortunately, the patch-wise fusion framework and model-agnostic supervision hinder the exploitation of informative cues, thereby limiting performance gains. To overcome this deficiency, we introduce PromptHub, a framework that holistically strengthens multi-prompting through locality-aware fusion, concentration and alignment. PromptHub exploits spatial priors to capture richer contextual information, employs complementary concentration, alignment, and prediction objectives to mutually guide training, and incorporates data augmentation to further reinforce supervision. Extensive experiments on three fundamental vision tasks demonstrate the superiority of PromptHub. Moreover, we validate its universality, transferability, and robustness across out-of-distribution settings, and various retrieval scenarios. This work establishes a reliable locality-aware paradigm for prompt fusion, moving beyond prior patch-wise approaches. Code is available at https://github.com/luotc-why/ICLR26-PromptHub.

cs.CV↗

arXiv2026

SAMA: Factorized Semantic Anchoring and Motion Alignment for Instruction-Guided Video Editing

Current instruction-guided video editing models struggle to simultaneously balance precise semantic modifications with faithful motion preservation. While existing approaches rely on injecting explicit external priors (e.g., VLM features or structural conditions) to mitigate these issues, this reliance severely bottlenecks model robustness and generalization. To overcome this limitation, we present SAMA (factorized Semantic Anchoring and Motion Alignment), a framework that factorizes video editing into semantic anchoring and motion modeling. First, we introduce Semantic Anchoring, which establishes a reliable visual anchor by jointly predicting semantic tokens and video latents at sparse anchor frames, enabling purely instruction-aware structural planning. Second, Motion Alignment pre-trains the same backbone on motion-centric video restoration pretext tasks (cube inpainting, speed perturbation, and tube shuffle), enabling the model to internalize temporal dynamics directly from raw videos. SAMA is optimized with a two-stage pipeline: a factorized pre-training stage that learns inherent semantic-motion representations without paired video-instruction editing data, followed by supervised fine-tuning on paired editing data. Remarkably, the factorized pre-training alone already yields strong zero-shot video editing ability, validating the proposed factorization. SAMA achieves state-of-the-art performance among open-source models and is competitive with leading commercial systems (e.g., Kling-Omni). Code, models, and datasets will be released.

cs.CV↗

arXiv2026

Unified and Semantically Grounded Domain Adaptation for Medical Image Segmentation

Most prior unsupervised domain adaptation approaches for medical image segmentation are narrowly tailored to either the source-accessible setting, where adaptation is guided by source-target alignment, or the source-free setting, which typically resorts to implicit adaptation mechanisms such as pseudo-labeling and network distillation. This substantial divergence in methodological designs between the two settings reveals an inherent flaw: the lack of an explicit, structured construction of anatomical knowledge that naturally generalizes across domains and settings. To bridge this longstanding divide, we introduce a unified, semantically grounded framework that supports both source-accessible and source-free adaptation. Fundamentally distinct from all prior works, our framework's adaptability emerges naturally as a direct consequence of the model architecture, without relying on explicit cross-domain alignment strategies. Specifically, our model learns a domain-agnostic probabilistic manifold as a global space of anatomical regularities, mirroring how humans establish visual understanding. Thus, the structural content in each image can be interpreted as a canonical anatomy retrieved from the manifold and a spatial transformation capturing individual-specific geometry. This disentangled, interpretable formulation enables semantically meaningful prediction with intrinsic adaptability. Extensive experiments on challenging cardiac and abdominal datasets show that our framework achieves state-of-the-art results in both settings, with source-free performance closely approaching its source-accessible counterpart, a level of consistency rarely observed in prior works. The results provide a principled foundation for anatomically informed, interpretable, and unified solutions for domain adaptation in medical imaging. The code is available at https://github.com/wxdrizzle/remind

cs.CV↗

arXiv2026

OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging

Foundation models update slowly due to resource-intensive training, whereas domain-specific models evolve rapidly between releases. Model merging seeks to combine multiple expert models into a single, more capable model, reducing storage and serving costs while supporting decentralized development. Despite its potential, previous studies have primarily focused on merging visual classification models or Large Language Models (LLMs) for code and math tasks. Recently, Multimodal LLMs (MLLMs) that extend LLMs through large-scale multimodal training have gained traction. However, there lacks a benchmark for model merging research that clearly divides the tasks for MLLM training and evaluation. In this paper, $\textbf{(i)}$ we introduce a model merging benchmark for MLLMs, which includes multiple tasks such as VQA, Geometry, Chart, OCR, and Grounding, studying both LoRA and full fine-tuning models. Moreover, we explore how model merging can combine different modalities (e.g., vision-language, audio-language, and video-language models), moving toward the Omni-language model. $\textbf{(ii)}$ We implement 10 model merging algorithms on the benchmark. Furthermore, we propose a novel method that removes noise from task vectors and robustly optimizes the merged vector based on a loss defined over task vector interactions, achieving an average performance gain of 2.48%. $\textbf{(iii)}$ We find that model merging offers a promising way for building improved MLLMs without requiring training data. Our results also demonstrate that the complementarity among multiple modalities outperforms individual modalities.

cs.AI↗

arXiv2026

Revisiting Global Token Mixing in Task-Dependent MRI Restoration: Insights from Minimal Gated CNN Baselines

Global token mixing, implemented via self-attention or state-space sequence models, has become a popular model design choice for MRI restoration. However, MRI restoration tasks differ substantially in how their degradations vary over image and k-space domains, and in the degree to which global coupling is already imposed by physics-driven data consistency terms. In this work, we ask the question whether global token mixing is actually beneficial in each individual task across three representative settings: accelerated MRI reconstruction with explicit data consistency, MRI super-resolution with k-space center cropping, and denoising of clinical carotid MRI data with spatially heteroscedastic noise. To reduce confounding factors, we establish a controlled testbed comparing a minimal local gated CNN and its large-field variant, benchmarking them directly against state-of-the-art global models under aligned training and evaluation protocols. For accelerated MRI reconstruction, the minimal unrolled gated-CNN baseline is already highly competitive compared to recent token-mixing approaches in public reconstruction benchmarks, suggesting limited additional benefits when the forward model and data-consistency steps provide strong global constraints. For super-resolution, where low-frequency k-space data are largely preserved by the controlled low-pass degradation, local gated models remain competitive, and a lightweight large-field variant yields only modest improvements. In contrast, for denoising with pronounced spatially heteroscedastic noise, token-mixing models achieve the strongest overall performance, consistent with the need to estimate spatially varying reliability. In conclusion, our results demonstrate that the utility of global token mixing in MRI restoration is task-dependent, and it should be tailored to the underlying imaging physics and degradation structure.

eess.IV↗

arXiv2026

ContextRL: Enhancing MLLM's Knowledge Discovery Efficiency with Context-Augmented RL

We propose ContextRL, a novel framework that leverages context augmentation to overcome these bottlenecks. Specifically, to enhance Identifiability, we provide the reward model with full reference solutions as context, enabling fine-grained process verification to filter out false positives (samples with the right answer but low-quality reasoning process). To improve Reachability, we introduce a multi-turn sampling strategy where the reward model generates mistake reports for failed attempts, guiding the policy to "recover" correct responses from previously all-negative groups. Experimental results on 11 perception and reasoning benchmarks show that ContextRL significantly improves knowledge discovery efficiency. Notably, ContextRL enables the Qwen3-VL-8B model to achieve performance comparable to the 32B model, outperforming standard RLVR baselines by a large margin while effectively mitigating reward hacking. Our in-depth analysis reveals the significant potential of contextual information for improving reward model accuracy and document the widespread occurrence of reward hacking, offering valuable insights for future RLVR research.

cs.LG↗

arXiv2026

Enhancing Geometric Perception in VLMs via Translator-Guided Reinforcement Learning

Vision-language models (VLMs) often struggle with geometric reasoning due to their limited perception of fundamental diagram elements. To tackle this challenge, we introduce GeoPerceive, a benchmark comprising diagram instances paired with domain-specific language (DSL) representations, along with an efficient automatic data generation pipeline. This design enables the isolated evaluation of geometric perception independently from reasoning. To exploit the data provided by GeoPerceive for enhancing the geometric perception capabilities of VLMs, we propose GeoDPO, a translator-guided reinforcement learning (RL) framework. GeoDPO employs an NL-to-DSL translator, which is trained on synthetic pairs generated by the data engine of GeoPerceive, to bridge natural language and DSL. This translator facilitates the computation of fine-grained, DSL-level scores, which serve as reward signals in reinforcement learning. We assess GeoDPO on both in-domain and out-of-domain datasets, spanning tasks in geometric perception as well as downstream reasoning. Experimental results demonstrate that, while supervised fine-tuning (SFT) offers only marginal improvements and may even impair performance in out-of-domain scenarios, GeoDPO achieves substantial gains: $+26.5\%$ on in-domain data, $+8.0\%$ on out-of-domain data, and $+39.0\%$ on downstream reasoning tasks. These findings underscore the superior performance and generalization ability of GeoDPO over SFT. All codes are released at https://github.com/Longin-Yu/GeoPerceive to ensure reproducibility.

cs.LG↗

arXiv2026

NarraScore: Bridging Visual Narrative and Musical Dynamics via Hierarchical Affective Control

Synthesizing coherent soundtracks for long-form videos remains a formidable challenge, currently stalled by three critical impediments: computational scalability, temporal coherence, and, most critically, a pervasive semantic blindness to evolving narrative logic. To bridge these gaps, we propose NarraScore, a hierarchical framework predicated on the core insight that emotion serves as a high-density compression of narrative logic. Uniquely, we repurpose frozen Vision-Language Models (VLMs) as continuous affective sensors, distilling high-dimensional visual streams into dense, narrative-aware Valence-Arousal trajectories. Mechanistically, NarraScore employs a Dual-Branch Injection strategy to reconcile global structure with local dynamism: a \textit{Global Semantic Anchor} ensures stylistic stability, while a surgical \textit{Token-Level Affective Adapter} modulates local tension via direct element-wise residual injection. This minimalist design bypasses the bottlenecks of dense attention and architectural cloning, effectively mitigating the overfitting risks associated with data scarcity. Experiments demonstrate that NarraScore achieves state-of-the-art consistency and narrative alignment with negligible computational overhead, establishing a fully autonomous paradigm for long-video soundtrack generation.

cs.SD↗

arXiv2026

PRIME: A Process-Outcome Alignment Benchmark for Verifiable Reasoning in Mathematics and Engineering

While model-based verifiers are essential for scaling Reinforcement Learning with Verifiable Rewards (RLVR), current outcome-centric verification paradigms primarily focus on the consistency between the final result and the ground truth, often neglecting potential errors in the derivation process. This leads to assigning positive rewards to correct answers produced from incorrect derivations. To bridge this gap, we introduce PRIME, a benchmark for evaluating verifiers on Process-Outcome Alignment verification in Mathematics and Engineering. Curated from a comprehensive collection of college-level STEM problems, PRIME comprises 2,530 high-difficulty samples through a consistency-based filtering pipeline. Through extensive evaluation, we find that current verifiers frequently fail to detect derivation flaws. Furthermore, we propose a process-aware RLVR training paradigm utilizing verifiers selected via PRIME. This approach substantially outperforms the outcome-only verification baseline, achieving absolute performance gains of 8.29%, 9.12%, and 7.31% on AIME24, AIME25, and Beyond-AIME, respectively, for the Qwen3-14B-Base model. Finally, we demonstrate a strong linear correlation ($R^2 > 0.92$) between verifier accuracy on PRIME and RLVR training effectiveness, validating PRIME as a reliable predictor for verifier selection.

cs.CL↗