Feng Qingyang had always hoped to launch his own company, but he never thought this would be how—or that the day would come this fast. 

Feng, a 27-year-old software engineer based in Beijing, started tinkering with OpenClaw, a popular new open-source AI tool that can take over a device and autonomously complete tasks for a user, in January. He was immediately hooked, and before long he was helping other curious tech workers with less technical proficiency install the AI agent.

Feng soon realized this could be a lucrative opportunity. By the end of January, he had set up a page on Xianyu, a secondhand shopping site, advertising “OpenClaw installation support.” “No need to know coding or complex terms. Fully remote,” reads the posting. “Anyone can quickly own an AI assistant, available within 30 minutes.” 

At the same time, the broader Chinese public was beginning to catch on—and the tool, which had begun as a niche interest among tech workers, started to evolve into a popular sensation.

Feng quickly became inundated with requests, and he started chatting with customers and managing orders late into the night. At the end of February, he quit his job. His side gig has now grown into a full-fledged professional operation with over 100 employees. So far, the store has handled 7,000 orders, each worth about 248 RMB or approximately $34. 

“Opportunities are always fleeting,” says Feng. “As programmers, we are the first to feel the winds shift.”

Feng is among a small cohort of savvy early adopters turning China’s OpenClaw craze into cash. As users with little technical background want in, a cottage industry of people offering installation services and preconfigured hardware has sprung up to meet them. The sudden rise of these tinkerers and impromptu consultants shows just how eager the general public in China is to adopt cutting-edge AI—even when there are huge security risks. 

A “lobster craze”

“Have you raised a lobster yet?” 

Xie Manrui, a 36-year-old software engineer in Shenzhen, says he has heard this question nonstop over the past month. “Lobster” is the nickname Chinese users have given to OpenClaw—a reference to its logo.

Xie, like Feng, has been experimenting with OpenClaw since January. He’s built new open-source tools on top of the ecosystem, including one that visualizes the agent’s progress as an animated little desktop worker and another that lets users voice-chat with it. 

“I’ve met so many new people through ‘lobster raising,’” says Xie. “Many are lawyers or doctors, with little technical background, but all dedicated to learning new things.”

Lobsters are indeed popping up everywhere in China right now—on and offline. In February, for instance, the entrepreneur and tech influencer Fu Sheng hosted a livestream showing off OpenClaw’s capabilities that got 20,000 views. And just last weekend, Xie attended three different OpenClaw events in Shenzhen, each drawing more than 500 people. These self-organized, unofficial gatherings feature power users, influencers, and sometimes venture capitalists as speakers. The biggest event Xie attended, on March 7, drew more than 1,000 people; in the packed venue, he says, people were shoulder to shoulder, with many attendees unable to even get a seat.

Now China’s AI giants are starting to piggyback on the trend too, promoting their models, APIs,  and cloud services (which can be used with OpenClaw), as well as their own OpenClaw-like agents. Earlier this month, Tencent held a public event offering free installation support for OpenClaw, drawing long lines of people waiting for help, including elderly users and children.

This sudden burst in popularity has even prompted local governments to get involved. Earlier this month the government of Longgang, a district in Shenzhen, released several policies to support OpenClaw-related ventures, including free computing credits and cash rewards for standout projects. Other cities, including Wuxi, have begun rolling out similar measures.

These policies only catalyze what’s already in the air. “It was not until my father, who is 77, asked me to help install a ‘lobster’ for him that I realized this thing is truly viral,” says Henry Li, a software engineer based in Beijing. 

A programmer gold rush

What’s making this moment particularly lucrative for people with technical skills, like Feng, is that so many people want OpenClaw, but not nearly as many have the capabilities to access it. Setting it up requires a level of technical knowledge most people do not possess, from typing commands into a black terminal window to navigating unfamiliar developer platforms. On the hardware side, an older or budget laptop may struggle to run it smoothly. And if the tool is not installed on a device separate from someone’s everyday computer, or if the data accessible to OpenClaw is not properly partitioned, the user’s privacy could be at risk—opening the door to data leaks and even malicious attacks. 

Chris Zhao, known as “Qi Shifu” online, organizes OpenClaw social media groups and events in Beijing. On apps like Rednote and Jike, Zhao routinely shares his thoughts on AI, and he asks other interested users to leave their WeChat ID so he can invite them to a semi-private group chat. The proof required to join is a screenshot that shows your “lobster” up and running. Zhao says that even in group chats for experienced users, hardware and cloud setup remain a constant topic of discussion.

The relatively high bar for setting up OpenClaw has generated a sense of exclusivity, creating a natural opening for a service industry to start unfolding around it. On Chinese e-commerce platforms like Taobao and JD, a simple search for “OpenClaw” now returns hundreds of listings, most of them installation guides and technical support packages aimed at nontechnical users, priced anywhere from 100 to 700 RMB (approximately $15 to $100). At the higher end, many vendors offer to come to help you in person. 

Like Feng, most providers of these services are early adopters with some technical ability who are looking for a side gig. But as demand has surged, some have found themselves overwhelmed. Xie, the developer in Shenzhen who created tools to layer on OpenClaw, was asked by a friend who runs one such business to help out over the weekend; the friend had a customer who worked in e-commerce and had little technical experience, so Xie had to show up in person to get it done. He walked away with 600 RMB ($87) for the afternoon.

The growing demand has also pushed vendors like Feng to expand quickly. He has now standardized his operation into tiers: a basic installation, a custom package where users can make specific requests like configuring a preferred chat app, and an ongoing tutoring service for those who want a hand to hold as they find their footing with the technology.

Other vendors in China are making money combining OpenClaw with hardware. Li Gong, a Shenzhen-based seller of refurbished Mac computers, was among the first online sellers to do this—offering Mac minis and MacBooks with OpenClaw preinstalled. Because OpenClaw is designed to operate with deep access to a hard drive and can run continuously in the background unattended, many users prefer to install it on a separate device rather than on the one they use every day. This would help prevent bad actors from infiltrating the program and immediately gaining access to a wide swathe of someone’s personal information. Many turn to secondhand or refurbished options to keep the cost down. Li says that in the last two weeks, orders have increased eightfold.

Though OpenClaw itself is a new technology, the general practice of buying software bundles, downloading third-party packages, and seeking out modified devices is nothing new for many Chinese internet users, says Tianyu Fang, a PhD candidate studying the history of technology at Harvard University. Many users pay for one-off IT support services for tasks from installing Adobe software to jailbreaking a Kindle.

Still, not everyone is getting swept up. Jiang Yunhui, a tech worker based in Ningbo, worries that ordinary users who struggle with setup may not be the right audience for a technology that is still effectively in testing. 

“The hype in first-tier cities can be a little overblown,” he says. “The agent is still a proof of concept, and I doubt it would be of any life-changing use to the average person for now.” He argues that using it safely and getting anything meaningful out of it requires a level of technical fluency and independent judgment that most new users simply don’t have yet.

He’s not alone in his concerns. On March 10, the Chinese cybersecurity regulator CNCERT issued a warning about the security and data risks tied to OpenClaw, saying it heightens users’ exposure to data breaches.

Despite the potential pitfalls, though, China’s enthusiasm for OpenClaw doesn’t seem to be slowing.

Feng, now flush with the earnings from his operation, wants to use the momentum—and the capital—to keep building out his own venture with AI tools at the center of it.

“With OpenClaw and other AI agents, I want to see if I can run a one-person company,” he says. “I’m giving myself one year.”

arXiv:2603.07853v1 Announce Type: new Abstract: Research Agents enable models to gather information from the web using tools to answer user queries, requiring them to dynamically interleave internal reasoning with tool use. While such capabilities can in principle be learned via reinforcement learning with verifiable rewards (RLVR), we observe that agents often exhibit poor exploration behaviors, including premature termination and biased tool usage. As a result, RLVR alone yields limited improvements. We propose SynPlanResearch-R1, a framework that synthesizes tool-use trajectories that encourage deeper exploration to shape exploration during cold-start supervised fine-tuning, providing a strong initialization for subsequent RL. Across seven multi-hop and open-web benchmarks, \framework improves performance by up to 6.0% on Qwen3-8B and 5.8% on Qwen3-4B backbones respectively compared to SOTA baselines. Further analyses of tool-use patterns and training dynamics compared to baselines shed light on the factors underlying these gains. Our code is publicly available at https://github.com/HansiZeng/syn-plan-research.

arXiv:2603.07915v1 Announce Type: new Abstract: Modern agents powered by thinking LLMs achieve high accuracy through long chain-of-thought reasoning but incur substantial inference costs. While many LLMs now support configurable reasoning levels (e.g., high/medium/low), static strategies are often ineffective: using low-effort modes at every step leads to significant performance degradation, while random selection fails to preserve accuracy or provide meaningful cost reduction. However, agents should reserve high reasoning effort for difficult steps like navigating complex website structures, while using lower-effort modes for simpler steps like opening a target URL. In this paper, we propose Ares, a framework for per-step dynamic reasoning effort selection tailored for multi-step agent tasks. Ares employs a lightweight router to predict the lowest appropriate reasoning level for each step based on the interaction history. To train this router, we develop a data generation pipeline that identifies the minimum reasoning effort required for successful step completion. We then fine-tune the router to predict these levels, enabling plug-and-play integration for any LLM agents. We evaluate Ares on a diverse set of agent tasks, including TAU-Bench for tool use agents, BrowseComp-Plus for deep-research agents, and WebArena for web agents. Experimental results show that Ares reduces reasoning token usage by up to 52.7% compared to fixed high-effort reasoning, while introducing minimal degradation in task success rates.

arXiv:2603.07972v1 Announce Type: new Abstract: While scaling individual Large Language Models (LLMs) has delivered remarkable progress, the next frontier lies in scaling collaboration through multi-agent systems (MAS). However, purely autonomous MAS remain ''closed-world'' systems, constrained by the static knowledge horizon of pre-trained models. This limitation makes them brittle on tasks requiring knowledge beyond training data, often leading to collective failure under novel challenges. To address this, we propose the Human-In-the-Loop Multi-Agent Collaboration (HILA) framework, a principled paradigm for human--agent collaboration. HILA trains agents to learn a metacognitive policy that governs when to solve problems autonomously and when to defer to a human expert. To operationalize this policy, we introduce Dual-Loop Policy Optimization, which disentangles immediate decision-making from long-term capability growth. The inner loop applies Group Relative Policy Optimization (GRPO) with a cost-aware reward to optimize deferral decisions, while the outer loop implements continual learning, transforming expert feedback into high-quality supervised signals that strengthen the agent's reasoning ability. Experiments on challenging mathematical and problem-solving benchmarks show that HILA, equipped with Dual-Loop Policy Optimization, consistently outperforms advanced MAS, establishing a principled foundation for collaborative and continually improving agentic systems.

arXiv:2603.08035v1 Announce Type: new Abstract: Reward modeling is essential for aligning Large Language Models(LLMs) with human preferences, yet conventional reward models suffer from poor interpretability and heavy reliance on costly expert annotations. While recent rubric-based approaches enhance evaluation transparency, they lack systematic quality control, yielding noisy and redundant criteria, failing to mitigate persistent biases (e.g., verbosity, position) in LLM evaluators, and creating a scalability-reliability trade-off. To address these limitations, we propose CDRRM (Contrast-Driven Rubric Reward Model), a framework built on a novel Contrast-then-Synthesis paradigm for high-quality rubric generation and guided preference judgment. CDRRM first conducts multi-dimensional contrastive profiling on preference pairs to identify causal discriminative factors, then synthesizes these insights into compact, context-aware rubrics to guide preference judg- ments. Extensive experiments on three authoritative benchmarks (RewardBench, RMBench, RMB) demonstrate that CDRRM achieves state-of-the-art performance across diverse domains and effectively mitigates aforementioned evaluation biases. Notably, our approach delivers exceptional data efficiency: training the rubric generator on only 3k high-quality samples empowers a frozen pre-trained judge model to outperform fully fine-tuned baselines. This work offers a scalable, interpretable, and data-efficient path for reward modeling.

arXiv:2603.08262v1 Announce Type: new Abstract: The integration of Large Language Models (LLMs) into the financial domain is driving a paradigm shift from passive information retrieval to dynamic, agentic interaction. While general-purpose tool learning has witnessed a surge in benchmarks, the financial sector, characterized by high stakes, strict compliance, and rapid data volatility, remains critically underserved. Existing financial evaluations predominantly focus on static textual analysis or document-based QA, ignoring the complex reality of tool execution. Conversely, general tool benchmarks lack the domain-specific rigor required for finance, often relying on toy environments or a negligible number of financial APIs. To bridge this gap, we introduce FinToolBench, the first real-world, runnable benchmark dedicated to evaluating financial tool learning agents. Unlike prior works limited to a handful of mock tools, FinToolBench establishes a realistic ecosystem coupling 760 executable financial tools with 295 rigorous, tool-required queries. We propose a novel evaluation framework that goes beyond binary execution success, assessing agents on finance-critical dimensions: timeliness, intent type, and regulatory domain alignment. Furthermore, we present FATR, a finance-aware tool retrieval and reasoning baseline that enhances stability and compliance. By providing the first testbed for auditable, agentic financial execution, FinToolBench sets a new standard for trustworthy AI in finance. The tool manifest, execution environment, and evaluation code will be open-sourced to facilitate future research.

arXiv:2603.08652v1 Announce Type: new Abstract: Recent advancements in Unified Multimodal Models (UMMs) have significantly advanced text-to-image (T2I) generation, particularly through the integration of Chain-of-Thought (CoT) reasoning. However, existing CoT-based T2I methods largely rely on abstract natural-language planning, which lacks the precision required for complex spatial layouts, structured visual elements, and dense textual content. In this work, we propose CoCo (Code-as-CoT), a code-driven reasoning framework that represents the reasoning process as executable code, enabling explicit and verifiable intermediate planning for image generation. Given a text prompt, CoCo first generates executable code that specifies the structural layout of the scene, which is then executed in a sandboxed environment to render a deterministic draft image. The model subsequently refines this draft through fine-grained image editing to produce the final high-fidelity result. To support this training paradigm, we construct CoCo-10K, a curated dataset containing structured draft-final image pairs designed to teach both structured draft construction and corrective visual refinement. Empirical evaluations on StructT2IBench, OneIG-Bench, and LongText-Bench show that CoCo achieves improvements of +68.83%, +54.8%, and +41.23% over direct generation, while also outperforming other generation methods empowered by CoT. These results demonstrate that executable code is an effective and reliable reasoning paradigm for precise, controllable, and structured text-to-image generation. The code is available at: https://github.com/micky-li-hd/CoCo

arXiv:2603.08706v1 Announce Type: new Abstract: Training large language models (LLMs) as autonomous agents often begins with imitation learning, but it only teaches agents what to do without understanding why: agents never contrast successful actions against suboptimal alternatives and thus lack awareness of action quality. Recent approaches attempt to address this by introducing self-reflection supervision derived from contrasts between expert and alternative actions. However, the training paradigm fundamentally remains imitation learning: the model imitates pre-constructed reflection text rather than learning to reason autonomously. We propose Agentic Critical Training (ACT), a reinforcement learning paradigm that trains agents to identify the better action among alternatives. By rewarding whether the model's judgment is correct, ACT drives the model to autonomously develop reasoning about action quality, producing genuine self-reflection rather than imitating it. Across three challenging agent benchmarks, ACT consistently improves agent performance when combined with different post-training methods. It achieves an average improvement of 5.07 points over imitation learning and 4.62 points over reinforcement learning. Compared to approaches that inject reflection capability through knowledge distillation, ACT also demonstrates clear advantages, yielding an average improvement of 2.42 points. Moreover, ACT enables strong out-of-distribution generalization on agentic benchmarks and improves performance on general reasoning benchmarks without any reasoning-specific training data, highlighting the value of our method. These results suggest that ACT is a promising path toward developing more reflective and capable LLM agents.

arXiv:2507.11202v1 Announce Type: cross Abstract: Multimodal Emotion Recognition (MER) often encounters incomplete multimodality in practical applications due to sensor failures or privacy protection requirements. While existing methods attempt to address various incomplete multimodal scenarios by balancing the training of each modality combination through additional gradients, these approaches face a critical limitation: training gradients from different modality combinations conflict with each other, ultimately degrading the performance of the final prediction model. In this paper, we propose a unimodal decoupled dynamic low-rank adaptation method based on modality combinations, named MCULoRA, which is a novel framework for the parameter-efficient training of incomplete multimodal learning models. MCULoRA consists of two key modules, modality combination aware low-rank adaptation (MCLA) and dynamic parameter fine-tuning (DPFT). The MCLA module effectively decouples the shared information from the distinct characteristics of individual modality combinations. The DPFT module adjusts the training ratio of modality combinations based on the separability of each modality's representation space, optimizing the learning efficiency across different modality combinations. Our extensive experimental evaluation in multiple benchmark datasets demonstrates that MCULoRA substantially outperforms previous incomplete multimodal learning approaches in downstream task accuracy.

arXiv:2603.06600v1 Announce Type: cross Abstract: Vision Language Models (VLMs) are prone to errors, and identifying where these errors occur is critical for ensuring the reliability and safety of AI systems. In this paper, we propose an approach that automatically generates questions designed to deliberately induce incorrect responses from VLMs, thereby revealing their vulnerabilities. The core of this approach lies in fuzz testing and reinforcement finetuning: we transform a single input query into a large set of diverse variants through vision and language fuzzing. Based on the fuzzing outcomes, the question generator is further instructed by adversarial reinforcement fine-tuning to produce increasingly challenging queries that trigger model failures. With this approach, we can consistently drive down a target VLM's answer accuracy -- for example, the accuracy of Qwen2.5-VL-32B on our generated questions drops from 86.58\% to 65.53\% in four RL iterations. Moreover, a fuzzing policy trained against a single target VLM transfers to multiple other VLMs, producing challenging queries that degrade their performance as well.

arXiv:2603.06616v1 Announce Type: cross Abstract: Efficiently routing queries to the optimal large language model (LLM) is crucial for optimizing the cost-performance trade-off in multi-model systems. However, most existing routers rely on single-model selection, making them susceptible to misrouting. In this work, we formulate LLM routing as the $\alpha$-VOR problem to minimize expected set size while controlling the misrouting risk, and propose a novel method -- RACER, extending base routers to output model sets that can be subsequently aggregated for improved output. In particular, RACER constructs nested model sets via augmented scoring and utilizes finite-sample concentration bounds to calibrate a threshold that allows for both variable set sizes and abstention. We theoretically prove that RACER achieves rigorous distribution-free risk control on unseen test data in a post-hoc and model-agnostic manner. Extensive experiments verify our theoretical guarantees and demonstrate that RACER consistently enhances downstream accuracy across a wide range of benchmarks.

arXiv:2603.06638v1 Announce Type: cross Abstract: The rise of large language models (LLMs) has shifted time series analysis from narrow analytics to general-purpose reasoning. Yet, existing benchmarks cover only a small set of health time series modalities and tasks, failing to reflect the diverse domains and extensive temporal dependencies inherent in real-world physiological modeling. To bridge these gaps, we introduce HEARTS (Health Reasoning over Time Series), a unified benchmark for evaluating hierarchical reasoning capabilities of LLMs over general health time series. HEARTS integrates 16 real-world datasets across 12 health domains and 20 signal modalities, and defines a comprehensive taxonomy of 110 tasks grouped into four core capabilities: Perception, Inference, Generation, and Deduction. Evaluating 14 state-of-the-art LLMs on more than 20K test samples reveals intriguing findings. First, LLMs substantially underperform specialized models, and their performance is only weakly related to general reasoning scores. Moreover, LLMs often rely on simple heuristics and struggle with multi-step temporal reasoning. Finally, performance declines with increasing temporal complexity, with similar failure modes within model families, indicating that scaling alone is insufficient. By making these gaps measurable, HEARTS provides a standardized testbed and living benchmark for developing next-generation LLM agents capable of reasoning over diverse health signals.

arXiv:2603.06651v1 Announce Type: cross Abstract: Euclidean gradient descent algorithms barely capture the geometry of objective function-induced hypersurfaces and risk driving update trajectories off the hypersurfaces. Riemannian gradient descent algorithms address these issues but fail to represent complex hypersurfaces via a single classic manifold. We propose geodesic gradient descent (GGD), a generic and learning-rate-free Riemannian gradient descent algorithm. At each iteration, GGD uses an n-dimensional sphere to approximate a local neighborhood on the objective function-induced hypersurface, adapting to arbitrarily complex geometries. A tangent vector derived from the Euclidean gradient is projected onto the sphere to form a geodesic, ensuring the update trajectory stays on the hypersurface. Parameter updates are performed using the endpoint of the geodesic. The maximum step size of the gradient in GGD is equal to a quarter of the arc length on the n-dimensional sphere, thus eliminating the need for a learning rate. Experimental results show that compared with the classic Adam algorithm, GGD achieves test MSE reductions ranging from 35.79% to 48.76% for fully connected networks on the Burgers' dataset, and cross-entropy loss reductions ranging from 3.14% to 11.59% for convolutional neural networks on the MNIST dataset.

arXiv:2603.06697v1 Announce Type: cross Abstract: Vision--language models (VLMs) process images as visual tokens, yet their intermediate reasoning is often carried out in text, which can be suboptimal for visually grounded radiology tasks. Radiologists instead diagnose via sequential visual search; eye-tracking captures this process as time-ordered gaze trajectories that reveal how evidence is acquired over time. We use eye-gaze as supervision to guide VLM reasoning by introducing a small set of dedicated gaze tokens. These tokens are trained to predict gaze-selected image patch indices in temporal order, encouraging the model to follow human-like evidence acquisition and integration. Experiments on MIMIC-EYE and multiple external zero-shot benchmarks show consistent gains over baselines, achieving state-of-the-art in-domain performance and improved out-of-domain robustness. These results highlight temporally ordered gaze as an effective supervision signal for learning visually grounded medical reasoning.

arXiv:2603.06726v1 Announce Type: cross Abstract: Electricity market prices exhibit extreme volatility, nonlinearity, and non-stationarity, making accurate forecasting a significant challenge. While cutting-edge time series foundation models (TSFMs) effectively capture temporal dependencies, they typically underutilize cross-variate correlations and non-periodic patterns that are essential for price forecasting. Conversely, regression models excel at capturing feature interactions but are limited to future-available inputs, ignoring crucial historical drivers that are unavailable at forecast time. To bridge this gap, we propose FutureBoosting, a novel paradigm that enhances regression-based forecasts by integrating forecasted features generated from a frozen TSFM. This approach leverages the TSFM's ability to model historical patterns and injects these insights as enriched inputs into a downstream regression model. We instantiate this paradigm into a lightweight, plug-and-play framework for electricity price forecasting. Extensive evaluations on real-world electricity market data demonstrate that our framework consistently outperforms state-of-the-art TSFMs and regression baselines, achieving reductions in Mean Absolute Error (MAE) of more than 30% at most. Through ablation studies and explainable AI (XAI) techniques, we validate the contribution of forecasted features and elucidate the model's decision-making process. FutureBoosting establishes a robust, interpretable, and effective solution for practical market participation, offering a general framework for enhancing regression models with temporal context.

arXiv:2603.06859v1 Announce Type: cross Abstract: Cooperative multi-agent reinforcement learning (MARL) systems powered by large language models (LLMs) are frequently optimized via sparse terminal-only feedback. This shared signal entangles upstream decisions, obstructing accurate decision-level credit assignment. To address this trajectory-level diffusion, we introduce Contextual Counterfactual Credit Assignment (\textbf{\texttt{C3}}). Instead of distributing rewards across an entire episode, \textbf{\texttt{C3}} isolates the causal impact of individual messages by freezing the exact transcript-derived context, evaluating context-matched alternatives via fixed-continuation replay, and applying a leave-one-out (LOO) baseline. This localized intervention extracts unbiased, low-variance marginal advantages for standard policy-gradient optimization. Evaluated across five mathematical and coding benchmarks under matched budgets, \textbf{\texttt{C3}} improves terminal performance over established baselines. Mechanistic diagnostics further show that these gains are accompanied by higher credit fidelity, lower contextual variance, and stronger inter-agent causal dependence. Our code is available at https://github.com/EIT-EAST-Lab/C3.

arXiv:2603.06981v1 Announce Type: cross Abstract: Controllable diffusion generation often relies on various heuristics that are seemingly disconnected without a unified understanding. We bridge this gap with Diffusion Controller (DiffCon), a unified control-theoretic view that casts reverse diffusion sampling as state-only stochastic control within (generalized) linearly-solvable Markov Decision Processes (LS-MDPs). Under this framework, control acts by reweighting the pretrained reverse-time transition kernels, balancing terminal objectives against an $f$-divergence cost. From the resulting optimality conditions, we derive practical reinforcement learning methods for diffusion fine-tuning: (i) f-divergence-regularized policy-gradient updates, including a PPO-style rule, and (ii) a regularizer-determined reward-weighted regression objective with a minimizer-preservation guarantee under the Kullback-Leibler (KL) divergence. The LS-MDP framework further implies a principled model form: the optimal score decomposes into a fixed pretrained baseline plus a lightweight control correction, motivating a side-network parameterization conditioned on exposed intermediate denoising outputs, enabling effective gray-box adaptation with a frozen backbone. Experiments on Stable Diffusion v1.4 across supervised and reward-driven finetuning show consistent gains in preference-alignment win rates and improved quality-efficiency trade-offs versus gray-box baselines and even the parameter-efficient white-box adapter LoRA.

arXiv:2603.07048v1 Announce Type: cross Abstract: Although large vision-language models (LVLMs) have demonstrated remarkable capabilities, they are prone to hallucinations in multi-image tasks. We attribute this issue to limitations in existing attention mechanisms and insufficient cross-image modeling. Inspired by this, we propose a structured hallucination mitigation framework involving Cross-Image Attention calibration and Preference Learning (CAPL). CAPL explicitly enhances inter-image interactions at the architectural level while reinforcing reliance on genuine cross-image evidence during training, thereby improving the model's perception and modeling of cross-image associations. Specifically, we (i) introduce a selectable image token interaction attention mechanism to establish fine-grained cross-image entity alignment and information flow; (ii) design a cross-image modeling-based preference optimization strategy that contrasts reasoning outcomes under full inter-image interaction and those obtained when images are mutually invisible, encouraging the model to ground its predictions in authentic visual evidence and mitigating erroneous inferences driven by textual priors. Experimental results demonstrate that CAPL consistently improves performance across multiple model architectures, achieving stable gains on both multi-image hallucination and general benchmarks. Notably, performance on single-image visual tasks remains stable or slightly improves, indicating strong generalization capability.

arXiv:2603.07107v1 Announce Type: cross Abstract: Generative models offer a promising paradigm for the final stage reranking in multi-stage recommender systems, with the ability to capture inter-item dependencies within reranked lists. However, their practical deployment still faces two key challenges: (1) an inherent conflict between achieving high generation quality and ensuring low-latency inference, making it difficult to balance the two, and (2) insufficient interaction between user and item features in existing methods. To address these challenges, we propose a novel Personalized Semi-Autoregressive with online knowledge Distillation (PSAD) framework for reranking. In this framework, the teacher model adopts a semi-autoregressive generator to balance generation quality and efficiency, while its ranking knowledge is distilled online into a lightweight scoring network during joint training, enabling real-time and efficient inference. Furthermore, we propose a User Profile Network (UPN) that injects user intent and models interest dynamics, enabling deeper interactions between users and items. Extensive experiments conducted on three large-scale public datasets demonstrate that PSAD significantly outperforms state-of-the-art baselines in both ranking performance and inference efficiency.

arXiv:2603.07131v1 Announce Type: cross Abstract: Large Vision Language Models (LVLMs) show immense potential for automated ophthalmic diagnosis. However, their clinical deployment is severely hindered by lacking domain-specific knowledge. In this work, we identify two structural deficiencies hindering reliable medical reasoning: 1) the Perception Gap, where general-purpose visual encoders fail to resolve fine-grained pathological cues (e.g., microaneurysms); and 2) the Reasoning Gap, where sparse visual evidence is progressively overridden by massive language priors in deeper transformer layers, leading to ungrounded hallucinations. To bridge these gaps, we propose EyExIn, a data-efficient framework designed to anchor retinal VLMs with expert knowledge via a Deep Expert Injection mechanism. Our architecture employs an Expert-Aware Dual-Stream encoding strategy that decouples visual representation into a general stream for anatomical context and a specialized expert stream for pathological semantics. To ensure high-fidelity integration, we design a Semantic-Adaptive Gated Fusion module, which dynamically amplifies subtle lesion signals while filtering irrelevant background noise. Furthermore, we introduce Adaptive Deep Expert Injection to embed persistent "Vision Anchors" by integrating fused visual features as residual biases directly into intermediate LLM layers. This mechanism creates a visual shortcut that forces the reasoning stack to remain strictly grounded in visual evidence. Extensive experiments across four benchmarks demonstrate that our model consistently outperforms massive proprietary systems. EyExIn significantly enhances domain-specific knowledge embedding and achieves state-of-the-art precision in ophthalmic visual question answering, advancing the development of trustworthy ophthalmic AI.