arXiv subjects

Zhuohao Li

Publications and source records attributed to Zhuohao Li.

15 recordsLinked to original sources

arXiv2026

MedQA-MM: Shortcuts Behind Medical Visual Reasoning

A benchmark score credits final answers, but not the route by which an item can be answered. In medical multimodal multiple-choice questions (MCQs), this distinction matters because a correct answer can be supported by the intended image finding or by benchmark-preserved cues in the wording of answers, non-visual clinical text, visible image text, artificial annotations, or device/context artifacts. We call the resulting score-level overinterpretation reasoning inflation. Here, a route is an observable input path that can support answer selection, not a claim about the model's hidden cognition. Across six medical multimodal MCQ datasets, we separate candidate cues from behavioral evidence through prompt- and image-side audits, modality ablations, and matched repairs that preserve the medical target and answer key. In a 13-configuration open-model panel, full-input accuracy is 62.63%, while text-only and options-only settings achieve 53.96% and 29.71%, respectively. Removing length-gap, absolute/conspicuous, and spatial/prepositional cues lowers accuracy by 6.58, 3.50, and 4.77 percentage points. We also construct MedQA-MM, a 1,000-item shortcut-mitigated subset, where text-only and options-only accuracy fall to 5.21% and 12.33%. This does not imply that models never use images; it shows that medical image-reasoning claims require route-level evidence.

cs.CV↗

arXiv2026

TrapVLA: Trapping Vision-Language-Action Models in Configured Failure Modes

This work introduces Configured Failure Trapping, a novel backdoor attack task against Vision-Language-Action (VLA) models, which aims to activate attacks through stealthy textual triggers and induce configured failure modes. Unlike prior backdoor attacks that treat any task failure as a successful attack, Configured Failure Trapping requires the attacker to control how the robot fails (e.g., causing the robot to grasp with a specified positional offset), making it substantially more challenging and hard to detect. To support the new task, we propose an effective data engine for synthesizing high-quality target trajectories and an automated suite for measuring configured-failure fidelity. Then, based on this foundation, we construct two new benchmarks, namely Trap-LIBERO and Trap-RoboTwin, that instantiate Configured Failure Trapping across four representative failure modes. To address this task, we identify sparse action deviation as a critical challenge and accordingly propose a novel method named TrapVLA, which explicitly learns trigger-induced action residuals to steer the policy toward the configured failure behavior. Extensive experiments across simulation benchmarks and real-world robotic settings show that TrapVLA effectively injects configured failure modes into VLA models while largely preserving performance on clean data. Project page: https://john-liua.github.io/TrapVLA/

cs.RO↗

arXiv2026

UT-ACA: Uncertainty-Triggered Adaptive Context Allocation for Long-Context Inference

Long-context inference remains challenging for large language models due to attention dilution and out-of-distribution degradation. Context selection mitigates this limitation by attending to a subset of key-value cache entries, yet most methods allocate a fixed context budget throughout decoding despite highly non-uniform token-level contextual demands. To address this issue, we propose Uncertainty-Triggered Adaptive Context Allocation (UT-ACA), an inference-time framework that dynamically adjusts the context window based on token-wise uncertainty. UT-ACA learns an uncertainty detector that combines semantic embeddings with logit-based confidence while accounting for uncertainty accumulation across decoding steps. When insufficient evidence is indicated, UT-ACA selectively rolls back, expands the context window, and regenerates the token with additional support. Experiments show that UT-ACA substantially reduces average context usage while preserving generation quality in long-context settings. Code and dataset are available at https://github.com/Tommy307/UT-ACA.

cs.CL↗

arXiv2026

A Closed-Loop Multi-Agent Framework for Robust Multi-Robot Manipulation

Multi-robot systems provide the parallelism and redundancy necessary for long-horizon tasks, while Large Language Models (LLMs) offer the reasoning capabilities to decompose these objectives into actionable plans. However, effectively grounding this high-level reasoning in physical multi-robot execution remains an open challenge. Existing LLM-based approaches fall mainly into two categories: Single-robot methods achieve robust contact-rich manipulation but lack the coordination mechanisms required for tasks spanning multiple workspaces. Current multi-robot frameworks focus on high-level planning, often treating manipulation as an idealized primitive that fails to account for real-world execution uncertainties. To address this, we propose a hierarchical closed-loop agentic LLM-based framework to ensure robust multi-robot manipulation. Our system consists of three specialized agents: the Planning Agent decomposes instructions into allocated sub-tasks, the Manipulation Agent for each robot executes actions via adaptive tool use, and the Verification Agent closes the loop by monitoring physical outcomes and feeding back semantic corrections. Extensive real-world experiments demonstrate that our framework achieves superior success rates, ensures robust adaptability ranging from single to cross workspace manipulation, and offers a generalizable approach for diverse manipulation tasks.

cs.RO↗

arXiv2026

RoboFlow4D: A Lightweight Flow World Model Toward Real-Time Flow-Guided Robotic Manipulation

Planning and acting in 3D environments is a fundamental capability for robotic manipulation in the real world. Although prior work has explored predictive flow planners to guide 3D manipulation, existing approaches often rely on modular pipelines stacking multiple submodels, resulting in high computational overhead and limited real-time performance. To address these challenges, we introduce RoboFlow4D, a lightweight flow world model that unifies perception and planning by estimating temporal motion in physical 3D space. As an end-to-end framework, RoboFlow4D directly predicts multi-frame 3D flows from visual observations and textual instructions, providing explicit flow-based planning to guide action generation. This design allows seamless integration with general action policies, forming an efficient observation-planning-execution closed loop. Through slow-fast collaboration between flow prediction and action control, RoboFlow4D enables real-time and resource-efficient manipulation. Extensive experiments in both simulation and real-world settings demonstrate that RoboFlow4D consistently improves manipulation success rates and computational efficiency, advancing flow-guided planning for embodied intelligence.

cs.RO↗

arXiv2026

ReViP: Mitigating False Completion in Vision-Language-Action Models with Vision-Proprioception Rebalance

Vision-Language-Action (VLA) models have advanced robotic manipulation by combining vision, language, and proprioception to predict actions. However, previous methods fuse proprioceptive signals directly with vision-language features, resulting in state-dominant bias and \textbf{false completions} despite visible execution failures. We systematically analyze this failure mode, attributing it to modality imbalance, where policies overly rely on internal state progression and underuse visual evidence. To address this, we introduce the first \textbf{False-Completion Benchmark Suite}, featuring eight tasks with three controlled perturbations (\emph{Object Drop}, \emph{Distractor Swap}, \emph{Relayout}) to comprehensively evaluate false completion. Moreover, we propose \textbf{ReViP}, a novel VLA framework with \textbf{Vi}sion-\textbf{P}roprioception \textbf{Re}balance to enhance visual grounding and robustness under perturbations. The key insight is to introduce auxiliary \emph{progress-aware visual cues} to adaptively modulate the coupling between semantic perception and proprioceptive dynamics. Specifically, progress-aware visual cues are extracted by an external Task-Stage Observer, which performs task-relevant reasoning on real-time observations to drive task-stage feature-wise linear modulation, enhancing environmental awareness and mitigating state-driven errors. Extensive experiments show that ReViP effectively mitigates false completion and improves success rates over strong VLA baselines, achieving a \textbf{26\%} gain over $π_0$ model on our suite, with gains extending to LIBERO, RoboTwin 2.0, and real-world evaluations.

cs.RO↗

arXiv2025

Decomposed Object Manipulation via Dual-Actor Policy

Object manipulation, which focuses on learning to perform tasks on similar parts across different types of objects, can be divided into an approaching stage and a manipulation stage. However, previous works often ignore this characteristic of the task and rely on a single policy to directly learn the whole process of object manipulation. To address this problem, we propose a novel Dual-Actor Policy, termed DAP, which explicitly considers different stages and leverages heterogeneous visual priors to enhance each stage. Specifically, we introduce an affordance-based actor to locate the functional part in the manipulation task, thereby improving the approaching process. Following this, we propose a motion flow-based actor to capture the movement of the component, facilitating the manipulation process. Finally, we introduce a decision maker to determine the current stage of DAP and select the corresponding actor. Moreover, existing object manipulation datasets contain few objects and lack the visual priors needed to support training. To address this, we construct a simulated dataset, the Dual-Prior Object Manipulation Dataset, which combines the two visual priors and includes seven tasks, including two challenging long-term, multi-stage tasks. Experimental results on our dataset, the RoboTwin benchmark and real-world scenarios illustrate that our method consistently outperforms the SOTA method by 5.55%, 14.7% and 10.4% on average respectively.

cs.RO↗

arXiv2025

Hybrid Reward Normalization for Process-supervised Non-verifiable Agentic Tasks

Large Language Models (LLMs) increasingly rely on external tools such as search engines to solve complex agentic tasks that require reasoning and external knowledge retrieval. Recently, reinforcement learning with verifiable rewards (RLVR) has demonstrated its effectiveness in advancing capabilities of LLMs by rewarding the final answers via outcome rewards. While straightforward to supervise, outcome rewards only provide sparse signals and delayed feedback, which limits their effectiveness on long trajectories. Process rewards address this by evaluating intermediate steps, providing fine-grained supervision and encouraging grounded problem solving. However, it is notoriously hard to annotate step-wise labels, especially in non-verifiable process without "golden" answers. Furthermore, step-wise judgment requires the balance between local quality with contribution to the final outcome, as optimizing towards higher process reward may not always align with better final outcomes. To address the above challenges, we introduce Principle Process Reward (PPR), an RL approach that unifies principled step-level assessment and outcome verification. We train a principle-based reward model to improve the transparency and reliability of process evaluation, and further introduce a Reward Normalization (ReNorm) strategy to calibrate outcome and process rewards. Experiment results show that PPR achieves state-of-the-art performance across a wide range of benchmarks, demonstrating its impressive robustness and generalization. Our code and model collection is available in this link.

cs.AI↗

arXiv2025

Functional Consistency of LLM Code Embeddings: A Self-Evolving Data Synthesis Framework for Benchmarking

Embedding models have demonstrated strong performance in tasks like clustering, retrieval, and feature extraction while offering computational advantages over generative models and cross-encoders. Benchmarks such as MTEB have shown that text embeddings from large language models (LLMs) capture rich semantic information, but their ability to reflect code-level functional semantics remains unclear. Existing studies largely focus on code clone detection, which emphasizes syntactic similarity and overlooks functional understanding. In this paper, we focus on the functional consistency of LLM code embeddings, which determines if two code snippets perform the same function regardless of syntactic differences. We propose a novel data synthesis framework called Functionality-Oriented Code Self-Evolution to construct diverse and challenging benchmarks. Specifically, we define code examples across four semantic and syntactic categories and find that existing datasets predominantly capture syntactic properties. Our framework generates four unique variations from a single code instance, providing a broader spectrum of code examples that better reflect functional differences. Extensive experiments on three downstream tasks-code clone detection, code functional consistency identification, and code retrieval-demonstrate that embedding models significantly improve their performance when trained on our evolved datasets. These results highlight the effectiveness and generalization of our data synthesis framework, advancing the functional understanding of code.

cs.SE↗

arXiv2025

Hawkeye:Efficient Reasoning with Model Collaboration

Chain-of-Thought (CoT) reasoning has demonstrated remarkable effectiveness in enhancing the reasoning abilities of large language models (LLMs). However, its efficiency remains a challenge due to the generation of excessive intermediate reasoning tokens, which introduce semantic redundancy and overly detailed reasoning steps. Moreover, computational expense and latency are significant concerns, as the cost scales with the number of output tokens, including those intermediate steps. In this work, we observe that most CoT tokens are unnecessary, and retaining only a small portion of them is sufficient for producing high-quality responses. Inspired by this, we propose HAWKEYE, a novel post-training and inference framework where a large model produces concise CoT instructions to guide a smaller model in response generation. HAWKEYE quantifies redundancy in CoT reasoning and distills high-density information via reinforcement learning. By leveraging these concise CoTs, HAWKEYE is able to expand responses while reducing token usage and computational cost significantly. Our evaluation shows that HAWKEYE can achieve comparable response quality using only 35% of the full CoTs, while improving clarity, coherence, and conciseness by approximately 10%. Furthermore, HAWKEYE can accelerate end-to-end reasoning by up to 3.4x on complex math tasks while reducing inference cost by up to 60%. HAWKEYE will be open-sourced and the models will be available soon.

cs.AI↗

arXiv2025

Torpor: GPU-Enabled Serverless Computing for Low-Latency, Resource-Efficient Inference

Serverless computing offers a compelling cloud model for online inference services. However, existing serverless platforms lack efficient support for GPUs, hindering their ability to deliver high-performance inference. In this paper, we present Torpor, a serverless platform for GPU-efficient, low-latency inference. To enable efficient sharing of a node's GPUs among numerous inference functions, Torpor maintains models in main memory and dynamically swaps them onto GPUs upon request arrivals (i.e., late binding with model swapping). Torpor uses various techniques, including asynchronous API redirection, GPU runtime sharing, pipelined model execution, and efficient GPU memory management, to minimize latency overhead caused by model swapping. Additionally, we design an interference-aware request scheduling algorithm that utilizes high-speed GPU interconnects to meet latency service-level objectives (SLOs) for individual inference functions. We have implemented Torpor and evaluated its performance in a production environment. Utilizing late binding and model swapping, Torpor can concurrently serve hundreds of inference functions on a worker node with 4 GPUs, while achieving latency performance comparable to native execution, where each model is cached exclusively on a GPU. Pilot deployment in a leading commercial serverless cloud shows that Torpor reduces the GPU provisioning cost by 70% and 65% for users and the platform, respectively.

cs.DC↗

arXiv2025

FSSUWNet: Mitigating the Fragility of Pre-trained Models with Feature Enhancement for Few-Shot Semantic Segmentation in Underwater Images

Few-Shot Semantic Segmentation (FSS), which focuses on segmenting new classes in images using only a limited number of annotated examples, has recently progressed in data-scarce domains. However, in this work, we show that the existing FSS methods often struggle to generalize to underwater environments. Specifically, the prior features extracted by pre-trained models used as feature extractors are fragile due to the unique challenges of underwater images. To address this, we propose FSSUWNet, a tailored FSS framework for underwater images with feature enhancement. FSSUWNet exploits the integration of complementary features, emphasizing both low-level and high-level image characteristics. In addition to employing a pre-trained model as the primary encoder, we propose an auxiliary encoder called Feature Enhanced Encoder which extracts complementary features to better adapt to underwater scene characteristics. Furthermore, a simple and effective Feature Alignment Module aims to provide global prior knowledge and align low-level features with high-level features in dimensions. Given the scarcity of underwater images, we introduce a cross-validation dataset version based on the Segmentation of Underwater Imagery dataset. Extensive experiments on public underwater segmentation datasets demonstrate that our approach achieves state-of-the-art performance. For example, our method outperforms the previous best method by 2.8% and 2.6% in terms of the mean Intersection over Union metric for 1-shot and 5-shot scenarios in the datasets, respectively. Our implementation is available at https://github.com/lizhh268/FSSUWNet.

cs.CV↗

arXiv2025

VulBinLLM: LLM-powered Vulnerability Detection for Stripped Binaries

Recognizing vulnerabilities in stripped binary files presents a significant challenge in software security. Although some progress has been made in generating human-readable information from decompiled binary files with Large Language Models (LLMs), effectively and scalably detecting vulnerabilities within these binary files is still an open problem. This paper explores the novel application of LLMs to detect vulnerabilities within these binary files. We demonstrate the feasibility of identifying vulnerable programs through a combined approach of decompilation optimization to make the vulnerabilities more prominent and long-term memory for a larger context window, achieving state-of-the-art performance in binary vulnerability analysis. Our findings highlight the potential for LLMs to overcome the limitations of traditional analysis methods and advance the field of binary vulnerability detection, paving the way for more secure software systems. In this paper, we present Vul-BinLLM , an LLM-based framework for binary vulnerability detection that mirrors traditional binary analysis workflows with fine-grained optimizations in decompilation and vulnerability reasoning with an extended context. In the decompilation phase, Vul-BinLLM adds vulnerability and weakness comments without altering the code structure or functionality, providing more contextual information for vulnerability reasoning later. Then for vulnerability reasoning, Vul-BinLLM combines in-context learning and chain-of-thought prompting along with a memory management agent to enhance accuracy. Our evaluations encompass the commonly used synthetic dataset Juliet to evaluate the potential feasibility for analysis and vulnerability detection in C/C++ binaries. Our evaluations show that Vul-BinLLM is highly effective in detecting vulnerabilities on the compiled Juliet dataset.

cs.CR↗

arXiv2025

ShadowMaskFormer: Mask Augmented Patch Embeddings for Shadow Removal

Transformer recently emerged as the de facto model for computer vision tasks and has also been successfully applied to shadow removal. However, these existing methods heavily rely on intricate modifications to the attention mechanisms within the transformer blocks while using a generic patch embedding. As a result, it often leads to complex architectural designs requiring additional computation resources. In this work, we aim to explore the efficacy of incorporating shadow information within the early processing stage. Accordingly, we propose a transformer-based framework with a novel patch embedding that is tailored for shadow removal, dubbed ShadowMaskFormer. Specifically, we present a simple and effective mask-augmented patch embedding to integrate shadow information and promote the model's emphasis on acquiring knowledge for shadow regions. Extensive experiments conducted on the ISTD, ISTD+, and SRD benchmark datasets demonstrate the efficacy of our method against state-of-the-art approaches while using fewer model parameters.g fewer model parameters. Our implementation is available at https://github.com/lizhh268/ShadowMaskFormer.

cs.CV↗

arXiv2022

RealNet: Combining Optimized Object Detection with Information Fusion Depth Estimation Co-Design Method on IoT

Depth Estimation and Object Detection Recognition play an important role in autonomous driving technology under the guidance of deep learning artificial intelligence. We propose a hybrid structure called RealNet: a co-design method combining the model-streamlined recognition algorithm, the depth estimation algorithm with information fusion, and deploying them on the Jetson-Nano for unmanned vehicles with monocular vision sensors. We use ROS for experiment. The method proposed in this paper is suitable for mobile platforms with high real-time request. Innovation of our method is using information fusion to compensate the problem of insufficient frame rate of output image, and improve the robustness of target detection and depth estimation under monocular vision.Object Detection is based on YOLO-v5. We have simplified the network structure of its DarkNet53 and realized a prediction speed up to 0.01s. Depth Estimation is based on the VNL Depth Estimation, which considers multiple geometric constraints in 3D global space. It calculates the loss function by calculating the deviation of the virtual normal vector VN and the label, which can obtain deeper depth information. We use PnP fusion algorithm to solve the problem of insufficient frame rate of depth map output. It solves the motion estimation depth from three-dimensional target to two-dimensional point based on corner feature matching, which is faster than VNL calculation. We interpolate VNL output and PnP output to achieve information fusion. Experiments show that this can effectively eliminate the jitter of depth information and improve robustness. At the control end, this method combines the results of target detection and depth estimation to calculate the target position, and uses a pure tracking control algorithm to track it.

cs.CV↗