arXiv subjects

Ziang Li

Publications and source records attributed to Ziang Li.

At least 19 recordsLinked to original sources

arXiv2026

WorldReward: Reward Modeling for Camera-Conditioned World Models

Camera-conditioned world models generate interactive videos in which commanded actions should induce the expected scene changes while appearance, geometry, and temporal dynamics remain coherent. Existing rewards assess these requirements separately: geometry-based rewards estimate trajectory execution but cannot judge the visual quality of the executed motion, whereas image-based rewards measure frame quality without capturing action execution or temporal dynamics. We posit that a vision-language model (VLM) offers a shared reasoning space for relating actions to their visual outcomes. However, judging a complete long video against its full action sequence creates a lengthy, noisy context in which short-lived local action evidence can be missed or diluted. We present WorldReward, a VLM-based pairwise preference reward model that unifies action-consistency and visual-quality evaluation for camera-conditioned world models. WorldReward decomposes paired videos into action-aligned chunks, organizes each chunk into structured visual evidence, and aggregates chunk-level decisions by voting into separate video-level action and visual-quality preferences. To train it, we construct a large-scale reasoning-augmented preference dataset using structured judgments generated by a frontier VLM and refined through tool-based agent auditing and targeted human review. We further introduce WorldReward-Bench, a human-annotated benchmark measuring reward-model agreement with human preferences across action consistency, appearance quality, and motion quality. WorldReward achieves the highest agreement on all three dimensions, exceeding GPT-5.5 by 3.42, 1.45, and 3.56 percentage points, respectively. When used for RL post-training of HY-WorldPlay 1.5, it consistently improves both action execution and visual quality across short- to long-term horizons.

cs.CV↗

arXiv2026

InstructMove: A Text-Indispensable Benchmark for Instruction-Following Manipulation

Vision-language-action (VLA) models have made general-purpose robot manipulation increasingly plausible by conditioning robot actions on natural-language instructions. A key test of such generality is whether policies actually follow language instructions. Yet many manipulation benchmarks leave this ability underdetermined: the intended object or destination is often visually salient or uniquely feasible, allowing policies to succeed without grounding the instruction. We argue that instruction-following evaluation should be text-indispensable: multiple actions should be visually and physically plausible, while only one should be consistent with the language instruction. We introduce InstructMove, a text-indispensable benchmark for instruction-following manipulation. InstructMove instantiates this principle in pick-and-place scenes with semantic distractors, decomposing instruction following into category identification, attribute discrimination, spatial reasoning, and compositional pick-and-place. InstructMove supports a train-eval protocol with InstructMove training data and held-out evaluation tasks, with additional diagnostics for language dependence. Experiments with representative VLA policies show that InstructMove provides a controlled testbed for diagnosing visual shortcuts and that InstructMove simulation data can improve real-world instruction-following manipulation performance. Code: https://github.com/HorizonRobotics/RoboOrchardSim

cs.RO↗

arXiv2026

Robust Bimanual Vision-Language-Action Models via Embarrassingly Simple Modality Masking

Query-based Vision-Language-Action (VLA) models offer low-latency inference that is attractive for bimanual robotic manipulation, but we observe that they can still exhibit discontinuous actions and execution failures in complex dual-arm tasks. We hypothesize that unstable multi-view and language fusion is one contributing factor in these failures, often coinciding with attention spreading to distracting regions. To improve robustness, we introduce the Modality Masking Mechanism (M3), an embarrassingly simple, training-only strategy that requires no architectural changes or large-scale robot pretraining. M3 stochastically masks subsets of modality channels during training, exposing the policy to controlled partial observations and encouraging it to rely less on distracting cues and more on evidence that remains reliable. We evaluate M3 on ten bimanual tasks from RoboTwin 2.0 and on three long-horizon real-world tasks. Compared with the Adapter baseline, M3 improves average success by 21.7% in the Clean setting and 11.4% in Clean2Rand, where policies are trained on clean demonstrations and evaluated on randomized scenes, while also improving averaged real-world full-task success by over 30%. These results suggest that structured training-time masking is a practical way to improve the robustness of query-based VLA policies for bimanual manipulation.

cs.RO↗

arXiv2026

From Long to Short: How Interest Rates Shape Life Insurance Markets

This paper explores how financial institutions pass interest rate risk through to product markets using the life insurance industry as a setting. We show theoretically that it is optimal for insurers to distort product issuance across maturities to offset duration gaps. We examine insurers exogenously exposed to interest rate risk through their variable annuity liabilities after the 2008 financial crisis. Consistent with our mechanism, exposed insurers developed negative duration gaps, increased markups on long-duration products, and rebalanced product issuance toward shorter-duration products to hedge. This response reduced long-duration life insurance coverage by 12.1% of GDP between 2005 and 2023.

econ.GN↗

arXiv2026

EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI

We present EmbodiedGen V2, a generative 3D world engine for building executable policy-ready environments for embodied intelligence. Sim-ready 3D asset generation has advanced rapidly, yet assembling such assets into policy-ready task environments remains largely manual, limiting scalable closed-loop learning. EmbodiedGen V2 addresses this gap through a unified sim-ready representation that connects cross-simulator assets, interaction affordances, task-driven worlds, large-scale multi-room scenes, and stateful Vibe Coding into a generative, editable, and reusable simulation pipeline. The generated environments support manipulation, navigation, mobile manipulation, cross-simulator deployment, and embodied policy training. In evaluation, the asset pipeline achieves 96.5% human acceptance and 98.6% collision success, and 83.3% of task-driven worlds are directly usable for downstream simulation without manual modification. Online reinforcement learning with generated environments further improves simulation success from 9.7% to 79.8%, and transfers to real robots with task success increasing from 21.7% to 75.0%. These results establish EmbodiedGen V2 as scalable simulation infrastructure for training, evaluating, and deploying embodied policies.

cs.RO↗

arXiv2026

ManiSplat: Manipulation Trajectory Synthesis from Monocular Video via Decoupled 3D Gaussian Splatting

Reconstructing dynamic and interactive 3D scenes from real-world observations remains a fundamental challenge in computer vision and robotics. While recent advances in 3D Gaussian Splatting have enabled high-fidelity static reconstruction, extending it to interactive environments with articulated robots and manipulable objects remains difficult due to complex contact interactions and abrupt pose changes. To address these challenges, we introduce ManiSplat, a unified framework that reconstructs controllable and decoupled Gaussian digital twins directly from monocular ego-view robotic videos. Our method introduces a Graph-Structured Disentangled Representation that separates the robot, objects, and background into independently optimizable Gaussian subfields organized within a scene graph. To ensure stability, we propose a Task-Oriented Spatio-Temporal Alignment module that leverages the inherent logic of manipulation tasks-alternating between Motion and Skill phases-to construct accurate pseudo-ground-truth trajectories. Finally, a joint photometric-geometric optimization ensures the reconstructed scenes are temporally coherent, physically consistent, and simulation-ready. Extensive experiments demonstrate that our approach reconstructs interaction-driven dynamic scenes with high fidelity and controllability, effectively supporting downstream robotic tasks and policy learning.

cs.CV↗

arXiv2026

Light-WAM: Efficient World Action Models with State-Fusion Action Decoding

World Action Models (WAMs) extend robot policy learning by incorporating future prediction as an additional training objective, encouraging the policy to encode task-relevant temporal structure in its representations. Current WAMs often rely on large-scale generative architectures that incur high training costs and inference latency, making them difficult to deploy as efficient closed-loop policies. We propose Light-WAM, a lightweight World Action Model for efficient robot manipulation. Specifically, it is built with a compact video backbone and performs future-video supervision in a downsampled latent space, reducing the cost of video co-training while retaining its benefits for representation learning. For action prediction, Light-WAM introduces the StateFusionActionExpert, which reads adapted states from multiple backbone layers, fuses them through learned-query pooling, and directly predicts action chunks in a single forward pass. This design provides an efficient interface between video backbone representations and robot actions, avoiding the need for heavy generative action experts. Experiments demonstrate that Light-WAM maintains strong performance on LIBERO and achieves usable multi-task performance on RoboTwin 2.0, while using only 0.44B trainable parameters. It also achieves 72.03ms inference latency with 4.1GiB peak GPU memory and improved training throughput.

cs.CV↗

arXiv2026

Not All Tokens Are Worth Caching: Learning Semantic-Aware Eviction for LLM Prefix Caches

Prefix caching is a key optimization in Large Language Model (LLM) serving, reusing attention Key-Value (KV) states across requests with shared prompt prefixes to reduce expensive prefill computation. However, its benefit depends critically on the eviction policy as GPU memory is scarce, and existing policies such as LRU largely treat cached blocks uniformly. This view ignores a fundamental property of LLM prompts: not all tokens are equally worth caching. We show that different token types within a prompt, including system prompts, user queries, tool outputs, model responses, and chain-of-thought reasoning, exhibit up to 756x variation in reuse rates, yet no existing eviction policy exploits this signal. In this paper, we present SAECache (Semantic-Adaptive Eviction for prefix caches), a semantic-adaptive prefix cache eviction policy that addresses this gap through three innovations: (1) a multi-queue architecture that routes KV blocks to task-specific queues with tailored priority metrics, capturing both session reuse in multi-turn requests and structural reuse in templated single-turn requests; (2) a semantic-aware token weighting mechanism that learns the reuse value of different token types online through eviction feedback; and (3) a fully adaptive online learning schema for all parameter updates, including log-normal timing parameters, position decay power, queue weights, and meta-parameters, which eliminates manual tuning and enables automatic adaptation to deployment-specific workload characteristics. Through extensive evaluation across heterogeneous workloads, we demonstrate that SAECache achieves 1.4x-2.7x TTFT improvement over production-style baselines, while fixed-parameter alternatives can degrade by up to 2.7x under workload mismatch -- a failure mode our adaptive approach avoids entirely.

cs.LG↗

arXiv2026

VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models

Scalable Vector Graphics (SVG) are essential for technical illustration and digital design, offering resolution independence and semantic editability. In practice, original vector files are frequently lost, leaving only rasterized versions (e.g., PNG, JPEG) that resist modification, while manual reconstruction is prohibitively expensive. Progress on automating raster-to-SVG conversion has been bottlenecked by two gaps: existing SVG datasets are dominated by icons and decorative graphics that lack the complexity of professional diagrams, and existing benchmarks rely on pixel- or embedding-level similarity that fails to capture structural correctness (e.g., broken connectivity, misplaced arrows). We close both gaps with paired contributions targeting diagram-centric figures (e.g., model architectures, flowcharts, schematics). For training, we introduce VFIG-Data, the largest figure-to-SVG dataset of its kind at 66K pairs, combining real paper figures converted via a describe-and-generate pipeline with programmatic diagrams that supply noise-free supervision over arrow styles, fonts, and geometry. For evaluation, we introduce VFIG-Bench, a structure-aware evaluation suite, paired with VFIG-Bench-OOD, an out-of-distribution set of figures manually curated from highly cited arXiv papers. Beyond pixel and embedding similarity, our protocol reports rubric-based VLM-Judge scores and Elo ratings from pairwise human preference evaluation. Built on these contributions, VFIG is a VLM family trained with a simple-to-complex SFT curriculum followed by RL with rendering-aware rewards. VFIG achieves state-of-the-art open-source performance, outperforming the best open-source VLM baseline by over 30%, and matches Claude Sonnet 4.6 on VFIG-BENCH: Gemini-Judge 78.2% vs. 76.7% and GPT-Judge 87.5% vs. 87.4%. It remains slightly behind the strongest proprietary models GPT-5.2 and Gemini-3.

cs.CV↗

arXiv2026

HoloBrain-0 Technical Report

In this work, we introduce HoloBrain-0, a comprehensive Vision-Language-Action (VLA) framework that bridges the gap between foundation model research and reliable real-world robot deployment. The core of our system is a novel VLA architecture that explicitly incorporates robot embodiment priors, including multi-view camera parameters and kinematic descriptions (URDF), to enhance 3D spatial reasoning and support diverse embodiments. We validate this design through a scalable ``pre-train then post-train" paradigm, achieving state-of-the-art results on simulation benchmarks such as RoboTwin 2.0, LIBERO, and GenieSim, as well as strong results on challenging long-horizon real-world manipulation tasks. Notably, our efficient 0.2B-parameter variant rivals significantly larger baselines, enabling low-latency on-device deployment. To further accelerate research and practical adoption, we fully open-source the entire HoloBrain ecosystem, which includes: (1) powerful pre-trained VLA foundations; (2) post-trained checkpoints for multiple simulation suites and real-world tasks; and (3) RoboOrchard, a full-stack VLA infrastructure for data curation, model training and deployment. Together with standardized data collection protocols, this release provides the community with a complete, reproducible path toward high-performance robotic manipulation.

cs.RO↗

arXiv2025

Rethinking Human Preference Evaluation of LLM Rationales

Large language models (LLMs) often generate natural language rationales -- free-form explanations that help improve performance on complex reasoning tasks and enhance interpretability for human users. However, evaluating these rationales remains challenging. While recent work has relied on binary preference judgments from humans or LLM judges, such evaluations are often opaque and coarse-grained, offering limited insight into what makes one rationale better than another. In this work, we rethink preference evaluation for LLM-generated rationales by asking: (1) What attributes define good rationales? (2) Can human preferences be explained by these attributes? (3) Can attribute-based evaluation overcome the limitations of binary comparisons? We identify a set of key rationale attributes from prior literature and assess them using automatic metrics, LLM judgments, and human annotations. We then analyze two standard human preference datasets MT Bench and Chatbot Arena using SHAP to identify which attributes best explain human preference outcomes. Finally, we re-evaluate model-generated rationales using attribute-specific ELO scores, revealing more nuanced model comparisons and insights. Our findings suggest that fine-grained attribute evaluations can better characterize rationale quality and guide future research toward more interpretable and reliable evaluation practices.

cs.AI↗

arXiv2025

Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations

Spatial cognition is essential for human intelligence, enabling problem-solving through visual simulations rather than solely relying on verbal reasoning. However, existing AI benchmarks primarily assess verbal reasoning, neglecting the complexities of non-verbal, multi-step visual simulation. We introduce STARE(Spatial Transformations and Reasoning Evaluation), a benchmark designed to rigorously evaluate multimodal large language models on tasks better solved through multi-step visual simulation. STARE features 4K tasks spanning foundational geometric transformations (2D and 3D), integrated spatial reasoning (cube net folding and tangram puzzles), and real-world spatial reasoning (perspective and temporal reasoning), reflecting practical cognitive challenges like object assembly, mechanical diagram interpretation, and everyday spatial navigation. Our evaluations show that models excel at reasoning over simpler 2D transformations, but perform close to random chance on more complex tasks like 3D cube net folding and tangram puzzles that require multi-step visual simulations. Humans achieve near-perfect accuracy but take considerable time (up to 28.9s) on complex tasks, significantly speeding up (down by 7.5 seconds on average) with intermediate visual simulations. In contrast, models exhibit inconsistent performance gains from visual simulations, improving on most tasks but declining in specific cases like tangram puzzles (GPT-4o, o1) and cube net folding (Claude-3.5, Gemini-2.0 Flash), indicating that models may not know how to effectively leverage intermediate visual information.

cs.CV↗

arXiv2025

From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning

Model Inversion Attacks (MIAs) aim to reconstruct private training data from models, leading to privacy leakage, particularly in facial recognition systems. Although many studies have enhanced the effectiveness of white-box MIAs, less attention has been paid to improving efficiency and utility under limited attacker capabilities. Existing black-box MIAs necessitate an impractical number of queries, incurring significant overhead. Therefore, we analyze the limitations of existing MIAs and introduce Surrogate Model-based Inversion with Long-tailed Enhancement (SMILE), a high-resolution oriented and query-efficient MIA for the black-box setting. We begin by analyzing the initialization of MIAs from a data distribution perspective and propose a long-tailed surrogate training method to obtain high-quality initial points. We then enhance the attack's effectiveness by employing the gradient-free black-box optimization algorithm selected by NGOpt. Our experiments show that SMILE outperforms existing state-of-the-art black-box MIAs while requiring only about 5% of the query overhead.

cs.CR↗

arXiv2025

DancingBoard: Streamlining the Creation of Motion Comics to Enhance Narratives

Motion comics, a digital animation format that enhances comic book narratives, have wide applications in storytelling, education, and advertising. However, their creation poses significant challenges for amateur creators, primarily due to the need for specialized skills and complex workflows. To address these issues, we conducted an exploratory survey (N=58) to understand the challenges associated with creating motion comics, and an expert interview (N=4) to identify a typical workflow for creation. We further analyzed $95$ online motion comics to gain insights into the design space of character and object actions. Based on our findings, we proposed DancingBoard, an integrated authoring tool designed to simplify the creation process. This tool features a user-friendly interface and a guided workflow, providing comprehensive support throughout each step of the creation process. A user study involving 23 creators showed that, compared to professional tools, DancingBoard is easily comprehensible and provides improved guidance and support, requiring less effort from users. Additionally, a separate study with $18$ audience members confirmed the tool's effectiveness in conveying the story to its viewers.

cs.HC↗

arXiv2025

Prefer2SD: A Human-in-the-Loop Approach to Balancing Similarity and Diversity in In-Game Friend Recommendations

In-game friend recommendations significantly impact player retention and sustained engagement in online games. Balancing similarity and diversity in recommendations is crucial for fostering stronger social bonds across diverse player groups. However, automated recommendation systems struggle to achieve this balance, especially as player preferences evolve over time. To tackle this challenge, we introduce Prefer2SD (derived from Preference to Similarity and Diversity), an iterative, human-in-the-loop approach designed to optimize the similarity-diversity (SD) ratio in friend recommendations. Developed in collaboration with a local game company, Prefer2D leverages a visual analytics system to help experts explore, analyze, and adjust friend recommendations dynamically, incorporating players' shifting preferences. The system employs interactive visualizations that enable experts to fine-tune the balance between similarity and diversity for distinct player groups. We demonstrate the efficacy of Prefer2SD through a within-subjects study (N=12), a case study, and expert interviews, showcasing its ability to enhance in-game friend recommendations and offering insights for the broader field of personalized recommendation systems.

cs.HC↗

arXiv2025

SteROI-D: System Design and Mapping for Stereo Depth Inference on Regions of Interest

Machine learning algorithms have enabled high quality stereo depth estimation to run on Augmented and Virtual Reality (AR/VR) devices. However, high energy consumption across the full image processing stack prevents stereo depth algorithms from running effectively on battery-limited devices. This paper introduces SteROI-D, a full stereo depth system paired with a mapping methodology. SteROI-D exploits Region-of-Interest (ROI) and temporal sparsity at the system level to save energy. SteROI-D's flexible and heterogeneous compute fabric supports diverse ROIs. Importantly, we introduce a systematic mapping methodology to effectively handle dynamic ROIs, thereby maximizing energy savings. Using these techniques, our 28nm prototype SteROI-D design achieves up to 4.35x reduction in total system energy compared to a baseline ASIC.

cs.CV↗

arXiv2024

Double-Ended Synthesis Planning with Goal-Constrained Bidirectional Search

Computer-aided synthesis planning (CASP) algorithms have demonstrated expert-level abilities in planning retrosynthetic routes to molecules of low to moderate complexity. However, current search methods assume the sufficiency of reaching arbitrary building blocks, failing to address the common real-world constraint where using specific molecules is desired. To this end, we present a formulation of synthesis planning with starting material constraints. Under this formulation, we propose Double-Ended Synthesis Planning (DESP), a novel CASP algorithm under a bidirectional graph search scheme that interleaves expansions from the target and from the goal starting materials to ensure constraint satisfiability. The search algorithm is guided by a goal-conditioned cost network learned offline from a partially observed hypergraph of valid chemical reactions. We demonstrate the utility of DESP in improving solve rates and reducing the number of search expansions by biasing synthesis planning towards expert goals on multiple new benchmarks. DESP can make use of existing one-step retrosynthesis models, and we anticipate its performance to scale as these one-step model capabilities improve.

cs.AI↗

arXiv2024

A Stealthy Wrongdoer: Feature-Oriented Reconstruction Attack against Split Learning

Split Learning (SL) is a distributed learning framework renowned for its privacy-preserving features and minimal computational requirements. Previous research consistently highlights the potential privacy breaches in SL systems by server adversaries reconstructing training data. However, these studies often rely on strong assumptions or compromise system utility to enhance attack performance. This paper introduces a new semi-honest Data Reconstruction Attack on SL, named Feature-Oriented Reconstruction Attack (FORA). In contrast to prior works, FORA relies on limited prior knowledge, specifically that the server utilizes auxiliary samples from the public without knowing any client's private information. This allows FORA to conduct the attack stealthily and achieve robust performance. The key vulnerability exploited by FORA is the revelation of the model representation preference in the smashed data output by victim client. FORA constructs a substitute client through feature-level transfer learning, aiming to closely mimic the victim client's representation preference. Leveraging this substitute client, the server trains the attack model to effectively reconstruct private data. Extensive experiments showcase FORA's superior performance compared to state-of-the-art methods. Furthermore, the paper systematically evaluates the proposed method's applicability across diverse settings and advanced defense strategies.

cs.CR↗