Search arXiv⌕ Search

arXiv subjects

Haoyu Zhang

Publications and source records attributed to Haoyu Zhang.

At least 19 recordsLinked to original sources
arXiv2026

VibeVideoBench: Evaluating Multimodal Agents on Code-Driven Video Generation

Diffusion video models produce striking clips, but they cannot reliably render an exact formula, a correct chart, a rule-consistent game or narration synchronized to on-screen events. A different route is emerging: multimodal agents that write code to render every frame and synthesize every sound. We introduce VibeVideoBench, a benchmark for evaluating agents on code-driven video generation. Its 239 self-contained tasks in English and Chinese span 12 categories, from data stories and music-driven pieces to classroom lessons and paper explainers; each asks an agent, running in its native harness inside a sandbox with no image or video generation model, to deliver a finished MP4 with sound within one hour. Evaluation is hierarchical and evidence-based: program checks of the delivery spec, about 13k task-specific rubric checks traced to the requirements they enforce, and five quality dimensions anchored per task, scored by three VLM judge panels. Across nine frontier agent systems and 2,151 scored rollouts, GPT-6 Astra and Claude Opus 5.5 rank first and second with both the judges and human annotators, and the strongest system satisfies every core requirement on 76.3% of tasks. Delivery is a second axis: counting missing videos as zero changes the ranking, and three quarters of the lost videos are final renders that started too late, ran too slowly or never started, a failure of how agents plan their hour. Agents build their own rendering pipelines and inspect their own frames, yet how much they look does not predict quality. The judges' ranking of the systems closely matches blind human preferences.

cs.CV↗
arXiv2026

Separating Angular Alignment from Radial Ordering in Hyperbolic Vision-Language Models

Hyperbolic vision-language models are well suited to modeling the hierarchical structure of multimodal data, because hyperbolic volume grows exponentially with radius, enabling low-distortion representations of tree-like taxonomies from general to specific concepts. However, existing pretraining objectives for hyperbolic vision-language models entangle semantic alignment with entailment modeling, forcing radial coordinates to simultaneously encode semantic matching and hierarchical structure, leading to conflicting geometric constraints. In this paper, we introduce AURORA, a hyperbolic vision-language framework that decouples the objectives through angular alignment and a novel Radial-Angular Order Cone (RAOC) for entailment modeling. Angular alignment captures semantic similarity while preserving the radial structure for representing hierarchies. Unlike conventional entailment cones, RAOC determines angular allowance from relative parent-child radial gaps. It is incorporated into hierarchical entailment contrastive training to capture partial-order relations. We further introduce a norm regularization term that prevents undesirable radial degeneration for stable training. We train models on millions of image-text pairs and evaluate them across zero-shot classification, image-text retrieval, hierarchical reasoning, and compositional understanding. Our model consistently outperforms strong Euclidean and hyperbolic vision-language baselines, demonstrating the effectiveness of decoupling semantic alignment from hierarchical entailment modeling.

cs.CV↗
arXiv2026

DRAMA-Eval: A Benchmark for Context-Aware, Multi-Turn, Multi-Speaker Speech Generation

Recent advances in speech generation are shifting toward context-aware, multi-turn, and multi-speaker synthesis. While these capabilities are in high demand for applications such as audiobook and audio drama production, evaluation has not kept pace with these advances. Beyond traditional zero-shot Text-to-Speech (TTS), modern systems in this domain primarily fall into three categories: instruction-controllable generation models, multi-turn dialogue models, and the audio branches of audio-visual generation models. However, existing benchmarks evaluate these paradigms in isolation, failing to provide a unified assessment of context modeling, instruction following, and multi-speaker generation. Meanwhile, recent progress in audio understanding models has made it feasible to evaluate such complex scenarios using LLM-as-a-Judge approaches. To bridge this gap, we introduce DRAMA-Eval, a comprehensive benchmark for evaluating such systems. DRAMA-Eval utilizes structured script contexts and offers a holistic evaluation framework covering objective, local, and global aspects of generation quality, enabling the assessment of acoustic fidelity, sentence-level execution, and long-range coherence. We benchmark leading open-source and commercial models on a human?verified dataset. The results reveal critical weaknesses in long-range speaker consistency and implicit instruction following that require deep contextual understanding. To facilitate reproducible research, we release a 12-hour bilingual (Chinese-English) test set of manually corrected real-world scripts, accompanied by ground-truth audio, evaluation code, and benchmark results.

eess.AS↗
arXiv2026

GeoBridge: Evolving Transferable Decoder Topologies for Geometric Vision

Depth estimation and Gaussian reconstruction share a need for multi-scale feature processing, yet use different prediction interfaces and training objectives. Whether a decoder's internal computation can be optimized once and reused across these systems remains a structural design question. Search over primitive configurations alone leaves the organization of transformations fixed. To address this gap, we propose GeoBridge, an evolutionary framework for transferable decoder topologies. First, GeoBridge makes stage-internal topology searchable through residual, sequential, and parallel composition of configurable primitives. Shape-preserving composition retains the stage interfaces of a fixed multi-scale scaffold. Second, a functional-tree encoding supports stage-aware crossover and hierarchical mutation, while a gradient sensitivity-alignment proxy evaluates candidates without training. Multi-objective evolution balances this proxy with computation and parameter cost. Third, selected topologies are instantiated and trained in target systems with task-specific prediction interfaces. A single source search takes approximately 12 GPU hours. GeoBridgeNet-A3 achieves the best depth accuracy among four adapted decoder baselines under a common encoder and training protocol on NYU Depth V2 and KITTI. The same topology reaches 27.11 dB PSNR in Gaussian novel-view synthesis on RealEstate10K, and its refined representations support open-vocabulary segmentation on LERF. These results support decoder topology as a reusable design object across the evaluated geometric vision systems. Code and pretrained weights are available at https://github.com/EMI-Group/GeoBridge-NAS.

cs.CV↗
arXiv2026

Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs

Meta deploys a heterogeneous fleet comprising NVIDIA GPUs, AMD GPUs, and proprietary MTIA accelerators to reduce costs and mitigate vendor lock-in. However, the extensive history of NVIDIA-specific optimizations hinders performance portability to other hardware. This paper presents our methods for optimizing alternative accelerators, using AMD Instinct MI350X versus NVIDIA B200 for recommendation model training as a case study. We propose three reusable principles. First, hardware-model co-design is essential: rather than directly porting models from NVIDIA GPUs, revising model structures and software recipes allows full exploitation of accelerator-specific strengths (e.g., MI350X's 1.6x HBM capacity relative to B200). Second, our analytical performance model enables zero-warmup optimal kernel selection across diverse datatypes and matrix shapes in production models, drastically simplifying tuning. Third, increasingly mature agentic workflows have lowered the barrier to porting from the CUDA ecosystem; we extend these workflows for microarchitecture-aware kernel optimizations and custom shim generation. Compared to the directly ported baseline, these methods deliver up to 1.5x improvement in end-to-end training throughput. To our knowledge, this is the first hyperscaler report on productionizing AMD GPUs for large-scale training. Our contributions have been upstreamed to the open-source PyTorch, Triton, and ROCm ecosystems.

cs.DC↗
arXiv2026

Code2Worlds-2: Physically Plausible Agent Harness for Embodied World Simulation

Recent coding large language models (LLMs) can synthesize dynamic 4D worlds from natural language, but the resulting content remains a visual artifact. It lacks physical semantics such as the collision geometry, mass, friction, inertia, and joint structure, required to enable it to simulate the physical world, and its dynamics are scripted to look plausible rather than designed to execute. What is needed is an executable world: one that a physics engine can load and step, resolving contact physically. We present Code2Worlds-2, an extension of Code2Worlds that turns 4D world generation into an embodied world simulation. We keep the original dual-stream generation pipeline intact and append two modules. First, a SimReady Scene Compiler that turns a generated visual scene into a simulator-compatible representation: it assigns collision meshes via convex decomposition, estimates physical attributes with a vision-language model, infers a lightweight support graph, instantiates the engine's contact, friction, and dynamics constraints, and applies a gravity settling pass, before exporting the scene as MuJoCo MJCF. Second, a Simulator-in-the-Loop 4D Refiner that implements an automated agent harness-a compile, simulate, critique and refine loop-that converts visual dynamics into executable rigid and articulated dynamics and evolves the physical settings rather than only the code, replacing the purely visual motion critic with a simulator critic that inspects penetration depth, collision pairs, stability, contact forces, joint limits, and robot reachability. We further propose SimCode4D, a state-based benchmark with five task templates for evaluating embodied simulation. Experiments on SimCode4D and across natural, indoor, industrial, urban, and service scenes show that Code2Worlds-2 produces physically plausible, simulator-compatible, and robot-interactable scenes.

cs.RO↗
arXiv2026

Depth, Not Breadth: Best-of-N Jailbreaking Beyond Surface Noise

Best-of-N jailbreaking spends a query budget on surface variation, scrambling and recasing a request until one draw lands. We ask what the same search buys when it wraps a structural encoding instead. Against SAGE, a published self-check defense, best-of-N over a code-completion encoding reaches 67, 22 and 15% of behaviors on three open-weight targets, where that encoding fired once reaches at most 4.7% and the published character search at full budget at most 3.0%: 9 to 68 times the sum of the parts, with bootstrap intervals clearing both ingredients on every target. We report the operative figure beside the headline rather than the headline alone: at the actionable severity threshold those cells read 24, 8 and 1 behaviors (95% interval [13, 28], [3, 13], [0, 3]). A 2x2 on Llama-3.1-8B holding encoding and variation apart shows that SAGE and a classifier gate fail to different factors: SAGE is broken by the depth of the encoding (7 -> 67 behaviors at fixed variation) and the gate by the breadth of the variation (13 -> 57 at fixed encoding). Per-draw reporting also overstates robustness, because an attacker who may try N times experiences the maximum over draws while safety results are reported as means: on one target SAGE stops 99.8% of individual draws yet loses at least as many behaviors to a repeat attacker (12) as a classifier gate that stops 95.6% (10). Removing the target's sampling shrinks SAGE's coverage, relative to its own level, by 59, 76, 82 and 29 percentage points more than an undefended control's, against 25, -5, 8 and 2 for a defense whose verdict comes from a fixed shadow model. The design that loses is the one fusing screening and answering into a single generation, so every attacker draw redraws the safety decision as well. The prescription is architectural, not free: do not fuse screening with generation.

cs.CR↗
arXiv2026

Unread or Unenforced? Separating Representation from Enforcement Failure in Content Guards

When an encoded attack passes a content guard, the guard either never represented the payload's harmful content or represented it and failed to act. End-to-end attack success rate reports one number for both, yet the two have opposite remedies: one is a representational limit that more verdict-level safety training cannot reach, the other is a decision rule that it can. We separate them by reading a guard's own residual stream, using a content probe fitted on plaintext and transferred without refitting to the encoded condition, alongside the verdict logits from the same pass. Licensing that read honestly is most of the problem and is our main contribution. A conventional permutation test admits the decode measurement on most of a 19-condition encoding ladder for each of two open guards. A length-matched null and a floor calibrated on conditions the guard's base model decodes at rate zero reduce it to four conditions each; holding out the items the probe was fitted on removes one more. A further screen constrains the block axis by running plaintext content inside each condition's own wrapper. It removes the largest surviving cell. What remains is a policy failure that survives an item-level holdout on the zero-width encoding for both guards, at 8 and 7 per 100 prompts, against 17 and 23 when the probe is allowed to have seen the prompt it is scoring. It is also confined to one family of surface encodings: where an encoding leaves content linearly recoverable we can separate the two failures, and on genuine ciphers we report the cells as unmeasured rather than as evidence that nothing was decoded. Across every guard and condition pair except two, both unmeasured on the decode axis, blocked without decoding is near zero, so we find little evidence for a pure encoding-format detector under the conditions we test. That cell is the one read we do not repeat under the holdout.

cs.CR↗
arXiv2026

SHARE: Sharing Human Actions across Robot Embodiments through Unified Retargeting and Tracking

The central challenge in reusing human actions across robots is the need to realize shared intent through incompatible kinematics, dynamics, and control interfaces. We present SHARE, which decomposes action sharing into two connected stages---reference generation and feedback execution---and uses native robot references as their interface to achieve shared structure, native realization. In the reference-generation stage, AdaMorph extracts human-motion intent with a shared temporal backbone and generates embodiment-compatible references through robot-conditioned decoding. In the feedback-execution stage, AdaTrack combines these references with proprioception, action history, and morphology, then uses variable-length dynamic representations, morphology-conditioned attention, expert distillation, and closed-loop reinforcement learning to emit native actions from one shared Actor. On G1, HI, AD, and T1, AdaMorph attains seen/unseen root-speed Pearson correlation coefficients (PCCs) of 0.951/0.894; AdaTrack reaches 94.00%/91.96% closed-loop Success Rate with 21.15% of the total parameters in four specialist Actors, while cross-embodiment pretraining improves seen/unseen adaptation by 9.67/13.62 points with only 1% target data.

cs.RO↗
arXiv2026

Inline Memory Meets Reusable Skills: Memory-centric Framework for Vision-Language-Action Model

Vision-Language-Action (VLA) models have shown strong promise for general-purpose robotic manipulation, yet adapting them to new tasks and domains remains inefficient: existing methods often rely on parameter tuning, incurring substantial costs and risking catastrophic forgetting of previously learned tasks. To address this, we propose \textbf{Optimus-R}, a memory-centric VLA framework that formulates robotic adaptation as explicit query-skill memory tuning. Optimus-R introduces: (i) An \textbf{Inline Memory Interface for skill extraction}. It inserts learnable memory tokens into the VLA prefix stream, allowing the backbone to derive control-aware query and skill representations within the native action-conditioning pathway. (ii) A \textbf{Query-Skill Memory Bank for skill learning}. It externalizes skills into query prototypes for deciding \emph{what} to retrieve and skill values for specifying \emph{how} to act, supporting skill reuse and expansion with limited parameter updates. (iii) A lightweight \textbf{Bridge-and-Adapt mechanism for skill updating}. It aligns target-domain queries and skills with the existing memory space through a lightweight adapter and residual memory updates. Experiments on in-domain adaptation, cross-domain adaptation, and lifelong learning show that Optimus-R enables data-efficient skill learning while mitigating catastrophic forgetting.

cs.CV↗
arXiv2026

WenetSpeech-Min: A Large-Scale Minnan Speech Corpus with Dual Transcriptions for Dialectal Speech Processing

Progress in dialectal speech technology is hindered by the scarcity of large-scale, real-world corpora. For Minnan speech, existing resources remain limited, and few provide paired Minnan and Mandarin transcripts at scale. To address these gaps, we introduce WenetSpeech-Min, an open-source corpus comprising around 10,000 hours of Minnan speech collected from diverse online media, with paired Minnan and Mandarin transcripts for every utterance. We further establish an automatic speech recognition (ASR) benchmark covering both Minnan and Mandarin transcripts and a text-to-speech synthesis (TTS) benchmark using Minnan transcripts, with manually verified evaluation sets for both tasks. To assess the effectiveness of the corpus, we train ASR and TTS models on WenetSpeech-Min and compare them with representative systems on the proposed benchmarks. The resulting models outperform the evaluated open-source models on most metrics and achieve competitive performance against commercial systems. We will release the corpus, benchmarks, and models to facilitate reproducible research on Minnan speech technology.

eess.AS↗
arXiv2026

When Sparse Reward Meets Dense Distillation: Training Dynamics of On-Policy Distillation

Reinforcement learning with verifiable rewards provides a sparse post-training signal: a single binary outcome evaluates the entire rollout, and every token receives the same sequence-level advantage regardless of its individual contribution. To complement this sparse supervision, a growing family of methods adds a scalar-weighted teacher KL term to the policy-gradient objective, providing dense token-level guidance that may be unreliable at some positions. Despite the benefits of combining these signals, their interaction during optimization can destabilize joint training. To understand how this instability develops, we study the learning dynamics of hybrid reward--distillation training through a neural tangent kernel (NTK) analysis. We introduce the cross-signal NTK $K_{DR}(n)$, a token-level statistic that measures the alignment between reward and distillation gradients at position n. Through this analysis, we identify two failure modes: 1 Magnitude drowning, where the reward gradient exceeds the distillation gradient by orders of magnitude, so that even weak directional conflict can cause the distillation loss to rise despite its explicit inclusion in the training objective; and 2 Localized directional conflict, where the sequence-level advantage and the teacher's position-specific distribution induce opposing updates at the same token ($K_{DR}(n)\!<\!0$). The severity of these effects depends on the optimization regime: the gradient-norm ratio $κ\!=\!\|\nabla\mathcal{L}_R\|/\|\nabla\mathcal{L}_D\|$ varies by roughly an order of magnitude across tasks, and our experiments reveal an empirical threshold beyond which naive mixing can lead to persistent training collapse. Motivated by these findings, we introduce the M3 family, which combines magnitude normalization with three strategies...

cs.LG↗
arXiv2026

An interacting particle consensus method for constrained global optimization

This paper presents a particle-based optimization method designed for addressing minimization problems with equality constraints, particularly in cases where the loss function exhibits non-differentiability or non-convexity. The proposed method combines components from consensus-based optimization algorithm with a newly introduced forcing term directed at the constraint set. A rigorous mean-field limit is derived for the anisotropic particle system, while convergence to the constrained minimizer is established for the isotropic mean-field dynamics. Additionally, we introduce a linearly implicit discretized algorithm, analyze its well-posedness, stability properties, and consistency with the continuous particle dynamics, and conduct various numerical experiments to demonstrate the performance of the proposed method.

math.OC↗
arXiv2026

ProxiCBO: A Provably Convergent Consensus-Based Method for Composite Optimization

This paper introduces an interacting-particle optimization method tailored to possibly non-convex composite optimization problems, which arise widely in signal processing. The proposed method, \emph{ProxiCBO}, integrates consensus-based optimization (CBO) with proximal gradient techniques to handle challenging optimization landscapes and exploit the composite structure of the objective function. We establish global convergence guarantees for the continuous-time finite-particle dynamics and develop an alternating update scheme for efficient practical implementation. Simulation results for signal processing tasks, including signal recovery from one-bit quantized measurements and parameter estimation from single-photon lidar data, demonstrate that ProxiCBO outperforms existing proximal gradient methods and CBO methods in terms of both accuracy and particle-efficiency.

math.OC↗
arXiv2026

Blind, Not Weak: A Best-of-Suite Safety-Utility Frontier for Recover-and-Reguard Defenses Against Encoded VLM Jailbreaks

Safety classifiers ("guards") are the dominant black-box defense for vision-language models, yet a guard judges an input's surface form, not its meaning: a harmful request re-encoded as set theory, formal logic, a classical language, code, or text rendered inside an image slips past a guard that would block it in plain language - the decode gap. The standard fix is a preprocessor that recovers image content and decodes the encoding before the guard. We build one and evaluate it against an ensemble of eleven encoding attacks - six published implementations, one standard encoding baseline, one adapted and three author-constructed renders - counting a behavior as broken if any attack succeeds. Restoring a view the guard never had is what buys coverage - block rates on image renders go from exactly zero to 67-90% - and what it costs in benign traffic is set by the guard, not by the mechanism: one guard pays 9 benign blocking points for the same 70-point gain another pays 69 for. It still does not make the system safer: against an attacker free to choose among eleven encodings, closing one channel relocates the success rather than removing it, and no ensemble contrast for that step survives multiple-comparison correction. What does lower ensemble attack success is a reguard step that re-screens the recovered pre-decode surface, and it is the one every guard pays for: it raises benign over-refusal on all ten guard-target pairs, where restoring a single channel raises it on some and not others. Across the full guard x target x condition factorial, no configuration reaches an ensemble attack-success rate at or below 40% while holding benign over-refusal under 70%. That empty region is a property of the configurations we sample, not a bound on what recovery-based defenses can reach, and we breach its safety half ourselves.

cs.CR↗
arXiv2026

Motion-Omni: End-to-End Joint Speech and Full-Body Motion for Spoken Dialogue

An avatar that holds a conversation should decide what to say and to move while saying it, yet these abilities live in separate model families: spoken dialogue models produce speech without motion, and co-speech motion models produce motion only from audio handed to them. The standard remedy is a cascade that first generates the spoken response and then runs a motion model over the finished audio, which requires a second full inference pass and precludes any joint optimisation between the two. We present Motion-Omni, an end-to-end framework in which a spoken dialogue model natively outputs explicit facial expression together with hand, upper-body and lower-body motion, generated directly from the hidden states that produce the speech. Joint training is not optional here: with the speech pathway frozen, motion remains misaligned with the audio, and co-adapting the LLM, Speech Generator and Motion Generator under both objectives is what recovers alignment while retaining spoken-dialogue ability. Supervision comes from a scalable, model-agnostic pipeline that pseudo-labels consistent-voice speech responses with a replaceable motion teacher, yielding 422,856 quality-ranked pairs (1,402 hours). We further release SwDA-500 and, to our knowledge, the first public evaluation protocol for stochastic open-ended full-body spoken dialogue, matching audio across motion systems while unifying rendering, automatic metrics, human evaluation, and latency measurement. Instantiated with a Qwen2.5-7B-Instruct backbone, Motion-Omni-Q7 matches the same-audio teacher cascade to within 2% on reference-free motion metrics while responding 5.4 x faster (RTF=0.78, faster than real time), surpasses all non-teacher cascades on beat correlation and diversity, and reaches a 2.62% word error rate, the lowest among the omni-modal systems compared.

cs.SD↗
arXiv2026

Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network

Text-guided image editing must introduce the requested changes while preserving unrelated source content. In training-free editing, diffusion editors often use spatial controls whose inaccuracies can leave edits incomplete or alter unrelated regions. Causal autoregressive editors face a further constraint: their fixed decoding order limits revision of earlier decisions. As the first to explore training-free image editing with Generative Refinement Networks (GRN), we observe that its refinement process is inherently suitable for editing and offers a promising way to address these limitations. Motivated by this observation, we introduce RefineEdit, a training-free prompt-to-prompt image editing framework built on the GRN. Our key idea is to couple edit localization with content generation through the global refinement of binary image codes, allowing editing evidence to be revised as the image evolves. More specifically, RefineEdit combines bit routing with two stabilization mechanisms: adaptive spatial freezing and finite bit locking. Bit routing starts from an intermediate source state and uses signed probability differences between the two branches to identify editable positions and bits. It directs selected bits toward editing refinement while anchoring the rest to the evolving source trajectory. Adaptive spatial freezing limits unnecessary expansion of the editing region, while finite bit locking maintains recent bit activations to support continued editing. The overall framework requires no additional training, external masks, or attention control. Across nine editing categories of PIE-Bench, RefineEdit achieves the best background-preservation scores in PSNR, LPIPS, MSE, and SSIM, together with the highest whole-image and edited-region CLIP scores among the evaluated methods. Code is available at https://github.com/mura1n/RefineEdit.

cs.CV↗
arXiv2026

The Uncontrolled Variable: Vision-Language Refusal Is Conditioned on the Image-Attachment Interface, and Not Robust to Irrelevant Image Properties

We show that aligned vision-language models also condition refusal on a property of a request's form: whether an image is attached, holding everything the request asks fixed. Attaching a blank canvas, an image that cannot be read, cannot relate to the request, and is byte-identical across every prompt in its condition, shifts benign refusal by tens of points. The shift is not blanket caution but a threshold shift: genuinely neutral instructions are almost unaffected (<=2 percentage points on three of four hosted models) while borderline-benign prompts move +23 to +51 points, so the cost falls on sensitivity-adjacent traffic, meaning benign questions about privacy, self-harm, violence and illegal activity. There is a benign reading of such a threshold, namely that attachment correlates with risk in real traffic, and we take it seriously; a black-box study cannot measure that correlation and we do not claim to. What it can test is whether the response to attachment is robust to variation carrying no information about the request, and on four independent measurements it is not. It varies with canvas colour and pixel count. Its sign inverts across checkpoints. It survives an explicit instruction to disregard the image. And on one model it fires on a bare assertion that an attachment exists, with nothing attached and the modality word contributing none of it. Finally we price it. On a matched harmful set the same canvas does lower attack success, so the cue buys something. But the charge is decoupled from the purchase: the checkpoint with the least harmful headroom we measure, completing only 2% of plain harmful requests, still pays the benign cost in full, and across our models the harmful-side denominator falls as alignment improves while the benign cost does not track it down. Image presence is not a conservative default that a deployer chose and priced. It is an uncontrolled variable.

cs.CR↗