arXiv subjects

Zhongyi Han

Publications and source records attributed to Zhongyi Han.

At least 19 recordsLinked to original sources

arXiv2026

Useful to Whom? The Easy-versus-Coverage Boundary Moves with the Learner

Is the same data equally useful to every learner? The question is whether the value of a training sample belongs to the sample or arises in its interaction with the learner. We test it with a selection experiment. Under a per-class budget, two kinds of samples compete for the slots: easy samples, which a short training run learns first, and coverage samples, which spread over the class in feature space. Easy samples win at small budgets and coverage samples at large ones, and the budget at which the winner flips, the crossover m0, can be measured. We ask whether m0 is set by the data or by the learner. To separate the two, we freeze the selected subsets and change only the learner that trains on them. On low-resolution ImageNet-100, doubling the width of a ResNet-18 moves the crossover from 57 to 85 samples per class on the same subsets. Enlarging the input grid from 32 to 64 pixels at fixed image information moves it from 85 to 57, and a single stride change reproduces or removes this shift. The grid shift shrinks as width grows. Under the native ImageNet-1k protocol, a 4x range of width moves the boundary by at most a few samples per class. On a Vision Transformer, coverage wins at every measured budget, even with easy subsets that work on the convolutional learner. The budget at which one kind of sample stops being the more useful kind therefore belongs to the learner, and the preferred regime of a selection strategy has to be read against the learner that will use the data.

cs.LG↗

arXiv2026

DVLA-RL++: Dual-Level Vision-Language Alignment with Reinforcement Learning Gating for Few-Shot Learning

Few-shot learning aims to recognize novel categories from limited labeled examples. Recent studies incorporate textual semantics to compensate for limited visual observations and improve class representations. However, high image-text agreement may reflect both intrinsic object properties and incidental context, making support prototypes susceptible to contextual contamination. To address this problem, we propose DVLA-RL++, which extends DVLA-RL with complementary semantic purification (CSP) and counterfactual reinforcement-learning gating (CRG). Specifically, CSP generates intrinsic and nuisance descriptions from labeled supports and compares their agreement with each support token. An ambiguity-dependent rejection margin guides sparse evidence allocation, while an intrinsic semantic anchor fills the unassigned mass to provide a fallback when visual evidence is unreliable. CRG learns layer-wise semantic fusion strengths using a reward that balances recognition performance and nuisance exposure. An independently executed reference trajectory on the same episode provides a paired learning signal. Theoretical analysis relates retained evidence and anchor quality to prototype stability and establishes conditions for unbiased on-policy gradient estimation. Experiments on standard, fine-grained, and cross-domain benchmarks show state-of-the-art accuracy, with an average gain of 1.4% over DVLA-RL. The project page is available at https://peacelwh.github.io/TPAMI27-DVLA-RLpp/.

cs.CV↗

arXiv2026

Asking Earns Nothing: Scoring the Decision to Act in BFCL Multi-Turn

An agent that lacks the information it needs should ask rather than act, and the task definitions of agent leaderboards say so. BFCL multi-turn builds two of its four categories around a turn on which the model is supposed to ask, and its scorer never looks at that turn: the gold trajectory there is empty, the checker skips it, and the scripted user cannot answer, so asking earns nothing, guessing costs nothing on that turn, and asking twice loses the item. The benchmark also contains the control experiment for that decision. A should-ask item is a base item with one piece of information removed from one turn, so the same request appears twice at the same turn index, once complete and once not: on the first the model should make the call that changes the world, on the second it should ask. We score one decision per pair, whether the model attempted a world-changing call on that turn, read off the stored trajectories with no LLM judge; acting always and asking always both score 50. On the 223 pairs that pose this decision, gpt-5.4 attempts the call on 83.4% of the complete turns and holds back on 78.0% of the incomplete ones, the best decision accuracy of seven models at 80.7%; on the same items the official score ranks it sixth and puts first a model that lands in the middle here. One added line telling gpt-5.4 not to ask pushes it toward acting on both sides of the pair, so its decision accuracy shows no detectable change, while its official score rises by 13.5 to 23.5 points on the two should-ask categories and on the base twins; the opposite line, telling gemma-4-31B-it to ask first, improves its decision by 4.5 points and gains no official score. The score moves with the push toward action, not with the decision. We release the pairs, a turn-level scorer that runs on any BFCL output directory without an API key, and 31 manually verified bad items.

cs.AI↗

arXiv2026

What Does a Harness Buy? Tokens, Mostly

A coding agent is a language model wrapped in a harness: the system prompt, the tool set, and the context management that turn a chat model into something that can work inside a repository. Production harnesses ship releases daily, vendors advertise pass-rate gains from harness changes, and leaderboards mix harnesses freely. What is rarely measured is how much the harness itself moves the score when the model is held fixed. We run five models through three production harnesses, Claude Code, mini-SWE-agent, and OpenCode, on SWE-bench Verified, and rerun the same configurations to calibrate how much a score moves when nothing changes but the run. On 447 tasks and the two models we ran there, Claude Code and mini-SWE-agent, the heaviest and the lightest harness, are equivalent within five points. On a 45-task hard subset and five models, swapping the harness flips as many tasks as rerunning the same harness, 13% in both cases, and the tasks a harness wins in one run are not the tasks it wins in the next. The one harness effect that clears the noise is a loss, not a gain: OpenCode trails by up to 9 points on the large pool, and on one model half of that gap sits in runs its output cap cut short. What the harness does decide is the bill. With the same model, the same tasks, and one price list, cost per task differs by up to 3x across harnesses. The gap is set at the first call, by the preamble of system prompt and tool schemas each harness sends with every step, and scaled by the number of steps; per-step growth and per-call tool output differ far less. The provider's price for cached input scales the bill and does not reorder it. The rerun data also give the resolution a harness comparison needs: at the discordance we observe, 45 tasks catch a 13-point gap only half the time and no gap with 80% power, and 447 tasks resolve 5 points, still coarser than the gains many harness changes claim.

cs.AI↗

arXiv2026

A Dominant Supplier Slows Recursive Drift More Than It Steers It

More and more of the text future language models learn from is written by a few of today's models. If one supplier writes most of a shared corpus, does it pull the models trained on it toward its own writing, or change how fast they drift? We retrain eight open models from their base weights on a shared pool of each other's text for five generations, varying the part written by one model, Phi-2, from an equal share to 90%. The models drift together toward a style with fewer function words, and none starts repeating itself. No share of Phi-2 brings the other models closer to its text than the equal share does. We split each ecosystem's separation from the equal-share one into a delay along its route and a departure from that route, both counted beyond the difference between two equal-share runs. With Phi-2 at 90%, delay outweighs departure 72 to 28 and 64 to 36 in two runs, and the ecosystem falls 2.7 and 2.5 generations behind. With Phi-2 at half the pool the two parts are about equal. When SmolLM2 or Qwen3-1.7B writes half instead, the ecosystem slows less or not at all. The departure leans toward Phi-2 more as its share grows, but more than toward every other model only at 90%. Human text filling a quarter or half of the pool slows the models along the same route.

cs.AI↗

arXiv2026

Break Step: Recursive Training Resonates with Replayed Sampling Noise

How fast does a language model degrade when trained on its own outputs? Theory traces it to gradually accumulating errors, while experiments report repeated phrases within ten generations. Under a fixed sampling seed in vLLM, the fast loss of lexical diversity comes from the sampler. When vLLM serves a batch from one seeded sampling configuration, every request receives the same random draws, and a fixed seed replays them every generation. Fine-tuning raises the tokens that won, and the replayed draws let them win by more. Sharing across requests and replay across generations matter only together. Remove either one, by changing the shared seed every generation or by giving each request its own seed that repeats every generation, and the unique-4-gram fraction of two StableLM checkpoints stays near its starting value of about 0.98 through generation 3. Keep both, and the replayed shared seed takes seven checkpoints from five families to between 0.045 and 0.38 by then. Three generations of replay write the favoured phrases into the weights: decoded with one seed per request, the generation-3 weights of the replayed StableLM-2-1.6B chain recover most of their diversity, yet the phrase that filled every sample under the shared seed still opens 46% of them. Without replay, five checkpoints drift slowly, consistent with the gradual accumulation that theory describes, and three turn incoherent though their diversity scores stay high. One peer-reviewed model-collapse pipeline that fine-tunes Gemma-2-27B samples identical prompts under one seeded configuration, and three quarters of the rows it released for one iteration repeat nearly as often as one such batch copies them. A seed per request restores the fresh sample that stability analyses assume.

cs.CL↗

arXiv2026

Are Coreset Selection Methods Worth Their Cost?

Coreset selection picks a representative subset of the labeled training set to make training cheaper. However, it is usually evaluated by downstream accuracy at a fixed subset size, ignoring both the time spent selecting the subset and the training recipe behind each reported number. We introduce an end-to-end benchmark that standardizes downstream training and charges selection and training to the same auditable wall-clock budget, spanning 4 datasets from CIFAR-10 to ImageNet-1K, 11 selectors, 5 fractions, and 3 seeds, with over 1,500 released runs. Repeated-sampling work has shown that budget-aware evaluation already favors random strategies. Our two budget studies test whether that verdict survives when every selector is granted its most favorable operating point. Across eight wall-clock budget anchors on each of CIFAR-10 and Tiny ImageNet, no anchor is won by a sophisticated selector: every winner is class-balanced random sampling, repeated random sampling, or full-data training. In fixed-budget duels on ImageNet-1K, training on all data for fewer epochs beats every selection strategy we probe while also costing the least. A per-dataset cost audit shows that selection cost is dominated at every scale by a fixed full-dataset scan, so it cannot be amortized away by selecting a smaller fraction, and its absolute size does not extrapolate from one dataset to another. We further quantify when selection does pay back through subset reuse, and document 9 correctness fixes to a widely used codebase, one of which shifts a standard Herding baseline by nearly 6 points. Selection time is not free preprocessing, and an evaluation that ignores it measures the wrong quantity.

cs.LG↗

arXiv2026

OpenClawBench: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories

Task success can hide process anomalies in real-world agent executions. An agent may pass the final task oracle while still accumulating unresolved ambiguity, unsafe external writes, ignored errors, weakly grounded commitments, or capability-boundary overcommitment. We study this mismatch as the Outcome-Process Gap and introduce OpenClawBench, a large-scale dataset for measuring and supervising process-side anomalies in real agent execution processes. OpenClawBench is built from BFCL-driven OpenClaw sessions produced by 6 source models and contains 31,264 annotated trajectories. It aligns task-oracle outcomes with structured process evidence. FullTax converts the aligned trajectories into structured anomaly supervision: binary labels, supporting evidence, onset/span localization, severity, recoverability, and a 5-class anomaly taxonomy. Using OpenClawBench, we make the Outcome-Process Gap measurable. Among 31,135 oracle-passing executions, 2,904 are still labeled process-anomalous under FullTax. These results show that success-only evaluation misses a concrete class of process-side failures in real agent executions. A LoRA-fine-tuned Gemma 3 12B detector trained on the high-confidence FullTax supervised pool reaches binary F1=0.729 on the cleaner-labels held-out test split. Together, OpenClawBench turns real agent execution logs into auditable and reusable supervision for studying, diagnosing, and operationally monitoring runtime agent reliability.

cs.AI↗

arXiv2026

EchoAlign: Bridging Generative and Discriminative Learning under Noisy Labels

Noisy labels severely hinder the accuracy and generalization of machine learning models, especially when ambiguous instance features make reliable annotation difficult. Existing approaches, including transition-matrix-based label correction, struggle to capture complex relationships between instances and noisy labels, limiting their effectiveness in such settings. We present EchoAlign, a framework that bridges generative and discriminative learning under noisy labels. Instead of correcting labels, EchoAlign treats noisy labels as supervision targets and modifies the corresponding instances to align with them. The framework has two components: EchoMod uses controllable generative models to adjust instance features while preserving key instance-level structural cues, such as shape and edges, and avoiding excessive distortion; EchoSelect mitigates distribution shifts by retaining a reliable subset of original instances, guided by feature similarity between original and modified samples. This generative-discriminative interplay enables robust learning in highly noisy settings. Experiments on three benchmark datasets show that EchoAlign outperforms state-of-the-art methods in most evaluated settings. Under 30% instance-dependent noise, EchoSelect retains nearly twice as many correctly labeled samples as competing approaches while maintaining 99% selection accuracy, demonstrating the robustness and effectiveness of EchoAlign.

cs.LG↗

arXiv2026

From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space

Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.

cs.CL↗

arXiv2026

Agentic Unlearning: When LLM Agent Meets Machine Unlearning

In this paper, we introduce \textbf{agentic unlearning} which removes specified information from both model parameters and persistent memory in agents with closed-loop interaction. Existing unlearning methods target parameters alone, leaving two critical gaps: (i) parameter-memory backflow, where retrieval reactivates parametric remnants or memory artifacts reintroduce sensitive content, and (ii) the absence of a unified strategy that covers both parameter and memory pathways. We present Synchronized Backflow Unlearning (SBU), a framework that unlearns jointly across parameter and memory pathways. The memory pathway performs dependency closure-based unlearning that prunes isolated entities while logically invalidating shared artifacts. The parameter pathway employs stochastic reference alignment to guide model outputs toward a high-entropy prior. These pathways are integrated via a synchronized dual-update protocol, forming a closed-loop mechanism where memory unlearning and parametric suppression reinforce each other to prevent cross-pathway recontamination. Experiments on medical QA benchmarks show that SBU reduces traces of targeted private information across both pathways with limited degradation on retained data.

cs.LG↗

arXiv2026

Revisiting Data Scaling in Medical Image Segmentation via Topology-Aware Augmentation

Understanding how segmentation performance scales with training data is fundamental for developing data-efficient medical AI systems. In this study, we systematically revisit data scaling behavior across 15 anatomical segmentation tasks spanning four imaging modalities. We observe that medical segmentation follows a structurally stable power-law-like relationship between predictive error and dataset size, characterized by rapid improvement in low-data regimes. However, unlike classical large-scale vision or language tasks, segmentation exhibits earlier and task-dependent performance saturation, with a persistent error floor emerging even as data increases. This behavior suggests that segmentation scaling is not purely data-constrained but is influenced by intrinsic geometric and anatomical structure. To further probe this geometry-constrained regime, we investigate whether topology-aware deformation-based augmentation can modify effective scaling dynamics. We compare random elastic deformation with registration-guided and generative deformation-field modeling strategies. While the overall functional form of the scaling law remains preserved, topology-aware augmentation systematically lowers the effective error scale and reshapes convergence behavior in a task-dependent manner, leading to improved sample efficiency without overturning the underlying scaling principle. These findings indicate that medical segmentation obeys a geometry-limited scaling law, and that anatomically grounded augmentation enhances data efficiency by expanding effective topological coverage rather than altering the fundamental scaling structure. Our results provide a principled empirical perspective on data-efficient learning in medical image segmentation. The code will be released after acceptance.

cs.CV↗

arXiv2026

DVLA-RL: Dual-Level Vision-Language Alignment with Reinforcement Learning Gating for Few-Shot Learning

Few-shot learning (FSL) aims to generalize to novel categories with only a few samples. Recent approaches incorporate large language models (LLMs) to enrich visual representations with semantic embeddings derived from class names. However, they overlook progressive and adaptive alignment between vision and language from low-level to high-level semantics, resulting in limited semantic gains. To address these challenges, we propose Dual-level Vision-Language Alignment with Reinforcement Learning gating (DVLA-RL), which consists of Dual-level Semantic Construction (DSC) and RL-gated Attention (RLA). Specifically, DSC conditions LLMs on both class names and support samples to generate discriminative attributes, progressively selects the most relevant ones, and then synthesizes them into coherent class descriptions. This process provides complementary low-level attributes and high-level descriptions, enabling both fine-grained grounding and holistic class understanding. To dynamically integrate dual-level semantics along with the visual network layers, RLA formulates cross-modal fusion as a sequential decision process. A lightweight policy trained with episodic REINFORCE adaptively adjusts the contributions of self-attention and cross-attention to integrate textual and visual tokens. As a result, shallow layers refine local attributes and deep layers emphasize global semantics, enabling more precise cross-modal alignment. This achieves class-specific discrimination and generalized representations with merely a few support samples. DVLA-RL achieves new state-of-the-art performance across nine benchmarks in three diverse FSL scenarios.

cs.CV↗

arXiv2026

TrajAD: Trajectory Anomaly Detection for Trustworthy LLM Agents

We address the problem of runtime trajectory anomaly detection, a critical capability for enabling trustworthy LLM agents. Current safety measures predominantly focus on static input/output filtering. However, we argue that ensuring LLM agents reliability requires auditing the intermediate execution process. In this work, we formulate the task of Trajectory Anomaly Detection. The goal is not merely detection, but precise error localization. This capability is essential for enabling efficient rollback-and-retry. To achieve this, we construct TrajBench, a dataset synthesized via a perturb-and-complete strategy to cover diverse procedural anomalies. Using this benchmark, we investigate the capability of models in process supervision. We observe that general-purpose LLMs, even with zero-shot prompting, struggle to identify and localize these anomalies. This reveals that generalized capabilities do not automatically translate to process reliability. To address this, we propose TrajAD, a specialized verifier trained with fine-grained process supervision. Our approach outperforms baselines, demonstrating that specialized supervision is essential for building trustworthy agents.

cs.CR↗

arXiv2026

HCVP: Leveraging Hierarchical Contrastive Visual Prompt for Domain Generalization

Domain Generalization (DG) endeavors to create machine learning models that excel in unseen scenarios by learning invariant features. In DG, the prevalent practice of constraining models to a fixed structure or uniform parameterization to encapsulate invariant features can inadvertently blend specific aspects. Such an approach struggles with nuanced differentiation of inter-domain variations and may exhibit bias towards certain domains, hindering the precise learning of domain-invariant features. Recognizing this, we introduce a novel method designed to supplement the model with domain-level and task-specific characteristics. This approach aims to guide the model in more effectively separating invariant features from specific characteristics, thereby boosting the generalization. Building on the emerging trend of visual prompts in the DG paradigm, our work introduces the novel \textbf{H}ierarchical \textbf{C}ontrastive \textbf{V}isual \textbf{P}rompt (HCVP) methodology. This represents a significant advancement in the field, setting itself apart with a unique generative approach to prompts, alongside an explicit model structure and specialized loss functions. Differing from traditional visual prompts that are often shared across entire datasets, HCVP utilizes a hierarchical prompt generation network enhanced by prompt contrastive learning. These generative prompts are instance-dependent, catering to the unique characteristics inherent to different domains and tasks. Additionally, we devise a prompt modulation network that serves as a bridge, effectively incorporating the generated visual prompts into the vision transformer backbone. Experiments conducted on five DG datasets demonstrate the effectiveness of HCVP, outperforming both established DG algorithms and adaptation protocols.

cs.CV↗

arXiv2025

SkinCaRe: A Multimodal Dermatology Dataset Annotated with Medical Caption and Chain-of-Thought Reasoning

With the widespread application of artificial intelligence (AI), particularly deep learning (DL) and vision large language models (VLLMs), in skin disease diagnosis, the need for interpretability becomes crucial. However, existing dermatology datasets are limited in their inclusion of concept-level meta-labels, and none offer rich medical descriptions in natural language. This deficiency impedes the advancement of LLM-based methods in dermatologic diagnosis. To address this gap and provide a meticulously annotated dermatology dataset with comprehensive natural language descriptions, we introduce \textbf{SkinCaRe}, a comprehensive multimodal resource that unifies \textit{SkinCAP} and \textit{SkinCoT}. \textbf{SkinCAP} comprises 4,000 images sourced from the Fitzpatrick 17k skin disease dataset and the Diverse Dermatology Images dataset, annotated by board-certified dermatologists to provide extensive medical descriptions and captions. In addition, we introduce \textbf{SkinCoT}, a curated dataset pairing 3,041 dermatologic images with clinician-verified, hierarchical chain-of-thought (CoT) diagnoses. Each diagnostic narrative is rigorously evaluated against six quality criteria and iteratively refined until it meets a predefined standard of clinical accuracy and explanatory depth. Together, SkinCAP (captioning) and SkinCoT (reasoning), collectively referred to as SkinCaRe, encompass 7,041 expertly curated dermatologic cases and provide a unified and trustworthy resource for training multimodal models that both describe and explain dermatologic images. SkinCaRe is publicly available at https://huggingface.co/datasets/yuhos16/SkinCaRe.

cs.CV↗

arXiv2025

Facial Foundational Model Advances Early Warning of Coronary Artery Disease from Live Videos with DigitalShadow

Global population aging presents increasing challenges to healthcare systems, with coronary artery disease (CAD) responsible for approximately 17.8 million deaths annually, making it a leading cause of global mortality. As CAD is largely preventable, early detection and proactive management are essential. In this work, we introduce DigitalShadow, an advanced early warning system for CAD, powered by a fine-tuned facial foundation model. The system is pre-trained on 21 million facial images and subsequently fine-tuned into LiveCAD, a specialized CAD risk assessment model trained on 7,004 facial images from 1,751 subjects across four hospitals in China. DigitalShadow functions passively and contactlessly, extracting facial features from live video streams without requiring active user engagement. Integrated with a personalized database, it generates natural language risk reports and individualized health recommendations. With privacy as a core design principle, DigitalShadow supports local deployment to ensure secure handling of user data.

cs.CV↗

arXiv2025

Improving Representation of High-frequency Components for Medical Visual Foundation Models

Foundation models have recently attracted significant attention for their impressive generalizability across diverse downstream tasks. However, these models are demonstrated to exhibit great limitations in representing high-frequency components and fine-grained details. In many medical imaging tasks, the precise representation of such information is crucial due to the inherently intricate anatomical structures, sub-visual features, and complex boundaries involved. Consequently, the limited representation of prevalent foundation models can result in significant performance degradation or even failure in these tasks. To address these challenges, we propose a novel pretraining strategy, named Frequency-advanced Representation Autoencoder (Frepa). Through high-frequency masking and low-frequency perturbation combined with adversarial learning, Frepa encourages the encoder to effectively represent and preserve high-frequency components in the image embeddings. Additionally, we introduce an innovative histogram-equalized image masking strategy, extending the Masked Autoencoder approach beyond ViT to other architectures such as Swin Transformer and convolutional networks. We develop Frepa across nine medical modalities and validate it on 32 downstream tasks for both 2D images and 3D volume data. Without fine-tuning, Frepa can outperform other self-supervised pretraining methods and, in some cases, even surpasses task-specific trained models. This improvement is particularly significant for tasks involving fine-grained details, such as achieving up to a +15% increase in DSC for retina vessel segmentation and a +7% increase in IoU for lung nodule detection. Further experiments quantitatively reveal that Frepa enables superior high-frequency representations and preservation in the embeddings, underscoring its potential for developing more generalized and universal medical image foundation models.

eess.IV↗