arXiv subjects

Haoze Li

Publications and source records attributed to Haoze Li.

12 recordsLinked to original sources

arXiv2026

KINESYS: Test-Time Adaptation for Reference-Based Effect Editing in Images

Visual effects are often easier to demonstrate via reference than to describe in words. We study effect-reference editing, where a reference image shows a desired visual effect and language identifies which properties should guide the edit. Existing single-pass editing techniques often fail to make fine-grained edits, resulting in unintended changes to source content or deviations from the reference effect. We introduce KINESYS, a test-time adaptation framework for effect-reference editing that combines reference-aware and source-aware feedback to transfer the desired effect while preserving the composition and content of the original shot. To ensure KINESYS can produce a diverse set of edits that preserve the source, we develop a robust optimization strategy in which the generative model is fine-tuned across multiple noise samples and optimization runs. Evaluation with VLM judges, source preservation metrics (LPIPS, TPIPS), and a user study shows that KINESYS outperforms state-of-the-art image editors in source preservation; on effect transfer, it remains competitive with Nano Banana 2 and outperforms open-weight editors.

cs.CV↗

arXiv2026

VH-GPO: Verified Handoff for LLM Agent Reinforcement Learning

Group-relative reinforcement learning methods such as GRPO train large language model (LLM) agents by comparing rollouts for the same task. When every trajectory in a group fails with the same reward, relative advantages collapse toward zero, leaving no useful learning signal. A frozen teacher can recover from a student-visited state, but letting it complete the task makes the recovered suffix entirely teacher-generated. We ask: after repairing a student failure, when should control return to the student? We propose outcome-verified teacher-student handoff, instantiated in Verified Handoff for Group Policy Optimization (VH-GPO). The teacher provides enough support to reach a state from which the student can succeed, then returns control so that successful experience is again student-generated. Given an environment-verified teacher recovery, VH-GPO selects the earliest independently confirmed handoff point found within a bounded search budget. With Qwen2.5-1.5B, VH-GPO increases success rates over GRPO from 59.38% to 94.79% on ALFWorld and from 71.62% to 77.34% on WebShop. Improvements also hold with the 7B student, for which additional teacher and student output tokens amount to 1.77% and 2.61% of regular student rollout output tokens on ALFWorld and WebShop, respectively. Ablation studies demonstrate the effectiveness of the key components. These results support using verified teacher assistance to repair failures while reinforcing the student's ability to complete tasks independently.

cs.LG↗

arXiv2026

Distributed Optimal Resource Allocation Search: A Dynamic Event-Triggered Algorithm

This paper investigates an equality-coupled distributed resource allocation problem with smooth general convex local objective functions. A discrete-time residual-aware dynamic event-triggered algorithm is proposed over time-varying switching undirected graphs. Unlike existing event-triggered resource allocation algorithms that rely on strong convexity, the proposed method establishes convergence for general convex costs without using strong monotonicity or contraction arguments. The key idea is to co-design a resource-allocation search recursion with a dynamic triggering rule that incorporates both local gradient-estimation errors and local gradient-disagreement residuals. The resulting triggering mechanism reduces unnecessary communication and generates a summable error bound, which is embedded into a Mirror-EXTRA-type Lyapunov analysis. Under suitable step-size conditions, the proposed algorithm is proved to converge to an optimal solution. Further, when the local objective functions are strongly convex, a linear convergence result is established. Numerical simulations and comparative tests with related event-triggered methods verify the effectiveness and communication efficiency of the proposed algorithm.

math.OC↗

arXiv2026

VisualDeltas: Learning Preferences from Visual Quality Perturbations

We present VisualDeltas, a lightweight preference-learning framework that extracts supervision from visual quality variations in multimodal data. By leveraging the systematic impact of image quality on visual perception and reasoning, VisualDeltas induces informative preference signals without relying on human annotations or external teachers. The framework supports both label-free and label-based regimes, enabling flexible use of available supervision when present. Across diverse multimodal benchmarks and model scales, VisualDeltas consistently outperforms rejection-sampling fine-tuning and improves generalization, and extends naturally to a range of visual degradations.

cs.AI↗

arXiv2025

Steering Vision-Language Pre-trained Models for Incremental Face Presentation Attack Detection

Face Presentation Attack Detection (PAD) demands incremental learning (IL) to combat evolving spoofing tactics and domains. Privacy regulations, however, forbid retaining past data, necessitating rehearsal-free IL (RF-IL). Vision-Language Pre-trained (VLP) models, with their prompt-tunable cross-modal representations, enable efficient adaptation to new spoofing styles and domains. Capitalizing on this strength, we propose \textbf{SVLP-IL}, a VLP-based RF-IL framework that balances stability and plasticity via \textit{Multi-Aspect Prompting} (MAP) and \textit{Selective Elastic Weight Consolidation} (SEWC). MAP isolates domain dependencies, enhances distribution-shift sensitivity, and mitigates forgetting by jointly exploiting universal and domain-specific cues. SEWC selectively preserves critical weights from previous tasks, retaining essential knowledge while allowing flexibility for new adaptations. Comprehensive experiments across multiple PAD benchmarks show that SVLP-IL significantly reduces catastrophic forgetting and enhances performance on unseen domains. SVLP-IL offers a privacy-compliant, practical solution for robust lifelong PAD deployment in RF-IL settings.

cs.CV↗

arXiv2025

ALPS: Attention Localization and Pruning Strategy for Efficient Alignment of Large Language Models

Aligning general-purpose large language models (LLMs) to downstream tasks often incurs significant training adjustment costs. Prior research has explored various avenues to enhance alignment efficiency, primarily through minimal-data training or data-driven activations to identify key attention heads. However, these approaches inherently introduce data dependency, which hinders generalization and reusability. To address this issue and enhance model alignment efficiency, we propose the Attention Localization and Pruning Strategy (ALPS), an efficient algorithm that localizes the most task-sensitive attention heads and prunes by restricting attention training updates to these heads, thereby reducing alignment costs. Experimental results demonstrate that our method activates only 10% of attention parameters during fine-tuning while achieving a 2% performance improvement over baselines on three tasks. Moreover, the identified task-specific heads are transferable across datasets and mitigate knowledge forgetting. Our work and findings provide a novel perspective on efficient LLM alignment. The code is available at https://github.com/VoiceBeer/ALPS.

cs.CL↗

arXiv2025

Kimi k1.5: Scaling Reinforcement Learning with LLMs

Language model pretraining with next token prediction has proved effective for scaling compute but is limited to the amount of available training data. Scaling reinforcement learning (RL) unlocks a new axis for the continued improvement of artificial intelligence, with the promise that large language models (LLMs) can scale their training data by learning to explore with rewards. However, prior published work has not produced competitive results. In light of this, we report on the training practice of Kimi k1.5, our latest multi-modal LLM trained with RL, including its RL training techniques, multi-modal data recipes, and infrastructure optimization. Long context scaling and improved policy optimization methods are key ingredients of our approach, which establishes a simplistic, effective RL framework without relying on more complex techniques such as Monte Carlo tree search, value functions, and process reward models. Notably, our system achieves state-of-the-art reasoning performance across multiple benchmarks and modalities -- e.g., 77.5 on AIME, 96.2 on MATH 500, 94-th percentile on Codeforces, 74.9 on MathVista -- matching OpenAI's o1. Moreover, we present effective long2short methods that use long-CoT techniques to improve short-CoT models, yielding state-of-the-art short-CoT reasoning results -- e.g., 60.8 on AIME, 94.6 on MATH500, 47.3 on LiveCodeBench -- outperforming existing short-CoT models such as GPT-4o and Claude Sonnet 3.5 by a large margin (up to +550%).

cs.AI↗

arXiv2025

A Cooperative Aerial System of A Payload Drone Equipped with Dexterous Rappelling End Droid for Cluttered Space Pickup

In cluttered spaces, such as forests, drone picking up a payload via an abseil claw is an open challenge, as the cable is likely tangled and blocked by the branches and obstacles. To address such a challenge, in this work, a cooperative aerial system is proposed, which consists of a payload drone and a dexterous rappelling end droid. The two ends are linked via a Kevlar tether cable. The end droid is actuated by four propellers, which enable mid-air dexterous adjustment of clawing angle and guidance of cable movement. To avoid tanglement and rappelling obstacles, a trajectory optimization method that integrates cable length constraints and dynamic feasibility is developed, which guarantees safe pickup. A tether cable dynamic model is established to evaluate real-time cable status, considering both taut and sagging conditions. Simulation and real-world experiments are conducted to demonstrate that the proposed system is capable of picking up payload in cluttered spaces. As a result, the end droid can reach the target point successfully under cable constraints and achieve passive retrieval during the lifting phase without propulsion, which enables effective and efficient aerial manipulation.

cs.RO↗

arXiv2025

Feature Matching Intervention: Leveraging Observational Data for Causal Representation Learning

A major challenge in causal discovery from observational data is the absence of perfect interventions, making it difficult to distinguish causal features from spurious ones. We propose an innovative approach, Feature Matching Intervention (FMI), which uses a matching procedure to mimic perfect interventions. We define causal latent graphs, extending structural causal models to latent feature space, providing a framework that connects FMI with causal graph learning. Our feature matching procedure emulates perfect interventions within these causal latent graphs. Theoretical results demonstrate that FMI exhibits strong out-of-distribution (OOD) generalizability. Experiments further highlight FMI's superior performance in effectively identifying causal features solely from observational data.

stat.ML↗

arXiv2024

The role of inhibitory control in garden-path sentence processing: A Chinese-English bilingual perspective

In reading garden-path sentences, people must resolve competing interpretations, though initial misinterpretations can linger despite reanalysis. This study examines the role of inhibitory control (IC) in managing these misinterpretations among Chinese-English bilinguals. Using self-paced reading tasks, we investigated how IC influences recovery from garden-path sentences in Chinese (L1) and its interaction with language proficiency during English (L2) processing. Results indicate that IC does not affect garden-path recovery in Chinese, suggesting reliance on semantic context may reduce the need for IC. In contrast, findings for English L2 learners reveal a complex relationship between language proficiency and IC: Participants with low L2 proficiency but high IC showed lingering misinterpretations, while those with high proficiency exhibited none. These results support and extend the Model of Cognitive Control (Ness et al., 2023). Moreover, our comparison of three Stroop task versions identifies L1 colour-word Stroop task as the preferred measure of IC in bilingual research.

cs.CL↗

arXiv2024

TableGPT2: A Large Multimodal Model with Tabular Data Integration

The emergence of models like GPTs, Claude, LLaMA, and Qwen has reshaped AI applications, presenting vast new opportunities across industries. Yet, the integration of tabular data remains notably underdeveloped, despite its foundational role in numerous real-world domains. This gap is critical for three main reasons. First, database or data warehouse data integration is essential for advanced applications; second, the vast and largely untapped resource of tabular data offers immense potential for analysis; and third, the business intelligence domain specifically demands adaptable, precise solutions that many current LLMs may struggle to provide. In response, we introduce TableGPT2, a model rigorously pre-trained and fine-tuned with over 593.8K tables and 2.36M high-quality query-table-output tuples, a scale of table-related data unprecedented in prior research. This extensive training enables TableGPT2 to excel in table-centric tasks while maintaining strong general language and coding abilities. One of TableGPT2's key innovations is its novel table encoder, specifically designed to capture schema-level and cell-level information. This encoder strengthens the model's ability to handle ambiguous queries, missing column names, and irregular tables commonly encountered in real-world applications. Similar to visual language models, this pioneering approach integrates with the decoder to form a robust large multimodal model. We believe the results are compelling: over 23 benchmarking metrics, TableGPT2 achieves an average performance improvement of 35.20% in the 7B model and 49.32% in the 72B model over prior benchmark-neutral LLMs, with robust general-purpose capabilities intact.

cs.LG↗

arXiv2023

PyXAB -- A Python Library for $\mathcal{X}$-Armed Bandit and Online Blackbox Optimization Algorithms

We introduce a Python open-source library for $\mathcal{X}$-armed bandit and online blackbox optimization named PyXAB. PyXAB contains the implementations for more than 10 $\mathcal{X}$-armed bandit algorithms, such as HOO, StoSOO, HCT, and the most recent works GPO and VHCT. PyXAB also provides the most commonly-used synthetic objectives to evaluate the performance of different algorithms and the various choices of the hierarchical partitions on the parameter space. The online documentation for PyXAB includes clear instructions for installation, straight-forward examples, detailed feature descriptions, and a complete reference of the API. PyXAB is released under the MIT license in order to encourage both academic and industrial usage. The library can be directly installed from PyPI with its source code available at https://github.com/WilliamLwj/PyXAB

stat.ML↗