Searcharxiv⌕ Search

arXiv subjects

Zilong Wang

Publications and source records attributed to Zilong Wang.

At least 37 records · Page 2Linked to original sources
arXiv2026

CocoaBench: Evaluating Unified Digital Agents in the Wild

LLM agents now perform strongly in software engineering, deep research, GUI automation, and various other applications, while recent agent scaffolds and models are increasingly integrating these capabilities into unified systems. Yet, most evaluations still test these capabilities in isolation, which leaves a gap for more diverse use cases that require agents to combine different capabilities. We introduce CocoaBench, a benchmark for unified digital agents built from human-designed, long-horizon tasks that require flexible composition of vision, search, and coding. Tasks are specified only by an instruction and an automatic evaluation function over the final output, enabling reliable and scalable evaluation across diverse agent infrastructures. We also present CocoaAgent, a lightweight shared scaffold for controlled comparison across model backbones. Experiments show that current agents remain far from reliable on CocoaBench, with the best evaluated system achieving only 45.1% success rate. Our analysis further points to substantial room for improvement in reasoning and planning, tool use and execution, and visual grounding.

cs.CL↗
arXiv2026

Controllable and Verifiable Tool-Use Data Synthesis for Agentic Reinforcement Learning

Existing synthetic tool-use corpora are primarily designed for offline supervised fine-tuning, yet reinforcement learning (RL) requires executable environments that support reward-checkable online rollouts. We propose COVERT, a two-stage pipeline that first generates reliable base tool-use trajectories through self-evolving synthesis with multi-level validation, and then applies oracle-preserving augmentations that systematically increase environmental complexity. These augmentations introduce distractor tools, indirect or ambiguous user queries, and noisy, multi-format, or erroneous tool outputs, while strictly preserving oracle tool calls and final answers as ground truth. This design enables automatic reward computation via reference matching for standard cases and lightweight judge-assisted verification for special behaviors such as error detection, supporting RL optimization of tool-calling policies. On Qwen2.5-Instruct-14B, COVERT-RL improves overall accuracy on BFCL v3 from 56.5 to 59.9 and on ACEBench from 53.0 to 59.3, with minimal regressions on general-ability benchmarks; when stacked on SFT, it further reaches 62.1 and 61.8, confirming additive gains. These results suggest that oracle-preserving synthetic environments offer a practical RL refinement stage, complementary to SFT, for improving tool-use robustness under ambiguity and unreliable tool feedback.

cs.AI↗
arXiv2026

Transfer Learning for Meta-analysis Under Covariate Shift

Randomized controlled trials often do not represent the populations where decisions are made, and covariate shift across studies can invalidate standard IPD meta-analysis and transport estimators. We propose a placebo-anchored transport framework that treats source-trial outcomes as abundant proxy signals and target-trial placebo outcomes as scarce, high-fidelity gold labels to calibrate baseline risk. A low-complexity (sparse) correction anchors proxy outcome models to the target population, and the anchored models are embedded in a cross-fitted doubly robust learner, yielding a Neyman-orthogonal, target-site doubly robust estimator for patient-level heterogeneous treatment effects when target treated outcomes are available. We distinguish two regimes: in connected targets (with a treated arm), the method yields target-identified effect estimates; in disconnected targets (placebo-only), it reduces to a principled screen--then--transport procedure under explicit working-model transport assumptions. Experiments on synthetic data and a semi-synthetic IHDP benchmark evaluate pointwise CATE accuracy, ATE error, ranking quality for targeting, decision-theoretic policy regret, and calibration. Across connected settings, the proposed method is best or near-best and improves substantially over proxy-only, target-only, and transport baselines at small target sample sizes; in disconnected settings, it retains strong ranking performance for targeting while pointwise accuracy depends on the strength of the working transport condition.

stat.ML↗
arXiv2026

Learning to Optimize Multi-Objective Alignment Through Dynamic Reward Weighting

Prior work in multi-objective reinforcement learning typically uses linear reward scalarization with fixed weights, which provably fails to capture non-convex Pareto fronts and thus yields suboptimal results. This limitation becomes especially critical in online preference alignment for large language models. Here, stochastic trajectories generated by parameterized policies create highly non-linear and non-convex mappings from parameters to objectives that no single static weighting scheme can find optimal trade-offs. We address this limitation by introducing dynamic reward weighting, which adaptively adjusts reward weights during the online reinforcement learning process. Unlike existing approaches that rely on fixed-weight interpolation, our dynamic weighting continuously balances and prioritizes objectives in training, facilitating effective exploration of Pareto fronts in objective space. We introduce two approaches of increasing sophistication and generalizability: hypervolume-guided weight adaptation and gradient-based weight optimization, offering a versatile toolkit for online multi-objective alignment. Our extensive experiments demonstrate their compatibility with commonly used online reinforcement learning algorithms, effectiveness across multiple datasets, and applicability to different model families, consistently achieving Pareto dominant solutions with fewer training steps than fixed-weight linear scalarization baselines.

cs.LG↗
arXiv2026

ReMe: Scaffolding Personalized Cognitive Training via Controllable LLM-Mediated Conversations

Global aging calls for scalable and engaging cognitive interventions. Computerized cognitive training (CCT) is a promising non-pharmacological approach, yet many unsupervised programs rely on rigid, hand-authored puzzles that are difficult to personalize and can hinder adherence. Large language models (LLMs) offer more natural interaction, but their open-ended generation complicates the controlled task structure required for cognitive training. We present ReMe, a web-based framework that scaffolds cognitive training through controllable LLM-mediated conversations, addressing both rigidity in conventional CCT content and the need for conversational controllability. ReMe features a modular Puzzle Engine that represents training activities as reusable puzzle groups specified by structured templates and constraint rules, enabling rapid development of dialogue-based word games and personalized tasks grounded in user context. By integrating personal life logs, ReMe supports Life Recall activities for episodic-memory practice through guided retrieval and progressive cues. A community pilot with 32 adults aged 50+ provides initial feasibility signals.

cs.AI↗
arXiv2026

VIBEVOICE-ASR Technical Report

This report presents VibeVoice-ASR, a general-purpose speech understanding framework built upon VibeVoice, designed to address the persistent challenges of context fragmentation and multi-speaker complexity in long-form audio (e.g., meetings, podcasts) that remain despite recent advancements in short-form speech recognition. Unlike traditional pipelined approaches that rely on audio chunking, VibeVoice-ASRsupports single-pass processing for up to 60 minutes of audio. It unifies Automatic Speech Recognition, Speaker Diarization, and Timestamping into a single end-to-end generation task. In addition, VibeVoice-ASR supports over 50 languages, requires no explicit language setting, and natively handles code-switching within and across utterances. Furthermore, we introduce a prompt-based context injection mechanism that allows users to supply customized conetxt, significantly improving accuracy on domain-specific terminology and polyphonic character disambiguation.

cs.SD↗
arXiv2026

Reasoning-Driven Multimodal LLM for Domain Generalization

This paper addresses the domain generalization (DG) problem in deep learning. While most DG methods focus on enforcing visual feature invariance, we leverage the reasoning capability of multimodal large language models (MLLMs) and explore the potential of constructing reasoning chains that derives image categories to achieve more robust predictions under domain shift. To this end, we systematically study the role of reasoning in DG using DomainBed-Reasoning, a newly constructed extension of DomainBed dataset, in which each sample is paired with class-relevant reasoning chains. Our analysis reveals two key challenges: (i) fine-tuning MLLMs with reasoning chains for classification is more challenging than direct label supervision, since the model must optimize complex reasoning sequences before label prediction; and (ii) mismatches in reasoning patterns between supervision signals and fine-tuned MLLMs lead to a trade-off between semantic richness (informative but harder to optimize) and optimization efficiency (easier to optimize but less informative). To address these issues, we propose RD-MLDG (Reasoning-Driven Multimodal LLM for Domain Generalization), a framework with two components: (i) MTCT (Multi-Task Cross-Training), which introduces an additional direct classification pathway to guide reasoning supervision; and (ii) SARR (Self-Aligned Reasoning Regularization), which preserves the semantic richness of reasoning chains while mitigating reasoning-pattern mismatches via iterative self-labeling. Experiments on standard DomainBed datasets (PACS, VLCS, OfficeHome, TerraInc) demonstrate that RD-MLDG achieves state-of-the-art performances, highlighting reasoning as a promising complementary signal for robust out-of-domain generalization.

cs.AI↗
arXiv2026

Exploring Interpretability for Visual Prompt Tuning with Cross-layer Concepts

Visual prompt tuning offers significant advantages for adapting pre-trained visual foundation models to specific tasks. However, current research provides limited insight into the interpretability of this approach, which is essential for enhancing AI reliability and enabling AI-driven knowledge discovery. In this paper, rather than learning abstract prompt embeddings, we propose the first framework, named Interpretable Visual Prompt Tuning (IVPT), to explore interpretability for visual prompts by introducing cross-layer concept prototypes. Specifically, visual prompts are linked to human-understandable semantic concepts, represented as a set of category-agnostic prototypes, each corresponding to a specific region of the image. IVPT then aggregates features from these regions to generate interpretable prompts for multiple network layers, allowing the explanation of visual prompts at different network depths and semantic granularities. Comprehensive qualitative and quantitative evaluations on fine-grained classification benchmarks show its superior interpretability and performance over visual prompt tuning methods and existing interpretable methods.

cs.CV↗
arXiv2026

OMGs: A multi-agent system supporting MDT decision-making across the ovarian tumour care continuum

Ovarian tumour management has increasingly relied on multidisciplinary tumour board (MDT) deliberation to address treatment complexity and disease heterogeneity. However, most patients worldwide lack access to timely expert consensus, particularly in resource-constrained centres where MDT resources are scarce or unavailable. Here we present OMGs (Ovarian tumour Multidisciplinary intelligent aGent System), a multi-agent AI framework where domain-specific agents deliberate collaboratively to integrate multidisciplinary evidence and generate MDT-style recommendations with transparent rationales. To systematically evaluate MDT recommendation quality, we developed SPEAR (Safety, Personalization, Evidence, Actionability, Robustness) and validated OMGs across diverse clinical scenarios spanning the care continuum. In multicentre re-evaluation, OMGs achieved performance comparable to expert MDT consensus ($4.45 \pm 0.30$ versus $4.53 \pm 0.23$), with higher Evidence scores (4.57 versus 3.92). In prospective multicentre evaluation (59 patients), OMGs demonstrated high concordance with routine MDT decisions. Critically, in paired human-AI studies, OMGs most substantially enhanced clinicians' recommendations in Evidence and Robustness, the dimensions most compromised when multidisciplinary expertise is unavailable. These findings suggest that multi-agent deliberative systems can achieve performance comparable to expert MDT consensus, with potential to expand access to specialized oncology expertise in resource-limited settings.

cs.CL↗
arXiv2026

Mirror: A Multi-Agent System for AI-Assisted Ethics Review

Ethics review is a foundational mechanism of modern research governance, yet contemporary systems face increasing strain as ethical risks arise as structural consequences of large-scale, interdisciplinary scientific practice. The demand for consistent and defensible decisions under heterogeneous risk profiles exposes limitations in institutional review capacity rather than in the legitimacy of ethics oversight. Recent advances in large language models (LLMs) offer new opportunities to support ethics review, but their direct application remains limited by insufficient ethical reasoning capability, weak integration with regulatory structures, and strict privacy constraints on authentic review materials. In this work, we introduce Mirror, an agentic framework for AI-assisted ethical review that integrates ethical reasoning, structured rule interpretation, and multi-agent deliberation within a unified architecture. At its core is EthicsLLM, a foundational model fine-tuned on EthicsQA, a specialized dataset of 41K question-chain-of-thought-answer triples distilled from authoritative ethics and regulatory corpora. EthicsLLM provides detailed normative and regulatory understanding, enabling Mirror to operate in two complementary modes. Mirror-ER (expedited Review) automates expedited review through an executable rule base that supports efficient and transparent compliance checks for minimal-risk studies. Mirror-CR (Committee Review) simulates full-board deliberation through coordinated interactions among expert agents, an ethics secretary agent, and a principal investigator agent, producing structured, committee-level assessments across ten ethical dimensions. Empirical evaluations demonstrate that Mirror significantly improves the quality, consistency, and professionalism of ethics assessments compared with strong generalist LLMs.

cs.AI↗
arXiv2026

From Struggle to Success: Context-Aware Guidance for Screen Reader Users in Computer Use

Equal access to digital technologies is critical for education, employment, and social participation. However, mainstream interfaces are visually oriented, creating steep learning curves and frequent obstacles for screen reader users, and limiting their independence and opportunities. Existing support is inadequate -- tutorials mainly target sighted users, while human assistance lacks real-time availability. We introduce AskEase, an on-demand AI assistant that provides step-by-step, screen reader user-friendly guidance for computer use. AskEase manages multiple sources of context to infer user intent and deliver precise, situation-specific guidance. Its seamless interaction design minimizes disruption and reduces the effort of seeking help. We demonstrated its effectiveness through representative usage scenarios and robustness tests. In a within-subjects study with 12 screen reader users, AskEase significantly improved task success while reducing perceived workload, including physical demand, effort, and frustration. These results demonstrate the potential of LLM-powered assistants to promote accessible computing and expand opportunities for users with visual impairments.

cs.HC↗
arXiv2026

Conjugate Relation Modeling for Few-Shot Knowledge Graph Completion

Few-shot Knowledge Graph Completion (FKGC) infers missing triples from limited support samples, tackling long-tail distribution challenges. Existing methods, however, struggle to capture complex relational patterns and mitigate data sparsity. To address these challenges, we propose a novel FKGC framework for conjugate relation modeling (CR-FKGC). Specifically, it employs a neighborhood aggregation encoder to integrate higher-order neighbor information, a conjugate relation learner combining an implicit conditional diffusion relation module with a stable relation module to capture stable semantics and uncertainty offsets, and a manifold conjugate decoder for efficient evaluation and inference of missing triples in manifold space. Experiments on three benchmarks demonstrate that our method achieves superior performance over state-of-the-art methods.

cs.CL↗
arXiv2026

Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces

AI agents may soon become capable of autonomously completing valuable, long-horizon tasks in diverse domains. Current benchmarks either do not measure real-world tasks, or are not sufficiently difficult to meaningfully measure frontier models. To this end, we present Terminal-Bench 2.0: a carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows. Each task features a unique environment, human-written solution, and comprehensive tests for verification. We show that frontier models and agents score less than 65\% on the benchmark and conduct an error analysis to identify areas for model and agent improvement. We publish the dataset and evaluation harness to assist developers and researchers in future work at https://www.tbench.ai/ .

cs.SE↗
arXiv2026

A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5

The rapid evolution of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has driven major gains in reasoning, perception, and generation across language and vision, yet whether these advances translate into comparable improvements in safety remains unclear, partly due to fragmented evaluations that focus on isolated modalities or threat models. In this report, we present an integrated safety evaluation of six frontier models--GPT-5.2, Gemini 3 Pro, Qwen3-VL, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5--assessing each across language, vision-language, and image generation using a unified protocol that combines benchmark, adversarial, multilingual, and compliance evaluations. By aggregating results into safety leaderboards and model profiles, we reveal a highly uneven safety landscape: while GPT-5.2 demonstrates consistently strong and balanced performance, other models exhibit clear trade-offs across benchmark safety, adversarial robustness, multilingual generalization, and regulatory compliance. Despite strong results under standard benchmarks, all models remain highly vulnerable under adversarial testing, with worst-case safety rates dropping below 6%. Text-to-image models show slightly stronger alignment in regulated visual risk categories, yet remain fragile when faced with adversarial or semantically ambiguous prompts. Overall, these findings highlight that safety in frontier models is inherently multidimensional--shaped by modality, language, and evaluation design--underscoring the need for standardized, holistic safety assessments to better reflect real-world risk and guide responsible deployment.

cs.AI↗
arXiv2025

Embedding Samples Dispatching for Recommendation Model Training in Edge Environments

Training deep learning recommendation models (DLRMs) on edge workers brings several benefits, particularly in terms of data privacy protection, low latency and personalization. However, due to the huge size of embedding tables, typical DLRM training frameworks adopt one or more parameter servers to maintain global embedding tables, while leveraging the edge workers cache part of them. This incurs significant transmission cost for embedding transmissions between workers and parameter servers, which can dominate the training cycle. In this paper, we investigate how to dispatch input embedding samples to appropriate edge workers to minimize the total embedding transmission cost when facing edge-specific challenges such as heterogeneous networks and limited resources. We develop ESD, a novel mechanism that optimizes the dispatch of input embedding samples to edge workers based on expected embedding transmission cost. We propose HybridDis as the dispatch decision method within ESD, which combines a resource-intensive optimal algorithm and a heuristic algorithm to balance decision quality and resource consumption. We implement a prototype of ESD and compare it with state-of-the-art mechanisms on real-world workloads. Extensive experimental results show that ESD reduces the embedding transmission cost by up to 36.76% and achieves up to 1.74 times speedup in end-to-end DLRM training.

cs.DC↗
arXiv2025

Astraea: A State-Aware Scheduling Engine for LLM-Powered Agents

Large Language Models (LLMs) are increasingly being deployed as intelligent agents. Their multi-stage workflows, which alternate between local computation and calls to external network services like Web APIs, introduce a mismatch in their execution pattern and the scheduling granularity of existing inference systems such as vLLM. Existing systems typically focus on per-segment optimization which prevents them from minimizing the end-to-end latency of the complete agentic workflow, i.e., the global Job Completion Time (JCT) over the entire request lifecycle. To address this limitation, we propose Astraea, a service engine designed to shift the optimization from local segments to the global request lifecycle. Astraea employs a state-aware, hierarchical scheduling algorithm that integrates a request's historical state with future predictions. It dynamically classifies requests by their I/O and compute intensive nature and uses an enhanced HRRN policy to balance efficiency and fairness. Astraea also implements an adaptive KV cache manager that intelligently handles the agent state during I/O waits based on the system memory pressure. Extensive experiments show that Astraea reduces average JCT by up to 25.5\% compared to baseline methods. Moreover, our approach demonstrates strong robustness and stability under high load across various model scales.

cs.CL↗
arXiv2025

Screen Reader Programmers in the Vibe Coding Era: Adaptation, Empowerment, and New Accessibility Landscape

Generative AI agents are reshaping human-computer interaction, shifting users from direct task execution to supervising machine-driven actions, especially the rise of "vibe coding" in programming. Yet little is known about how screen reader programmers interact with AI code assistants in practice. We conducted a longitudinal study with 16 blind and low-vision programmers. Participants completed a GitHub Copilot tutorial, engaged with a programming task, and provided initial feedback. After two weeks of AI-assisted programming, follow-ups examined how their practices and perceptions evolved. Our findings show that code assistants enhanced programming efficiency and bridged accessibility gaps. However, participants struggled to convey intent, interpret AI outputs, and manage multiple views while maintaining situational awareness. They showed diverse preferences for accessibility features, expressed a need to balance automation with control, and encountered barriers when learning to use these tools. Furthermore, we propose design principles and recommendations for more accessible and inclusive human-AI collaborations.

cs.HC↗
arXiv2025

Optimal Analysis for Bandit Learning in Matching Markets with Serial Dictatorship

The problem of two-sided matching markets is well-studied in computer science and economics, owing to its diverse applications across numerous domains. Since market participants are usually uncertain about their preferences in various online matching platforms, an emerging line of research is dedicated to the online setting where one-side participants (players) learn their unknown preferences through multiple rounds of interactions with the other side (arms). Sankararaman et al. provide an $Ω\left( \frac{N\log(T)}{Δ^2} + \frac{K\log(T)}Δ \right)$ regret lower bound for this problem under serial dictatorship assumption, where $N$ is the number of players, $K (\geq N)$ is the number of arms, $Δ$ is the minimum reward gap across players and arms, and $T$ is the time horizon. Serial dictatorship assumes arms have the same preferences, which is common in reality when one side participants have a unified evaluation standard. Recently, the work of Kong and Li proposes the ET-GS algorithm and achieves an $O\left( \frac{K\log(T)}{Δ^2} \right)$ regret upper bound, which is the best upper bound attained so far. Nonetheless, a gap between the lower and upper bounds, ranging from $N$ to $K$, persists. It remains unclear whether the lower bound or the upper bound needs to be improved. In this paper, we propose a multi-level successive selection algorithm that obtains an $O\left( \frac{N\log(T)}{Δ^2} + \frac{K\log(T)}Δ \right)$ regret bound when the market satisfies serial dictatorship. To the best of our knowledge, we are the first to propose an algorithm that matches the lower bound in the problem of matching markets with bandits.

cs.LG↗