Zihan Dong
Publications
How Benchmarks Mis-Score Computer-Use Agents
Computer-use agents (CUA) are being deployed to browse the web and operate desktop software, yet their benchmark scores are still commonly produced by brittle scripted oracles. A score is the output of a pipeline in which tasks can be stale, trajectories can omit decisive visual evidence, evaluators can reject valid alternatives, and aggregate reports can hide the cause of failure. We organize these problems into a reliability framework spanning task construction, trajectory observation, scoring, and reporting. We then audit 150 public failure-scored trajectories from five web, enterprise-workflow, and desktop-control benchmarks, find that 15.3\% of FAIL verdicts are wrong: 10.7\% are evaluator false negatives and 4.7\% are broken tasks. For genuine failures, a three-tier diagnostic taxonomy shows that verification/feedback and planning failures dominate execution/grounding errors, while a single scalar success rate can not explain. We connect these findings to newer long-horizon CUA benchmarks and derive stage-specific design rules for CUA evaluation.
Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising
Slide design requires personalizing both deck themes and page layouts. Yet, current AI agent-based methods struggle with fine-grained, page-level design. Solely relying on prespecified templates or user verbose instructions, they fail to capture latent design intents, leaving Page-level Slide Personalization (PSP) unresolved. To close this gap, this work formulates PSP as an inverse planning problem. We propose to learn a design intent without assuming any knowledge of the specific executing tools (e.g., PowerPoint, Beamer) being used. However, relinquishing control over these tools makes the problem intractable to optimize end-to-end. To overcome this, we propose SPIRE, a principled framework to solve PSP approximately. By intentionally corrupting the visual structures of clean slides, SPIRE creates a verifiable task to denoise the corruption, whereby two agents learn to collaboratively refine executable designs via reinforcement learning (RL). We present a proof that structural denoising is a consistent surrogate for PSP, and that the multi-agent formulation strictly reduces policy gradient variance in RL. Extensive experiments demonstrate the superiority of SPIRE.
LUCID-SAE: Learning Unified Vision-Language Sparse Codes for Interpretable Concept Discovery
Sparse autoencoders (SAEs) offer a natural path toward comparable explanations across different representation spaces. However, current SAEs are trained per modality, producing dictionaries whose features are not directly understandable and whose explanations do not transfer across domains. In this study, we introduce LUCID (Learning Unified vision-language sparse Codes for Interpretable concept Discovery), a unified vision-language sparse autoencoder that learns a shared latent dictionary for image patch and text token representations, while reserving private capacity for modality-specific details. We achieve feature alignment by coupling the shared codes with a learned optimal transport matching objective without the need of labeling. LUCID yields interpretable shared features that support patch-level grounding, establish cross-modal neuron correspondence, and enhance robustness against the concept clustering problem in similarity-based evaluation. Leveraging the alignment properties, we develop an automated dictionary interpretation pipeline based on term clustering without manual observations. Our analysis reveals that LUCID's shared features capture diverse semantic categories beyond objects, including actions, attributes, and abstract concepts, demonstrating a comprehensive approach to interpretable multimodal representations.
Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals
Evaluating mathematical reasoning in LLMs is constrained by limited benchmark sizes and inherent model stochasticity, yielding high-variance accuracy estimates and unstable rankings across platforms. On difficult problems, an LLM may fail to produce a correct final answer, yet still provide reliable pairwise comparison signals indicating which of two candidate solutions is better. We leverage this observation to design a statistically efficient evaluation framework that combines standard labeled outcomes with pairwise comparison signals obtained by having models judge auxiliary reasoning chains. Treating these comparison signals as control variates, we develop a semiparametric estimator based on the efficient influence function (EIF) for the setting where auxiliary reasoning chains are observed. This yields a one-step estimator that achieves the semiparametric efficiency bound, guarantees strict variance reduction over naive sample averaging, and admits asymptotic normality for principled uncertainty quantification. Across simulations, our one-step estimator substantially improves ranking accuracy, with gains increasing as model output noise grows. Experiments on GPQA Diamond, AIME 2025, and GSM8K further demonstrate more precise performance estimation and more reliable model rankings, especially in small-sample regimes where conventional evaluation is pretty unstable.