arXiv subjects

Zihao Wan

Publications and source records attributed to Zihao Wan.

9 recordsLinked to original sources

arXiv2026

One Verdict Is Not Enough: Regression Testing Learned World Models and Their Test Harnesses

Learned world models ship as software components whose model versions, preprocessing, and test harnesses change between releases. Work on model updates already counts per-sample regressions, or negative flips, that an accuracy check hides. Two problems remain: a test can pass under one configuration and fail under another, and a harness fault can produce erroneous scores and wrong verdicts without a crash. We present TetraState, a regression-testing protocol for paired possible/impossible video tests. Each test runs under four fixed model-native configurations; its number of correct verdicts assigns one of four outcome classes. Inconclusive tests and failed executions are reported separately alongside the verdict rate. Revisions are compared by per-test outcome transitions, and the harness is versioned as part of the system under test. On all 506 IntPhys 2 Main tests and two heterogeneous systems, Cosmos 3 Nano and V-JEPA 2, two V-JEPA 2 harness revisions move modal accuracy by only 0.2 and 2.6 points yet change 11 and 214 test labels; in the second, accuracy rises although fewer tests are judged correctly, and 43 regressions keep the modal verdict, so a modal-verdict flip check misses them. These results outline a research agenda on configuration-level regression oracles for learned components.

cs.SE↗

arXiv2026

A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement

Recursive self-improvement requires models not only to acquire useful training experience, but also to decide when their current experience has ceased to be useful. Existing methods often focus on what experience to generate next, while paying less attention to when fresh experience should be introduced. Many also construct training targets through self-consistency or majority voting, making supervision depend on the model's own uncertain predictions. We address these two questions separately: training dynamics determine when to acquire new experience, while the answers to new tasks remain independently checkable. A-RSI monitors reward variation and accuracy in training logs to refresh a task pool when its learning signal declines, without consulting validation or test scores. A static library of model-authored programs produces multimodal questions and derives their reference answers; we specify the task-generation contract, automatically verify chosen properties of the reward and refresh rules, and separately audit the generated answers. The training loop uses no external training examples or human-written answer labels. On the primary eleven-benchmark suite, A-RSI raises the performance average from 57.77 to 61.05 with 800 fresh tasks.

cs.AI↗

arXiv2026

What Software Engineering Looks Like to AI Agents? -- An Empirical Study of AI-Only Technical Discourse on MoltBook

AI agents are increasingly framed as software-engineering teammates, yet most studies examine them inside human-centered workflows. Little is known about the discourse autonomous AI agents produce when they interact mainly with one another. This paper examines what autonomous agents discuss on MoltBook, how that discourse is organized, and how it differs from human developer discourse. We combine human open coding of a 500-post sample, a concentration-plus-check topic-analysis pipeline over 4,707 English-filtered MoltBook technology posts, and a matched comparison with 5,211 human-generated GitHub Discussions posts. MoltBook technology discourse spans 12 recurring themes, led by Security and Trust (27.4%). At the community level, activity is highly concentrated: the largest submolt accounts for 63.5% of posts (Gini = 0.88), yet a stability-aware BERTopic pipeline still identifies 32 non-outlier sub-topics. Relative to the GitHub Discussions baseline, MoltBook discourse contains fewer concrete, context-rich cues such as code-formatted artifacts, environment details, runtime failures, and reproduction steps. Social mimicry appears only in limited form, while idealization is reflected mainly through lower hedging. Overall, AI-only technical discourse is coherent but selective. It repeatedly returns to security and trust, memory and context management, tooling and APIs, debugging and error handling, workflow automation, and infrastructure/ops, while omitting much of the project-local and runtime detail common in human developer discourse. This may reflect fewer environment-specific failures, reproduction steps, and other grounding cues in MoltBook.

cs.SE↗

arXiv2026

Deep Pre-Alignment for VLMs

Most Vision Language Models (VLMs) directly map outputs from ViT encoders to the LLM via a lightweight projector. While effective, recent analysis suggests this architecture suffers from an alignment challenge: visual features remain distant from the text space in the initial layers of the LLM, forcing the model to waste critical depth~\cite{zhang-etal-2024-investigating,artzy-schwartz-2024-attend} on superficial modality alignment rather than deep understanding and complex reasoning. In this work, we propose Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder with a small VLM as perceiver, ensuring visual features are deeply aligned with the text space of the target large language model. Comprehensive experiments demonstrate the effectiveness of DPA. On the 4B parameter scale, DPA outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32B scale. Moreover, by offloading alignment to the perceiver, DPA achieves a 32.9\% reduction in language capability forgetting over 3 text benchmarks. We further demonstrate that these gains are consistent across different LLM families including Qwen3 and LLaMA 3.2, highlighting the generality of our approach. Beyond performance, DPA also offers a seamless upgrade path for current VLM development, requiring only a modular replacement for the visual encoder with marginal computation overhead.

cs.CV↗

arXiv2026

Enabling Stroke-Level Structural Analysis of Hieroglyphic Scripts without Language-Specific Priors

Hieroglyphs, as logographic writing systems, encode rich semantic and cultural information within their internal structural composition. Yet, current advanced Large Language Models (LLMs) and Multimodal LLMs (MLLMs) usually remain structurally blind to this information. LLMs process characters as textual tokens, while MLLMs additionally view them as raw pixel grids. Both fall short to model the underlying logic of character strokes. Furthermore, existing structural analysis methods are often script-specific and labor-intensive. In this paper, we propose Hieroglyphic Stroke Analyzer (HieroSA), a novel and generalizable framework that enables MLLMs to automatically derive stroke-level structures from character bitmaps without handcrafted data. It transforms modern logographic and ancient hieroglyphs character images into explicit, interpretable line-segment representations in a normalized coordinate space, allowing for cross-lingual generalization. Extensive experiments demonstrate that HieroSA effectively captures character-internal structures and semantics, bypassing the need for language-specific priors. Experimental results highlight the potential of our work as a graphematics analysis tool for a deeper understanding of hieroglyphic scripts. View our code at https://github.com/THUNLP-MT/HieroSA.

cs.CV↗

arXiv2025

Bridging Vision, Language, and Mathematics: Pictographic Character Reconstruction with Bézier Curves

While Vision-language Models (VLMs) have demonstrated strong semantic capabilities, their ability to interpret the underlying geometric structure of visual information is less explored. Pictographic characters, which combine visual form with symbolic structure, provide an ideal test case for this capability. We formulate this visual recognition challenge in the mathematical domain, where each character is represented by an executable program of geometric primitives. This is framed as a program synthesis task, training a VLM to decompile raster images into programs composed of Bézier curves. Our model, acting as a "visual decompiler", demonstrates performance superior to strong zero-shot baselines, including GPT-4o. The most significant finding is that when trained solely on modern Chinese characters, the model is able to reconstruct ancient Oracle Bone Script in a zero-shot context. This generalization provides strong evidence that the model acquires an abstract and transferable geometric grammar, moving beyond pixel-level pattern recognition to a more structured form of visual understanding.

cs.LG↗

arXiv2024

StreamingBench: Assessing the Gap for MLLMs to Achieve Streaming Video Understanding

The rapid development of Multimodal Large Language Models (MLLMs) has expanded their capabilities from image comprehension to video understanding. However, most of these MLLMs focus primarily on offline video comprehension, necessitating extensive processing of all video frames before any queries can be made. This presents a significant gap compared to the human ability to watch, listen, think, and respond to streaming inputs in real time, highlighting the limitations of current MLLMs. In this paper, we introduce StreamingBench, the first comprehensive benchmark designed to evaluate the streaming video understanding capabilities of MLLMs. StreamingBench assesses three core aspects of streaming video understanding: (1) real-time visual understanding, (2) omni-source understanding, and (3) contextual understanding. The benchmark consists of 18 tasks, featuring 900 videos and 4,500 human-curated QA pairs. Each video features five questions presented at different time points to simulate a continuous streaming scenario. We conduct experiments on StreamingBench with 13 open-source and proprietary MLLMs and find that even the most advanced proprietary MLLMs like Gemini 1.5 Pro and GPT-4o perform significantly below human-level streaming video understanding capabilities. We hope our work can facilitate further advancements for MLLMs, empowering them to approach human-level video comprehension and interaction in more realistic scenarios.

cs.CV↗

arXiv2024

Results of the Big ANN: NeurIPS'23 competition

The 2023 Big ANN Challenge, held at NeurIPS 2023, focused on advancing the state-of-the-art in indexing data structures and search algorithms for practical variants of Approximate Nearest Neighbor (ANN) search that reflect the growing complexity and diversity of workloads. Unlike prior challenges that emphasized scaling up classical ANN search ~\cite{DBLP:conf/nips/SimhadriWADBBCH21}, this competition addressed filtered search, out-of-distribution data, sparse and streaming variants of ANNS. Participants developed and submitted innovative solutions that were evaluated on new standard datasets with constrained computational resources. The results showcased significant improvements in search accuracy and efficiency over industry-standard baselines, with notable contributions from both academic and industrial teams. This paper summarizes the competition tracks, datasets, evaluation metrics, and the innovative approaches of the top-performing submissions, providing insights into the current advancements and future directions in the field of approximate nearest neighbor search.

cs.IR↗

arXiv2024

CODIS: Benchmarking Context-Dependent Visual Comprehension for Multimodal Large Language Models

Multimodal large language models (MLLMs) have demonstrated promising results in a variety of tasks that combine vision and language. As these models become more integral to research and applications, conducting comprehensive evaluations of their capabilities has grown increasingly important. However, most existing benchmarks fail to consider that, in certain situations, images need to be interpreted within a broader context. In this work, we introduce a new benchmark, named as CODIS, designed to assess the ability of models to use context provided in free-form text to enhance visual comprehension. Our findings indicate that MLLMs consistently fall short of human performance on this benchmark. Further analysis confirms that these models struggle to effectively extract and utilize contextual information to improve their understanding of images. This underscores the pressing need to enhance the ability of MLLMs to comprehend visuals in a context-dependent manner. View our project website at https://thunlp-mt.github.io/CODIS.

cs.CV↗