🧑‍🔬 Yiwei Wang

🏛️ Affiliations

🎓University of California, MercedMerced · United States2024–2026 · 15 paperscurrent🎓University of California, Los AngelesLos Angeles · United States2024–2026 · 2 paperscurrent🎓Rutgers, The State University of New JerseyNew Brunswick · United States2016–2018 · 6 papers🎓Shandong Normal UniversityJinan · China2022–2025 · 3 papersEducation 2022–2025 · 硕士, School of Information Science and Engineering

From the bylines of their papers (OpenAlex, 2026-10-14), ORCID and Wikidata; institutions on papers are matched by machine. 7 more institutions on single bylines are not listed.

🧭 Fields, topics and concepts

Of its works in our library, with the number of works on each.

📅 Works by year

2020s

📄 Papers

The most cited first (OpenAlex and our library).

1. Mitigating Lost-in-Retrieval Problems in Retrieval Augmented Multi-Hop Question Answering (2025 · 16 citations)
   In retrieval-augmented multi-hop QA, LLMs often split a complex question into sub-questions. Later sub-questions then refer to an earlier result through a pronoun such as "this author" instead of naming the entity. The authors call this failure "lost-in-retrieval." Without the key entity, retrieval for these…
2. METAL: A Multi-Agent Framework for Chart Generation with Test-Time Scaling (2025 · 15 citations)
   The paper addresses chart-to-code generation: given an image of a reference chart, a model must write plotting code (e.g., Python) whose rendered output reproduces the chart's type, layout, colors and text.
3. Can Graph Descriptive Order Affect Solving Graph Problems with LLMs? (2025 · 10 citations)
   Studies of LLM graph reasoning usually turn a graph into text by listing its edges in an arbitrary sequence.
4. Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models (2026 · 9 citations)
   Most earlier reviews of mechanistic interpretability (MI) in LLMs describe what researchers have discovered inside models.
5. DiMo-GUI: Advancing Test-time Scaling in GUI Grounding via Modality-Aware Visual Reasoning (2025 · 8 citations)
   Locating the on-screen element that a natural-language command refers to (GUI grounding) is difficult. High-resolution screenshots are crowded with distractors, and vision-language models tend to favour visible text over icons, even when an icon is the real target. DiMo-GUI is a training-free, plug-and-play inference…
6. Thinking with Sound: Audio Chain-of-Thought Enables Multimodal Reasoning in Large Audio-Language Models (2025 · 6 citations)
   Large Audio-Language Models (LALMs) turn a recording into a fixed set of embeddings once. After that step they cannot go back to the signal. Because of this, they perform poorly when audio is noisy or degraded, or when a task depends on fine acoustic detail. The authors propose Thinking-with-Sound (TwS), a framework…
7. Mapping the Minds of LLMs: A Graph-Based Analysis of Reasoning LLMs (2025 · 5 citations)
   Large Reasoning Models (LRMs) such as DeepSeek-R1 produce long Chain-of-Thought traces, yet they often behave erratically.
8. Vulnerability of LLMs to Vertically Aligned Text Manipulations (2025 · 5 citations)
   The paper asks whether decoder-based LLMs are as vulnerable as encoder models like BERT when some words in a text are written vertically, one character per line.
9. Gated Differentiable Working Memory for Long-Context Language Modeling (2026 · 4 citations)
   Long inputs remain difficult for transformer LLMs. Attention spreads thin over thousands of tokens, and evidence placed mid-document is often overlooked. Test-time training methods such as qTTT address this by briefly fine-tuning "fast weights" on the given context. However, they sample positions uniformly, so they…
10. HiddenGuard: Fine-Grained Safe Generation with Specialized Representation Router (2026 · 4 citations)
    Safety alignment in LLMs usually works as a yes/no gate: a prompt is either answered or refused. This tends to block harmless requests, such as a "killer slideshow," and to miss subtle or context-dependent harm. The authors propose HiddenGuard, which moderates output token by token instead of refusing whole answers.
11. a1: Steep Test-time Scaling Law via Environment Augmented Generation (2026 · 2 citations)
    LLMs often fail at long multi-step reasoning. They usually generate a whole solution in one pass with no intermediate checks, so an early mistake carries through to the final answer. The authors propose Environment Augmented Generation (EAG), which casts reasoning as a Markov Decision Process. While generating, the…
12. Hard to Read, Easy to Jailbreak: How Visual Degradation Bypasses MLLM Safety Alignment (2026)
    Vision-based context compression, as in DeepSeek-OCR and Glyph, renders long text as images so that a model needs fewer tokens.
13. Mitigating Coordinate Prediction Bias from Positional Encoding Failures (2026)
    Multimodal LLMs that output click coordinates as digit strings often drift on high-resolution screenshots, because visual positional encodings (VPEs) lose grounding strength in long visual contexts.

🤝 Co-authors here

12 of 29 co-authors with a page, the most papers together first.

• Yujun Cai ×5
• Baolong Bi ×4
• Lingrui Mei ×4
• Shenghua Liu ×4
• Xueqi Cheng ×4
• Yuyao Ge ×3
• Zhecheng Li ×3
• Zhen Xiong ×3
• Jun Wan ×2
• Kai-Wei Chang ×2
• Nanyun Peng ×2
• Bingxuan Li

Venues

Where their works here were published, the most works first, with the editions of those that have a page.

• Annual Meeting of the Association for Computational Linguistics (ACL) ×6: ACL 2026 ×2, ACL 2025 ×4
• Findings of the Association for Computational Linguistics ×4: Findings of the Association for Computational Linguistics 2026 ×4
• Conference on Empirical Methods in Natural Language Processing (EMNLP) ×2: EMNLP 2025 ×2

🔗 Cited by

Papers of our library cite their works 71 times (of the 149,439 papers whose bibliographies we read); 20 citing papers with a page here →.

Updated 1 November 2026

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/authors/yiwei-wang/