--- title: "Kai-Wei Chang — papers, page 2 — Elamur Research" description: "Kai-Wei Chang, University of California, Los Angeles: 50 papers on Elamur Research (2011–2026), h-index 53 — affiliations, topics…" canonical: "https://research.elamur.ai/free/en/authors/kai-wei-chang/page/2/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/authors/kai-wei-chang/page/2/) [home](https://md.research.elamur.ai/free/en/) / [authors](https://md.research.elamur.ai/free/en/authors/) / [Kai-Wei Chang](https://md.research.elamur.ai/free/en/authors/kai-wei-chang/) / page 2 # Kai-Wei Chang · page 2 50 papers in our library · 2011–2026 · page 2 of 2 [The profile, affiliations and co-authors →](https://md.research.elamur.ai/free/en/authors/kai-wei-chang/) ## Papers The most cited first (OpenAlex and our library). [← prev](https://md.research.elamur.ai/free/en/authors/kai-wei-chang/) [1](https://md.research.elamur.ai/free/en/authors/kai-wei-chang/) **2** 26. **[A Constrained Latent Variable Model for Coreference Resolution](https://md.research.elamur.ai/free/en/papers/chang-2013-constrained-latent-variable-model-coreference/)** *(2013 · 81 citations)* Coreference resolution clusters noun-phrase mentions according to which real-world entity each one denotes. 27. **[A Joint Framework for Coreference Resolution and Mention Head Detection](https://md.research.elamur.ai/free/en/papers/peng-2015-joint-framework-coreference-resolution-mention/)** *(2015 · 75 citations)* Coreference systems lose much of their accuracy when they must find mentions in raw text rather than receive gold-annotated ones. 28. **[Model Editing Harms General Abilities of Large Language Models: Regularization to the Rescue](https://md.research.elamur.ai/free/en/papers/gu-2024-model-editing-harms-general-abilities/)** *(2024 · 67 citations)* Model editing updates facts stored in large language models without full retraining. However, it is usually judged only on whether the edit sticks, generalizes to paraphrases and leaves unrelated facts intact. This EMNLP 2024 paper asks whether such edits quietly damage the models' broader capabilities. The authors… 29. **[Red Teaming Language Model Detectors with Language Models](https://md.research.elamur.ai/free/en/papers/shi-2024-red-teaming-language-model-detectors/)** *(2024 · 62 citations)* Detectors that flag LLM-generated text are proposed as a safeguard against misuse such as plagiarism or mass misinformation, but it was unclear how well they resist deliberate evasion. 30. **[Inference Protocols for Coreference Resolution](https://md.research.elamur.ai/free/en/papers/chang-2011-inference-protocols-coreference-resolution/)** *(2011 · 50 citations)* This paper presents Illinois-Coref, the system the University of Illinois submitted to the closed-track competition of CoNLL-2011, a shared task on resolving coreference in OntoNotes-4.0. 31. **[FLIRT: Feedback Loop In-context Red Teaming](https://md.research.elamur.ai/free/en/papers/mehrabi-2024-flirt-feedback-loop-context-red/)** *(2024 · 42 citations)* The paper tackles automated red teaming: finding inputs that push generative models into producing unsafe content, without expensive human probing. 32. **[Syntax-augmented Multilingual BERT for Cross-lingual Transfer](https://md.research.elamur.ai/free/en/papers/ahmad-2021-syntax-augmented-multilingual-bert-cross/)** *(2021 · 42 citations)* The authors test whether explicitly added universal syntax helps the multilingual encoder mBERT in zero-shot cross-lingual transfer. 33. **[ParaAMR: A Large-Scale Syntactically Diverse Paraphrase Dataset by AMR Back-Translation](https://md.research.elamur.ai/free/en/papers/huang-2023-paraamr-large-scale-syntactically-diverse/)** *(2023 · 24 citations)* ParaAMR builds about 15 million paraphrase pairs by AMR back-translation. Each source sentence is parsed to an AMR graph; a different node is chosen as the focus (the main assertion); the modified graph is linearised and a sentence is generated from it. Candidates with high perplexity under a language model are… 34. **[METAL: A Multi-Agent Framework for Chart Generation with Test-Time Scaling](https://md.research.elamur.ai/free/en/papers/li-2025-metal-multi-agent-framework-chart/)** *(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. 35. **[Unsupervised Syntactically Controlled Paraphrase Generation with Abstract Meaning Representations](https://md.research.elamur.ai/free/en/papers/huang-2022-unsupervised-syntactically-controlled-paraphrase-generation/)** *(2022 · 15 citations)* AMRPG generates a paraphrase that follows a requested syntactic form without any paraphrase pairs. It encodes the input twice: an AMR graph gives a semantic embedding, and the constituency parse gives a syntactic embedding. A decoder learns to reconstruct the input from the two. At test time the syntactic embedding… 36. **[When To Solve, When To Verify: Compute-Optimal Problem Solving and Generative Verification for LLM Reasoning](https://md.research.elamur.ai/free/en/papers/singhi-2025-solve-verify-compute-optimal-problem/)** *(2025 · 12 citations)* Large language models can spend extra test-time compute on reasoning in two main ways. The first is Self-Consistency (SC), which samples many candidate solutions and takes a majority vote. The second samples fewer solutions and ranks them with a Generative Reward Model (GenRM), which writes several verification… 37. **[PLUE: Language Understanding Evaluation Benchmark for Privacy Policies in English](https://md.research.elamur.ai/free/en/papers/chi-2023-plue-language-understanding-evaluation-benchmark/)** *(2023 · 12 citations)* Privacy policies are long, legally worded documents that few users read. Earlier NLP work on them treated each task separately, with its own expert-annotated dataset and model. This makes progress hard to compare, and generic pretrained models fit the domain poorly. The authors built PLUE, a unified English benchmark… 38. **[Using Item Response Theory to Measure Gender and Racial Bias of a BERT-based Automated English Speech Assessment System](https://md.research.elamur.ai/free/en/papers/kwako-2022-item-response-theory-measure-gender/)** *(2022 · 10 citations)* The paper asks whether a transformer-based scorer for spoken English treats children differently by gender or by race. 39. **[Does BERT Exacerbate Gender or L1 Biases in Automated English Speaking Assessment?](https://md.research.elamur.ai/free/en/papers/kwako-2023-does-bert-exacerbate-gender-or/)** *(2023 · 8 citations)* Pretrained language models like BERT can score spoken English test answers about as accurately as human raters. 40. **[QUDSELECT: Selective Decoding for Questions Under Discussion Parsing](https://md.research.elamur.ai/free/en/papers/suvarna-2024-qudselect-selective-decoding-questions-under/)** *(2024 · 7 citations)* Questions Under Discussion (QUD) parsing links each sentence of a text to an earlier "anchor" sentence through an implicit question that the later sentence answers. 41. **[Vulnerability of LLMs to Vertically Aligned Text Manipulations](https://md.research.elamur.ai/free/en/papers/li-2025-vulnerability-llms-vertically-aligned-text/)** *(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. 42. **[BILLY: Steering Large Language Models via Merging Persona Vectors for Creative Generation](https://md.research.elamur.ai/free/en/papers/pai-2026-billy-steering-large-language-models/)** *(2026 · 4 citations)* Multi-LLM setups, where several role-playing models debate an idea, tend to produce more creative output than a single model. 43. **[BLUR: A Bi-Level Optimization Approach for LLM Unlearning](https://md.research.elamur.ai/free/en/papers/reisizadeh-2026-blur-bi-level-optimization-approach/)** *(2026 · 4 citations)* Most LLM unlearning methods minimize a weighted sum of a forget loss and a retain loss with a fixed coefficient. 44. **[DialectGen: Benchmarking and Improving Dialect Robustness in Multimodal Generation](https://md.research.elamur.ai/free/en/papers/zhou-2025-dialectgen-benchmarking-improving-dialect-robustness/)** *(2025 · 3 citations)* Text-to-image and text-to-video models are trained mostly on Standard American and British English. This paper asks how they cope when a prompt uses another English dialect. The authors built DialectGen, a benchmark covering SAE plus five other dialects: AAE, BrE, ChE, InE and SgE. 45. **[VisRet: Visualization Improves Knowledge-Intensive Text-to-Image Retrieval](https://md.research.elamur.ai/free/en/papers/wu-2026-visret-visualization-improves-knowledge-intensive/)** *(2026 · 1 citation)* Text-to-image retrieval often fails when a query demands specific visual configurations, such as an animal's posture or the camera angle. 46. **[Reinforcing Stereotypes of Anger: Emotion AI on African American Vernacular English](https://md.research.elamur.ai/free/en/papers/dorn-2025-reinforcing-stereotypes-anger-emotion-ai/)** *(2025 · 1 citation)* The paper examines whether text-based emotion recognition treats African American Vernacular English (AAVE) differently from General American English (GAE). 47. **[AutoSUIT Bench - Automated Security UnIt Test Benchmark for LLM Coding](https://md.research.elamur.ai/free/en/papers/osebe-2026-autosuit-bench-automated-security-unit/)** *(2026)* Benchmarks that check whether LLM-written code is secure tend to rely on one of two approaches. Some use static analyzers such as CodeQL, which make many errors. Others use hand-written security tests, which limits them to a few dozen of the most common CWE categories. AutoSUIT addresses both problems with an… 48. **[From Narrow Unlearning to Emergent Misalignment in LLMs](https://md.research.elamur.ai/free/en/papers/mushtaq-2026-narrow-unlearning-emergent-misalignment-llms/)** *(2026)* Earlier work found that fine-tuning an aligned LLM on a narrow task, such as writing insecure code, can cause emergent misalignment (EMA): harmful behaviour that spreads to unrelated domains. 49. **[InsideOut: Measuring and Mitigating Insider–Outsider Bias in Interview Script Generation](https://md.research.elamur.ai/free/en/papers/wan-2026-insideout-measuring-mitigating-insider-outsider/)** *(2026)* The paper examines a subtle cultural bias in LLMs: when writing text "about" a culture, models tend to speak as members of dominant cultures, especially the US, while treating other cultures as foreign objects of curiosity. 50. **[SWAN: Semantic Watermarking with Abstract Meaning Representation](https://md.research.elamur.ai/free/en/papers/ye-2026-swan-semantic-watermarking-abstract-meaning/)** *(2026)* Watermarks for LLM-generated text are usually planted at the token level by biasing which tokens the decoder picks. [← prev](https://md.research.elamur.ai/free/en/authors/kai-wei-chang/) [1](https://md.research.elamur.ai/free/en/authors/kai-wei-chang/) **2**