William Yang Wang

Oversensitivity occurs when language models defensively reject prompts that are actually benign. This behavior not only disrupts user interactions but also obscures the boundary between harmful and harmless content. Existing benchmarks rely on static datasets that degrade overtime as models evolve, leading to data contamination and diminished evaluative power. To address this, we develop a framework that dynamically generates model-specific challenging datasets, capturing emerging defensive patterns and aligning with each model's unique behavior. Building on this approach, we construct OVERBENCH, a benchmark that aggregates these datasets across diverse LLM families, encompassing 450,000 samples from 25 models. OVERBENCH provides a dynamic and evolving perspective on oversensitivity, allowing for continuous monitoring of defensive triggers as models advance, highlighting vulnerabilities that static datasets overlook.

Sophia Xiao Pu, Sitao Cheng, Xin Eric Wang, William Yang Wang. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025.

Large language models (LLMs) encode vast amounts of knowledge during pre-training (parametric knowledge, or PK) and can further be enhanced by incorporating contextual knowledge (CK). Can LLMs effectively integrate their internal PK with external CK to solve complex problems? In this paper, we investigate the dynamic interaction between PK and CK, categorizing their relationships into four types: Supportive, Complementary, Conflicting, and Irrelevant. To support this investigation, we introduce ECHOQA, a benchmark spanning scientific, factual, and commonsense knowledge. Our results show that LLMs tend to suppress their PK when contextual information is available, even when it is complementary or irrelevant. While tailored instructions can encourage LLMs to rely more on their PK, they still struggle to fully leverage it. These findings reveal a key vulnerability in LLMs, raising concerns about their reliability in knowledge-intensive tasks. Resources are available at https://github.com/sitaocheng/Knowledge_Interplay

Kexun Zhang, Yee Choi, Zhenqiao Song, Taiqi He, William Yang Wang, Lei Li. Findings of the Association for Computational Linguistics ACL 2024. 2024.

Large language models (LLMs) excel at implementing code from functionality descriptions but struggle with algorithmic problems that require not only implementation but also identification of the suitable algorithm. Moreover, LLM-generated programs lack guaranteed correctness and require human verification. To address these challenges, we propose ALGO, a framework that synthesizes Algorithmic programs with LLM-Generated Oracles to guide the generation and verify their correctness. ALGO first generates a reference oracle by prompting an LLM to exhaustively enumerate all the combinations of relevant variables. This oracle is then utilized to guide an arbitrary search strategy in exploring the algorithm space and to verify the synthesized algorithms. Our study shows that the LLM-generated oracles are correct for 88% of the cases. With the oracles as verifiers, ALGO can be integrated with any existing code generation model in a model-agnostic manner to enhance its performance. Experiments show that when equipped with ALGO, we achieve an 8x better one-submission pass rate over the Codex model and a 2.6x better one-submission pass rate over CodeT, the current state-of-the-art model on CodeContests. We can also get 1.3x better pass rate over the ChatGPT Code Interpreter on unseen problems. The problem set we used for testing, the prompts we used, the verifier and solution programs, and the test cases generated by ALGO are available at https://github.com/zkx06111/ALGO.

The field of text-to-image (T2I) generation has garnered significant attention both within the research community and among everyday users. Despite the advancements of T2I models, a common issue encountered by users is the need for repetitive editing of input prompts in order to receive a satisfactory image, which is time-consuming and labor-intensive. Given the demonstrated text generation power of large-scale language models, such as GPT-k, we investigate the potential of utilizing such models to improve the prompt editing process for T2I generation. We conduct a series of experiments to compare the common edits made by humans and GPT-k, evaluate the performance of GPT-k in prompting T2I, and examine factors that may influence this process. We found that GPT-k models focus more on inserting modifiers while humans tend to replace words and phrases, which includes changes to the subject matter. Experimental results show that GPT-k are more effective in adjusting modifiers rather than predicting spontaneous changes in the primary subject matters. Adopting the edit suggested by GPT-k models may reduce the percentage of remaining edits by 20-30%.

Prompt tuning is a new few-shot transfer learning technique that only tunes the learnable prompt for pre-trained vision and language models such as CLIP. However, existing prompt tuning methods tend to learn spurious or entangled representations, which leads to poor generalization to unseen concepts. Towards non-spurious and efficient prompt learning from limited examples, this paper presents a novel \underline{\textbf{C}}ounterfactual \underline{\textbf{P}}rompt \underline{\textbf{L}}earning (CPL) method for vision and language models, which simultaneously employs counterfactual generation and contrastive learning in a joint optimization framework. Particularly, CPL constructs counterfactual by identifying minimal non-spurious feature change between semantically-similar positive and negative samples that causes concept change, and learns more generalizable prompt representation from both factual and counterfactual examples via contrastive learning. Extensive experiments demonstrate that CPL can obtain superior few-shot performance on different vision and language tasks than previous prompt tuning methods on CLIP. On image classification, we achieve 3.55\% average relative improvement on unseen classes across seven datasets; on image-text retrieval and visual question answering, we gain up to 4.09\% and 25.08\% relative improvements across three few-shot scenarios on unseen test sets respectively.

The problem of verifying whether a textual hypothesis holds based on the given evidence, also known as fact verification, plays an important role in the study of natural language understanding and semantic representation. However, existing studies are mainly restricted to dealing with unstructured evidence (e.g., natural language sentences and documents, news, etc), while verification under structured evidence, such as tables, graphs, and databases, remains under-explored. This paper specifically aims to study the fact verification given semi-structured data as evidence. To this end, we construct a large-scale dataset called TabFact with 16k Wikipedia tables as the evidence for 118k human-annotated natural language statements, which are labeled as either ENTAILED or REFUTED. TabFact is challenging since it involves both soft linguistic reasoning and hard symbolic reasoning. To address these reasoning challenges, we design two different models: Table-BERT and Latent Program Algorithm (LPA). Table-BERT leverages the state-of-the-art pre-trained language model to encode the linearized tables and statements into continuous vectors for verification. LPA parses statements into programs and executes them against the tables to obtain the returned binary value for verification. Both methods achieve similar accuracy but still lag far behind human performance. We also perform a comprehensive analysis to demonstrate great future opportunities. The data and code of the dataset are provided in \url{https://github.com/wenhuchen/Table-Fact-Checking}.

While social media empowers freedom of expression and individual voices, it also enables anti-social behavior, online harassment, cyberbullying, and hate speech. In this paper, we deepen our understanding of online hate speech by focusing on a largely neglected but crucial aspect of hate speech -- its target: either "directed" towards a specific person or entity, or "generalized" towards a group of people sharing a common protected characteristic. We perform the first linguistic and psycholinguistic analysis of these two forms of hate speech and reveal the presence of interesting markers that distinguish these types of hate speech. Our analysis reveals that Directed hate speech, in addition to being more personal and directed, is more informal, angrier, and often explicitly attacks the target (via name calling) with fewer analytic words and more words suggesting authority and influence. Generalized hate speech, on the other hand, is dominated by religious hate, is characterized by the use of lethal words such as murder, exterminate, and kill; and quantity words such as million and many. Altogether, our work provides a data-driven analysis of the nuances of online-hate speech that enables not only a deepened understanding of hate speech and its social implications but also its detection.