Papers by Jackie CK Cheung
Collaborative Rational Speech Act: Pragmatic Reasoning for Multi-Turn Dialog (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing extensions of Rational Speech Act face challenges in scaling to multi-turn, collaborative scenarios. |
| Approach: | They propose a Rational Speech Act extension that optimizes a gain function adapted from rate-distortion theory to model multi-turn dialog by optimizing a model gain . they demonstrate the effectiveness of CRSA on referential games and template-based doctor–patient dialogs in the medical domain. |
| Outcome: | The proposed model yields more consistent, interpretable, and collaborative behavior than baselines, paving the way for more pragmatic and socially aware language agents. |
Beyond the Safety Bundle: Auditing the Helpful and Harmless Dataset (2025.naacl-long)
Copied to clipboard
| Challenge: | Learning from human feedback (LHF) has been used to mitigate the harms of large language models (LLMs) but the quality of this feedback and its effectiveness as a safety mitigation technique remain unclear. |
| Approach: | They audit the Helpful and Harmless (HH) dataset by Anthropic and examine how conceptualization failures and quality issues identified in the dataset can create additional harms . |
| Outcome: | The findings highlight the need for more nuanced, context-sensitive approaches to safety mitigation in large language models. |
Rethinking Full Finetuning from Pretraining Checkpoints in Active Learning for African Languages (2025.acl-srw)
Copied to clipboard
| Challenge: | Existing approaches to improve model performance are finetuning on all acquired data after each round, which is computationally expensive in multilingual and low-resource settings. |
| Approach: | They evaluate continual finetuning (CF) against full finetuned (FA) across 28 African languages using MasakhaNEWS and SIB-200. |
| Outcome: | The proposed approach outperforms full finetuning (FA) in 28 African languages, achieving up to 35% reductions in GPU memory, FLOPs, and training time. |
(RSA)²: A Rhetorical-Strategy-Aware Rational Speech Act Framework for Figurative Language Understanding (2025.acl-long)
Copied to clipboard
| Challenge: | Existing implementations of the Rational Speech Act (RSA) framework do not account for figurative expressions or require modeling the implicit motivations behind using figurativ language in a setting-specific way. |
| Approach: | They propose a framework which models figurative language use by considering a speaker's employed rhetorical strategy and a computational model which incorporates rhetorical strategies to support non-literal interpretation. |
| Outcome: | The proposed framework enables human-compatible interpretations of non-literal utterances without modeling speaker's motivations for being non-lative. |
Can LLMs Reason Abstractly Over Math Word Problems Without CoT? Disentangling Abstract Formulation From Arithmetic Computation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) are often evaluated on math word problems . however, such metrics conflate two distinct sub-skills: abstract formulation and arithmetic computation. |
| Approach: | They propose to use Final-answer-based metrics to evaluate large language models on math word problems to conflate two distinct sub-skills: abstract formulation and arithmetic computation. |
| Outcome: | The proposed model performance is bottlenecked by arithmetic computation and not abstract formulation, the study shows. |
PreSumm: Predicting Summarization Performance Without Summarizing (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in summarization models do not produce all documents in the same way, despite their inherent design principles and operational mechanisms. |
| Approach: | They propose a task where a system predicts summarization performance based solely on the source document. |
| Outcome: | The proposed task identifies documents that require manual summarization and improves dataset quality by filtering outliers and noisy documents. |
Improving the Calibration of Confidence Scores in Text Generation Using the Output Distribution’s Characteristics (2025.acl-short)
Copied to clipboard
| Challenge: | Existing methods for estimating confidence in text generation do not account for many valid answers in generation tasks. |
| Approach: | They propose task-agnostic confidence metrics that rely solely on model probabilities without the need for further fine-tuning or heuristics. |
| Outcome: | The proposed models improve the accuracy of BART and Flan-T5 on summarization, translation, and question answering datasets. |
Where Did That Come From? Sentence-Level Error-Tolerant Attribution (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing task definitions exclude unsupported or hallucinated content leaving them unattributed . authors propose a new definition for sentence-level error-tolerant attribution . |
| Approach: | They propose a new definition for sentence-level error-tolerant attribution that extends attribution to include incorrect or hallucinated content. |
| Outcome: | The proposed approach reduces annotation time and facilitates hallucination fixing. |
Stochastic Chameleons: Irrelevant Context Hallucinations Reveal Class-Based (Mis)Generalization in LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that LLMs reproduce training artifacts, exploit spurious correlations, and fail when faced with distribution shifts. |
| Approach: | They examine irrelevant context hallucinations in which models integrate misleading contextual cues into their predictions. |
| Outcome: | The proposed model errors are reflected in the model's internal computations, and they are consistent with previous studies. |