Papers by Zhixue Zhao
On the Impact of Temporal Concept Drift on Model Explanations (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Explanation faithfulness of model predictions is typically evaluated on held-out data from the same temporal distribution as the training data. |
| Approach: | They examine the impact of temporal variation on model explanations extracted by eight feature attribution methods and three select-then-predict models across six text classification tasks. |
| Outcome: | The proposed method shows the most robust faithfulness scores across datasets and in asynchronous settings. |
Seeing All Sides: Multi-Perspective In-Context Learning for Subjective NLP (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern language models excel at factual reasoning but struggle with value diversity, authors say . task-sensitive tasks such as hate speech expose this limitation . human disagreement captures the diversity of plausible human perspectives, authors argue . |
| Approach: | They evaluate four large language models with human disagreements on five datasets . they find multi-perspective in-context learning outperforms standard prompting . |
| Outcome: | The proposed approach outperforms standard prompting on English labels while disaggregated soft predictions better align with human judgments in Arabic and Italian datasets. |
Revisiting Pruning vs Quantization for Small Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Compressing Small Language Models (SLMs) is particularly suited for resource-constrained devices, but their compression dynamics remain underexplored compared to Large Language Model (LLMs). |
| Approach: | They evaluated post-training pruning and quantization methods across six SLMs from 0.5 to 3.8B, seven languages, and seven downstream tasks. |
| Outcome: | The proposed methods outperform pruning and quantization on six SLMs from 0.5 to 3.8B, seven languages, and seven downstream tasks. |
Explainable Hallucination through Natural Language Inference Mapping (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) often generate hallucinated content, making it crucial to identify and quantify inconsistencies in their outputs. |
| Approach: | They propose a framework that maps entailment and contradiction relations between inputs and outputs using a natural language inference model. |
| Outcome: | The proposed framework outperforms state-of-the-art methods by five percentage points while providing clear, interpretable explanations. |
Analysing Chain of Thought Dynamics: Active Guidance or Unfaithful Post-hoc Rationalisation? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent work has demonstrated that using chain of thought (CoT) on soft-reasoning tasks can yield limited or even negative performance gains. |
| Approach: | They investigate how chain of thought (CoT) is used in soft-reasoning tasks across instruction-tuned, reasoning and reasoning-distilled models. |
| Outcome: | The proposed model can steer predictions without faithfully reflecting reasoning, indicating a disconnect between CoT influence and faithfulness. |
How to Make LLMs Forget: On Reversing In-Context Knowledge Edits (2025.naacl-long)
Copied to clipboard
| Challenge: | In-context knowledge editing (IKE) is an efficient and efficient knowledge editing method (Zheng et al., 2022b; Gangadhar and Stratos, 2024) it can be misused to manipulate responses opaquely, e.g., insert misinformation or offensive content. |
| Approach: | They propose to detect and reverse IKE-edits using only the top-10 output probabilities of the next token, even in a black-box setting. |
| Outcome: | The proposed method can be detected with high accuracy even in a black-box setting, achieving over 80% accuracy in recovering original, unedited outputs across multiple LLMs. |
KidsArtBench: Multi-Dimensional Children’s Art Evaluation with Attribute-Aware MLLMs (2026.eacl-long)
Copied to clipboard
| Challenge: | Multimodal Large Language Models (MLLMs) show impressive capabilities across visual–language tasks, but their capacity to evaluate artistic expression remains limited. |
| Approach: | They propose an attribute-specific multi-LoRA approach where each attribute corresponds to a distinct evaluation dimension in the scoring rubric. |
| Outcome: | The proposed approach increases correlation from 0.468 to 0.653 on Qwen2.5-VL-7B, with the largest gains on perceptual dimensions and narrowed gaps on higher-order attributes. |
Incorporating Attribution Importance for Improving Faithfulness Metrics (2023.acl-long)
Copied to clipboard
| Challenge: | Feature attribution methods (FAs) are popular for providing insights into the model reasoning process of making predictions. |
| Approach: | They propose a simple yet effective criterion that randomly masks tokens proportionately to their FA importance. |
| Outcome: | The proposed method is more faithful than hard sufficiency and comprehensiveness metrics. |
Knowledge Image Matters: Improving Knowledge-Based Visual Reasoning with Multi-Image Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Knowledge-based visual reasoning (KB-VR) is a challenging task, as it requires machines not only to understand concepts and relationships of visual scenes, but also to associate them with external world knowledge to perform chain of reasoning on open-world questions. |
| Approach: | They propose a visual knowledge card (VKC) that integrates internal visual knowledge and external world knowledge produced by a knowledge generator into an image. |
| Outcome: | The proposed model achieves new state-of-the-art results compared to previous top-performing models on three popular KB-VR benchmarks. |
Has this Fact been Edited? Detecting Knowledge Edits in Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Knowledge editing methods (KEs) can update language models’ obsolete or inaccurate knowledge learned from pre-training. |
| Approach: | They propose to detect knowledge edits in language models by using four KEs, two large language models and two datasets to classify the knowledge as unedited (based on pre-training) and edited (based upon subsequent editing). |
| Outcome: | The proposed method detects whether an output is based on edited knowledge or first-hand knowledge from pre-training. |
Comparing Explanation Faithfulness between Multilingual and Monolingual Fine-tuned Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Previous studies have investigated how different factors affect faithfulness of model explanations . |
| Approach: | They find that the larger the multilingual model, the less faithful FAs are compared to its counterpart monolingual models. |
| Outcome: | The results show that the larger the multilingual model, the less faithful the FAs are compared to its counterpart monolingual models. |
It’s All About In-Context Learning! Teaching Extremely Low-Resource Languages to LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Low-resource languages, especially those written in rare scripts, remain unsupported by large language models due to lack of training data. |
| Approach: | They evaluate 20 under-represented languages across three state-of-the-art multilingual LLMs and compare their methods to parameter-efficient fine-tuning. |
| Outcome: | The proposed methods compare with parameter-efficient fine-tuning (PEFT) on low-resource languages. |
Label Set Optimization via Activation Distribution Kurtosis for Zero-Shot Classification with Generative Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | In-context learning (ICL) performance is highly sensitive to prompt design, yet the impact of class label options (e.g. lexicon or order) in zero-shot classification remains underexplored. |
| Approach: | They propose a post-hoc method for selecting optimal label sets in zero-shot ICL with large language models. |
| Outcome: | The proposed method consistently achieves performance gains of 0.54 to 0.76 compared to the conventional method. |
On the Limitations of Language-targeted Pruning: Investigating the Calibration Language Impact in Multilingual LLM Pruning (2026.tacl-1)
Copied to clipboard
| Challenge: | Recent advances in large language model pruning have shown high predictive performance in post-training settings. |
| Approach: | They conduct an empirical study on the performance and internal representation changes associated with pruning multilingual models for monolingual applications. |
| Outcome: | The proposed pruning methods retain perplexity and yield high signal-to-noise ratios, but not consistently improve downstream tasks. |
Minimal, Local, and Robust: Embedding-Only Edits for Implicit Bias in T2I Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | EmbEdit is a text-to-image editing method that only fine-tunes the word token embedding (WTE) of the target object. |
| Approach: | They propose a method to edit implicit assumptions and priors in text-to-image models without affecting unrelated objects or degrading overall performance. |
| Outcome: | The proposed method outperforms previous methods in various models, tasks, and editing scenarios. |