Papers by Runzhe Zhan
Yu Sheng: Human-in-Loop Classical Chinese Poetry Generation System (2023.eacl-demo)
Copied to clipboard
| Challenge: | Existing systems for poetry generation are not flexible in polishing and customization. |
| Approach: | They propose a web-based poetry generation system that provides customization options for users with different backgrounds to engage in the process of poetry composition. |
| Outcome: | The proposed system can generate and polish classical Chinese poetry compared to other vanilla models. |
Prefix Text as a Yarn: Eliciting Non-English Alignment in Foundation Language Model (2024.findings-acl)
Copied to clipboard
| Challenge: | supervised fine-tuning (SFT) has been a straightforward approach for tailoring the output of foundation large language models (LLMs) to specific preferences. |
| Approach: | They propose a training-free alignment method that uses minimal prior tokens to bridge the foundation LLM and the SFT LLM. |
| Outcome: | The proposed method achieves comparable performance without training on machine translation and part-of-speech tagging across seven languages. |
Difficulty-Aware Machine Translation Evaluation (2021.acl-short)
Copied to clipboard
| Challenge: | Current MT evaluation measures pay the same attention to each sentence component . in real-world examinations, the questions vary in difficulty and weightings . |
| Approach: | They propose a difficulty-aware MT evaluation metric that takes translation difficulty into account . they propose to use this metric to evaluate machine translation (MT) results . |
| Outcome: | The proposed method outperforms most MT evaluation metrics in terms of human correlation. |
Let’s Focus on Neuron: Neuron-Level Supervised Fine-tuning for Large Language Model (2025.coling-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are composed of neurons that exhibit diverse behaviors and roles. |
| Approach: | They propose a novel approach that refines the granularity of parameter training down to the individual neuron, enabling a more parameter-efficient fine-tuning model. |
| Outcome: | The proposed approach exceeds the performance of full-parameter fine-tuning and PEFT and provides insights into the analysis of neurons. |
Intrinsic Model Weaknesses: How Priming Attacks Unveil Vulnerabilities in Large Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have significant impact on various industries and societal functions due to advanced instruction-following capabilities. |
| Approach: | They developed and tested novel attack strategies on popular LLMs to expose their vulnerabilities in generating harmful content. |
| Outcome: | The proposed attacks achieved an ASR of 100% on open-source models, including Meta’s Llama-3.2, Google’s Gemma-2, Mistral’s Mistral-NeMo, Falcon’s Falcon-mamba, Apple’s DCLM, Microsoft’s Phi3, and Qwen’s Qwend2.5, among others. |
Can LLMs Learn Uncertainty on Their Own? Expressing Uncertainty Effectively in A Self-Training Manner (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit excessive, random, and uninformative uncertainty rendering them unsuitable for decision-making in human-computer interactions. |
| Approach: | They propose an uncertainty-aware instruction tuning method that aligns LLMs’ perception with the probabilistic uncertainty of the generation. |
| Outcome: | The proposed method improves LLMs' performance by 45.2%, with reasonably good out-of-domain generalization capabilities. |
Who Wrote This? The Key to Zero-Shot LLM-Generated Text Detection Is GECScore (2025.coling-main)
Copied to clipboard
| Challenge: | Existing methods for detecting LLM-generated text require no training data. |
| Approach: | They propose a black-box zero-shot detection approach that calculates the Grammar Error Correction Score for a given text to differentiate between human-written and LLM-generated texts. |
| Outcome: | The proposed method outperforms current state-of-the-art zero-shot and supervised methods, achieving an average AUROC of 98.62% across XSum and Writing Prompts datasets. |
VisAidMath: Benchmarking Visual-Aided Mathematical Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing Large Multi-modal Models lack a robust visual processing capability that is often masked by evaluation metrics that prioritize final-answer accuracy. |
| Approach: | They propose a three-layer evaluation framework that scrutinizes the generation of valid visual aids and the soundness of subsequent reasoning steps. |
| Outcome: | The proposed framework examines the generation of valid visual aids and the soundness of subsequent reasoning steps on state-of-the-art models. |
TransGEC: Improving Grammatical Error Correction with Translationese (2023.findings-acl)
Copied to clipboard
| Challenge: | Experimental results show that data augmentation improves accuracy over strong baselines. |
| Approach: | They propose to use translationese as input for GEC data augmentation to overcome stylistic discrepancies . they propose to obtain human-translated texts with a more similar style to non-native texts . |
| Outcome: | The proposed method improves correction accuracy over strong baselines on four GEC benchmarks. |
Path Drift in Large Reasoning Models: How First-Person Commitments Override Safety (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on prompt injection and jailbreak attacks primarily target the surface structure of input prompts. |
| Approach: | They propose a three-stage approach to mitigate the risk of Long-CoT reasoning drift . they propose 'path-level defense' strategy that incorporates role attribution correction and metacognitive reflection . |
| Outcome: | The proposed framework reduces refusal rates and ethical evaporation, while ethical escalation and layered disclaimers progressively steer models toward unsafe completions. |
Revisiting Commonsense Reasoning in Machine Translation: Training, Evaluation and Challenge (2023.acl-long)
Copied to clipboard
| Challenge: | CR is the ability to understand and navigate the world using basic knowledge and understanding shared by most people. |
| Approach: | They propose to incorporate pretrained knowledge into NMT models and use them as robust testbeds for investigating CR in NMT. |
| Outcome: | The proposed method improves the training of NMT models with high CR abilities and provides accurate evaluation metrics. |
Rethinking Prompt-based Debiasing in Large Language Model (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing prompt-based methods for debiasing are often superficial and lack a thorough understanding of complex bias concepts. |
| Approach: | They analyze a BBQ and stereoSet benchmarks to examine the assumption that large language models understand biases. |
| Outcome: | The proposed model misclassified 90% of unbiased content as biased despite high accuracy on BBQ dataset . the proposed model may have been flawed in previous attempts to debiase . |
G-IdiomAlign: A Gloss-Pivoted Benchmark for Cross-Lingual Idiom Alignment (2026.acl-long)
Copied to clipboard
| Challenge: | Existing tools for cross-lingual idiom-to-idiom equivalence evaluation are limited . figurative meanings are non-compositional and culturally grounded, making literal mappings unreliable. |
| Approach: | They propose a gloss-pivoted benchmark where each idiom is anchored by an English gloss from Wiktionary. |
| Outcome: | The proposed benchmark is based on a dictionary-anchored English idiom . a bias to literal translation is a dominant failure mode across diverse LLMs, the study shows . |
Test-time Adaptation for Machine Translation Evaluation by Uncertainty Minimization (2023.acl-long)
Copied to clipboard
| Challenge: | evaluators of machine translation systems often use text-based metrics to evaluate performance . however, these metrics lack semantic-level information and exhibit poor correlation with human ratings . authors propose a method to reduce inference bias of neural metrics in out-of-distribution data . |
| Approach: | They propose to reduce inference bias by using uncertainty estimation, test-time adaptation, and inference to reduce model uncertainty. |
| Outcome: | The proposed method reduces model uncertainty and improves correlation performance across models. |