Papers by Xinpeng Wang
Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers (2026.findings-acl)
Copied to clipboard
Van Yang, Shouren Wang, Debargha Ganguly, Xinpeng Li, Chaoda Song, Vikash Singh, Vipin Chaudhary, Xiaotian Han
| Challenge: | Explicit /think> tags are used to expose intermediate reasoning and enable hybrid thinking behaviors. |
| Approach: | They propose a training-free prompting format that combines these triggers to achieve intermediate-budget reasoning, outperforming fixed-token and prompt-based baselines in terms of the accuracy–length trade-off. |
| Outcome: | The proposed method outperforms fixed-token and prompt-based prompts in accuracy–length trade-offs while improving Qwen3-8B on AIME from 69.8% to 72.4% and on GPQA from 58.5% to 61.1%. |
“Seeing the Big through the Small”: Can LLMs Approximate Human Judgment Distributions on NLI from a Few Explanations? (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Human label variation arises when multiple human annotators provide different labels for valid reasons. |
| Approach: | They propose to use crowd workers to represent human judgment distributions or expert linguists to provide detailed explanations for their chosen labels. |
| Outcome: | The proposed model can approximate human judgment distributions using a small number of expert labels and explanations. |
“My Answer is C”: First-Token Probabilities Do Not Match Text Answers in Instruction-Tuned Language Models (2024.findings-acl)
Copied to clipboard
Xinpeng Wang, Bolei Ma, Chengzhi Hu, Leon Weber-Genzel, Paul Röttger, Frauke Kreuter, Dirk Hovy, Barbara Plank
| Challenge: | Multiple choice questions are one of the most popular evaluation formats for understanding the capabilities of autoregressive large language models (LLMs). |
| Approach: | They evaluated how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. |
| Outcome: | The proposed evaluation methods are misaligned on all dimensions, reaching mismatch rates over 60%. |
How to Distill your BERT: An Empirical Study on the Impact of Weight Initialisation and Distillation Objectives (2023.acl-short)
Copied to clipboard
| Challenge: | Recent studies show that intermediate layer distillation (ILD) objectives improve model compression, but a comprehensive evaluation of distillation objectives in both task-specific and task-agnostic settings is lacking. |
| Approach: | They propose to use knowledge distillation to improve model compression by transferring knowledge from one model to another. |
| Outcome: | The proposed framework improves on the task of QNLI with lower teacher layers and higher teacher layers. |
CHAE: Fine-Grained Controllable Story Generation with Characters, Actions and Emotions (2022.coling-1)
Copied to clipboard
| Challenge: | Existing studies on story generation focus on coarse-grained control of the story, neglecting the details of the narrative. |
| Approach: | They propose a model for fine-grained control on the story that allows the generation of customized stories with characters, corresponding actions and emotions arbitrarily assigned. |
| Outcome: | The proposed method has strong controllability to generate customized stories according to the fine-grained personalized guidance. |
The Potential and Challenges of Evaluating Attitudes, Opinions, and Values in Large Language Models (2024.findings-emnlp)
Copied to clipboard
Bolei Ma, Xinpeng Wang, Tiancheng Hu, Anna-Carolina Haensch, Michael Hedderich, Barbara Plank, Frauke Kreuter
| Challenge: | Recent advances in Large Language Models have sparked interest in validating human-like cognitive-behavioral traits. |
| Approach: | They examine whether LLM outputs reflect human-like cognitive-behavioral traits . they find that measuring AOVs embedded within LLMs remains opaque . |
| Outcome: | The proposed model can be used to evaluate human-like cognitive-behavioral traits . the proposed model could be used in writing assistants and other applications . |
ToViLaG: Your Visual-Language Generative Model is Also An Evildoer (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent large-scale Visual-Language Generative Models (VLGMs) generate toxic content, e.g., offensive text and pornography images, raising significant ethical risks. |
| Approach: | They propose a bottleneck-based detoxification method to reduce toxicity while maintaining comparable generation quality. |
| Outcome: | The proposed method could reduce toxicity while maintaining comparable generation quality. |
Algorithmic Fidelity of Large Language Models in Generating Synthetic German Public Opinions: A Case Study (2025.acl-long)
Copied to clipboard
Bolei Ma, Berk Yoztyurk, Anna-Carolina Haensch, Xinpeng Wang, Markus Herklotz, Frauke Kreuter, Barbara Plank, Matthias Aßenmacher
| Challenge: | Recent advances in large language models have generated significant interest in their potential for synthetic data generation across various domains. |
| Approach: | They use open-ended survey data from the German Longitudinal Election Studies to prompt different LLMs to generate synthetic public opinions reflective of German subpopulations by incorporating demographic features into the persona prompts. |
| Outcome: | The LLM performs better for supporters of left-leaning parties like The Greens and The Left compared to other parties, and matches the least with the right-party AfD. |
DAMRO: Dive into the Attention Mechanism of LVLM to Reduce Object Hallucination (2024.emnlp-main)
Copied to clipboard
| Challenge: | Despite the success of Large Vision-Language Models, they suffer from hallucination. |
| Approach: | They propose a training-free strategy that "D**ive into" the attention of LVLMs to "R**educe" object hallucination by using classification tokens of ViT. |
| Outcome: | The proposed method reduces the impact of outlier tokens on LVLMs . the proposed method is based on LLaVA-1.5, LLvaVA-NeXT and InstructBLIP . |
Benchmarking and Enabling Efficient Chinese Medical Retrieval via Asymmetric Encoders (2026.acl-long)
Copied to clipboard
| Challenge: | Effective medical text retrieval requires high accuracy and low latency. |
| Approach: | They propose a benchmark for medical text retrieval in Chinese using a symmetric architecture . CARE is a lightweight encoder with an LLM-based encoder for offline document encoding . |
| Outcome: | The proposed benchmark surpasses state-of-the-art symmetric models on CMedTEB . it matches high retrieval quality without increasing latency, and it performs well on a single GPU . |
Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for opinion summarization are deficient in epitomizing extensive reviews and offering opinion summaries from various angles. |
| Approach: | They propose a supervised opinion summarization framework that takes sentiment orientation into account and trains the summarizer to learn from sub-optimal and optimal review subsets. |
| Outcome: | The proposed framework generates pros, cons, and verdict summaries from hundreds of input reviews. |
ACTOR: Active Learning with Annotator-specific Classification Heads to Embrace Human Label Variation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies indicate that learning from individual annotations outperforms learning from aggregated labels, though they require a considerable amount of annotation. |
| Approach: | They propose to use a multi-head model to learn from disagreements in an active learning setting to identify annotators with a single head. |
| Outcome: | The proposed model outperforms a single-head model in terms of uncertainty estimation and prediction while saving 70% of the annotation budget. |
Towards Reverse Engineering of Language Models: A Survey (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Due to the vast amounts of data and computational resources required for model development, protecting the model’s parameters and training data has become an urgent and crucial concern. |
| Approach: | They define "reverse engineering" techniques as attacks on large language models and provide an in-depth analysis of them. |
| Outcome: | The proposed attacks are described as “reverse engineering” techniques on LMs and provide an introduction to existing protective strategies. |