PromptEng'25 - Workshop on Prompt Engineering for Pre-Trained Language Models

The recent achievements and availability of Large Language Models has paved the road to a new range of applications and use-cases. Pre-trained language models are now being involved at-scale in many fields where they were until now absent from. More specifically, the progress made by causal generative models has opened the door to using them through textual instructions aka. prompts. Unfortunately, the performances of these prompts are highly dependent on the exact phrasing used and therefore practitioners need to adopt fail-retry strategies.

In a nutshell, PromptEng provides the research community with a forum to discuss, exchange and design advanced prompting techniques for LLM applications.

This second international workshop on prompt engineering aims at gathering practitioners (both from Academia and Industry) to exchange good practices, optimizations, results and novel paradigms about the designing of efficient prompts to make use of LLMs.

Undoubtedly, the recent Large Language Models (LLMs) are becoming more and more omnipotent in many tasks. Different sub-fields from the Semantic Web such as Knowledge Graph construction, knowledge verbalization, Web pages summarization have considerably benefited from such a prompting mechanism. The ability to query and interact with them using prompts is crucial to generate high-quality output in the desired format. While existing contributions have been made towards prompt engineering, several difficulties and challenges remain to gain a better understanding of how those LLMs respond to different prompts. Typically, the way instructions are conveyed in prompts can lead to either distinct or similar output from the models.

Moreover, some instructions are better respected while others are simply ignored for some tasks. So far, LLM-practitioners have been mainly working on their own, developing and testing bespoke techniques to achieve their goals, re-starting the prompt-design tasks for each new model they have been using. Such an approach often leads to tackle problems which have already been explored by other researchers.

This workshop aims to investigate and analyze these behaviors, through experimental analysis and probing of LLMs, in order to gain insights into the models' sensitivity to different prompts. By uncovering significant findings, the community can greatly benefit in utilizing LLMs more effectively while also preventing the generation of harmful content. Ultimately, this workshop endeavors to compile and index successful and unavailing prompts with respect to both tasks and models.

Topics of interest include, but are not limited to themes related to the techniques of prompt engineering:

• Prompts & Chain-of-Thought Prompts Design
• Theoretical and Experimental Analysis of Prompting
• Prompts Transferability
• Specific prompt techniques for Web crawling
• Ontology generation combining LLM and Web data
• Semantic and Syntactic comparison of prompt performances
• Structured Prediction with Prompts
• Prompt Retrieval and Generation
• Visualization with Prompt Techniques

We envision five types of submissions covering the entire workshop topics spectrum:

1. Research Papers (max 10 pages), presenting novel scientific research addressing topics of the workshop.
2. Position & Demo papers (max 5 pages), encouraging papers describing significant work in progress, late breaking results or ideas of the domain, as well as functional systems relevant to the community.
3. Industry & Use Case Presentations (max 5 pages), in which industry experts can present and discuss practical solutions, use case prototypes, best practices, etc. at any stage of implementation.
4. Expression of Interest (max 2 pages), presenting a research topic, a work in progress, practical applications or needs, etc.
5. Technical prompting technique (max 2 pages), describing practically a prompt together with a minimal working example and an associated use-case motivating it.

In order to ease the reviewing process, authors may add the track they are submitting to directly in their titles, for instance: "Article Title [Industry]".

Submissions must be in double-column format, and must adhere to the ACM template and format (also available in Overleaf). The recommended setting for LaTeX is: \documentclass[sigconf, anonymous, review]{acmart}. The PDF files must have all non-standard fonts embedded. Workshop submissions must be self-contained and in English. Note: The review process is single-blind, no need for authors to submit anonymous articles.

All papers should be submitted to https://easychair.org/conferences/?conf=prompteng2025.

• Submission: January 13^th 15^th, 2025
• Notification: January 24^th, 2025
• Camera-ready: February 19^th, 2025
• Presentation: April 29^th, 2025

Attending the workshop

PromptEng will take place on 29^th of April 2025 afternoon from 1:30pm to 5:00pm. All hours are local Sydney time.
(In the meantime, please, don't forget to register and attend... ☺)

Keynote #1

Title: Release the Powers of Prompt Tuning: Cross-Modality Prompt Transfer
Dr. Zhen Fang, Lecturer/Assistant Professor, University of Technology Sydney, Australia
https://fang-zhen.github.io/index.html

Abstract: Prompt tuning has redefined the paradigm of adapting pre-trained transformers, achieving parameter-efficient adaptation while preserving competitive performance—a breakthrough particularly impactful for deploying large models in resource-constrained scenarios. However, emerging research uncovers a deeper capability: these learned prompts exhibit cross-task transferability. This discovery catalyses prompt transfer, where prompts pretrained on data-rich source tasks are reused on target tasks within the same modality, effectively mitigating prompt tuning's sensitivity to data scarcity and accelerating adaptation efficiency. Yet, a critical limitation persists: conventional prompt transfer assumes intra-modality transfer, failing when target tasks reside in data-scarce modalities (e.g., medical imaging or satellite visuals) that lack sufficient data to train transferrable prompts.
In this talk, we unveil a pioneering framework: cross-modality prompt transfer, where source prompts can be trained on a different but data-rich modality. We focus on demonstrating its feasibility and efficiency, and present novel criteria for selecting optimal source prompts when multiple candidate tasks are available.

Bio: Dr. Zhen Fang is a lecturer at the University of Technology Sydney, Australia. He received his Ph.D. from the University of Technology Sydney in 2021. His research focuses on the algorithmic and theoretical foundations of transfer learning and out-of-distribution learning. He has served as an Area Chair for NeurIPS and ACM MM. He is the recipient of the Australasian AI Emerging Researcher Award, the NeurIPS 2022 Outstanding Reviewer Award, and the Australian Research Council Discovery Early Career Researcher Award (DECRA 2025). His first author work on out-of-distribution learning received the Outstanding Paper Award at NeurIPS 2022.

Keynote #2

Title: Agent Learning from Interactions: Unifying Grounding and Large Language Models for Generalizable Agency
Dr. Wei Peng, Principal Research Fellow, RMIT University, Australia
https://www.rmit.edu.au/profiles/p/wei-peng3

Abstract: The pursuit of generalizable agency, where agents can adapt to unseen tasks by grounding intelligence in environmental and social contexts, has evolved through distinct eras. Early pioneers in embodied AI and situated agency emphasized that intelligence emerges from real-world interaction. In contrast, modern advancements in large language models (LLMs) have shifted the focus toward data-driven pretraining and instruction fine-tuning. Recently, there has been a resurgence of interest in grounding and interaction as key components of intelligence. This talk traces both foundational paradigms and recent breakthroughs in agent learning through interaction. We revisit landmark literature while highlighting emerging research on trajectory learning in LLM-based agents and, at a broader scale, how LLM-powered multi-agent systems handle collective decision-making. By synthesizing historical insights with state-of-the-art techniques, this talk outlines a potential path toward generalizable AI, one that unifies interaction-driven grounding with linguistic adaptability of language models for robust and generalizable agency.

Bio: Dr. Wei Peng is an AI expert with 15+ years of experience in research, innovation, and commercialization across both industry and academia. He received a PhD from The University of Sydney in 2007, specializing in cognitive agents. Currently, as a Principal Research Fellow at School of Engineering, RMIT University, he leads AI research and the translation of research into real-world solutions. His research focuses on Large Language Models, Knowledge Representation, Cognitive Agents, and Embodied AI, along with their applications. Before RMIT, Wei held principal research leadership roles in the technology sector, leading teams focused on generative AI and enterprise AI solutions. His previous experience includes serving as a Senior Lecturer at La Trobe University and holding senior technical roles at Telstra, Lenovo, and AUSTRAC, where he applied AI across various domains. He has published over 70 peer-reviewed papers, holds multiple patents, and actively contributes to top AI conferences such as ACL, AAAI, EMNLP, NeurIPS, ICLR.

Accepted Articles

• Engineering Prompts for Spatial Questions by Nicole Schneider, Nandini Ramachandran, Kent O'Sullivan and Hanan Samet
• The Iterative Proof-Driven Development LLM Prompt by Aneesha Bakharia
• EdgePrompt: Engineering Guardrail Techniques for Offline LLMs in K-12 Educational Settings by Riza Alaudin Syah, Christoforus Yoga Haryanto, Emily Lomempow, Krishna Malik and Irvan Putra
• A Concept for Integrating an LLM-Based Natural Language Interface for Visualizations Grammars by Adrian Jobst, Daniel Atzberger, Mariia Tytarenko, Willy Scheibel, Jürgen Döllner and Tobias Schreck
• Empirical Evaluation of Prompting Strategies for Fact Verification Tasks by Mohna Chakraborty, Adithya Kulkarni and Qi Li
• LLM Shots: Best Fired at System or User Prompts? by Umut Halil, Jin Huang, Damien Graux and Jeff Z. Pan
• Analyzing the Sensitivity of Prompt Engineering Techniques in Natural Language Interfaces for 2.5D Software Visualization by Daniel Atzberger, Adrian Jobst, Mariia Tytarenko, Willy Scheibel, Jürgen Döllner and Tobias Schreck
• Leveraging Prompt Engineering with Lightweight Large Language Model to Label and Extract Clinical Information from Radiology Report by Chayan Mondal, Duc-Son Pham, Ashu Gupta, Tele Tan and Tom Gedeon
• From Tables to Triples: A Prompt Engineering Approach by Maria Angela Pellegrino and Gabriele Tuozzo

Agenda

Organisers

• Damien Graux (Huawei Ltd., UK) is a principal research scientist at the Huawei Research Center. He has been contributing to research efforts in Semantic Web technologies: focusing on query evaluation and designing complex pipelines for heterogeneous Big Data. Prior, he had research positions at Inria (France), Trinity College Dublin (Ireland) and Fraunhofer IAIS (Germany).
• Sebastien Montella (Huawei Ltd., UK) is a research scientist at the Huawei Edinburgh Research Center. During his Ph.D., he specialized in Natural Language Generation and Knowledge Graph Embeddings research areas. Additionally, he has a keen interest in statistical learning, geometric deep learning, natural language processing, and computer vision. In the past, Sebastien as co-organized the 18th Workshop on Spoken Dialogue Systems for PhDs, PostDocs & New Researchers (YRRSDS) in Edinburgh, Scotland (2022).
• Hajira Jabeen (UniKlinik Cologne, Germany) leads the 'AI in Research Data Management' team at the Institute for Biomedical Informatics. Her team leverages artificial intelligence and large language models (LLMs) to enhance research data management practices, particularly in the biomedical field. They focus on developing scalable, AI-driven tools and workflows that improve data organization, integration, and analysis, driving innovative, data-centric solutions. Hajira has a diverse background in research and teaching, with prior affiliations at the University of Bonn, the University of Cologne, and ITU Copenhagen. She has also organized numerous workshops and conferences in data science and informatics.
• Claire Gardent (CNRS/LORIA, France) is a senior research scientist at the French National Center for Scientific Research (CNRS), based at the LORIA Computer Science research unit in Nancy, France. She works in the field of Natural Language Processing with a particular interest for Natural Language Generation. In 2017, she launched the WebNLG challenge, a shared task where the goal is to generate text from Knowledge Base fragments. She has proposed neural models for simplification and summarisation; for the generation of long form documents such as multi-document summaries and Wikipedia articles; for multilingual generation from Abstract Meaning Representations and for response generation in dialog. She currently heads the AI XNLG Chair on multi-lingual, multi-source NLG and the CNRS LIFT Research Network on Computational, Formal and Field Linguistics. In 2022, she was awarded the CNRS Silver Medal and was selected as ACL (Association of Computational Linguistics) Fellow.
• Dr Jeff Z. Pan is Professor of Knowledge Computing at the School of Informatics at The University of Edinburgh. He is a Chair of the Knowledge Graphs group at the Alan Turing Institute. His recent research explores interactions between explicit and parametric knowledge, with influential survey papers on knowledge graph reasoning and embeddings. Recently, he led a global collaboration on a visionary paper examining large language models and knowledge graphs.

Program Committee

Important Dates

• Submission (EasyChair): January 15^th, 2025
• Notification: January 24^th, 2025
• Camera-ready: February 19^th, 2025
• Presentation: April 29^th, 2025