16 Papers Accepted at ICML 2025

The International Conference on Machine Learning (ICML) is a premier conference that brings together experts committed to advancing the field of machine learning – a core area of artificial intelligence.

A total of 16 papers from A*STAR Centre for Frontier AI Research (A*STAR CFAR) have been accepted at ICML 2025. Congratulations to the following scientists on this achievement:

Prof Ivor Tsang, Director, A*STAR CFAR
Prof Ong Yew Soon, Chief Artificial Intelligence (AI) Scientist and Advisor
Dr Li Xiaoli, Senior Principal Scientist
Dr Joey Zhou, Deputy Director and Principal Scientist
Dr Foo Chuan Sheng, Principal Scientist
Dr Atsushi Nitanda, Principal Scientist
Dr David Bossens, Senior Scientist
Dr Guo Qing, Senior Scientist
Dr He Tiantian, Senior Scientist
Dr Emadeldeen Eldele, Scientist
Dr Flint Fan, Scientist
Dr Lyu Yueming, Scientist
Dr Yu Xingrui, Scientist
Dr Zhang Jie, Scientist
Dr Zhang Mengmi, Adjunct Member
Mr Neoh Tzeh Yuan, Research Engineer

List of accepted papers:

1.	Deep Unsupervised Hashing via External Guidance Qihong Song, XitingLiu, Hongyuan Zhu, Joey Tianyi Zhou, Xi Peng, Peng Hu In this work, we propose a novel method called Deep Unsupervised Hashing with External Guidance (DUH-EG), which incorporates external textual knowledge as semantic guidance to enhance discrete representation learning.
2.	Counterfactual Contrastive Learning with Normalising Flows for Robust Treatment Effect Estimation Jiaxuan Zhang, Emadeldeen Eldele, Fuyuan CAO, Yang Wang, Xiaoli Li, Jiye Liang Estimating Individual Treatment Effects (ITE) from observational data is challenging due to covariate shift and counterfactual absence. We propose FCCL, a novel method designed to effectively capture the nuances of potential outcomes under different treatments by (i) generating diffeomorphic counterfactuals that adhere to the data manifold while maintaining high semantic similarity to their factual counterparts, and (ii) mitigating distribution shift via sample-level alignment grounded in our derived generalisation-error bound, which considers factual-counterfactual similarity and category consistency.
3.	NICE: Non-differentiable Evaluation Metric-based Data Selection for Instruction Tuning Jingtan Wang, Xiaoqiang Lin, Rui Qiao, Pang Wei Koh, Chuan-Sheng Foo, Bryan Kian Hsiang Low We introduce a novel Non-differentiable evaluation metric-based InfluenCe Estimation (NICE), which leverages the policy gradient to select the training data that improves performance on a metric of choice, beyond next-token prediction loss that is used to train LLMs. NICE can perform data selection in the absence of labels (ground-truth responses) when the evaluation metrics do not require labels (e.g., a reward model in alignment settings that can output reward scores without supervision from labels).
4.	Fraud-Proof Revenue Division on Subscription Platforms Abheek Ghosh, Tzeh Yuan Neoh, Nicholas Teh, Giannis Tyrovolas We study a model of subscription-based platforms where users pay a fixed fee for unlimited access to content, and creators receive a share of the revenue. We perform axiomatic analysis of existing revenue division rule in literature as well as deriving a new rule with strong theoretical and empirical guarantees.
5.	Diversifying Robot Locomotion Behaviours with Extrinsic Behavioural Curiosity Zhenglin Wan, Xingrui Yu, David Mark Bossens, Yueming Lyu, Qing Guo, Flint Xiaofeng Fan, Yew-Soon Ong, Ivor Tsang* We introduce Quality Diversity Inverse Reinforcement Learning (QD-IRL), a novel framework that integrates quality-diversity optimisation with IRL methods, enabling agents to learn diverse behaviours from limited demonstrations. In addition, we introduce Extrinsic Behavioural Curiosity (EBC) that allows agents to receive additional curiousity rewards from an external critic based on how novel the behaviours are with respect to a large behavioral archive.
6.	Unveiling AI’s Blind Spots: An Oracle for In-Domain, Out-of-Domain, and Adversarial Errors Shuangpeng Han, Mengmi Zhang Our work tackles the challenge of predicting errors of AI models through extensive empirical evaluations using an end-to-end trainable “mentor model”, paving the way for promising research directions in safe and trustworthy AI.
7.	Voronoi-grid-based Pareto Front Learning and its Application to Collaborative Federated Learning Mengmeng Chen, Xiaohu Wu, Qiqi Liu, Tiantian He, Yew-Soon Ong, Yaochu Jin, Qicheng Lao, Han Yu We introduce a novel Pareto-front learning (PFL) framework that decomposes the design space into Voronoi grids and deploys a genetic algorithm (GA) for Voronoi grid partitioning within high-dimensional space. This novel framework has been shown to be effective in solving multiple machine-learning tasks of multi-objective optimisation.
8.	Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal Link Prediction Yadong Sun, Xiaofeng Cao, Ivor Tsang, Heng Tao Shen We propose a unified cross-geometric learning framework for dynamic systems that integrates Euclidean and hyperbolic spaces. This approach aligns embedding spaces with structural properties through fine-grained substructure modelling. It incorporates a temporal state aggregation mechanism and an evolution-driven optimisation objective, enabling comprehensive modelling of both nodal and relational dynamics over time.
9.	EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers Daiheng Gao, Shilin Lu, Wenbo Zhou, Jiaming Chu, Jie Zhang, Mengxi Jia, Bang Zhang, Zhaoxin Fan, Weiming Zhang We introduce the first approach for controllable concept erasure in Flow-matching Transformers (e.g., Flux), using bi-level optimisation and adversarial regularisation to selectively remove sensitive concepts while preserving model capability.
10.	Cowpox: Towards the Immunity of VLM-based Multi-Agent Systems Yutong Wu, Jie Zhang, Yiming Li, Chao Zhang, Qing Guo, Han Qiu, Nils Lukas, Tianwei Zhang We propose COWPOX, a novel immunity mechanism for VLM-based multi-agent systems against contagious jailbreak attacks. A small number of healing agents can propagate cure samples to enable system-wide self-repair.
11.	Propagation of Chaos for Mean-Field Langevin Dynamics and its Application to Model Ensemble *Atsushi Nitanda*, Anzelle Lee, Damian Tan Xing Kai, Mizuki Sakaguchi, Taiji Suzuki We establish an improved approximation and optimisation guarantees for two-layer neural networks trained with the noisy gradient descent. As an application, we propose a model ensemble strategy with theoretical guarantees.
12.	Provable In-Context Vector Arithmetic via Retrieving Task Concepts Dake Bu, Wei Huang, Andi Han, Atsushi Nitanda, Taiji Suzuki, Qingfu Zhang, Hau-San Wong We develop an optimisation theory, showing how nonlinear residual transformers trained via gradient descent on cross-entropy loss perform factual-recall ICL tasks via task vector arithmetic. We prove 0-1 loss convergence and highlight their superior generalisation capabilities, adeptly handling concept recombinations and data shifts.
13.	Defending LVLMs Against Vision Attacks through Partial-Perception Supervision Qi Zhou, Tianlin Li, Dongxia Wang, Yun Lin, Yang Liu, Jin Song Dong, Qing Guo We propose a black-box, training-free method called DPS (Defense through Partial-Perception Supervision) to defend against diverse adversarial attacks. With DPS, the model can adjust its response based on partial image understanding when under attack, while confidently maintaining its original response for clean input.
14.	Improving Zero-Shot Adversarial Robustness in VLMs by Closed-form Alignment of Adversarial Path Simplices Junhao Dong, Piotr Koniusz, Yifei Zhang, Hao Zhu, Weiming Liu, Xinghua Qu, Yew-Soon Ong We propose a closed-form prediction alignment-based adversarial fine-tuning method to improve the robustness of foundation vision-language models across diverse downstream tasks. We further derive an upper bound of prediction alignment by modeling simplices with the vertices of adversarial examples.
15.	Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets Ning Lu, Shengcai Liu, Jiahao Wu, Weiyu Chen, Zhirui Zhang, Yew-Soon Ong, Qi Wang, Ke Tang We propose Safe Delta, a safety-aware framework that can consistantly preserve the safely of LLMs when finetuning them on diverse datasets.
16.	A Mixed-Curvature based Pre-training Paradigm for Multi-Task Vehicle Routing Solver Suyu Liu, Zhiguang Cao, Shanshan Feng, *Yew-Soon Ong* We propose a new pre-training paradigm based on the mixed-curvature space for establishing a neural solver for various kinds of vehicle routing problems. The designed geometric spaces have shown their effectiveness in capturing fine-grained geometric structures in the datasets.

More information on ICML 2025.

News

16 Papers Accepted at ICML 2025

A*STAR celebrates International Women's Day