7 Papers Accepted at KDD 2026

Congratulations to the following scientists from A*STAR Centre for Frontier AI Research (A*STAR CFAR) on having their papers accepted at the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2026):

Prof Ivor Tsang, Director, A*STAR CFAR
Prof Ong Yew Soon, Chief Artificial Intelligence (AI) Scientist and Advisor
Dr Ooi Chin Chun, Investigator
Dr Pan Yuangang, Early Career Investigator
Dr Du Jiawei, Senior Scientist
Dr Feng Zeyu, Senior Scientist
Dr Lyu Yueming, Senior Scientist
Dr Yin Haiyan, Senior Scientist
Dr Qian Hangwei, Scientist
Dr Yao Yinghua, Scientist
Mr Tang Tianyi, Research Engineer

Held from 9 – 13 August 2026 in Jeju, South Korea, KDD 2026 is a premier Data Science and AI conference featuring a Research track, an Applied Data Science Track, a new Datasets & Benchmarks Track, and a new AI for Sciences Track.

List of accepted papers:

SkillTracer: Structural Failure Attribution and Refinement of Agentic Skills in Long-Horizon Web Tasks
Yuyang Li***, Yiran Dou*, Jie-Jing Shao*, Yueming Lyu, Ivor Tsang, Haiyan Yin

SkillTracer enhances autonomous web agents by replacing monolithic skill macros with editable, programmatically verifiable plan graphs. This explicit topology allows the agent to localise execution breakdowns to specific nodes and perform targeted structural repairs instead of discarding the entire workflow.
Causal Scaffolding for Physical Reasoning: A Benchmark for Causally-Informed Physical World Understanding in VLMs (Datasets and Benchmarks Track)
Tianyi Tang, Zhuoyi Lin, Zeyu Feng, Tianyi Ma*, Yew-Soon Ong, Ivor Tsang, Haiyan Yin

CausalPhys is a benchmark of over 3,000 multimodal problems designed to evaluate physical and causal reasoning in vision-language models. Every scenario is paired with a ground-truth directed acyclic graph that maps out physical relationships to enforce verifiable, step-by-step evaluation. To fix widespread model failures, a Causal Rationale-informed Fine-Tuning (CRFT) strategy is formulated to structure the model's chain-of-thought using these causal graphs to boost accuracy and interpretability.
Geometry-Aware OOD Generalisation for Composite Materials (AI4Sciences Track)
Abhiroop Bhattacharya*, Hangwei Qian, Sylvain G. Cloutier, Ivor Tsang

Our proposed method improves OOD generalisation in composite materials by using geodesic reachability on a JEPA-learned manifold to distinguish supported variations from true distribution shift.
MindAdapter: Few-Shot Parameter-Efficient Residual Calibration of Cross-Subject Brain-to-Visual Decoding Models (AI4Sciences Track)
Jiaxiang Liu***, Jiawei Du, Xupeng Chen, Guoqi Li, Jiang Cai, Simon James Fong, Mingkun Xu

Cross-subject brain-to-visual decoding is fundamentally challenged by severe inter-individual variability that causes subject-specific functional misalignment. We propose MindAdapter, a parameter-efficient few-shot calibration framework that combines frozen coarse functional alignment with lightweight residual adaptation and topology-anchored dual-stream constraints to preserve global representational geometry while enabling fine-grained subject-specific calibration.
Ready to Sim? Inverse Parametric Building Modelling via Universal Anchor Representation for Urban Simulation (AI4Sciences Track)
Jaeyeon Kim***, Po-Yen Lai, Jian Cheng Wong, Chin Chun Ooi, Yew-Soon Ong, Ivor Tsang

Physics-based urban simulation needs editable, topology-valid buildings, but urban data often exist only as images or masks. We introduce Universal Anchor Representation, a normalised hierarchical parameterisation that bounds coordinates and preserves connectivity by construction. Coupled with a domain-specific language, compiler, and VLM-guided refinement pipeline, UAR recovers parametric procedural graphs from top-view images.
From Structure to Function: Preference Alignment for Function-aware Protein Inverse Folding (AI4Sciences Track)
Tamatgar Nilufer***, Soobin Park*, Yinghua Yao, Xixian Chen, Yuangang Pan

Our FPA framework guides protein sequence design models toward generating sequences that better preserve functional integrity, while remaining compatible with existing inverse folding pipelines such as GNN-based ProteinMPNN, transformer-based ESM-IF and all-atom model ADFLIP.
Baldwinian PINNs: Evolutionary Physics-Informed Meta-Learning for Test-Time Adaptation (AI4Sciences Track)
Jian Cheng Wong, Chin Chun Ooi, Abhishek Gupta, Pao-Hsiung Chiu, Joshua Low Shao Zheng, My Ha Dao, Yew-Soon Ong

We frame scientific solving as learning across a distribution of PDE tasks. We present Baldwinian-PINNs, an evolutionary meta-learning approach that improves cross-task generalisation by evolving PINN initialisations that are predisposed to fast test-time adaptation while maintaining physics compliance.

Learn more about KDD 2026.