12 Papers Accepted at EMNLP 2026

Congratulations to the following scientists from A*STAR Centre for Frontier AI Research (A*STAR CFAR) on having their papers accepted at the Empirical Methods in Natural Language Processing (EMNLP) 2026:

Prof Ivor Tsang, Director, A*STAR CFAR
Prof Ong Yew Soon, Chief Artificial Intelligence (AI) Scientist and Advisor
Dr Li Jing, Scientist
Mr Liu Zhengyuan, Team Lead, Ethical AI, Ethical & Trust AI (ETAI)
Dr Lyu Yueming, Senior Scientist
Dr Nancy Chen, Group Lead, ETAI
Dr Qu Bohao, Scientist
Mr Tang Tianyi, Research Engineer
Dr Wei Chengwei, Scientist
Dr Xu Xun, Senior Scientist
Dr Yin Haiyan, Team Lead, Agentic AI, Agentic Super Intelligence (ASI)
Dr Yu Xingrui, Scientist
Dr Zheng Weihua, Senior Research Engineer

Held from 24 – 29 October 2026 in Budapest, Hungary, EMNLP 2026 brings together researchers and practitioners from academia and industry to present advances in language technologies and data-driven approaches to language understanding and generation.

List of accepted papers:

PrivAgentFlow: Agentic Workflow for Distributed Privacy Control in Web Agents
Tianyi Tang, Tianyi Ma*, Yueming Lyu, Haiyan Yin, Yew-Soon Ong, Ivor Tsang

PrivAgentFlow is a privacy-aware web agent framework that decomposes a user task into a workflow graph and assigns each sub-task node a privacy level that gates its access to sensitive data, enforcing fine-grained data minimisation. A persistent memory of past privacy decisions further guides policy assignment across tasks.
Forgetting Across Modalities: Unlearning in Multimodal Large Language Models
Yan Li*, Xingrui Yu, Zhiqiang Wang, Zhenyi Wang, Ivor Tsang

EASE (Efficient Alignment-aware Selective Erasing) resolves the severe gradient conflicts caused by cross-modal coupling in multimodal large language model unlearning by dynamically aligning opposing task gradients within a two-dimensional subspace. This lightweight optimisation achieves faster convergence and superior efficiency without requiring manually tuned trade-off coefficients.
SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale
Tong Bai, Zhenglin Wan*, Pengfei Zhou, Xingrui Yu, Yang You, Ivor Tsang

SkillDAG models skill library relationships as a typed directed graph that LLM agents can query and evolve during execution, outperforming fixed retrieval pipelines on complex tasks. This structural approach improves candidate ranking as libraries scale and enhances ground-truth recall through online edits.
DySD: Dynamic Speculative Decoding for Evolving LLMs under Rapid Distributional Shifts
Xinglin Liu, Bohao Qu, Chenyu Li, Yongjie Wang, Qing Guo**, Ying Wang

DySD adapts speculative decoding to evolving LLMs through hierarchical feature alignment, distribution-level distillation, and selective online updates, avoiding costly offline retraining. This dynamic approach maintains draft-target alignment under model and input shifts, improving token acceptance while achieving 1.6× to 2.0× inference speedups.
TypoCD: Training-Free Contrastive Decoding for Typographic Defense in Large Vision-Language Models
Xia Sihan***, Jing Li, Junhao Dong, Ivor Tsang, Yew-Soon Ong

TypoCD is a training-free contrastive decoding defense that removes visible attack text and subtracts its induced logit shifts to improve LVLM robustness against typographic attacks. When combined with a simple prompt prefix, TypoCD also substantially reduces jailbreak attack success on safety-oriented typographic attacks.
InfoDensity: Rewarding Information-Dense Traces for Efficient Reasoning
Chengwei Wei, Jung-jae Kim, Longyin Zhang, Shengkai Chen, Nancy F. Chen

Info-Density is a reinforcement learning reward framework that improves LLM reasoning efficiency by encouraging fast uncertainty reduction and low final uncertainty, rather than simply shortening reasoning traces. Experiments demonstrate improved accuracy-efficiency trade-offs across mathematical and general reasoning benchmarks.
RegioZH: Unveiling Comprehension Disparities across Regional Chinese Varieties in LLMs
Weihua Zheng, Wu Kui, Thong T. Doan, Lin Jingxia, AiTi Aw, Nancy F. Chen, Roy Ka-Wei Lee

RegioZH, a dataset designed to evaluate region-sensitive Chinese language understanding, reveals that LLMs’ apparent Chinese proficiency masks substantial disparities in lexical and sociocultural understanding across Chinese-speaking communities.
CultureConverse: A Multilingual Multi-turn Simulation Harness for Culturally Grounded Assistance in East and Southeast Asia
Bryan Chen Zhengyu Tan***, Weihua Zheng, Thong T. Doan, Bich Ngoc Doan, Jia Wang Peh, Xiaoyuan Yi, Jing Yao, Xing Xie, Nancy F. Chen, Zhengyuan Liu, JinYeong Bak, Wafi Shamdi, Soo Kai Chie, Liew Yu Siong, Aina Azyyati Binti Mohamad Rezal, Lew Yan Yan Vanessa, Huadan Wu, Dylan Raharja, Nadya Yuki Wangsajaya, Akane Fukushige, Kazushi Kato, Koji Inoue, Tatsuya Kawahara, Jaehyung Seo, Dongjun Kim, Seungyoon Lee, Zi Haur Pang, Rui Yang Tan, Charibeth Ko Cheng, Maria Regina Justina Estuar, Jann Railey Montalan, Pham Minh Duc, Roy Ka-Wei Lee

CultureConverse is a multi-turn simulation harness and 289k-episode dataset for training and evaluating culturally grounded AI assistance across 10 East and Southeast Asian regions.
Small Changes, Big Impact: Demographic Bias in LLM-Based Hiring Through Subtle Sociocultural Markers in Anonymised Resumes
Bryan Chen Zhengyu Tan***, Shaun Khoo, Bich Ngoc Doan, Zhengyuan Liu, Nancy F. Chen, Roy Ka-Wei Lee

We introduce a stress-test framework for hiring fairness in Singapore, generating 4,100 resume variants from 100 neutral resumes across four ethnicities and two genders. Testing 18 LLMs using direct comparison and score-based shortlisting, we find that models infer demographics from job-irrelevant markers and exhibit systematic disparities.
Near-Optimal SFT–RL Budget Allocation for LLM Post-Training
Jingtan Wang***, Arun Verma, Xiaoqiang Lin, Zhengyuan Liu, Nancy F. Chen, Daniela Rus, Bryan Kian Hsiang Low

We study how to allocate fixed annotation budgets between SFT and RL. Rather than seeking a single optimal ratio, we characterise a near-optimal region of allocations. This region is broad, expands with model scale, and transfers reliably from proxy to target models, enabling efficient budget allocation without exhaustive large-scale search.
The Reward Model Selection Crisis in Personalised Alignment
Fady Rezk, Yuangang Pan**, Chuan-Sheng Foo**, Xun Xu, Nancy F. Chen, Henry Gouk, Timothy Hospedales

This work shows that reward-model ranking accuracy poorly predicts whether reward-guided decoding produces personalized responses, motivating policy accuracy and the Pref-LaMP benchmark for direct behavioural evaluation. Experiments reveal that simple in-context learning outperforms all reward-guided methods for models ≥3B parameters, gaining about 3 ROUGE-1 points at 7B scale.
Cluster-Level Attention-Guided Parallel Decoding for Masked Diffusion Language Models
Heqiang Qi, Wei Huang, Mingyuan Bai****, Xiangming Meng

We revisit token-level granularity for MDLMs and observe that reliable predictions often emerge as contiguous high-confidence spans, suggesting that the unit of parallel commitment can be larger than a single token. We propose CLAD (Cluster-Level Attention-Guided Decoding), a training-free cluster-level decoder for MDLMs that achieves significant improvement in efficiency.

* denotes former CFAR student
** denotes former CFAR researcher
*** denotes current CFAR student
**** denotes current CFAR post doc/visiting scientist
(accurate at time of posting)

>> More on EMNLP 2026.

Prof Ivor Tsang, Director, A*STAR CFAR
Prof Ong Yew Soon, Chief Artificial Intelligence (AI) Scientist and Advisor
Dr Li Jing, Scientist
Mr Liu Zhengyuan, Team Lead, Ethical AI, Ethical & Trust AI (ETAI)
Dr Lyu Yueming, Senior Scientist
Dr Nancy Chen, Group Lead, ETAI
Dr Qu Bohao, Scientist
Mr Tang Tianyi, Research Engineer
Dr Wei Chengwei, Scientist
Dr Xu Xun, Senior Scientist
Dr Yin Haiyan, Team Lead, Agentic AI, Agentic Super Intelligence (ASI)
Dr Yu Xingrui, Scientist
Dr Zheng Weihua, Senior Research Engineer

List of accepted papers:

PrivAgentFlow: Agentic Workflow for Distributed Privacy Control in Web Agents
Tianyi Tang, Tianyi Ma*, Yueming Lyu, Haiyan Yin, Yew-Soon Ong, Ivor Tsang

PrivAgentFlow is a privacy-aware web agent framework that decomposes a user task into a workflow graph and assigns each sub-task node a privacy level that gates its access to sensitive data, enforcing fine-grained data minimisation. A persistent memory of past privacy decisions further guides policy assignment across tasks.
Forgetting Across Modalities: Unlearning in Multimodal Large Language Models
Yan Li*, Xingrui Yu, Zhiqiang Wang, Zhenyi Wang, Ivor Tsang

EASE (Efficient Alignment-aware Selective Erasing) resolves the severe gradient conflicts caused by cross-modal coupling in multimodal large language model unlearning by dynamically aligning opposing task gradients within a two-dimensional subspace. This lightweight optimisation achieves faster convergence and superior efficiency without requiring manually tuned trade-off coefficients.
SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale
Tong Bai, Zhenglin Wan*, Pengfei Zhou, Xingrui Yu, Yang You, Ivor Tsang

SkillDAG models skill library relationships as a typed directed graph that LLM agents can query and evolve during execution, outperforming fixed retrieval pipelines on complex tasks. This structural approach improves candidate ranking as libraries scale and enhances ground-truth recall through online edits.
DySD: Dynamic Speculative Decoding for Evolving LLMs under Rapid Distributional Shifts
Xinglin Liu, Bohao Qu, Chenyu Li, Yongjie Wang, Qing Guo**, Ying Wang

DySD adapts speculative decoding to evolving LLMs through hierarchical feature alignment, distribution-level distillation, and selective online updates, avoiding costly offline retraining. This dynamic approach maintains draft-target alignment under model and input shifts, improving token acceptance while achieving 1.6× to 2.0× inference speedups.
TypoCD: Training-Free Contrastive Decoding for Typographic Defense in Large Vision-Language Models
Xia Sihan***, Jing Li, Junhao Dong, Ivor Tsang, Yew-Soon Ong

TypoCD is a training-free contrastive decoding defense that removes visible attack text and subtracts its induced logit shifts to improve LVLM robustness against typographic attacks. When combined with a simple prompt prefix, TypoCD also substantially reduces jailbreak attack success on safety-oriented typographic attacks.
InfoDensity: Rewarding Information-Dense Traces for Efficient Reasoning
Chengwei Wei, Jung-jae Kim, Longyin Zhang, Shengkai Chen, Nancy F. Chen

Info-Density is a reinforcement learning reward framework that improves LLM reasoning efficiency by encouraging fast uncertainty reduction and low final uncertainty, rather than simply shortening reasoning traces. Experiments demonstrate improved accuracy-efficiency trade-offs across mathematical and general reasoning benchmarks.
RegioZH: Unveiling Comprehension Disparities across Regional Chinese Varieties in LLMs
Weihua Zheng, Wu Kui, Thong T. Doan, Lin Jingxia, AiTi Aw, Nancy F. Chen, Roy Ka-Wei Lee

RegioZH, a dataset designed to evaluate region-sensitive Chinese language understanding, reveals that LLMs’ apparent Chinese proficiency masks substantial disparities in lexical and sociocultural understanding across Chinese-speaking communities.
CultureConverse: A Multilingual Multi-turn Simulation Harness for Culturally Grounded Assistance in East and Southeast Asia
Bryan Chen Zhengyu Tan***, Weihua Zheng, Thong T. Doan, Bich Ngoc Doan, Jia Wang Peh, Xiaoyuan Yi, Jing Yao, Xing Xie, Nancy F. Chen, Zhengyuan Liu, JinYeong Bak, Wafi Shamdi, Soo Kai Chie, Liew Yu Siong, Aina Azyyati Binti Mohamad Rezal, Lew Yan Yan Vanessa, Huadan Wu, Dylan Raharja, Nadya Yuki Wangsajaya, Akane Fukushige, Kazushi Kato, Koji Inoue, Tatsuya Kawahara, Jaehyung Seo, Dongjun Kim, Seungyoon Lee, Zi Haur Pang, Rui Yang Tan, Charibeth Ko Cheng, Maria Regina Justina Estuar, Jann Railey Montalan, Pham Minh Duc, Roy Ka-Wei Lee

CultureConverse is a multi-turn simulation harness and 289k-episode dataset for training and evaluating culturally grounded AI assistance across 10 East and Southeast Asian regions.
Small Changes, Big Impact: Demographic Bias in LLM-Based Hiring Through Subtle Sociocultural Markers in Anonymised Resumes
Bryan Chen Zhengyu Tan***, Shaun Khoo, Bich Ngoc Doan, Zhengyuan Liu, Nancy F. Chen, Roy Ka-Wei Lee

We introduce a stress-test framework for hiring fairness in Singapore, generating 4,100 resume variants from 100 neutral resumes across four ethnicities and two genders. Testing 18 LLMs using direct comparison and score-based shortlisting, we find that models infer demographics from job-irrelevant markers and exhibit systematic disparities.
Near-Optimal SFT–RL Budget Allocation for LLM Post-Training
Jingtan Wang***, Arun Verma, Xiaoqiang Lin, Zhengyuan Liu, Nancy F. Chen, Daniela Rus, Bryan Kian Hsiang Low

We study how to allocate fixed annotation budgets between SFT and RL. Rather than seeking a single optimal ratio, we characterise a near-optimal region of allocations. This region is broad, expands with model scale, and transfers reliably from proxy to target models, enabling efficient budget allocation without exhaustive large-scale search.
The Reward Model Selection Crisis in Personalised Alignment
Fady Rezk, Yuangang Pan**, Chuan-Sheng Foo**, Xun Xu, Nancy F. Chen, Henry Gouk, Timothy Hospedales

This work shows that reward-model ranking accuracy poorly predicts whether reward-guided decoding produces personalized responses, motivating policy accuracy and the Pref-LaMP benchmark for direct behavioural evaluation. Experiments reveal that simple in-context learning outperforms all reward-guided methods for models ≥3B parameters, gaining about 3 ROUGE-1 points at 7B scale.
Cluster-Level Attention-Guided Parallel Decoding for Masked Diffusion Language Models
Heqiang Qi, Wei Huang, Mingyuan Bai****, Xiangming Meng

We revisit token-level granularity for MDLMs and observe that reliable predictions often emerge as contiguous high-confidence spans, suggesting that the unit of parallel commitment can be larger than a single token. We propose CLAD (Cluster-Level Attention-Guided Decoding), a training-free cluster-level decoder for MDLMs that achieves significant improvement in efficiency.

* denotes former CFAR student
** denotes former CFAR researcher
*** denotes current CFAR student
**** denotes current CFAR post doc/visiting scientist
(accurate at time of posting)

>> More on EMNLP 2026.