Papers by Ning Ma
Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work has shown that safety behaviors are governed by low-rank structures . Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks . |
| Approach: | They propose a safety alignment system that disentangles safety-relevant directions into monosemantic features and constructs an interpretable safety subspace from SAE directions. |
| Outcome: | Empirically, the proposed model achieves 99.6% safety rates across multiple model families and scales . low-rank Adaptation consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks compared with previous methods . |
Protecting Language Models Against Unauthorized Distillation through Trace Rewriting (2026.acl-long)
Copied to clipboard
| Challenge: | Knowledge distillation is a widely adopted technique for transferring capabilities from LLMs to smaller, more efficient student models. |
| Approach: | They propose methods for modifying teacher-generated reasoning traces to achieve two objectives that deter unauthorized distillation: (1) anti-distillation, or degrading the training usefulness of query responses; and (2) API watermarking, which embeds verifiable signatures in student models. |
| Outcome: | The proposed method achieves strong anti-distillation effect while maintaining or even improving teacher performance. |
MINED: Probing and Updating with Multimodal Time-Sensitive Knowledge for Large Multimodal Models (2026.findings-acl)
Copied to clipboard
Kailin Jiang, Ning Jiang, Yuntao Du, Yuchen Ren, Yuchen Li, Yifan Gao, Jinhe Bi, Yunpu Ma, Bin Li, Lei Liu, Qing Li
| Challenge: | Existing benchmarks for Large Multimodal Models (LMMs) are constrained by static representations, inadequately evaluating their ability to understand time-sensitive knowledge. |
| Approach: | They propose a benchmark containing 2,104 time-sensitive knowledge samples spanning six knowledge types to evaluate temporal awareness along 6 key dimensions and 11 challenging tasks. |
| Outcome: | The proposed benchmark measures temporal awareness along 6 key dimensions and 11 tasks, while most open-source LMMs still lack time understanding ability. |
Mixture of Heterogeneous Grouped Experts for Language Modeling (2026.acl-industry)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) based on Mixture-of-Experts (MoE) enforce uniform expert sizes, creating a rigidity that fails to align computational costs with varying token-level complexity. |
| Approach: | They propose a mixture of heterogeneous grouped experts (MoHGE) that allows for flexible, resource-aware expert combinations. |
| Outcome: | The proposed model matches the performance of existing Mixture-of-Experts architectures while maintaining balanced GPU utilization. |
How Private are Language Models in Abstractive Summarization? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Effective protection of private information is essential for knowledge dissemination in sensitive domains such as medical and legal. |
| Approach: | They perform a comprehensive study of privacy risks in LM-based summarization using closed- and four-weight models of different sizes and families. |
| Outcome: | The proposed models show that they leak personally identifiable information in their summaries, compared to human-generated summary summators, which show significantly higher privacy protection levels. |
Hierarchy-Aware Global Model for Hierarchical Text Classification (2020.acl-main)
Copied to clipboard
| Challenge: | Existing methods for hierarchical text classification are limited and lack holistic structural information. |
| Approach: | They propose a hierarchy-aware global model with two variants that learn hierarchy-based label embeddings through an encoder and conduct inductive fusion of label-alike text features. |
| Outcome: | The proposed model improves on three benchmark datasets. |
Hybrid Transducer and Attention based Encoder-Decoder Modeling for Speech-to-Text Tasks (2023.acl-long)
Copied to clipboard
| Challenge: | Neural based end-to-end frameworks have achieved remarkable success in speech-totext tasks, such as automatic speech recognition (ASR) and speech- totext translation (ST). |
| Approach: | They propose to combine Transducer and Attention based Encoder-Decoder (TAED) for speech-to-text tasks and leverage AED's strength in non-monotonic sequence to sequence learning while retaining Transducers streaming property. |
| Outcome: | The proposed model outperforms Transducer and Attention based Encoder-Decoder (TAED) on the MuST-C dataset and shows that it is not bound by any specific language model. |
SNuC: The Sheffield Numbers Spoken Language Corpus (2022.lrec-1)
Copied to clipboard
| Challenge: | SNuC is the first published corpus of spoken alphanumeric identifiers . it contains recordings and transcriptions of over 50 native British English speakers . |
| Approach: | They present a corpus of spoken alphanumeric identifiers of the sort typically used as serial and part numbers in the manufacturing sector. |
| Outcome: | The proposed corpus can be used to improve spoken alphanumeric identifier recognition. |
CODIS: Benchmarking Context-dependent Visual Comprehension for Multimodal Large Language Models (2024.acl-long)
Copied to clipboard
Fuwen Luo, Chi Chen, Zihao Wan, Zhaolu Kang, Qidong Yan, Yingjie Li, Xiaolong Wang, Siyu Wang, Ziyue Wang, Xiaoyue Mi, Peng Li, Ning Ma, Maosong Sun, Yang Liu
| Challenge: | Multimodal large language models have demonstrated promising results in a variety of tasks that combine vision and language. |
| Approach: | They propose a benchmark to assess the ability of models to use contextual information in free-form text to enhance visual comprehension. |
| Outcome: | The proposed model fails to extract and utilize contextual information to improve understanding of images. |
From Behavioral Performance to Internal Competence: Interpreting Vision-Language Models with VLM-Lens (2025.emnlp-demos)
Copied to clipboard
Hala Sheta, Eric Haoran Huang, Shuyu Wu, Ilia Alenabi, Jiajun Hong, Ryker Lin, Ruoxi Ning, Daniel Wei, Jialin Yang, Jiawei Zhou, Ziqiao Ma, Freda Shi
| Challenge: | Existing vision-language models are based on exactmatch based accuracy and its derivations to evaluate performance. |
| Approach: | They propose a toolkit that supports systematic benchmarking, analysis, and interpretation of vision-language models by extracting intermediate outputs from any layer during the forward pass of open-source VLMs. |
| Outcome: | The proposed toolkit supports 16 state-of-the-art base VLMs and their over 30 variants and is extensible to accommodate new models without changing the core logic. |
PATCH: Mitigating PII Leakage in Language Models with Privacy-Aware Targeted Circuit PatcHing (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing defense mechanisms to mitigate PII leakage are limited by existing defenses . a new approach, PATCH, identifies and edits PI I circuits to reduce leakage . |
| Approach: | They propose to use PATCH: Privacy-Aware Targeted Circuit Patching to identify PII leakage circuits in language models to reduce leakage. |
| Outcome: | The proposed approach reduces leakage by up to 65% and can reduce residual leakage to as low as 0.01%. |