Papers by Bryan Li
Enabling Self-Improving Agents to Learn at Test Time With Human-In-The-Loop Guidance (2025.emnlp-industry)
Copied to clipboard
Yufei He, Ruoyu Li, Alex Chen, Yue Liu, Yulin Chen, Yuan Sui, Cheng Chen, Yi Zhu, Luca Luo, Frank Yang, Bryan Hooi
| Challenge: | Existing large language model (LLM) agents are unable to adapt to changing domain knowledge and rules. |
| Approach: | They propose an LLM agent framework that continuously learns updated domain knowledge at test time. |
| Outcome: | The proposed agent improves on a customer due diligence name screening task on . the agent learns updated domain knowledge at test time. |
Scaling Laws for Code: Every Programming Language Matters (2026.findings-acl)
Copied to clipboard
Jian Yang, Shuyue Guo, Linzheng Chai, Wei Zhang, Aishan Liu, Chuan Hao, Zhoujun Li, Xin Zhao, Xianglong Liu, Weifeng Lv, Bryan Dai
| Challenge: | Existing studies focus on language-agnostic settings, neglecting the inherently multilingual nature of modern software development. |
| Approach: | They propose a proportion-dependent scaling law that prioritizes high-utility languages . they propose PLs to have varying effects during pre-training that affect model performance . |
| Outcome: | The proposed scaling law is based on 1000+ experiments across multiple languages and models. |
Re-ViLM: Retrieval-Augmented Visual Language Model for Zero and Few-Shot Image Captioning (2023.findings-emnlp)
Copied to clipboard
Zhuolin Yang, Wei Ping, Zihan Liu, Vijay Korthikanti, Weili Nie, De-An Huang, Linxi Fan, Zhiding Yu, Shiyi Lan, Bo Li, Mohammad Shoeybi, Ming-Yu Liu, Yuke Zhu, Bryan Catanzaro, Chaowei Xiao, Anima Anandkumar
| Challenge: | Existing methods for image-to-text generation store all knowledge within parameters, thus requiring computational-expensive fine-tuning. |
| Approach: | They propose a Retrieval-augmented Visual Language Model that stores all the knowledge within parameters and can be used to retrieve it from the external database. |
| Outcome: | The proposed model significantly boosts performance for image-to-text generation tasks with 4x less parameters compared with baseline methods. |
LoopCoder: Scaling Code Intelligence via Looped Language Models (2026.findings-acl)
Copied to clipboard
Jian Yang, Wei Zhang, Shuyue Guo, Yizhi LI, Linzheng Chai, Zhengmao Ye, Shukai Liu, Yuyang Song, Jiajun Wu, Che Liu, Tianyu Zheng, Siwei Wu, Leo L, Xudong Ma, Chuan Hao, Ran Tao, Yan Xing, Jianzhou Wang, Mingjie Tang, Aishan Liu, Zhoujun Li, Xianglong Liu, Weifeng Lv, Bryan Dai
| Challenge: | Large language models have mastered syntax-level code generation, but complex algorithmic reasoning remains a challenge. |
| Approach: | They propose a recurrent inductive bias that aligns with the recursive nature of programming logic. |
| Outcome: | The proposed model achieves comparable performance to standard dense models with more parameters. |
Defense Against Prompt Injection Attack by Leveraging Attack Techniques (2025.acl-long)
Copied to clipboard
| Challenge: | Recent attacks leverage LLMs’ instruction-following abilities and their inabilities to distinguish instructions injected in the data content. |
| Approach: | They invert the intention of prompt injection methods to develop novel defense methods based on previous training-free attack methods by repeating the attack process with the original input instruction rather than the injected instruction. |
| Outcome: | The proposed methods outperform existing defense approaches, achieving state-of-the-art results. |
TopicAttack: An Indirect Prompt Injection Attack via Topic Transition (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent attacks have demonstrated potential, but their abrupt instruction injection often undermines their effectiveness. |
| Approach: | They propose a method that prompts the LLM to generate a fabricated conversational transition prompt that gradually shifts the topic toward the injected instruction. |
| Outcome: | The proposed method achieves state-of-the-art performance with an attack success rate (ASR) over 90% in most cases, even when various defense methods are applied. |
Robustness via Referencing: Defending against Prompt Injection Attacks by Referencing the Executed Instruction (2026.findings-acl)
Copied to clipboard
Yulin Chen, Haoran Li, Yuan Sui, Yue Liu, Yufei He, Xiaoling Bai, Chi Fei, Li Yabo, Haozhe Ma, Yangqiu Song, Bryan Hooi
| Challenge: | Prompt injection attacks manipulate large language models (LLMs) by misleading them to deviate from the original input instructions and execute maliciously injected instructions. |
| Approach: | They propose a prompt injection defense method that suppresses the model's instruction-following tendencies rather than suppressing them. |
| Outcome: | The proposed method outperforms prompt-engineering-based approaches and fine-tuning methods and reduces the ASR to nearly 0% in some scenarios. |
Can Indirect Prompt Injection Attacks Be Detected and Removed? (2025.acl-long)
Copied to clipboard
| Challenge: | Recent studies have developed various detection mechanisms to protect against prompt injection attacks. |
| Approach: | They investigate the feasibility of detecting and removing indirect prompt injection attacks . they use two methods to evaluate their performance and train detection models . |
| Outcome: | The proposed method is based on a benchmark dataset and is available on github . it evaluates the performance of existing models and open-source detection models . |
IndoNLG: Benchmark and Resources for Evaluating Indonesian Natural Language Generation (2021.emnlp-main)
Copied to clipboard
Samuel Cahyawijaya, Genta Indra Winata, Bryan Wilie, Karissa Vincentio, Xiaohong Li, Adhiguna Kuncoro, Sebastian Ruder, Zhi Yuan Lim, Syafri Bahar, Masayu Khodra, Ayu Purwarianti, Pascale Fung
| Challenge: | Lack of publicly available NLG benchmarks for low-resource languages poses a challenge . authors show that IndoBART and IndoGPT achieve competitive performance on all tasks . |
| Approach: | They propose a benchmark to measure natural language generation progress in three low-resource languages of Indonesia . they use a corpus of pretraining datasets to build their models . |
| Outcome: | The proposed benchmark measures progress in Indonesian, Javanese, and Sundanese . the results highlight the importance of pretraining on closely related, localized languages . |
Vulnerability of LLMs to Vertically Aligned Text Manipulations (2025.acl-long)
Copied to clipboard
| Challenge: | Recent research shows that vertical text input significantly degrades the accuracy of large language models (LLMs) in text classification tasks. |
| Approach: | They investigate the impact of vertical text input on the performance of LLMs . they find that chain of thought reasoning does not help LLM recognize vertical input . |
| Outcome: | The proposed model can significantly mislead models, posing a risk of bypassing detection in real-world scenarios involving harmful or sensitive information. |
Towards Practical and Knowledgeable LLMs for a Multilingual World: A Thesis Proposal (2025.naacl-srw)
Copied to clipboard
| Challenge: | a proposed thesis examines the role that multilinguality occupies in the development of practical and knowledgeable LLMs. |
| Approach: | They propose to use multilingual knowledge to improve LLM performance on NLP tasks . they extend the territorial disputes benchmark to retrieval-augmented generation setting . |
| Outcome: | The proposed methods improve LLMs' performance on standard natural language processing tasks by leveraging their existing multilingual knowledge. |
Backdoor-Powered Prompt Injection Attacks Nullify Defense Methods (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have shown that LLMs are vulnerable to prompt injection attacks because of their instruction-following abilities and inability to distinguish the instructions in the data content. |
| Approach: | They propose backdoor-powered prompt injection attacks that trick LLMs into deviating from the original input instruction and executing the attackers’ target instruction. |
| Outcome: | The proposed attacks trick the LLMs into deviating from the input instruction and executing the attackers’ target instruction. |
Safety in Large Reasoning Models: A Survey (2025.findings-emnlp)
Copied to clipboard
Cheng Wang, Yue Liu, Baolong Bi, Duzhen Zhang, Zhong-Zhi Li, Yingwei Ma, Yufei He, Shengju Yu, Xinfeng Li, Junfeng Fang, Jiaheng Zhang, Bryan Hooi
| Challenge: | Large Reasoning Models (LRMs) have a high level of advanced reasoning capabilities, but they are vulnerable and vulnerable. |
| Approach: | This paper presents the first comprehensive survey of Large Reasoning Models . it explores the new safety risks, attacks, and defense strategies specific to LRMs based on reasoning . |
| Outcome: | The proposed study examines the safety and security risks of large reasoning models. |
Scaling is Not All You Need: Clinical-Oriented Reinforcement Learning Makes Parameter-Efficient Clinical Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models are increasingly used in medicine, but expert-level clinical reasoning remains a high-complexity, high-stakes frontier. |
| Approach: | They propose to train clinical reasoning models using a Reasoning-Oriented Data Strategy based on topological synthesis and CoT cold-start. |
| Outcome: | The proposed pipeline outperforms existing models and outperformed the strongest open-source alternatives up to 671B in MedXpertQA. |
Eliciting Better Multilingual Structured Reasoning from LLMs through Code (2024.acl-long)
Copied to clipboard
| Challenge: | xSTREET exposes a gap in base LLM performance between English and non-English reasoning tasks. |
| Approach: | They propose a multilingual structured reasoning and explanation dataset that covers four tasks across six languages and extends the English STREET benchmark to 5 additional diverse languages. |
| Outcome: | The proposed models show improved multilingual performance on scientific commonsense reasoning subtasks and no regression on non-reasoning tasks. |
GEMv2: Multilingual NLG Benchmarking in a Single Line of Code (2022.emnlp-demos)
Copied to clipboard
Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran, Alex Wang, Alexandros Papangelis, Aman Madaan, Angelina Mcmillan-major, Anna Shvets, Ashish Upadhyay, Bernd Bohnet, Bingsheng Yao, Bryan Wilie, Chandra Bhagavatula, Chaobin You, Craig Thomson, Cristina Garbacea, Dakuo Wang, Daniel Deutsch, Deyi Xiong, Di Jin, Dimitra Gkatzia, Dragomir Radev, Elizabeth Clark, Esin Durmus, Faisal Ladhak, Filip Ginter, Genta Indra Winata, Hendrik Strobelt, Hiroaki Hayashi, Jekaterina Novikova, Jenna Kanerva, Jenny Chim, Jiawei Zhou, Jordan Clive, Joshua Maynez, João Sedoc, Juraj Juraska, Kaustubh Dhole, Khyathi Raghavi Chandu, Laura Perez Beltrachini, Leonardo F . R. Ribeiro, Lewis Tunstall, Li Zhang, Mahim Pushkarna, Mathias Creutz, Michael White, Mihir Sanjay Kale, Moussa Kamal Eddine, Nico Daheim, Nishant Subramani, Ondrej Dusek, Paul Pu Liang, Pawan Sasanka Ammanamanchi, Qi Zhu, Ratish Puduppully, Reno Kriz, Rifat Shahriyar, Ronald Cardenas, Saad Mahamood, Salomey Osei, Samuel Cahyawijaya, Sanja Štajner, Sebastien Montella, Shailza Jolly, Simon Mille, Tahmid Hasan, Tianhao Shen, Tosin Adewumi, Vikas Raunak, Vipul Raheja, Vitaly Nikolaev, Vivian Tsai, Yacine Jernite, Ying Xu, Yisi Sang, Yixin Liu, Yufang Hou
| Challenge: | Evaluations in machine learning rarely use the latest metrics, datasets, or human evaluation in favor of remaining compatible with prior work. |
| Approach: | They propose to use the Generation, Evaluation, and Metrics Benchmark to integrate new evaluation methods into existing evaluations. |
| Outcome: | The proposed evaluation infrastructure bridges the gap between the advantages of leaderboards and in-depth and evolving evaluations by allowing model developers to benefit from each other's work. |
IndoNLU: Benchmark and Resources for Evaluating Indonesian Natural Language Understanding (2020.aacl-main)
Copied to clipboard
Bryan Wilie, Karissa Vincentio, Genta Indra Winata, Samuel Cahyawijaya, Xiaohong Li, Zhi Yuan Lim, Sidik Soleman, Rahmad Mahendra, Pascale Fung, Syafri Bahar, Ayu Purwarianti
| Challenge: | Despite the availability of data on Indonesian, progress on this language is slow . available datasets are scattered, with a lack of documentation and minimal community engagement. |
| Approach: | They propose a resource for training, evaluation, and benchmarking on Indonesian natural language understanding tasks. |
| Outcome: | The proposed resource includes 12 tasks ranging from single sentence classification to pair-sentences sequence labeling with different levels of complexity. |
Multilingual Retrieval Augmented Generation for Culturally-Sensitive Tasks: A Benchmark for Cross-lingual Robustness (2025.findings-acl)
Copied to clipboard
Bryan Li, Fiona Luo, Samar Haider, Adwait Agashe, Siyu Li, Runqi Liu, Miranda Muqing Miao, Shriya Ramakrishnan, Yuan Yuan, Chris Callison-Burch
| Challenge: | Existing approaches to retrieval-augmented generated (RAG) can be useful in multilingual settings, but they also introduce biases in the retrieved documents. |
| Approach: | They propose a dataset of territorial disputes paired with retrieved Wikipedia documents in 49 languages to evaluate cross-lingual robustness. |
| Outcome: | The proposed paradigm helps mitigate hallucinations of large language models (LLMs). |
Exploring Content Selection in Summarization of Novel Chapters (2020.acl-main)
Copied to clipboard
| Challenge: | We focus on extractive summarization, which requires the creation of a gold-standard set of extractive summary summaries. |
| Approach: | They propose a new metric for aligning summary sentences with chapter sentences to create gold extracts. |
| Outcome: | The proposed method improves on previous methods and automatic metrics and a crowd-sourced pyramid analysis. |
Beyond ’Aha!’: Toward Systematic Meta-Abilities Alignment in Large Reasoning Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior work has shown that outcome-based reinforcement learning (RL) can incidentally elicit advanced reasoning behaviors such as self-correction, backtracking, and verification. |
| Approach: | They explicitly align large reasoning models with three meta-abilities: deduction, induction, and abduction, using automatically generated, self-verifiable tasks. |
| Outcome: | The proposed model aligns models with deduction, induction, and abduction meta-abilities using automatically generated, self-verifiable tasks. |
PAXQA: Generating Cross-lingual Question Answering Examples at Training Scale (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing question answering systems rely on large, high-quality training data. |
| Approach: | They propose a synthetic data generation method which decomposes cross-lingual QA into two stages . they apply a question generation model to the English side and annotation projection to translate both questions and answers. |
| Outcome: | The proposed method outperforms existing methods on cross-lingual QA datasets. |
From 128K to 4M: Efficient Training of Ultra-Long Context Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Long-context capabilities are essential for document and video understanding, in-contact learning, and inference-time scaling. |
| Approach: | They propose an efficient training recipe for building ultra-long context LLMs from aligned instruct model, pushing the boundaries of context lengths from 128K to 1M, 2M, and 4M tokens. |
| Outcome: | The proposed model extends the context window while maintaining short context capabilities while maintaining the performance of the existing model. |
Shall We Pretrain Autoregressive Language Models with Retrieval? A Comprehensive Study (2023.emnlp-main)
Copied to clipboard
Boxin Wang, Wei Ping, Peng Xu, Lawrence McAfee, Zihan Liu, Mohammad Shoeybi, Yi Dong, Oleksii Kuchaiev, Bo Li, Chaowei Xiao, Anima Anandkumar, Bryan Catanzaro
| Challenge: | a recent study shows that retrieval-augmented LMs can improve text generation quality and accuracy. |
| Approach: | They propose a model that reproduces RETRO parameters while retrieving a text corpus . they find RETRO outperforms GPT on text generation with less repetition . |
| Outcome: | The proposed model outperforms standard retrieval-augmented GPT and retrieval augmented GTP on text generation and accuracy tasks. |
DRS: Deep Question Reformulation With Structured Output (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing models like GPT-3 and Instruct-GPT lack the ability to reformulate unanswerable questions. |
| Approach: | They propose a zero-shot method that combines the strengths of LLMs with a DFS-based algorithm to iteratively explore potential entity combinations and constrain outputs using predefined entities. |
| Outcome: | The proposed method outperforms all baselines, including the GPT-3.5 model, on the unanswerable question reformulation task. |
A Zero-Shot Language Agent for Computer Control with Structured Reflection (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent works require a model to learn from trace examples of a task via supervised learning or few/many-shot prompting. |
| Approach: | They propose a model that iteratively learns from its mistakes via self-reflection and structured thought management. |
| Outcome: | The proposed model outperforms previous models on easy tasks with more efficient reasoning and self-reflection. |
This Land is Your, My Land: Evaluating Geopolitical Bias in Language Models through Territorial Disputes (2024.naacl-long)
Copied to clipboard
| Challenge: | Pretrained large language models may answer differently in different languages . this contrasts with a multilingual human, who would likely answer consistently . |
| Approach: | They propose a dataset of territorial disputes which includes multiple-choice questions in 49 languages . they propose metrics to quantify bias and consistency in responses across different languages based on their data . |
| Outcome: | The proposed model recalls certain knowledge inconsistently when asked in different languages. |
Discovering Language Model Behaviors with Model-Written Evaluations (2023.findings-acl)
Copied to clipboard
Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, Andy Jones, Anna Chen, Benjamin Mann, Brian Israel, Bryan Seethor, Cameron McKinnon, Christopher Olah, Da Yan, Daniela Amodei, Dario Amodei, Dawn Drain, Dustin Li, Eli Tran-Johnson, Guro Khundadze, Jackson Kernion, James Landis, Jamie Kerr, Jared Mueller, Jeeyoon Hyun, Joshua Landau, Kamal Ndousse, Landon Goldberg, Liane Lovitt, Martin Lucas, Michael Sellitto, Miranda Zhang, Neerav Kingsland, Nelson Elhage, Nicholas Joseph, Noemi Mercado, Nova DasSarma, Oliver Rausch, Robin Larson, Sam McCandlish, Scott Johnston, Shauna Kravec, Sheer El Showk, Tamera Lanham, Timothy Telleen-Lawton, Tom Brown, Tom Henighan, Tristan Hume, Yuntao Bai, Zac Hatfield-Dodds, Jack Clark, Samuel R. Bowman, Amanda Askell, Roger Grosse, Danny Hernandez, Deep Ganguli, Evan Hubinger, Nicholas Schiefer, Jared Kaplan
| Challenge: | Prior work creates evaluations with crowdwork or existing data sources, which are not always available. |
| Approach: | They generate evaluations automatically with language models (LMs) using crowdwork or existing data sources to find out how they behave . |
| Outcome: | The results show that large LMs repeat back a dialog user’s preferred answer and express greater desire to pursue concerning goals like resource acquisition and goal preservation. |