Papers by Haofei Yu
TRAMS: Training-free Memory Selection for Long-range Language Modeling (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods like Transformer-XL are plagued by ineffective memory selections due to the high number of tokens involved in attention calculation. |
| Approach: | They propose a plug-and-play strategy that selects tokens participating in attention calculation based on one simple metric and ignores the other ones. |
| Outcome: | The proposed strategy keeps tokens with high attention scores and ignores the other ones on word-level and character-level benchmarks without additional training or adding additional parameters. |
SOTOPIA-π: Interactive Learning of Socially Intelligent Language Agents (2024.acl-long)
Copied to clipboard
Ruiyi Wang, Haofei Yu, Wenxin Zhang, Zhengyang Qi, Maarten Sap, Yonatan Bisk, Graham Neubig, Hao Zhu
| Challenge: | Existing studies on building language agents have not addressed this social learning gap. |
| Approach: | They propose an interactive learning method that improves the social intelligence of language agents by using behavior cloning and self-reinforcement based training on filtered social interaction data. |
| Outcome: | The proposed method allows a 7B LLM to reach the social goal completion ability of an expert model (GPT-4-based agent) without the loss of more generic abilities, such as the ability to answer knowledge-based questions. |
TinyScientist: An Interactive, Extensible, and Controllable Framework for Building Research Agents (2025.emnlp-demos)
Copied to clipboard
Haofei Yu, Keyang Xuan, Fenghai Li, Kunlun Zhu, Zijie Lei, Jiaxun Zhang, Ziheng Qi, Kyle Richardson, Jiaxuan You
| Challenge: | Existing research systems often design and use agentic workflows to perform research tasks such as ideation, scientific coding, review writing, and tree-based search. |
| Approach: | They propose an open-source codebase, an interactive web demonstration, and a PyPI Python package to make state-of-the-art auto-research pipelines broadly accessible to every researcher and developer. |
| Outcome: | The proposed framework adapts easily to new tools and supports iterative growth. |
In-Context Learning May Not Elicit Trustworthy Reasoning: A-Not-B Errors in Pretrained Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in artificial intelligence have led to the creation of highly capable large language models (LLMs) that can perform tasks in a human-like manner, but lack infant-level cognitive abilities in certain areas. |
| Approach: | They designed a text-based multi-choice QA scenario similar to the A-Not-B error to test their inhibitory control abilities. |
| Outcome: | The proposed model shows that state-of-the-art LLMs perform well with in-context learning but make errors and show a drop of as many as 83.3% in reasoning tasks when the context changes trivially. |
RFiD: Towards Rational Fusion-in-Decoder for Open-Domain Question Answering (2023.findings-acl)
Copied to clipboard
| Challenge: | Open-domain Question Answering (ODQA) systems rely on spurious features instead of genuine causal relationships to generate answers. |
| Approach: | They propose a model that leverages the encoders of FiD to distinguish between causal relationships and spurious features and guides the decoder to generate answers informed by this discernment. |
| Outcome: | The proposed model improves on two ODQA datasets and shows that it can identify causal relationships and identify spurious features. |
Beyond Facts: Evaluating Intent Hallucination in Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) produce unsatisfactory results when faced with complex queries containing multiple conditions. |
| Approach: | They propose a benchmark for intent hallucination that covers 20,068 problems and an automatic LLM generation evaluation metric for detecting intent hallucinosis. |
| Outcome: | The proposed benchmark covers query-only and retrieval-augmented generation (RAG) setups with varying topics and difficulty. |
ConsistencyChecker: Tree-based Evaluation of LLM Generalization Capabilities (2025.acl-long)
Copied to clipboard
| Challenge: | Traditional self-consistency methods fail to capture subtle semantic errors in multi-step tasks. |
| Approach: | They propose a tree-based evaluation framework that measures LLMs’ ability to preserve semantic consistency during reversible transformations. |
| Outcome: | The proposed framework measures generalization abilities across models from 1.5B to 72B and can be used to benchmark LLMs without constructing new datasets. |
MMoE: Enhancing Multimodal Models with Mixtures of Multimodal Interaction Experts (2024.emnlp-main)
Copied to clipboard
| Challenge: | Multimodal models focus on the correspondence between images and text, but this only covers a subset of real-world interactions. |
| Approach: | They propose an approach to enhance multimodal models by training separate expert models for each type of interaction, such as redundancy present in both modalities, uniqueness in one modality, or synergy that emerges when both . modality is used to capture overlaps in semantic content between images and text, making a strong multi-view redundancies assumption. |
| Outcome: | The proposed approach improves on a sarcasm detection and humor detection task. |
Counting the Bugs in ChatGPT’s Wugs: A Multilingual Investigation into the Morphological Capabilities of a Large Language Model (2023.emnlp-main)
Copied to clipboard
Leonie Weissweiler, Valentin Hofmann, Anjali Kantharuban, Anna Cai, Ritam Dutt, Amey Hengle, Anubha Kabra, Atharva Kulkarni, Abhishek Vijayakumar, Haofei Yu, Hinrich Schuetze, Kemal Oflazer, David Mortensen
| Challenge: | Existing studies on large language models (LLMs) ignore the remarkable ability of humans to generalize and focus only on English. |
| Approach: | They conduct the first rigorous analysis of the morphological capabilities of ChatGPT in four typologically varied languages. |
| Outcome: | The proposed model massively underperforms purpose-built systems, particularly in English. |
Uni-Encoder: A Fast and Accurate Response Selection Paradigm for Generation-Based Dialogue Systems (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to decode a given context-candidate pair are expensive and time-consuming. |
| Approach: | They propose a new paradigm that keeps full attention over each pair while only encoding the context once. |
| Outcome: | The proposed paradigm achieves new state-of-the-art results on four benchmark datasets with high computational efficiency. |
SafeScientist: Enhancing AI Scientist Safety for Risk-Aware Scientific Discovery (2025.emnlp-main)
Copied to clipboard
Kunlun Zhu, Jiaxun Zhang, Ziheng Qi, Nuoxing Shang, Zijia Liu, Peixuan Han, Yue Su, Haofei Yu, Jiaxuan You
| Challenge: | Recent advances in large language model (LLM) agents have significantly accelerated scientific discovery automation, yet raised critical ethical and safety concerns. |
| Approach: | They propose a framework to enhance safety and ethical responsibility in AI-driven scientific exploration. |
| Outcome: | The proposed framework significantly improves safety performance by 35% compared to traditional frameworks. |