Papers by Zhaofeng Wu
ABC: Attention with Bounded-memory Control (2022.acl-long)
Copied to clipboard
Hao Peng, Jungo Kasai, Nikolaos Pappas, Dani Yogatama, Zhaofeng Wu, Lingpeng Kong, Roy Schwartz, Noah A. Smith
| Challenge: | Existing approaches to attention with bounded-memory control (ABC) have a quadratic complexity in sequence lengths, making it prohibitive for long sequences. |
| Approach: | They propose a new abstraction that bounds memory size to improve efficiency . they propose bounded-memory control, which connects several efficient attention variants . |
| Outcome: | The proposed approach outperforms existing approaches on language modeling, machine translation, and masked language model finetuning. |
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 . |
Sparkle: Mastering Basic Spatial Capabilities in Vision Language Models Elicits Generalization to Spatial Reasoning (2025.findings-emnlp)
Copied to clipboard
Yihong Tang, Ao Qu, Zhaokai Wang, Dingyi Zhuang, Zhaofeng Wu, Wei Ma, Shenhao Wang, Yunhan Zheng, Zhan Zhao, Jinhua Zhao
| Challenge: | Currently, vision-language models excel in many downstream tasks but struggle with spatial reasoning, which is crucial for navigation and interaction with physical environments. |
| Approach: | They propose a framework that generates synthetic data to provide targeted supervision for VLMs across these basic spatial capabilities. |
| Outcome: | The proposed framework disentangles 2D spatial reasoning into three core components: direction comprehension, distance estimation, and localization. |
Select-Then-Decompose: From Empirical Analysis to Adaptive Selection Strategy for Task Decomposition in Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing task decomposition methods focus on memory, tool usage, and feedback mechanisms, but they often overlook the trade-off between performance and cost. |
| Approach: | They propose a strategy that selects the most suitable decomposition approach based on task characteristics and enhances the reliability of the results through a verification module. |
| Outcome: | The proposed strategy is based on categories of approaches, characteristics of tasks, and configuration of decomposition and execution models. |
Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks (2024.naacl-long)
Copied to clipboard
Zhaofeng Wu, Linlu Qiu, Alexis Ross, Ekin Akyürek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, Yoon Kim
| Challenge: | Recent language models possess impressive performance across a wide range of tasks . however, they often rely on narrow, non-transferable procedures for task-solving . |
| Approach: | They propose to evaluate language models using "counterfactual" task variants that deviate from standard tasks. |
| Outcome: | The proposed framework shows that language models perform better on a wide range of tasks compared to the default conditions. |
Infusing Finetuning with Semantic Dependencies (2021.tacl-1)
Copied to clipboard
| Challenge: | Several diagnostics help to localize the benefits of our approach. |
| Approach: | They apply convolutional graph encoders to integrate semantic parses into task-specific finetuning. |
| Outcome: | The proposed approach yields benefits to natural language understanding (NLU) tasks in the GLUE benchmark. |
Parallel-SFT: Improving Zero-Shot Cross-Programming-Language Transfer for Code RL (2026.findings-acl)
Copied to clipboard
Zhaofeng Wu, Shiqi Wang, Boya Peng, Anuj Kumar Goyal, Melanie Kambadur, Sebastian Ruder, Yoon Kim, Chloe Bi
| Challenge: | Modern language models demonstrate impressive coding capabilities in common programming languages (PLs) but their performance in lower-resource PLs is often limited by training data availability. |
| Approach: | They propose a zero-shot cross-programming-language transfer task for code RL . they propose RL training in a source PL fails to improve performance on other target PLs . |
| Outcome: | The proposed approach improves transferability in Llama-3.1 code generation on parallel-stack model . it also improves performance on other target PLs, compared to single-PL SFT . |
HyperMoE: Towards Better Mixture of Experts via Transferring Among Experts (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods for enhancing performance through increased use of expert knowledge often result in diminishing sparsity during expert selection. |
| Approach: | They propose a framework that integrates the computational processes of MoE with the concept of knowledge transferring in multi-task learning. |
| Outcome: | The proposed framework outperforms existing methods under identical conditions concerning the number of experts. |
Translation or Recitation? Calibrating Evaluation Scores for Machine Translation of Extremely Low-Resource Languages (2026.acl-short)
Copied to clipboard
| Challenge: | Existing studies show that performance across low-resource settings is variable, resulting in a significant barrier for the MT community. |
| Approach: | They propose to use FRED Difficulty Metrics to contextualize reported performance across different language pairs to determine whether breakthroughs reported in other contexts are artifacts of benchmark collection. |
| Outcome: | The proposed metrics explain a significant portion of result variability rather than model capability. |
Reuse Your Rewards: Reward Model Transfer for Zero-Shot Cross-Lingual Alignment (2024.emnlp-main)
Copied to clipboard
| Challenge: | Multilingual human preference data are difficult to obtain at scale, making it challenging to extend this framework to diverse languages. |
| Approach: | They propose a method where a reward model is trained on preference data in one source language and applied to other target languages. |
| Outcome: | The proposed approach is effective under comprehensive evaluation settings, including human evaluation. |
An Adaptive Prompt Generation Framework for Task-oriented Dialogue System (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing black-box large language models (LLMs) have excellent performance in task-oriented dialogue (TOD) tasks, but obtaining suitable prompts for specific tasks is challenging. |
| Approach: | They propose a black-box large language model that generates domain and slot information in the belief state, which serves as prior knowledge for subsequent prompt generation. |
| Outcome: | The proposed framework outperforms existing prompting methods on the MultiWOZ 2.0 dataset. |
Modeling Context With Linear Attention for Scalable Document-Level Translation (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Document-level machine translation models lack quadratic complexity in the sequence length due to their attention layers. |
| Approach: | They evaluate a recent linear attention model with a sentential gate to promote a recency inductive bias and compare it to open-source document translation. |
| Outcome: | The proposed model significantly improves translation quality on IWSLT 2015 and OpenSubtitles 2018 with similar or better BLEU scores. |
Can You Learn Semantics Through Next-Word Prediction? The Case of Entailment (2024.findings-acl)
Copied to clipboard
| Challenge: | et al. argued that sentence co-occurrence probabilities should reflect entailment . but it is unclear whether probabilities predicted by neural LMs encode enanglement based on their theory . |
| Approach: | They propose a test that decodes entailment relations between natural sentences . they argue that the test that predicts a flipped test does not account for redundancy . |
| Outcome: | The proposed test can decode entailment relations between natural sentences, but not perfectly. |
Transparency Helps Reveal When Language Models Learn Meaning (2023.tacl-1)
Copied to clipboard
| Challenge: | Existing language models are trained to optimize unsupervised objectives on text . despite their centrality, current models do not represent natural language semantics well . |
| Approach: | They show that autoregressive and masked language models learn to emulate semantic relations between expressions when context-dependent . they argue that a learner that has access to all Java code can never learn execution . |
| Outcome: | a new study shows that language models fail to represent natural language semantics well . the authors show that the model learning fails when denotations are changed to be context-dependent . |
We’re Afraid Language Models Aren’t Modeling Ambiguity (2023.emnlp-main)
Copied to clipboard
Alisa Liu, Zhaofeng Wu, Julian Michael, Alane Suhr, Peter West, Alexander Koller, Swabha Swayamdipta, Noah Smith, Yejin Choi
| Challenge: | Ambiguity is an intrinsic feature of natural language, allowing us to anticipate misunderstandings and revise our interpretations as listeners. |
| Approach: | They use AmbiEnt to capture ambiguity in a sentence and analyze it to evaluate pretrained LMs. |
| Outcome: | The proposed model can flag political claims in the wild that are misleading due to ambiguity. |
reWordBench: Benchmarking and Improving the Robustness of Reward Models with Transformed Inputs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing reward models have a high performance on benchmarks, but performance degradation is often due to overfitting. |
| Approach: | They propose to explicitly train reward models to assign similar scores to paraphrases to improve their robustness. |
| Outcome: | The proposed model reduces degradation by half for the Chat Hard subset in RewardBench. |
SecDecoding: Steerable Decoding for Safer LLM Generation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing decoding-time defense methods suffer from limited generalization, high computational overhead, or significant utility degradation. |
| Approach: | They propose a decoding-time defense framework that leverages a pair of small contrastive models to estimate token-level safety signals by measuring divergence in their output distributions. |
| Outcome: | The proposed framework achieves near-zero attack success rates against a wide spectrum of advanced jailbreak attacks while maintaining the model’s helpfulness with minimal degradation. |
ItiNera: Integrating Spatial Optimization with Large Language Models for Open-domain Urban Itinerary Planning (2024.emnlp-industry)
Copied to clipboard
Yihong Tang, Zhaokai Wang, Ao Qu, Yihao Yan, Zhaofeng Wu, Dingyi Zhuang, Jushi Kai, Kebing Hou, Xiaotong Guo, Jinhua Zhao, Zhan Zhao, Wei Ma
| Challenge: | Existing urban itinerary planning studies focus on traditional tourism, but they lack the precision and accuracy needed to create a personalized itinerary. |
| Approach: | They propose an open-domain urban itinerary planning system that integrates spatial optimization with large language models to provide customized urban itineraries based on user needs. |
| Outcome: | The proposed system can generate personalized urban itineraries based on user needs and scale with existing methods. |
Continued Pretraining for Better Zero- and Few-Shot Promptability (2022.emnlp-main)
Copied to clipboard
Zhaofeng Wu, Robert L Logan IV, Pete Walsh, Akshita Bhagia, Dirk Groeneveld, Sameer Singh, Iz Beltagy
| Challenge: | Recent language model prompting methods can achieve high accuracy in zero- and few-shot settings while requiring few to no learned task-specific parameters. |
| Approach: | They propose to use a dedicated pretraining stage to improve promptability in zero-shot settings and few-shot tuning. |
| Outcome: | The proposed method improves promptability in zero- and few-shot settings, while the existing method yields subpar performance. |
Implicit Representations of Grammaticality in Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Pretrained language models generate grammatically well-formed text and discriminate well between grammatical and ungrammatically sentences in tightly controlled minimal pairs. |
| Approach: | They propose a method to train pretrained LMs for representations of grammaticality by applying perturbations to a naturalistic text corpus. |
| Outcome: | The proposed model outperforms probability-based models on human-curated grammaticality judgment benchmarks and performs worse than string probabilities on plausibility benchmarks. |