Papers by Xiting Wang
Masked Thought: Simply Masking Partial Reasoning Steps Can Improve Mathematical Reasoning Learning of Language Models (2024.acl-long)
Copied to clipboard
Changyu Chen, Xiting Wang, Ting-En Lin, Ang Lv, Yuchuan Wu, Xin Gao, Ji-Rong Wen, Rui Yan, Yongbin Li
| Challenge: | Despite the advances in large language models, they still face difficulties with multi-step reasoning tasks. |
| Approach: | They propose a method that randomly masks certain tokens within the chain of thought to improve model accuracy by 5% over standard supervised fine-tuning. |
| Outcome: | The proposed method improves accuracy and accuracy by 5% over standard fine-tuning with a few codes modified. |
BSharedRAG: Backbone Shared Retrieval-Augmented Generation for the E-commerce Domain (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work adopts separate modules for retrieval and generation, which may be suboptimal since the retrieval task and generation task cannot benefit from each other to improve performance. |
| Approach: | They propose a backbone-shared RAG framework that uses a domain-specific corpus to continuously pre-train a model and then trains two plug-and-play Low-Rank Adaptation modules based on the shared backbone to minimize retrieval and generation losses respectively. |
| Outcome: | The proposed framework outperforms baseline models by 5% and 13% in Hit@3 upon two datasets in retrieval evaluation and by 23% in terms of BLEU-3 in generation evaluation. |
Evaluating Readability and Faithfulness of Concept-based Explanations (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for evaluating concepts from different perspectives lack a unified formalization. |
| Approach: | They propose a formal definition of concepts generalizing to diverse concept-based explanations’ settings and apply it to other types of explanations or tasks. |
| Outcome: | Extensive experimental analysis was carried out to determine the evaluation measures for explanation evaluation measures. |
Select, Read, and Write: A Multi-Agent Framework of Full-Text-based Related Work Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for related work generation (RWG) suffer from shallow comprehension due to taking the limited portions of references as input and isolated explanation for each reference due to ineffective capturing the relationships among them. |
| Approach: | They propose a multi-agent framework that takes the limited portions of references papers as input and isolates the relationships between them. |
| Outcome: | The proposed framework outperforms other selectors and improves reading order with constrains of the graph structure. |
Understanding Conflicts in Multi-Objective Alignment through Reward Consistency (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing training pipelines still face alignment conflicts where optimizing for one objective degrades performance on others. |
| Approach: | They propose a reward-based criterion that approximates alignment conflicts via reward models. |
| Outcome: | The proposed framework improves harmlessness and helpfulness scores by 23.07% over the vanilla dataset. |
Optimal Transport-Based Token Weighting scheme for Enhanced Preference Optimization (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for direct preference optimization assign equal importance to all tokens while humans focus on more meaningful parts. |
| Approach: | They propose to use a transport-based token weighting scheme to enhance direct preference optimization by emphasizing meaningful token pairs and de-emphasizing less relevant ones to yield a more contrastive reward difference estimate. |
| Outcome: | Extensive experiments have validated the proposed method in improving instruction-following ability across various settings. |
Distillation with Explanations from Large Language Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) like ChatGPT and GPT-4 have made remarkable progress in various NLP tasks while also providing explanations alongside their answers. |
| Approach: | They propose to use Large language models (LLMs) to generate more accurate answers and corresponding free-text explanations by combining ground truth labels and answers-explanations generated by LLMs. |
| Outcome: | The proposed method achieves improved predictive performance and generates explanations that exhibit greater alignment with the model’s task outputs. |
DuNST: Dual Noisy Self Training for Semi-Supervised Controllable Text Generation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to augment self-training (ST) in attribute-controllable language generation are limited and limited. |
| Approach: | They propose a new ST framework that integrates self-generated pseudo text into attribute-controllable language generation. |
| Outcome: | The proposed framework can be applied to semi-supervised controllable language generation. |
Value FULCRA: Mapping Large Language Models to the Multidimensional Spectrum of Basic Human Value (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing work specifies values as risk criteria formulated in the AI community, e.g., fairness and privacy protection, suffering from poor clarity, adaptability and transparency. |
| Approach: | They propose a value alignment paradigm based on Schwartz's Theory of Basic Values as an instantiation and propose 'BaseAlign' to support this paradigm. |
| Outcome: | The proposed model covers existing risks and anticipates unidentified ones with a low-data set. |
Internal Value Alignment in Large Language Models through Controlled Value Vector Activation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing LLMs do not possess consistent values, but many have been developed to align them at the behavioral level, including supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). |
| Approach: | They propose a Controlled Value Vector Activation method that directly aligns the internal values of Large Language Models by interpreting how a value is encoded in their latent representations. |
| Outcome: | The proposed method achieves highest success rate across 10 basic values without hurting model performance and fluency, and ensures target values even with opposite and potentially malicious input prompts. |
Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework (2026.findings-acl)
Copied to clipboard
Jiaqi Weng, Han Zheng, Hanyu Zhang, Ej Zhou, Qinqin He, Jialing Tao, Hui Xue, Zhixuan Chu, Xiting Wang
| Challenge: | Existing studies on how SAEs derive most fine-grained latent features for safety remain unexplored. |
| Approach: | They propose a framework for interpreting SAE features in safety-critical domains . they train a suite of SAEs with human-readable explanations and systematic evaluations based on pornography, politics, violence, and terror . |
| Outcome: | The proposed framework reduces interpretation cost by 55% and improves safety-critical features. |
Persuading across Diverse Domains: a Dataset and Persuasion Large Language Model (2024.acl-long)
Copied to clipboard
| Challenge: | Persuasive dialogue requires multi-turn following and planning abilities to achieve the goal of persuating users. |
| Approach: | They propose a general method to learn a persuasive model based on LLMs through intent-to-strategy reasoning, which summarizes the intent of user’s utterance and reasons next strategy to respond. |
| Outcome: | The proposed method outperforms baselines on automatic evaluation metric Win-Rate and human evaluation on two datasets. |
Influence-based Online Experience Selection for Effective RLHF (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for RL fail to establish an interpretable connection between data and optimization objectives. |
| Approach: | They propose a data selection method that dynamically estimates the influence of individual training samples on policy optimization. |
| Outcome: | The proposed method significantly improves training effectiveness with fewer optimization steps. |
Prototypical Reward Network for Data-Efficient Model Alignment (2024.acl-long)
Copied to clipboard
| Challenge: | Reinforcement Learning from Human Feedback (RLHF) is a reward model that fine-tunes Large Language Models (LLMs) by utilizing Prototypical Networks. |
| Approach: | They propose a framework utilizing Prototypical Networks to enhance reward models under limited human feedback, enabling more stable and reliable structural learning from fewer samples. |
| Outcome: | The proposed framework improves reward models under limited human feedback, surpassing traditional methods, especially in data-limited scenarios. |
Generating Multiple-Length Summaries via Reinforcement Learning for Unsupervised Sentence Summarization (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models to summarize texts without ground-truth summaries are extractive, which remove words from texts and thus are less flexible than abstractive models. |
| Approach: | They propose an unsupervised model that extracts words from texts and makes them mutually enhance each other. |
| Outcome: | The proposed model outperforms both abstractive and extractive models, while generating new words not contained in input texts. |
Entropy-based Exploration Conduction for Multi-step Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to automatically decide the depth of exploration of the reasoning procedure lead to high cost and a lack of flexibility. |
| Approach: | They propose a method that dynamically adjusts the exploration depth during multi-step reasoning by monitoring LLM’s output entropy and variance entropic. |
| Outcome: | The proposed method captures the uncertainty of the current step and the fluctuation of uncertainty across consecutive reasoning steps and then selects whether to deepen, expand, or stop exploration according to the probability. |
SOAR: Supervision from Observation for Agentic Reinforcement Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work assigns supervision based on outcome rewards or external reward models, but ignores environment observations, a critical source of learning. |
| Approach: | They propose a supervision-based agentic reinforcement learning system that integrates environment observations as an explicit supervision signal. |
| Outcome: | The proposed model improves performance on reasoning and deep research tasks while reducing erroneous and inefficient tool usage. |
Unlocking Implicit Experience: Synthesizing Tool-Use Trajectories from Text (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) can be used to effectively utilize tools in multi-turn interactions, but acquiring diverse and realistic multi-step tool-use data remains a challenge. |
| Approach: | They propose a text-based data synthesis pipeline that generates multi-turn tool-use trajectories from text corpora using relevance filtering, workflow tool extraction, trajectory grounding, and complexity refinement. |
| Outcome: | The proposed model achieves 14.9% improvement on the BFCL V3 Multi-turn benchmark while significantly reducing inference latency and costs. |
PENS: A Dataset and Generic Framework for Personalized News Headline Generation (2021.acl-long)
Copied to clipboard
| Challenge: | Using a dataset of Microsoft News, we propose a generic framework to personalize a text generator and establish personalized headlines. |
| Approach: | They propose a generic framework to personalize a news headline generator and establish personalized headlines by leveraging user behavioral data. |
| Outcome: | The proposed framework is based on user preference data and user preference injections to personalize a text generator and establish personalized headlines. |
PAM: Enhancing General Alignment of Large Reasoning Models through Priority-Aware Metacognition (2026.acl-long)
Copied to clipboard
| Challenge: | Existing studies indicate that System-2 thinking alone does not transfer to the general alignment domain. |
| Approach: | They propose to use priority-aware metacognition to help LRMs understand human preferences and monitor and regulate their thinking process. |
| Outcome: | The proposed model improves general alignment performance by 10 points on helpfulness and harmless benchmarks. |
DPWriter: Reinforcement Learning with Diverse Planning Branching for Creative Writing (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for enhancing large language models (LLMs) lack explicit mechanisms for guiding diverse exploration and instead prioritize efficiency and performance over diversity. |
| Approach: | They propose a reinforcement learning-based framework that decomposes the generation process into explicitly planned intermediate steps and introduces divergence at the planning phase based on diversity variation. |
| Outcome: | The proposed method significantly outperforms existing baselines on creative writing benchmarks on a semi-structured long chain-of-thought (CoT) it introduces divergence at the planning phase based on diversity variation, alongside a group-aware diversity reward to encourage distinct trajectories. |
PanoramaRAG: Enabling Consistent Global Topic Awareness in Graph-Based RAG (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing graph-based methods for enhancing Large Language Models (LLMs) with external knowledge are focusing on local relationships, resulting in suboptimal performance for tasks that require global context. |
| Approach: | They propose a "panorama"-guided paradigm that integrates a light yet comprehensive "panoramic" of the corpus to guide all stages of the retrieval process. |
| Outcome: | The proposed paradigm performs well across five datasets and a variety of tasks. |