Selected work

Representative Papers

Introducing OmniGEC: A Silver Multilingual Dataset for Grammatical Error Correction

Sep 17, 2025

Existing grammatical error correction (GEC) datasets suffer from severe multilingual undercoverage, hindering the development of non-English GEC models. To address this, we introduce OmniGEC—the first large-scale, silver-standard multilingual GEC dataset covering 11 languages. It integrates Wikipedia edit histories, Reddit user posts, and UberText 2.0, with high-quality automatic corrections generated by GPT-4o-mini. Unlike conventional sentence-level annotations, OmniGEC supports paragraph-level error correction modeling. Leveraging this dataset, we fine-tune Aya-Expanse and Gemma-3, achieving state-of-the-art performance on multilingual GEC benchmarks. All data and best-performing models are publicly released on Hugging Face, substantially alleviating the scarcity of non-English GEC resources and advancing multilingual grammatical error correction research.

0 citationsRead paper

Agent-Testing Agent: A Meta-Agent for Automated Testing and Evaluation of Conversational AI Agents

Aug 24, 2025

Current LLM agent evaluation relies on static benchmarks and small-scale manual assessments, lacking efficient, comprehensive, and automated testing methodologies. To address this, we propose an automated testing and evaluation meta-agent tailored for conversational AI agents. Our approach innovatively integrates role-driven adversarial test case generation, adaptive difficulty adjustment, multi-source evidence synthesis (including code analysis, design intent extraction, and literature retrieval), and dynamic LLM-as-a-Judge scoring, all within a closed-loop feedback framework. Evaluated on travel planning and Wikipedia writing agents, our method identifies defects with diversity and severity comparable to expert annotation, reduces testing time from days to 20–30 minutes, and concurrently produces interpretable quantitative metrics and qualitative diagnostic reports. This significantly enhances the precision and efficiency of capability gap identification while improving agent trustworthiness.

0 citationsRead paper

APIO: Automatic Prompt Induction and Optimization for Grammatical Error Correction and Text Simplification

Aug 12, 2025

This work addresses grammatical error correction (GEC) and text simplification (TS), proposing APIO—an automatic prompt induction and optimization framework that eliminates reliance on manually crafted seed prompts. APIO establishes an end-to-end, LLM-based prompt optimization pipeline, iteratively generating and selecting prompt templates guided by quantifiable evaluation metrics (e.g., GLEU, SARI). Crucially, it introduces the first fully seed-free, purely automated prompt optimization method—requiring no human-provided initial prompts. On standard benchmarks—including CoNLL-2014, BEA-2019, and TurkCorpus—APIO achieves state-of-the-art performance among prompt-only approaches and significantly outperforms existing automated prompt optimization techniques (e.g., PromptBreeder, AutoPrompt). By enabling effective LLM adaptation with minimal human intervention and low resource overhead, APIO establishes a novel paradigm for efficient, low-resource LLM customization.

0 citationsRead paper

Toward Evaluative Thinking: Meta Policy Optimization with Evolving Reward Models

Apr 28, 2025

Current reward-based LLM alignment methods face two key challenges: reward hacking and fragility of hand-crafted reward prompts. To address these, we propose Meta-Policy Optimization (MPO), a novel framework that introduces a meta-reward model capable of dynamically evolving reward prompts online—monitoring, diagnosing, and calibrating reward signals in real time. By elevating reward modeling to the meta-learning level, MPO eliminates reliance on task-specific prompt engineering. Integrating meta-learning, Reinforcement Learning from AI Feedback (RLAIF), and LLM-based reward modeling, MPO achieves adaptive, robust, and general-purpose alignment. Experiments across question-answering and mathematical reasoning tasks demonstrate that MPO matches or surpasses fine-tuned prompt baselines in performance, while significantly improving training stability and cross-task generalization.

0 citationsRead paper

Learning Explainable Dense Reward Shapes via Bayesian Optimization

Apr 22, 2025

Existing RLHF methods rely on sparse, sequence-level scalar rewards, leading to inaccurate token-level credit assignment and poor interpretability. To address this, we propose an explainable AI–inspired dense reward shaping framework. Our approach is the first to integrate attribution methods—such as SHAP and LIME—into reward shaping function design, formulating token-level credit assignment as a differentiable optimization problem. We further introduce a bilevel Bayesian optimization scheme for noise-robust parameter learning and provide theoretical guarantees that additive feature attribution preserves the optimal policy. Experiments demonstrate substantial improvements in token-level credit assignment fidelity, accelerated policy convergence, and superior performance over mainstream RLHF baselines across multiple downstream tasks.

0 citationsRead paper

Recent publications

Latest Papers

Introducing OmniGEC: A Silver Multilingual Dataset for Grammatical Error Correction

Sep 17, 2025

Existing grammatical error correction (GEC) datasets suffer from severe multilingual undercoverage, hindering the development of non-English GEC models. To address this, we introduce OmniGEC—the first large-scale, silver-standard multilingual GEC dataset covering 11 languages. It integrates Wikipedia edit histories, Reddit user posts, and UberText 2.0, with high-quality automatic corrections generated by GPT-4o-mini. Unlike conventional sentence-level annotations, OmniGEC supports paragraph-level error correction modeling. Leveraging this dataset, we fine-tune Aya-Expanse and Gemma-3, achieving state-of-the-art performance on multilingual GEC benchmarks. All data and best-performing models are publicly released on Hugging Face, substantially alleviating the scarcity of non-English GEC resources and advancing multilingual grammatical error correction research.

0 citationsRead paper

Agent-Testing Agent: A Meta-Agent for Automated Testing and Evaluation of Conversational AI Agents

Aug 24, 2025

Current LLM agent evaluation relies on static benchmarks and small-scale manual assessments, lacking efficient, comprehensive, and automated testing methodologies. To address this, we propose an automated testing and evaluation meta-agent tailored for conversational AI agents. Our approach innovatively integrates role-driven adversarial test case generation, adaptive difficulty adjustment, multi-source evidence synthesis (including code analysis, design intent extraction, and literature retrieval), and dynamic LLM-as-a-Judge scoring, all within a closed-loop feedback framework. Evaluated on travel planning and Wikipedia writing agents, our method identifies defects with diversity and severity comparable to expert annotation, reduces testing time from days to 20–30 minutes, and concurrently produces interpretable quantitative metrics and qualitative diagnostic reports. This significantly enhances the precision and efficiency of capability gap identification while improving agent trustworthiness.

0 citationsRead paper

APIO: Automatic Prompt Induction and Optimization for Grammatical Error Correction and Text Simplification

Aug 12, 2025

This work addresses grammatical error correction (GEC) and text simplification (TS), proposing APIO—an automatic prompt induction and optimization framework that eliminates reliance on manually crafted seed prompts. APIO establishes an end-to-end, LLM-based prompt optimization pipeline, iteratively generating and selecting prompt templates guided by quantifiable evaluation metrics (e.g., GLEU, SARI). Crucially, it introduces the first fully seed-free, purely automated prompt optimization method—requiring no human-provided initial prompts. On standard benchmarks—including CoNLL-2014, BEA-2019, and TurkCorpus—APIO achieves state-of-the-art performance among prompt-only approaches and significantly outperforms existing automated prompt optimization techniques (e.g., PromptBreeder, AutoPrompt). By enabling effective LLM adaptation with minimal human intervention and low resource overhead, APIO establishes a novel paradigm for efficient, low-resource LLM customization.

0 citationsRead paper

Toward Evaluative Thinking: Meta Policy Optimization with Evolving Reward Models

Apr 28, 2025

Current reward-based LLM alignment methods face two key challenges: reward hacking and fragility of hand-crafted reward prompts. To address these, we propose Meta-Policy Optimization (MPO), a novel framework that introduces a meta-reward model capable of dynamically evolving reward prompts online—monitoring, diagnosing, and calibrating reward signals in real time. By elevating reward modeling to the meta-learning level, MPO eliminates reliance on task-specific prompt engineering. Integrating meta-learning, Reinforcement Learning from AI Feedback (RLAIF), and LLM-based reward modeling, MPO achieves adaptive, robust, and general-purpose alignment. Experiments across question-answering and mathematical reasoning tasks demonstrate that MPO matches or surpasses fine-tuned prompt baselines in performance, while significantly improving training stability and cross-task generalization.

0 citationsRead paper

Learning Explainable Dense Reward Shapes via Bayesian Optimization

Apr 22, 2025

Existing RLHF methods rely on sparse, sequence-level scalar rewards, leading to inaccurate token-level credit assignment and poor interpretability. To address this, we propose an explainable AI–inspired dense reward shaping framework. Our approach is the first to integrate attribution methods—such as SHAP and LIME—into reward shaping function design, formulating token-level credit assignment as a differentiable optimization problem. We further introduce a bilevel Bayesian optimization scheme for noise-robust parameter learning and provide theoretical guarantees that additive feature attribution preserves the optimal policy. Experiments demonstrate substantial improvements in token-level credit assignment fidelity, accelerated policy convergence, and superior performance over mainstream RLHF baselines across multiple downstream tasks.

0 citationsRead paper