Selected work

Representative Papers

Guide-to-Explain for Controllable Summarization

Nov 19, 2024arXiv.org

Large language models (LLMs) exhibit limited precision in controlling numerical attributes—such as summary length and extractiveness—in controllable summarization, hindering practical deployment aligned with user preferences. To address this, we propose a Guided Reflection Framework featuring a novel two-stage self-reflective mechanism: “Guide–Explain.” First, attribute-aware bias detection identifies deviations between the initial summary and target constraints; second, an attributional error explanation is generated to guide conditional regeneration. Our approach integrates self-reflective prompt engineering with multi-attribute joint constraints, significantly improving control fidelity and optimization efficiency. Experiments on multidimensional controllable summarization demonstrate substantial gains: constraint satisfaction rates increase markedly, and average iteration counts decrease by over 40% compared to state-of-the-art pure-LLM iterative baselines.

1 citationsRead paper

MUDIDI: A Two-Stage Framework for Multilingual Dictionary Digitization with Language Models

Jun 08, 2026

Multilingual dictionaries are often available only as scanned documents, whose complex layouts, specialized scripts, and intricate entry structures hinder efficient conversion into structured digital formats. This work proposes MUDIDI, a two-stage framework that first evaluates the quality of character recognition and markup preservation, then performs lexical entry segmentation and maps the results to the SIL Dictionary Format. The study presents the first systematic evaluation of optical character recognition (OCR), large language models (LLMs), and vision-language models (VLMs) on this task, revealing that LLMs consistently outperform other approaches. Incorporating contextual information—such as dictionary prefaces—further enhances performance significantly. The project is publicly released with a manually annotated dataset spanning 30 dictionaries and a complete processing pipeline.

0 citationsRead paper

Segment-level Tree Search for Long Meeting Document Summarization

Jun 07, 2026

This work addresses the challenges of summarizing lengthy and structurally complex meeting documents, where existing approaches suffer from error propagation in multi-stage pipelines and are constrained by short, low-quality reference summaries. To overcome these limitations, the authors propose the S3 framework, which introduces Monte Carlo Tree Search (MCTS) to the meeting summarization task for the first time. S3 generates multiple candidate segment-level summaries as tree nodes and employs a self-reward mechanism to guide the search process toward optimal combinations, without requiring additional model training. This approach effectively mitigates error accumulation and, despite relying solely on a 7B-parameter large language model, produces summaries of moderate length and quality comparable to those generated by a 72B-parameter model.

0 citationsRead paper

GAIA-v2-LILT: Multilingual Adaptation of Agent Benchmark beyond Translation

Apr 27, 2026

This work addresses the limitations of existing agent benchmarks, which are predominantly English-centric and suffer from degraded evaluation validity when extended to other languages via machine translation due to query-answer misalignment and cultural-context mismatches. To overcome these issues, the study proposes a refined multilingual adaptation pipeline that, for the first time in agent benchmarking, systematically integrates task-level functional alignment and cultural alignment mechanisms. The approach combines automated validation, human expert review, and difficulty calibration to construct high-quality multilingual evaluation datasets. Experimental results demonstrate that, compared to naive translation baselines, the proposed method improves agent success rates by up to 32.7%, with performance in the best non-English setting falling only 3.1% short of the English baseline—substantially narrowing the multilingual evaluation gap.

0 citationsRead paper

From Binary to Bilingual: How the National Weather Service is Using Artificial Intelligence to Develop a Comprehensive Translation Program

Oct 16, 2025

To address the meteorological service gap affecting 68.8 million U.S. households with non-English home languages, this study develops the first scalable multilingual weather translation system for the National Weather Service (NWS). Methodologically, it integrates domain-adapted large language models, neural machine translation (NMT), and proprietary LILT training techniques, augmented by GIS-driven language-demand mapping, a culturally adapted risk communication framework, and human-in-the-loop review. The contribution includes the first comprehensive meteorological translation architecture—ensuring terminological accuracy, cultural relevance, and ethical compliance—and enables automated generation of official forecasts and warnings in Spanish, Simplified Chinese, Vietnamese, and other priority languages. Evaluation demonstrates substantial reductions in manual translation time, improved alert coverage across linguistically diverse populations, and accelerated service response. This system provides critical technical infrastructure for an equitable, universally accessible national warning system.

0 citationsRead paper

Recent publications

Latest Papers

MUDIDI: A Two-Stage Framework for Multilingual Dictionary Digitization with Language Models

Jun 08, 2026

Multilingual dictionaries are often available only as scanned documents, whose complex layouts, specialized scripts, and intricate entry structures hinder efficient conversion into structured digital formats. This work proposes MUDIDI, a two-stage framework that first evaluates the quality of character recognition and markup preservation, then performs lexical entry segmentation and maps the results to the SIL Dictionary Format. The study presents the first systematic evaluation of optical character recognition (OCR), large language models (LLMs), and vision-language models (VLMs) on this task, revealing that LLMs consistently outperform other approaches. Incorporating contextual information—such as dictionary prefaces—further enhances performance significantly. The project is publicly released with a manually annotated dataset spanning 30 dictionaries and a complete processing pipeline.

0 citationsRead paper

Segment-level Tree Search for Long Meeting Document Summarization

Jun 07, 2026

This work addresses the challenges of summarizing lengthy and structurally complex meeting documents, where existing approaches suffer from error propagation in multi-stage pipelines and are constrained by short, low-quality reference summaries. To overcome these limitations, the authors propose the S3 framework, which introduces Monte Carlo Tree Search (MCTS) to the meeting summarization task for the first time. S3 generates multiple candidate segment-level summaries as tree nodes and employs a self-reward mechanism to guide the search process toward optimal combinations, without requiring additional model training. This approach effectively mitigates error accumulation and, despite relying solely on a 7B-parameter large language model, produces summaries of moderate length and quality comparable to those generated by a 72B-parameter model.

0 citationsRead paper

GAIA-v2-LILT: Multilingual Adaptation of Agent Benchmark beyond Translation

Apr 27, 2026

This work addresses the limitations of existing agent benchmarks, which are predominantly English-centric and suffer from degraded evaluation validity when extended to other languages via machine translation due to query-answer misalignment and cultural-context mismatches. To overcome these issues, the study proposes a refined multilingual adaptation pipeline that, for the first time in agent benchmarking, systematically integrates task-level functional alignment and cultural alignment mechanisms. The approach combines automated validation, human expert review, and difficulty calibration to construct high-quality multilingual evaluation datasets. Experimental results demonstrate that, compared to naive translation baselines, the proposed method improves agent success rates by up to 32.7%, with performance in the best non-English setting falling only 3.1% short of the English baseline—substantially narrowing the multilingual evaluation gap.

0 citationsRead paper

From Binary to Bilingual: How the National Weather Service is Using Artificial Intelligence to Develop a Comprehensive Translation Program

Oct 16, 2025

To address the meteorological service gap affecting 68.8 million U.S. households with non-English home languages, this study develops the first scalable multilingual weather translation system for the National Weather Service (NWS). Methodologically, it integrates domain-adapted large language models, neural machine translation (NMT), and proprietary LILT training techniques, augmented by GIS-driven language-demand mapping, a culturally adapted risk communication framework, and human-in-the-loop review. The contribution includes the first comprehensive meteorological translation architecture—ensuring terminological accuracy, cultural relevance, and ethical compliance—and enables automated generation of official forecasts and warnings in Spanish, Simplified Chinese, Vietnamese, and other priority languages. Evaluation demonstrates substantial reductions in manual translation time, improved alert coverage across linguistically diverse populations, and accelerated service response. This system provides critical technical infrastructure for an equitable, universally accessible national warning system.

0 citationsRead paper

Adaptive Planning for Multi-Attribute Controllable Summarization with Monte Carlo Tree Search

Sep 30, 2025

This paper addresses the challenge of inconsistent attribute control in multi-attribute controllable summarization, arising from inter-attribute dependencies. We propose PACO, a training-free adaptive planning framework. Its core innovation is the first application of Monte Carlo Tree Search (MCTS) to controllable summarization, formulating multi-attribute optimization as a sequential planning problem: nodes represent summary states, actions correspond to incremental adjustments of individual attributes, and language-model-guided dynamic search automatically discovers the optimal control sequence. PACO requires no fine-tuning and flexibly adapts to arbitrary attribute combinations. Empirically, it significantly improves control accuracy and cross-domain generalization across diverse models and domains: using only Llama-3.2-1B, it matches the performance of 70B-scale baselines; with larger models, it consistently outperforms state-of-the-art methods.

0 citationsRead paper