Selected work

Representative Papers

Formula-One Prompting: Adaptive Reasoning Through Equations For Applied Mathematics

Jan 27, 2026

Existing prompting methods struggle to effectively generate or invoke governing equations when solving applied mathematical problems in domains such as finance and physics, thereby limiting the reasoning capabilities of large language models. This work proposes Formula-One Prompting (F-1), which introduces mathematical equations as an intermediate representation in prompt engineering for the first time. F-1 employs a two-stage framework: it first derives governing equations from the problem description and then adaptively selects among Chain-of-Thought (CoT), Program-of-Thought (PoT), or direct computation strategies for solution. The entire process—equation formulation and strategy selection—is completed within a single model call. Evaluated across five models and four benchmarks, F-1 outperforms CoT by 5.76% and PoT by 8.42% on average, with a notable 13.30% improvement on FinanceMath and particularly strong performance on physics-related problems.

0 citationsRead paper

ThaiOCRBench: A Task-Diverse Benchmark for Vision-Language Understanding in Thai

Nov 06, 2025

Existing vision-language model (VLM) benchmarks severely underrepresent low-resource languages such as Thai—particularly in document structure understanding tasks. To address this gap, we introduce ThaiOCRBench, the first multi-task vision-language benchmark dedicated to Thai, comprising 2,808 samples across 13 document understanding tasks. Our contributions are threefold: (1) We establish the first Thai-specific OCR evaluation framework, enabling zero-shot assessment; (2) Through systematic evaluation, we expose critical weaknesses of mainstream VLMs in handwritten text recognition and fine-grained text extraction, identifying root causes including linguistic bias and structural misalignment; (3) Rigorous human annotation and error analysis demonstrate that closed-source models (e.g., Gemini) outperform open-source counterparts on complex scripts. The benchmark is fully open-sourced, providing a reproducible standard and actionable insights for advancing Thai document intelligence.

0 citationsRead paper

Single Answer is Not Enough: On Generating Ranked Lists with Medical Reasoning Models

Sep 25, 2025

Clinical decision-making requires balancing multiple diagnostic hypotheses, yet existing medical reasoning models (MRMs) produce only single-answer outputs, leading to narrow, potentially unsafe reasoning. This work is the first systematic study on training MRMs to generate ranked answer lists—enabling safe, comprehensive diagnosis for open-ended clinical questions. We propose a novel ranking-aware reward function and integrate it into a unified training framework combining prompt engineering, supervised fine-tuning (SFT), and reinforcement fine-tuning (RFT). Experiments demonstrate that RFT substantially improves model robustness across diverse response formats; the resulting ranked lists not only recover ground-truth diagnoses but also surface clinically plausible alternatives, thereby enhancing decision support. Our approach advances medical large language models toward a multi-hypothesis, interpretable, and safety-critical reasoning paradigm.

0 citationsRead paper

Talk Less, Call Right: Enhancing Role-Play LLM Agents with Automatic Prompt Optimization and Role Prompting

Aug 30, 2025

In the CPDC 2025 API Track, role-playing dialogue agents suffer from two critical issues: verbose responses (over-speaking) and improper tool usage (e.g., invalid or premature function calls), undermining role consistency and behavioral controllability. To address these, we propose Rule-based Role Prompting (RRP), a novel method that introduces structured role cards and scene contracts, while strictly constraining the timing and validity of function calls. Compared to baseline prompting, handcrafted prompts, and automated prompt optimization (APO), RRP significantly improves role fidelity and tool-use accuracy—raising the overall score from 0.519 to 0.571, outperforming zero-shot baselines and multiple advanced prompting techniques. We open-source our optimal prompt templates and APO toolkit, establishing a reproducible benchmark and practical framework for controllable role-playing dialogue research.

0 citationsRead paper

FinCoT: Grounding Chain-of-Thought in Expert Financial Reasoning

Jun 19, 2025

Traditional chain-of-thought (CoT) prompting in Financial Natural Language Processing (FinNLP) suffers from insufficient domain-specific logical constraints, weak interpretability, and misalignment with expert reasoning patterns. To address this, we propose FinCoT—the first structured, CFA-level financial expert–driven CoT prompting framework. FinCoT systematically models domain experts’ decision-making pathways, formalizing a multi-stage, verifiable reasoning process that fundamentally departs from heuristic or unstructured CoT designs. Evaluated on models including Qwen-2.5-7B-Instruct, FinCoT achieves 80.5% accuracy across ten CFA-style financial tasks—a +17.3 percentage point improvement—while reducing inference token consumption by 8×. Moreover, its generated reasoning traces exhibit stronger alignment with real-world financial practice, enhanced traceability, and improved fidelity to expert judgment.

0 citationsRead paper

Recent publications

Latest Papers

Formula-One Prompting: Adaptive Reasoning Through Equations For Applied Mathematics

Jan 27, 2026

Existing prompting methods struggle to effectively generate or invoke governing equations when solving applied mathematical problems in domains such as finance and physics, thereby limiting the reasoning capabilities of large language models. This work proposes Formula-One Prompting (F-1), which introduces mathematical equations as an intermediate representation in prompt engineering for the first time. F-1 employs a two-stage framework: it first derives governing equations from the problem description and then adaptively selects among Chain-of-Thought (CoT), Program-of-Thought (PoT), or direct computation strategies for solution. The entire process—equation formulation and strategy selection—is completed within a single model call. Evaluated across five models and four benchmarks, F-1 outperforms CoT by 5.76% and PoT by 8.42% on average, with a notable 13.30% improvement on FinanceMath and particularly strong performance on physics-related problems.

0 citationsRead paper

ThaiOCRBench: A Task-Diverse Benchmark for Vision-Language Understanding in Thai

Nov 06, 2025

Existing vision-language model (VLM) benchmarks severely underrepresent low-resource languages such as Thai—particularly in document structure understanding tasks. To address this gap, we introduce ThaiOCRBench, the first multi-task vision-language benchmark dedicated to Thai, comprising 2,808 samples across 13 document understanding tasks. Our contributions are threefold: (1) We establish the first Thai-specific OCR evaluation framework, enabling zero-shot assessment; (2) Through systematic evaluation, we expose critical weaknesses of mainstream VLMs in handwritten text recognition and fine-grained text extraction, identifying root causes including linguistic bias and structural misalignment; (3) Rigorous human annotation and error analysis demonstrate that closed-source models (e.g., Gemini) outperform open-source counterparts on complex scripts. The benchmark is fully open-sourced, providing a reproducible standard and actionable insights for advancing Thai document intelligence.

0 citationsRead paper

Single Answer is Not Enough: On Generating Ranked Lists with Medical Reasoning Models

Sep 25, 2025

Clinical decision-making requires balancing multiple diagnostic hypotheses, yet existing medical reasoning models (MRMs) produce only single-answer outputs, leading to narrow, potentially unsafe reasoning. This work is the first systematic study on training MRMs to generate ranked answer lists—enabling safe, comprehensive diagnosis for open-ended clinical questions. We propose a novel ranking-aware reward function and integrate it into a unified training framework combining prompt engineering, supervised fine-tuning (SFT), and reinforcement fine-tuning (RFT). Experiments demonstrate that RFT substantially improves model robustness across diverse response formats; the resulting ranked lists not only recover ground-truth diagnoses but also surface clinically plausible alternatives, thereby enhancing decision support. Our approach advances medical large language models toward a multi-hypothesis, interpretable, and safety-critical reasoning paradigm.

0 citationsRead paper

Talk Less, Call Right: Enhancing Role-Play LLM Agents with Automatic Prompt Optimization and Role Prompting

Aug 30, 2025

In the CPDC 2025 API Track, role-playing dialogue agents suffer from two critical issues: verbose responses (over-speaking) and improper tool usage (e.g., invalid or premature function calls), undermining role consistency and behavioral controllability. To address these, we propose Rule-based Role Prompting (RRP), a novel method that introduces structured role cards and scene contracts, while strictly constraining the timing and validity of function calls. Compared to baseline prompting, handcrafted prompts, and automated prompt optimization (APO), RRP significantly improves role fidelity and tool-use accuracy—raising the overall score from 0.519 to 0.571, outperforming zero-shot baselines and multiple advanced prompting techniques. We open-source our optimal prompt templates and APO toolkit, establishing a reproducible benchmark and practical framework for controllable role-playing dialogue research.

0 citationsRead paper

FinCoT: Grounding Chain-of-Thought in Expert Financial Reasoning

Jun 19, 2025

Traditional chain-of-thought (CoT) prompting in Financial Natural Language Processing (FinNLP) suffers from insufficient domain-specific logical constraints, weak interpretability, and misalignment with expert reasoning patterns. To address this, we propose FinCoT—the first structured, CFA-level financial expert–driven CoT prompting framework. FinCoT systematically models domain experts’ decision-making pathways, formalizing a multi-stage, verifiable reasoning process that fundamentally departs from heuristic or unstructured CoT designs. Evaluated on models including Qwen-2.5-7B-Instruct, FinCoT achieves 80.5% accuracy across ten CFA-style financial tasks—a +17.3 percentage point improvement—while reducing inference token consumption by 8×. Moreover, its generated reasoning traces exhibit stronger alignment with real-world financial practice, enhanced traceability, and improved fidelity to expert judgment.

0 citationsRead paper