Institution profile

Megagon Labs

Industry researchasia · jp
Official website
Research library40linked papers
Opportunities0open roles
Selected work

Representative Papers

ARCS: Towards Precise Text-to-SQL via Structured Disambiguation

Oct 06, 2026

This study addresses the problem of Text-to-SQL execution errors caused by ambiguous user queries in real-world scenarios, where traditional conversational clarification proves inefficient. To this end, this work proposes a novel structured disambiguation paradigm that efficiently resolves semantic ambiguity through explicitly constrained interactions. The primary contributions include the construction of ARCS, the first real-world benchmark dataset featuring natural ambiguity annotations with comprehensive coverage of ambiguity points, clarifications, and SQL labels, alongside an end-to-end framework integrating large-scale semantic annotation with automated evaluation. Experiments reveal significant limitations of existing models in handling ambiguity: the best-performing closed-source model achieves only 51% accuracy, while open-source counterparts fall below 27%.

0 citationsRead paper

Confidence Reasoning Graphs: Structured Confidence Estimation for LLM Agents

Oct 06, 2026

This study addresses the reliance of inter-step confidence estimation for LLM agents on internal signals and training data by proposing an inference-time framework that requires no privileged access. The method constructs a Confidence Reasoning Graph (CRG) that decomposes task completion into evidence-based sub-claims, independently estimates the confidence of terminal claims, and aggregates them to produce an overall confidence score. Experiments demonstrate that this approach significantly outperforms sampling and white-box baselines across multiple benchmarks and models. It substantially improves calibration and risk-aware decision-making, thereby achieving interpretable and auditable confidence assessment.

0 citationsRead paper
Recent publications

Latest Papers

ARCS: Towards Precise Text-to-SQL via Structured Disambiguation

Oct 06, 2026

This study addresses the problem of Text-to-SQL execution errors caused by ambiguous user queries in real-world scenarios, where traditional conversational clarification proves inefficient. To this end, this work proposes a novel structured disambiguation paradigm that efficiently resolves semantic ambiguity through explicitly constrained interactions. The primary contributions include the construction of ARCS, the first real-world benchmark dataset featuring natural ambiguity annotations with comprehensive coverage of ambiguity points, clarifications, and SQL labels, alongside an end-to-end framework integrating large-scale semantic annotation with automated evaluation. Experiments reveal significant limitations of existing models in handling ambiguity: the best-performing closed-source model achieves only 51% accuracy, while open-source counterparts fall below 27%.

0 citationsRead paper

Confidence Reasoning Graphs: Structured Confidence Estimation for LLM Agents

Oct 06, 2026

This study addresses the reliance of inter-step confidence estimation for LLM agents on internal signals and training data by proposing an inference-time framework that requires no privileged access. The method constructs a Confidence Reasoning Graph (CRG) that decomposes task completion into evidence-based sub-claims, independently estimates the confidence of terminal claims, and aggregates them to produce an overall confidence score. Experiments demonstrate that this approach significantly outperforms sampling and white-box baselines across multiple benchmarks and models. It substantially improves calibration and risk-aware decision-making, thereby achieving interpretable and auditable confidence assessment.

0 citationsRead paper