Work.

From research questions to models, benchmarks, and working tools.

01 / Research into practice

Selected work.

RESEARCH & TOOL USE

Reasoning ↔ interaction

Agents & reasoning Apodex · 2025–present

From reasoning to getting work done.

How can an agent use files, search, code, and collaboration to carry a complex research task through to completion?

My work: Built the deep-research agent harness from scratch at Apodex, with evidence verification and multi-agent coordination.
Outcome: A model and workbench for complex research, alongside the open-source FrontierAgent harness.

AUDIO MODEL EVALUATION

From input to measurable behavior

Audio & evaluation I²R, A*STAR · NAACL 2025

Understanding what audio models understand.

How do we evaluate the capabilities of audio large language models?

My work: Led AudioBench: eight tasks and 26 datasets for evaluating audio-language models. First author.
Outcome: A universal audio-language benchmark, published at NAACL 2025.

ACROSS LANGUAGES & CULTURES

Shared knowledge. Different contexts.

Languages & culture I²R, A*STAR · NAACL 2024

AI that works across languages.

Do multilingual models carry knowledge and reasoning across languages and cultures?

My work: Led SeaEval to evaluate cross-lingual consistency and cultural reasoning. Co-first author.
Outcome: A multilingual evaluation benchmark, published at NAACL 2024.