Chentao Ye
Chentao Ye is a senior applied scientist on Amazon's AGI Responsible AI team, where he works on pretraining for large language models. He coauthored an art question-answering dataset and research on spoken-language domain classification. In October 2025, he coauthored an AWS post about adapting Amazon Nova for text moderation. His 2026 papers examined misaligned computer-use actions and attacks that exploit agent skills.
Timeline
-
2020
An art question dataset
In 2020, he coauthors AQUA, an art question-answering dataset whose questions draw on paintings and comments about art.
-
2021
Contextual dialogue classification
In 2021, he coauthors a study on spoken-language domain classification that uses time differences and turn order to combine recent and distant dialogue context.
-
October 2025
Customizing Nova moderation
In October 2025, he coauthors an AWS post on fine-tuning Amazon Nova for content moderation tailored to organization-specific policies.
-
February 2026
Misaligned computer-use actions
In February 2026, he coauthors a study that introduces MisActBench for action-level misalignment and DeAction, a guardrail that detects and corrects misaligned actions before execution.
-
June 2026
SkillHarm's attack benchmark
In June 2026, he coauthors SkillHarm, a benchmark for skill-based attacks across the skill-use lifecycle, with a taxonomy of related risks.
-
As of October 2026
Senior scientist at Amazon AGI
As of October 2026, he is a senior applied scientist on Amazon's AGI responsible-AI team, working on pretraining for large language models.
Most-cited works
Sources and credits
Photo: We have no published photo of Chentao Ye yet. Send us a free-licensed photo.
- A Dataset and Baselines for Visual Question Answering on Art
- Contextual Domain Classification with Temporal Representations
- Customizing text content moderation with Amazon Nova
- When Actions Go Off-Task: Detecting and Correcting Misaligned Actions in Computer-Use Agents
- SkillHarm: Lifecycle-Aware Skill-Based Attacks via Automated Construction
- Inside Amazon's responsible-AI pipeline