PeopleofSuperintelligence

Researcher

Yao Ma

Also known as Ma Yao

As of May 2026, Yao Ma is an applied scientist at Amazon whose work spans reinforcement learning from human feedback and news recommendation. He coauthored a 2024 news-ranking framework that combines pointwise relevance predictions with pairwise comparisons. His graph-routing fine-tuning paper appeared at CoLLAs in 2024 and in PMLR in 2025. In 2026, he coauthored LLMEvalRec for news-recommendation evaluation, ESRRSim for strategic-risk testing, and ARES for repairing joint failures in language and reward models. In May 2026, he explained how human rankings train auxiliary reward models for Amazon's policy-adherent responses.

Now at
Amazon

Timeline

  1. 2024

    Pointwise-pairwise news ranking

    In 2024, he coauthors an efficient framework for language-model news recommendation that combines pointwise relevance predictions with pairwise comparisons. Experiments outperform state-of-the-art methods on the MIND and Adressa datasets.

  2. 2024–2025

    Graph routing for fine-tuning

    At CoLLAs in 2024, he coauthors a gradual fine-tuning method that represents source domains as a weighted graph and uses graph routing to choose the training order. PMLR publishes it in 2025; its best strategy improves NLI accuracy by 2.3% over state-of-the-art and achieves competitive results on sentiment analysis.

  3. April 2026

    Testing strategic reasoning risks

    In April 2026, he coauthors ESRRSim, an agentic framework that tests deception, evaluation gaming, and reward hacking in language models. It groups seven risk categories into 20 subcategories and uses paired rubrics to assess model responses and reasoning traces.

  4. May 2026

    Explaining Amazon's RLHF

    In May 2026, he explains how human feedback and preference comparisons guide RLHF, while human-ranked responses train auxiliary reward models for Amazon policy-adherent outputs.

  5. 2026

    Simulating news readers

    In 2026, he coauthors LLMEvalRec, which uses language-model agents to simulate user behavior for news-recommendation evaluation. Its generated data reaches a 0.97 Spearman correlation with real evaluation rankings, and production validation tracks click-through-rate improvements.

  6. July 2026

    Repairing reward-model failures

    In July 2026, he coauthors ARES, which finds cases where the core language model and its reward model fail together. It generates adversarial prompts, fine-tunes the reward model to detect harmful content, and uses the improved model to optimize the core model.

  7. May 2026

    Applied scientist at Amazon

    In May 2026, he is an applied scientist at Amazon.

  1. Among employers with 30+ full-time cases, OpenAI leads primary technical role H-1B pay at $310,000, and Amazon, Google and Microsoft file 57% of all casesIncludes Amazon. Across 440,669 certified US H-1B filings from 44 employers (FY2020–FY2026), OpenAI leads primary technical role base pay at $310,000 and xAI leads overall base pay at $290,000.
  2. 20% of people at companies who shape AI also hold a university role, led by Google and Stanford UniversityIncludes Amazon. We estimate that 539 of the 6,766 people with a known current organization (8%) hold both a university role and a company role, including 20% of people at companies and 13% of people at universities.
  3. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Amazon. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.
  4. 22 of 133 people shaping AI with a former OpenAI role now work at AnthropicIncludes Amazon. Our estimates of where OpenAI recruits came from and where 133 people with former OpenAI roles work now, including the companies and labs they founded.

Sources and credits

Photo: We have no published photo of Yao Ma yet. Send us a free-licensed photo.

  1. Efficient pointwise-pairwise learning-to-rank for news recommendation amazon.science. Checked 22 October 2026.
  2. Gradual Fine-Tuning with Graph Routing for Multi-Source Unsupervised Domain Adaptation proceedings.mlr.press. Checked 22 October 2026.
  3. Gradual fine-tuning with graph routing for multi-source unsupervised domain adaptation amazon.science. Checked 22 October 2026.
  4. Emergent Strategic Reasoning Risks in AI: A Taxonomy-Driven Evaluation Framework arxiv.org. Checked 22 October 2026.
  5. Building trust into AI amazon.science. Checked 22 October 2026.
  6. LLMEvalRec: An agentic framework for simulating users to evaluate news recommendation systems amazon.science. Checked 22 October 2026.
  7. ARES: Adaptive red-teaming and end-to-end repair of policy-reward system amazon.science. Checked 22 October 2026.
  8. ARES: Adaptive Red-Teaming and End-to-End Repair of Policy-Reward System aclanthology.org. Checked 22 October 2026.