Hi, I'm Jiaxin Zhang 👋

I am an AI Researcher, working on reliable long-horizon AI agents, agentic reinforcement learning, and calibrated post-training.

My research asks a simple question:

How can AI agents know what they don’t know, act under uncertainty, and improve from their own prediction–reality gaps?

I build methods, environments, and evaluation frameworks that turn uncertainty, confidence, and consistency into first-class training signals for reliable and self-improving AI systems.

Homepage · Google Scholar · LinkedIn · X/Twitter · Email

Research Focus

• Agentic RL & Post-training
  Calibration-aware on-policy distillation, GRPO/RL training, self-evolving environments, synthetic feedback, and reward/evaluator design for long-horizon agents.
• Alignment, Calibration & Honesty
  Uncertainty-aware supervision, confidence calibration, hallucination detection, factuality, scalable oversight, and reliable model behavior.
• Long-horizon Agents & Evaluation
  Tool use, planning, trajectory-level evaluation, deep research agents, evidence grounding, failure attribution, and enterprise-scale agent benchmarks.

Selected Work

• [NeurIPS2026] Prospective Hindsight
  Self-calibrating reinforcement learning via prediction–reality gaps, aligning an agent’s action-time self-belief with verifier outcomes.
• CaOPD: Calibration-aware On-policy Distillation
  Decouples capability learning from honest confidence calibration in LLM post-training.
• [EMNLP2026] Agentic Uncertainty Quantification
  Turns verbalized uncertainty into active control signals for memory, reflection, and long-horizon execution.
• [ICML2026] Agentic Confidence Calibration
  A trajectory-level calibration framework for diagnosing and improving the reliability of long-horizon agents.
• [ACL2026] The Evolving Role of Uncertainty Quantification in Large Language Models
  The evolution of uncertainty from a passive diagnostic metric to an active control signal guiding real-time model behavior

For the full list of publications, please see my Google Scholar or homepage.

Contact

I am interested in reliable AI agents, agentic RL, post-training, calibration, uncertainty, scalable evaluation, and self-improving AI systems. Feel free to reach out via email or visit my homepage.