Blog — Jiajun Fan

Writing

Blog

Notes on reinforcement learning, generative models, and research observations — written to think through ideas, not to summarize papers.

What's happening in my field →
2026

Voice Agents Could Be Ranked, Never Improved. We Closed the Loop.

 ·  reinforcement learning · voice agents · agentic RL · audio LLMs

Voice agents could be measured, never improved. We closed the loop - two open models talking in native audio - and task success more than doubled.

Read →

For the First Time, We Can See an Audio LLM’s Brain Map and Consciousness Space

 ·  interpretability · audio LLMs · mechanistic interpretability · AI safety

A speech model answers a spoken question in one word. We read its middle layers while it listened - the answer was already there, in words it never says.

Read →
2025

Reasoning Made Audio LLMs Worse. We Flipped the Sign.

 ·  reinforcement learning · audio LLMs · reasoning · test-time scaling

Telling an Audio LLM to think cost it 3.40 points. Rewarding the reasoning process instead earns 3.40 - the same margin, mirrored.

Read →

One Subtraction Ends the Exploration–Exploitation Trade‑off

 ·  reinforcement learning · RLHF · generative models · flow matching

RL post-training runs on one fixed coefficient that must both protect the model and get out of its way. Subtract the advantage - one line.

Read →