a.k
hey! i'm abhinav.
I'm a researcher at Lossfunk ↗, an AI research lab in Bangalore. Most of my work is alignment research on LLMs and agents, with some AI4Science on the side. Most recently, our paper on authority bias ↗ was accepted at NeurIPS 2026.
sometimes i write things down.
some things i've worked on:
talks & slides.
Gleipnir under pressure: stress-testing efficient monitors for AI control
Reading a monitor's decision straight from its logits makes near-frontier monitoring ~200× cheaper. How we got there, and whether a monitor that cheap judges what an agent does or the story told around it.
Authority bias: the early results
Where authority bias started: Llama-3.1-8B defers more to authority-coded sources than to the user, in the order expert > note > user > someone online.
please say hi!
I check my mail and twitter ↗ daily, so drop a message if you'd like to chat about alignment, interpretability or anything else. Currently based in Bangalore, India.
abku2504 [at] gmail [dot] com