hey, I'm Armaan!I'm an MS student in computer science at UMass Amherst, studying the internals of language models — what they represent, and where that breaks down under pressure.
I co-founded curations.club, and previously worked at Google and Goldman Sachs.
Outside research, I'm usually making art or building something that feels a little more human. A few more things are below.
Selected for the Singapore AI Safety Fellowship, working with Tan Zhi-Xuan.
Our papers on position-confounded reward optimization and attention heads behind object hallucination in LLaVA were accepted to the AI Measurement Science and Actionable Interpretability workshops at COLM 2026.
Joined the Artificial Scientific Intelligence Lab at NUS as a research intern with Prof. Dianbo Liu.
Joined UMass's Resource Bounded Reasoning Lab to study influence in multi-agent LLM systems.
Started my MS in Computer Science at UMass Amherst.
I'm interested in how models represent beliefs about the people they talk to, and how their behavior holds up under distribution shift and optimization pressure — across interpretability, alignment, and RL.
AI Measurement Science (AIMS) Workshop at COLM · 2026 · * Equal contribution
A controlled study of goal misgeneralization under GRPO: models learn an answer-position shortcut from a correct but confounded reward, while their reasoning can remain correct even as their final answers follow the shortcut.
Actionable Interpretability (AIW) Workshop at COLM · 2026 · * Equal contribution
Causally localizes object hallucination in LLaVA-1.5-7B to a sparse set of attention heads, then combines targeted LoRA and DPO fine-tuning with inference-time grounding to reduce caption-level hallucination on COCO.
Layerwise linear probes on LLaMA-3-8B show the model encodes evolving beliefs about user expertise — decodable from early layers and shifting across a multi-turn dialogue. Activation patching didn't move outputs, suggesting the representation is distributed, not localized.
Fine-tuning an instruction-tuned model on tone-matched data (conspiratorial vs. passive-aggressive vs. neutral) raised misaligned outputs by 5–8% under conspiratorial framing across 3,200 generations per condition — evidence that tone is a surface-form alignment vulnerability.
An AI-native social platform where people connect through curated taste profiles, powered by embedding-based semantic retrieval. 600+ users; #2 Product of the Day on Product Hunt.
A merch store for Indian creators, born from hostel-room banter and incubated at Nirmaan, IIT Madras.
Books I recommend and art I make.
