Samad Syed
I’m a UW-Madison Computer Sciences alum working in AI Systems.
Check out my LinkedIn to see some things I’ve done.
Check out my Twitter for the latest on what I’m working on.
I have a ton of open source projects to check out on my GitHub, if they help you in any way please leave me a star :)
Send me an email if you’d like to chat: firstnamealastname@hotmail.com
news
| Oct 02, 2026 | StudyLoop was accepted at two NeurIPS 2026 workshops: Who Verifies the Agents? and Resource-Aware Agentic AI (RAAAI). |
|---|
selected research
-
StudyLoop: Can an AI Agent Study a Codebase?
Workshop paper Accepted at NeurIPS 2026 workshops Who Verifies the Agents? and Resource-Aware Agentic AI (RAAAI)
Can an agent study a codebase before it sees the questions, and does that preparation pay off? A controlled 540-trajectory StudyBench experiment comparing no study, a single-pass cheatsheet, and an iterative study loop across repository-search budgets and inference-time thinking.
-
Evaluating and Addressing Multimodal Hallucination
Course project CS 639, University of Wisconsin–Madison, 2026
Do multimodal LLMs actually use the audio, visual, and text inputs they claim to? A counterfactual-ablation framework on Qwen2.5-Omni-7B and Gemma-3n-E2B-IT finds strong modality bias and unreliable self-attribution: adversarial transcripts flipped 69% of Qwen's previously correct answers.
-
INSCONE: Unknown-Aware Detection of LLM-Generated Text via Informed Wild Data
Course project CS 639, University of Wisconsin–Madison, 2026
Detects LLM-generated text by treating human writing as semantic out-of-distribution data and unseen LLM families as covariate shift. Improves zero-shot FPR95 by 6.1 points over a competitive baseline on RAID, and introduces RAID+, an evaluation set regenerated with contemporary frontier models.
latest posts
| Aug 17, 2026 | StudyLoop |
|---|