Open to PhD · Fall 2027
Suresh Raghu
Independent AI Researcher · AI Lead at VFS Global
I'm an independent researcher studying reliable reasoning and scientific agents, with a focus on when and why language and vision-language models fail, and how better evaluation, tools, and environments can make them more reliable. My work spans uncertainty quantification, multimodal reasoning, and agentic systems, with a first-author paper accepted at NeurIPS 2026 and additional work accepted at the NeurIPS 2026, ICML 2026, and COLM 2026 workshops.
More recently, I've been working on scientific agents for small-molecule drug design, studying how harness design, test-time computation, and learning from long trajectories affect agent performance.
By day, I lead AI Engineering at VFS Global, building production AI systems deployed across 110+ countries.
Always happy to talk research or engineering. Feel free to reach out!
News
New blog post: Is Human Taste Overrated in Harness Engineering? We explore what automated harness search can improve in scientific agents, and where task-specific representation and state still matter.
Happy to share that Confidently Wrong, Silently So: Auditing Undetectable Failures of a Deployed On-Device Language Model has been accepted to the NeurIPS 2026 Workshop on On-Device Intelligence (ODI 2026)!
Happy to share that Proper Scoring Rules for Agentic Uncertainty Quantification has been accepted to NeurIPS 2026 main conference! See y'all in Sydney(?)!
New preprint: Confidently Wrong, Silently So: Auditing Undetectable Failures of a Deployed On-Device Language Model. We audit a deployed on-device language model, finding confident failures that user-visible signals cannot reliably detect while a black-box consistency wrapper substantially recovers reliability.
Happy to share that one of my papers, Don't Blink: Evidence Collapse during Multimodal Reasoning, has been accepted to the Actionable Interpretability and Scientific Understanding of Foundation Models workshops at the Conference on Language Modeling (COLM) 2026!
Two papers accepted at the Failure Modes in Agentic AI (FAGEN) workshop at ICML 2026: SELFDOUBT and Proper Scoring Rules for Agentic Uncertainty Quantification!
Proper Scoring Rules for Agentic Uncertainty Quantification was accepted as a poster at the Combining Theory and Benchmarks (CTB) workshop at ICML 2026. See y'all in Seoul!
New preprint: Proper Scoring Rules for Agentic Uncertainty Quantification. We introduce the Trajectory Proper Score (TPS), a predictor-agnostic family of strictly proper, trajectory-level scoring rules for evaluating uncertainty in LLM agents.
New preprint: SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio. We introduce the Hedge-to-Verify Ratio (HVR), a single-pass uncertainty signal for reasoning LLMs that outperforms Semantic Entropy at about 10x lower inference cost.
New preprint: Don't Blink: Evidence Collapse during Multimodal Reasoning. We identify evidence collapse, a decay of visual grounding during multimodal reasoning that text-only uncertainty signals cannot detect.
Published an article on prompt ordering in vision-language models: Why Your VLM Prompts Are Backwards (And How to Fix It)
Joined AI @ VFS as lead AI engineer
Selected publications
All researchProper Scoring Rules for Agentic Uncertainty Quantification
Accepted at NeurIPS 2026. Also accepted at the CTB and FAGEN Workshops at ICML 2026.
SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio
Accepted as a poster at the FAGEN Workshop at ICML 2026
Don't Blink: Evidence Collapse during Multimodal Reasoning
Accepted at the Actionable Interpretability and Scientific Understanding of Foundation Models Workshops at COLM 2026
Confidently Wrong, Silently So: Auditing Undetectable Failures of a Deployed On-Device Language Model
Accepted at the NeurIPS 2026 Workshop on On-Device Intelligence (ODI 2026).