LLM-ACES
NeurIPS 2026Closed-Loop Discovery of Dynamical Systems with LLM-Guided Adaptive Search
Nikhil Abhyankar · Virginia Tech
Building Safe LLM Agents for Open-Ended Scientific Discovery
LLMs possess scientific domain knowledge that can be leveraged to help scientists speed up the discovery process. We build agents that ground every proposal in this knowledge, validate it against real evidence, and run the next experiment when the data cannot yet tell competing hypotheses apart.
Explore the projectsThe Scientific Discovery Process
-
01
Propose candidates
The LLM drafts candidate hypotheses as equations, programs or crystal structures, guided by domain knowledge and priors.
-
02
Validate against evidence
Every proposal is executed, fitted and scored on data, simulators or property oracles. Nothing is accepted on the model’s word.
-
03
Choose the next experiment
The agent queries where its competing hypotheses disagree most, so each new observation rules something out.
-
04
Refine belief
Successes, failures and confidence checks update the agent’s belief about which hypotheses hold, steering later rounds away from spurious fits and memorized answers.
Accepted Publications
LLM-AutoSciLab
NeurIPS 2026Closed-Loop Scientific Discovery via Active Experimentation with LLMs
LLEMA
ICLR 2026Evolutionary Search with LLMs for Multi-Objective Materials Discovery
LLM-FE
TMLR 2026Automated Feature Engineering for Tabular Data with LLMs as Evolutionary Optimizers
About
Nikhil Abhyankar
PhD, CS at Virginia Tech
Nikhil works on making LLM-based systems reliable enough for decision-making over structured data and scientific workflows. His research sits where LLM reasoning, tabular data and AI for science meet, with a focus on methods and benchmarks that expose real failure modes and improve end-to-end performance.
- Advisor
- Dr. Chandan K. Reddy
- Research Interests
- Venues
- NeurIPS · ICLR · COLM · NAACL · TMLR