Radically Better Reasoning: Elicit's Andreas Stuhlmüller & Jungwon Byun on World Models for Research

2026-06-17 · Source: The Cognitive Revolution · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Advanced, extended

Summary

Elicit, led by Andreas Stuhlmüller and Jungwon Byun, is developing trusted reasoning workflows for scientific research, addressing the challenge of powerful yet opaque frontier AI models. Their approach integrates process supervision, domain-specific reasoning primitives, and inspectable "world models" to ensure reliable analysis of evidence, causality, and counterfactuals. Elicit's platform, which uses a domain-specific language to orchestrate agent calls, guarantees consistent application of reasoning processes across large datasets, serving seven of the top 20 life sciences companies for tasks like drug target ranking and regulatory defense. Internally, Elicit employs "The Line," an automated software engineering system delivering 30-50 code changes weekly. The company also explores external world models for continual learning and inspectable knowledge representations, while managing significant token costs (Andreas spends ~\$2,000/week) by dynamically dispatching tasks to appropriately sized models.

Key takeaway

For Research Scientists and Directors of AI/ML integrating AI into high-stakes scientific research, prioritize platforms that offer transparent, systematic reasoning over opaque "black box" outputs. Your teams should adopt tools like Elicit that implement process supervision and explicit world models, ensuring AI-generated conclusions are verifiable and consistently derived from evidence. This approach mitigates risks associated with models that are "easy to push around," fostering trust and improving the overall quality of decision-making.

Key insights

Elicit ensures trusted AI reasoning in scientific research via process supervision, domain-specific primitives, and inspectable world models.

Principles

• Process supervision validates AI reasoning steps, not just final answers.
• Evidence quality assessment should prioritize methodology over metadata.
• Explicit world models enable inspectable, continual AI learning.

Method

Elicit employs a domain-specific language (DSL) to orchestrate reasoning primitives, enabling frontier models to dynamically generate structured workflows guaranteed for systematic execution.

In practice

• Conduct systematic literature reviews with guaranteed process consistency across large datasets.
• Utilize AI for rigorous ranking of drug targets and justifying drug launch strategies.
• Automate software engineering for bug fixes and simple features via iterative AI workflows.

Topics

• AI for Science
• Reasoning Workflows
• Process Supervision
• World Models
• Life Sciences Research
• Automated Software Engineering

Code references

• oughtinc/ice
• badlogic/pi-mono
• danielmiessler/Personal_AI_Infrastructure

Best for: Executive, AI Architect, AI Product Manager, AI Scientist, Research Scientist, Director of AI/ML

Related on AIssential

• A Three-Layer Framework for AI in Scientific Discovery — True scientific discovery in AI hinges on qualitative reasoning for model formation, transcending mere search or execution.
• Interpretable and Explainable Surrogate Modeling for Simulations: A State-of-the-Art Survey and Perspectives on Explainable AI for Decision-Making — Integrating XAI with surrogate models enhances transparency and insight extraction from complex simulations.
• The uncritical adoption of AI in science is alarming — we urgently need guard rails — Uncritical AI adoption in science risks reduced research quality, narrowed scope, and deskilling future scientists.
• Teaching Language Models to Forecast Research Success Through Comparative Idea Evaluation — Fine-tuned 8B-parameter language models can objectively forecast research idea success by comparing empirical outcomes, outperforming larger frontier models.
• It's not the Language Model, it's the Tool: Deterministic Mediation for Scientific Workflows — Deterministic tools orchestrated by language models ensure reproducibility in scientific workflows, unlike direct code generation.
• Reasoning4Sciences: Bridging Reasoning Language Models to All Scientific Branches — Reasoning Language Model adoption is uneven across scientific disciplines, creating a significant research productivity gap.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Cognitive Revolution.