FirstResearch: Auditable Question Formation for LLM Scientific Discovery Agents

2026-07-06 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Research Methodology & Innovation · Depth: Expert, quick

Summary

FirstResearch introduces a novel framework for scientific LLM agents, addressing the challenge of auditing research questions proposed by these systems. Its core innovation is the "Research Question Certificate," a structured artifact that records primitive definitions, assumptions, a mechanism model, a tension, a falsifiable hypothesis, a minimal decisive test, and a failure update rule. This certificate makes proposed questions inspectable before execution. Evaluated on ten LLM-agent research topics, FirstResearch significantly outperformed controlled prompt-level baselines, including those inspired by AI co-scientist and Agent Laboratory. Under a Gemini-2.5-Flash independent-judge rescore, FirstResearch achieved a score of 4.86/5, compared to 4.38/5 for the strongest baseline, with a Pearson agreement of 0.865. An ablation study confirmed the certificate's critical role, with certificate-only scoring reaching 4.90/5 under DeepSeek. While preliminary and using LLM judges, these results suggest explicit derivation constraints enhance the auditability of LLM-generated scientific questions.

Key takeaway

For AI Scientists and Research Scientists developing LLM agents for scientific discovery, adopting structured question formation frameworks is crucial. If your goal is to ensure the auditability and scientific rigor of LLM-generated research questions, consider implementing explicit derivation constraints via structured certificates. This approach, exemplified by FirstResearch's 4.86/5 performance, can significantly improve the inspectability and reliability of proposed hypotheses, even if initial evaluations use LLM judges.

Key insights

FirstResearch uses structured "Research Question Certificates" to make LLM-generated scientific questions auditable and inspectable.

Principles

• Explicit derivation constraints enhance auditability.
• Structured certificates improve LLM output inspectability.

Method

FirstResearch forms scientific questions by generating a Research Question Certificate. This certificate details primitive definitions, assumptions, a mechanism model, tension, a falsifiable hypothesis, a decisive test, and a failure update rule.

In practice

• Implement structured certificates for LLM outputs.
• Integrate falsifiable hypotheses into LLM prompts.

Topics

• LLM Agents
• Scientific Discovery
• Auditable AI
• Research Question Certificate
• Question Formation
• Derivation Constraints

Code references

• louiswang524/FirstResearch

Best for: AI Scientist, Research Scientist

Related on AIssential

• PRL-Bench: A Comprehensive Benchmark Evaluating LLMs' Capabilities in Frontier Physics Research — LLMs struggle with autonomous, long-horizon physics research, primarily due to domain knowledge gaps and reasoning instability.
• OpenNovelty: An LLM-powered Agentic System for Verifiable Scholarly Novelty Assessment — OpenNovelty uses an LLM-powered agentic system for verifiable, evidence-based scholarly novelty assessment.
• Benchmarking LLM Agents on Meta-Analysis Articles from Nature Portfolio — LLM agents critically fail at criterion-based screening in scientific meta-analysis, despite high retrieval recall.
• Towards Generalist Agents for Accelerating Scientific Discovery171 — LLMs can accelerate scientific discovery by generating and refining hypotheses within agentic, iterative workflows.
• SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones? — Current LLMs show an optimism bias, failing to reliably judge research proposal soundness, making them unsuitable as first-gate evaluators.
• InfiniteScienceGym: An Unbounded, Procedurally-Generated Benchmark for Scientific Analysis — Procedurally generated benchmarks offer controlled, scalable evaluation for LLM scientific reasoning and abstention.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.