Breaking

HKUST Releases Curated Repo of LLM Research for Science

July 1, 2026 at 08:23 EDT

Research & Papers
AI Agents
Open Source

HKUST's KnowComp group has released a GitHub repository, "Awesome-LLM-Scientific-Discovery," that systematically compiles research papers and tools applying large language models (LLMs) to scientific discovery. Tied to the group's survey paper, it frames LLMs along a spectrum from "mere tools" to "autonomous scientists."

EMNLP 2025 · HKUST KnowComp

From Automation to Autonomy: Mapping LLMs in Scientific Discovery

A curated survey classifies the role of large language models across the research process into three levels of autonomy — from supervised tool, to analyst, to fully autonomous scientist.

THREE LEVELS OF AUTONOMY

Column height rises with the LLM's independence from human oversight.

Level 1

LLM as Tool

Specific tasks under human supervision

Level 2

LLM as Analyst

Autonomy in complex information processing

Level 3

LLM as Scientist

Autonomous execution of major stages

Dashed block = frontier: full Level-3 autonomy remains rare in practice today.

EXAMPLE SYSTEMS BY LEVEL

LEVEL 1 · TOOL

SCIMON · LitLLM · SciAgents · BioPlanner · SciCode

LEVEL 2 · ANALYST

MLAgentBench · MLR-Copilot · DiscoveryBench · DAgent

THE RESEARCH LOOP LLMs ARE ENTERING

Literature
Review

→

Idea
Generation

→

Hypothesis
Validation

→

Experiment
& Iterate

WHAT'S WORKING

Cuts researchers' working time on search and summarization
Offers multiple perspectives on a problem
Active progress in hypothesis-generation benchmarks and automated review

OPEN CONCERNS

Heavy token consumption
Accuracy risks in specialized domains and limited databases
Hallucination — content not grounded in fact — is especially critical for science

Autonomy is rising — but the human is still in the loop.

Full Level-3 systems that entrust entire research stages to an LLM remain limited; outputs still require human verification.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.