Research map / Large language models / Agents, code and math
Hierarchically Encapsulated Representation for Protocol Design in Self-Driving Labs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a hierarchically encapsulated, DSL-based representation (instance actions, operation interfaces, product-flow units) auto-generated from protocol corpora via non-parametric Bayesian modeling, which LLM designers use to plan, modify and adjust experimental protocols.
Problem
LLMs fail to fully elicit protocol design capability for self-driving laboratories, producing protocols that mirror existing ones and losing configuration details, because experimental knowledge is flat and unorganized.
Approach
A multi-faceted, multi-scale representation with three levels (instance actions with attributes, operation-centric function abstraction, product-flow-centric model abstraction) encapsulated as DSLs, with the dual operation/product-flow views enabling reciprocal verification; the abstractions are learned automatically from domain protocol corpora using a DPMM with Gaussian process value modeling.
Key results
- EE+ achieves IoU(Op) 0.607 and IoU(Prod) 0.605 on planning, versus 0.143 and 0.040 for the flatten-baseline FB.
- EE+ outperforms the instance-based II designer (t(278)=24.493, p<.0001) across protocol design tasks.
- Generated DSLs differ strongly across domains; e.g., Genetics has 304 operations and Ecology 100 operations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add product-flow view to the operation-centric DSL with external verification (ablation) | EE · Protocol design, all tasks and domains | Six-dimensional consistency scores: t(278)=8.007, p<.0001 | Helped |
| Add product-flow view to the operation-centric DSL with internal prompting (ablation) | EI · Protocol design, all tasks and domains | Six-dimensional consistency scores: t(278)=8.397, p<.0001 | Helped |
| Use dual-view DSL rather than instance-action prompting (ablation) | Protocol design · Protocol design, all tasks and domains | Six-dimensional consistency scores: t(278)=24.493, p<.0001 | Helped |
| Use operation-centric representation rather than instance-action representation (ablation) | Internal-prompt designer · Protocol design, all tasks and domains | Six-dimensional consistency scores: t(278)=15.259, p<.0001 | Helped |
| Use instance-action prompting rather than retrieval on original protocols (ablation) | Protocol design · Protocol design, all tasks and domains | Six-dimensional consistency scores: t(278)=8.340, p<.0001 | Helped |
Limitations
- Representation is highly dependent on domain-specific knowledge and manual DSL crafting is costly
- Testing set is small (140 protocols) and results have large standard errors
- Verification at DSL level is insufficient for true certification; human experts remain in loop
Experimental setup
Datasets: domain-specific corpora of protocols from top journals (Genetics, Medical, Bioengineering, Ecology)
Benchmarks: protocol design testing set (planning, modification, adjustment tasks)
Metrics: IoU(Op) IoU(Prod) IoU(Dev) Sim(Exec) Sim(Goal) Sim(Param)
Baselines: Flatten-Baseline (RAG over NL protocols) Instance-Baseline Instance-Internal (ISA prompting) Encapsulated-Internal Encapsulated-External
Builds on: retrieval-augmented generation chain-of-thought
Closest papers by meaning
- Expert-level protocol translation for self-driving labs (NeurIPS 2024 · similarity 0.92)
- AI-Researcher: Autonomous Scientific Innovation (NeurIPS 2025 · similarity 0.83)
- LLM-Powered Synthetic Environments for Self-Driving Scenarios (AAAI 2024 · similarity 0.83)
- Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.82)
- LaMPilot: An Open Benchmark Dataset for Autonomous Driving with Language Model Programs (CVPR 2024 · similarity 0.82)
- AutoSciLab: A Self-Driving Laboratory for Interpretable Scientific Discovery (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.