- The paper presents a dual-agent framework that translates natural-language protocols into device-level commands using semantic parsing and deterministic rule-based mapping.
- It demonstrates robust iterative correction with cross-model verification to recover accuracy in parameter extraction and command sequencing for ELISA protocols.
- Results indicate that the hybrid architecture effectively optimizes the accuracy-latency trade-off, enhancing reliability in robotic laboratory automation.
Motivation and Challenge
Automated laboratory platforms increasingly rely on robotic systems to execute high-throughput microplate-based experiments, yet the translation of natural-language protocols into device-level control commands remains a significant obstacle. This semantic gap is exacerbated by the variability and expressive diversity of biological protocols, which are readily written in unstructured text but require conversion into highly structured commands respecting instrument constraints, well mapping, reagent combinations, and operational parameters. Prior work has demonstrated limited direct translation capability, particularly for microplate-based workflows, and highlighted frequent execution and reasoning hallucinations in end-to-end LLM-driven approaches. The problem is compounded when complex tasks such as ELISA or protein quantification demand robust handling of procedural variants, step sequencing, and parameter extraction.
Dual-Agent Architecture
The framework introduces a dual-agent pipeline composed of a Parser Agent and a Validation Agent, augmented by a deterministic rule-based mapping engine. The Parser Agent leverages a LLM to convert natural-language protocols into a structured representation that separates user-performed and instrument-executed steps, applying a constrained prompt to minimize extraction errors and ordering inconsistencies. Each step is tagged according to operational domain (<MANUAL>, <INSTRUMENT>), enabling downstream restriction to instrument-executable actions.
The rule-based mapping engine transforms this intermediate representation into device-level command sequences, employing predefined rules for coordinate mapping, liquid handling, tip management, plate transport, and environmental control. Critical operational constraints, such as multi-channel pipetting optimization, tip usage based on contamination policy, and sequencing for environmental steps, are enforced deterministically. This layer absorbs the granularity gap between semantic units and execution units, providing reproducibility and reliability that LLM-only approaches lack.
The Validation Agent, heterogeneous by design, cross-verifies command sequences for completeness, parameter accuracy, and execution order against the original protocol. Upon error detection, structured feedback triggers a self-correction loop, regenerating the structured protocol and command sequence iteratively until verification passes or attempt limits are reached. This generator–verifier separation reduces self-confirmation bias and supports robust error recovery.
Experimental Evaluation
Protocol Dataset and Ground Truth Construction
A curated dataset of 30 ELISA protocols, spanning wide procedural and expressive variability, was selected from 1,000 candidates. Dual-layer ground truth construction—structured tags and device command sequences—enabled disaggregated evaluation of semantic and execution accuracy. Multiple expert reviewers ensured grounding, consensus, and reproducibility in labeling.
Parser and Validator Model Sweeps
Seven Parser models (GPT-5, GPT-4.1, o4-mini, GPT-4.1-mini, GPT-4.1-nano, llama-4-maverick, llama-3.3-70b) and three Validators (Claude Sonnet 4.6, GPT-5, llama-4-maverick) formed a 21-combination sweep. All models used identical prompt templates and call parameters for isolation. Models ranged from cloud-hosted (OpenAI) to on-premise (Meta open-weight).
Metrics
Parameter Accuracy measured exact matches across structured and command-sequence layers, penalizing over-generation and omission; Pass Rate decomposed outcome frequencies across attempts; Latency assessed processing time from LLM structuring through mapping, excluding validation stage latency.
Results
Iterative Correction and Validator Selection
Self-correction effectiveness was found to depend on Validator verification capability rather than model heterogeneity alone. Claude Sonnet 4.6, as an external model with strong critical verification capability, induced substantial accuracy recovery across regeneration attempts, even for mid- and small-sized Parsers—parameter accuracy rose by up to 0.4 for llama-3.3-70b and 0.7–0.8 for GPT-4.1-mini and llama-4-maverick. In contrast, GPT-5 and llama-4-maverick as Validators produced little recovery regardless of model size, their pass judgments being overly liberal. Self-validation combinations demonstrated near-zero correction effect. This result aligns with established findings on self-confirmation bias in multi-agent systems and highlights the necessity of explicit cross-model verification with strong external Validators.
Pass Rate Decomposition
Cross-verification with Claude Sonnet 4.6 consistently reduced Final Fail rates and recovered failing cases in 2nd and 3rd attempts. Large Parsers (GPT-5, GPT-4.1) already had high 1st Pass rates; smaller Parsers saw cumulative pass rates rise markedly post-regeneration, achieving practical translation accuracy for automation without the need for retraining or fine-tuning.
Accuracy–Latency Trade-off
The baseline trade-off revealed that larger Parsers achieve high accuracy at higher latency, while smaller Parsers are faster but less accurate. Applying iterative correction shifted the accuracy–latency Pareto frontier upward for smaller Parsers; they approached large-model accuracy while maintaining low processing times. GPT-4.1 emerged as a Pareto-dominant option, attaining 0.93 accuracy at one-quarter GPT-5's latency, suitable for mid-throughput deployment. On-premise models benefit privacy- and latency-sensitive environments, particularly when inference drives physical robot operations.
Hybrid Architecture vs. LLM End-to-End Mapping
The rule-based mapping engine outperformed direct LLM mapping in both accuracy and latency with the GPT-5 Parser fixed. Accuracy loss in the end-to-end condition was chiefly attributable to execution hallucinations—incorrect tip rules, coordinate errors, and step misordering—validating the hybrid design's deterministic enforcement of physical constraints. System modularity enables adaptation to other platforms by swapping rule sets, supporting generalizability and standardization.
End-to-End Demonstration
Microplate-based liquid handling and Bradford assay demonstrations, executed with GPT-5 × Claude Sonnet 4.6, confirmed successful translation and execution from protocol parsing through robotic operation. Parameter extraction, sequencing, and operational outcomes matched ground truth and validated standard curves, establishing reliable end-to-end play for typical experimental designs. The human-in-the-loop interaction paradigm transitions researchers from script writing to supervision and design, allowing cognitive emphasis on hypothesis generation over syntactic translation.
Implications and Future Directions
The dual-agent framework substantiates a practical path for narrowing the semantic gap in laboratory automation, supporting reproducibility and adaptability. The architecture achieves model-agnostic accuracy improvement for mid- and small-sized models via inference-time correction, enabling cost- and latency-effective deployment. Explicit cross-model verification mitigates hallucination propagation and self-confirmation bias, supporting the safety of LLM-driven robot agents.
Future work should pursue closed-loop integration where experimental feedback refines subsequent actions, incorporation of assay-specific ontologies for scientific validity, and extension to other experimental modalities. Refining protocol datasets and evaluation criteria into public benchmarks will foster standardization and comparative development in autonomous laboratory research.
Conclusion
The proposed dual-agent protocol translation framework, combining semantic LLM-based structuring, deterministic rule-based mapping, and heterogeneous cross-model verification, addresses the semantic gap between natural-language protocols and automated laboratory execution. Validator selection and iterative correction recover the translation accuracy of mid- and small-sized models to levels matching highly accurate cloud models. Rule-based mapping decisively outperforms direct LLM command translation. The modular, hybrid architecture admits adaptation to diverse deployment requirements and supports end-to-end operation in real-world experimental contexts. This work provides a formal mechanism for reliable, efficient translation in self-driving laboratories and presents foundational strategies for future AI-driven automation systems (2606.20120).