Dual-Agent Framework for Cross-Model Verified Translation of Natural-Language Protocols into Robotic Laboratory Platform¶

🕒 Published (v1): 2026-06-18 11:43 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

A dual-agent pipeline converts free-text biological laboratory protocols into executable robotic commands via an LLM-based Parser Agent, a deterministic rule-based mapping engine, and a heterogeneous LLM Validation Agent that runs cross-model verification with a structured self-correction loop. The framework is evaluated with a 7 Parser × 3 Validator sweep on 30 ELISA protocols and validated end-to-end on a physical robotic platform (KIMM BioForge-1) running a Bradford assay.

Problem¶

Biological protocols are written in natural language, but robotic laboratory platforms require deterministic device-level control commands. Single-LLM approaches suffer from two failure modes: execution hallucinations (incorrect commands that violate equipment specifications) and reasoning hallucinations (missing steps, ordering errors, misinterpreted conditions). No prior work systematically addressed NL-to-command translation for microplate-based experiments, which add complexity through simultaneous well mapping, multi-channel dispensing, sample–reagent combinatorics, and tip-management constraints.

Method¶

The framework is a three-stage pipeline:

Parser Agent — an LLM prompted with constrained output constraints produces a structured XML-like intermediate representation using tags (<KIT_TITLE>, <MANUAL>, <INSTRUMENT>), separating user-performed from instrument-executed steps. Seven candidate models (GPT-5, GPT-4.1, o4-mini, GPT-4.1-mini, GPT-4.1-nano, Llama-4-Maverick, Llama-3.3-70B) are evaluated as the Parser.
Rule-Based Mapping Engine — a deterministic compiler that ingests the structured protocol and expands each semantic step into device-level commands by applying rules for: coordinate system mapping (2D well grid → instrument coordinates), liquid handling (aspiration/dispensing/washing), tip management (cross-contamination prevention + 4-channel grouping), plate transport routing, and environmental control (incubation/shaking/temperature). A single semantic step can expand to dozens of commands (e.g., "Wash 4×" → 34 commands; "Read OD" → 12 commands).
Validation Agent — a heterogeneous LLM (Claude Sonnet 4.6, from a different model family than the Parser) that receives the original protocol, structured protocol, and command sequence simultaneously and verifies three criteria: completeness (no missing steps), parameter accuracy (quantitative values correctly converted), and execution order (tip/plate sequencing compliance). On failure, it emits structured feedback (error type, location, correction guidance) that drives a self-correction loop up to a predefined iteration limit.

Key Contributions¶

First framework explicitly targeting NL-to-command translation for microplate-based biological experiments with full multi-instrument coordination (plate transport + liquid handling + analytical instruments).
Hybrid LLM + deterministic architecture: LLM handles semantic variability; rule engine enforces physical constraints deterministically, eliminating execution hallucinations at the command level.
Cross-model (heterogeneous) verification architecture that structurally avoids self-verification bias by separating the generator and validator model families.
Structured two-layer ground truth (structured protocol GT + device-level command GT) enabling independent evaluation of the parsing and mapping stages.
End-to-end physical validation: Bradford assay protein quantification executed autonomously on KIMM BioForge-1.

Results¶

The paper text is truncated before quantitative results are reported. From the experimental design: - Evaluated on 30 ELISA protocols sampled from 1,000 collected commercial kit manuals. - 7 × 3 cross-product sweep (Parser model × Validator model) was conducted to measure translation accuracy and pass rate as a function of model scale and Validator type. - Accuracy–latency trade-off was compared between the proposed rule-based mapping and an LLM end-to-end direct mapping baseline. - Specific accuracy figures, pass rates, and latency numbers are not present in the provided text.

Limitations¶

Evaluation dataset is limited to ELISA protocols; generalizability to other assay families (PCR, Western blot, etc.) is not demonstrated.
Framework and rule engine are tightly coupled to the KIMM BioForge-1 platform; porting to other robotic systems requires re-authoring the rule base.
Self-correction loop has a hard iteration cap; protocols that cannot be corrected within the limit fail silently or require human intervention (exact cap not stated in the provided text).
Ground truth labeling required multiple expert reviewers with a consensus resolution process, making evaluation expensive to scale.
The Parser Agent requires a structured constrained prompt tailored to the tag schema; adapting to new protocol types may require prompt re-engineering.

Relevance to Harnesses / Meta-Harnesses¶

This paper is a direct instance of a multi-stage agentic harness with a self-correction meta-loop: a generator agent, a deterministic compilation layer, and a heterogeneous verifier agent are composed into a pipeline where the verifier can rewind and re-trigger upstream stages. The cross-model verification pattern — deliberately using a different model family for the validator to avoid self-verification bias — is an architectural principle directly applicable to general-purpose meta-harness design. The separation of an uncertain LLM stage from a deterministic rule-engine stage formalizes a reliability pattern (LLM for semantics, rules for correctness) that recurs in robust harness architectures. The structured feedback loop (error type + location + correction guidance → regeneration) is a concrete implementation of the reflect-and-retry loop central to harness reliability research.