BioProAgent: Neuro-Symbolic Grounding for Constrained Scientific Planning

Yuyang Liu; Jingya Wang; Liuzhenghao Lv; Yonghong Tian

cs.AIMar 1, 2026

BioProAgent: Neuro-Symbolic Grounding for Constrained Scientific Planning

Authors: Yuyang Liu, Jingya Wang, Liuzhenghao Lv, Yonghong Tian

Organizations: School of AI for Science, Peking University · School of Electronic and Computer Engineering, Peking University · School of Computer Science, Peking University

Abstract

Large language models (LLMs) have demonstrated significant reasoning capabilities in scientific discovery but struggle to bridge the gap to physical execution in wet-labs. In these irreversible environments, probabilistic hallucinations are not merely incorrect; they can cause equipment damage or experimental failure. We propose BioProAgent, a neuro-symbolic framework that anchors probabilistic planning in a deterministic Finite State Machine (FSM). We introduce a State-Augmented Planning mechanism that enforces a rigorous Design-Verify-Rectify workflow, ensuring hardware compliance before execution. Furthermore, we address the context bottleneck inherent in complex device schemas by Semantic Symbol Grounding, reducing token consumption by ~6* through symbolic abstraction. In the extended BioProBench benchmark, BioProAgent achieves 95.6% physical compliance (compared to 21.0% for ReAct), demonstrating that neuro-symbolic constraints are essential for reliable autonomy in irreversible physical environments. Code: https://github.com/YuyangSunshine/bioproagent | Website: https://yuyangsunshine.github.io/BioPro-Project.

Figures & tables

Figure 1: Scientific Reasoning vs. Automation Executability . Vanilla LLMs occupy the Theoretical Zone (high logic, weak code), while Neural Agents (like ReAct) often generate better code but fail in scientific reasoning. BioProAgent achieves Trustworthy Autonomy with superior performance in both dimensions.

Figure 2: Overview of BioProAgent. (1) Cognitive Memory utilizes Symbolic Grounding $\Phi$ to manage context efficiently; (2) Neural Planner $\pi_{\theta}$ is grounded in a Design-Verify-Rectify FSM $\Delta(\sigma)$ ; (3) Hierarchical Verification ( $\mathcal{K}_{s},\mathcal{K}_{p}$ ) acts as a safety interlock, enforcing physical compliance by deterministically triggering rectification.

Signal Condition ( $\sigma_{t}$ )	Priority	Target State ( $s_{t+1}$ )
$(\sigma_{phys}=-1)\land\sigma_{code}$	1	RECTIFY_CODE
$(\sigma_{sci}=-1)\land\sigma_{draft}$	2	RECTIFY_DRAFT
$\neg(\sigma_{draft})\land\sigma_{know}$	3	DESIGN_DRAFT
$\sigma_{draft}\land(\sigma_{sci}=1)$	4	DESIGN_CODE
$\sigma_{draft}\land(\sigma_{sci}=0)$	5	VERIFY_DRAFT
$\sigma_{code}\land(\sigma_{phys}=1)$	6	SUCCESS

Table 1: FSM Priority Decision Matrix. See Appendix C for the complete definition.

Dimension	Evaluator	Key Metrics
Scientific	Semantic Match	$S_{sem}$ ( $\uparrow$ ), ROUGE-L ( $\uparrow$ )
Scientific	LLM-as-a-Judge	$C_{s}$ ( $\uparrow$ )
Physical	Rule Engine	$S_{code}$ ( $\uparrow$ ), $C_{p}$ ( $\uparrow$ )
Physical	Ground Truth	$Acc_{seq}$ ( $\uparrow$ ), $Acc_{param}$ ( $\uparrow$ )
Efficiency	System Logs	$Succ.$ ( $\uparrow$ ), Tokens ( $\downarrow$ ), Loop Rate ( $\downarrow$ )

Table 2: The BioProAgent Evaluation Framework. Metrics are categorized by dimension, improvement direction (

\uparrow

\downarrow

), and evaluation source.

		Subset A: Scientific Reasoning				Subset B: Hardware Execution
Method	Backbone	ROUGE-L $\uparrow$	$S_{sem}\uparrow$	$C_{s}{\color[rgb]{0,1,0}\dagger}\uparrow$	Time (s) $\downarrow$	$S_{code}\uparrow$	$C_{p}\uparrow$	$Acc_{param}\uparrow$
Direct	GPT-4o	0.107	0.202	0.189	13.8	0.590	0.995	0.295
Direct	Gemini-3-Flash	0.130	0.247	0.322	12.1	0.576	0.996	0.287
Direct	DeepSeek-V3	0.123	0.260	0.285	52.1	0.495	0.995	0.205
Direct	Qwen3-30B-a3b	0.124	0.249	0.201	20.0	0.509	0.988	0.222
Direct	Kimi-k2-Instruct	0.097	0.271	0.286	83.1	0.499	0.971	0.215

Table 3: Main Results on Protocol Drafting (Subset A) and Code Generation (Subset B).

\dagger

: indicates statistical significance compared to all baselines (Biomni, ReAct, AutoGPT, Reflexion) with

p<0.001

(paired t-test).

		Subset C: Long-Horizon Stability			Subset D: Error Correction
Method	Backbone	Succ. $\uparrow$	$Acc_{param}$ $\uparrow$	$C_{p}$ $\uparrow$	Correction Succ. $\uparrow$	$C_{p}$ $\uparrow$	Loop Rate $\downarrow$
ReAct	Gemini-3-Flash	88.9%	0.114	0.217	0.0%	0.000	40.0%
Reflexion	Gemini-3-Flash	33.3%	0.000	0.000	0.0%	0.000	0.0%
AutoGPT	Gemini-3-Flash	66.7%	0.409	0.644	0.0%	0.000	0.0%
BioProAgent	Gemini-3-Flash	100.0%	0.718	0.950	90.0%	0.925	0.0%

Table 4: Main Results on Long-Horizon Stability (Subset C) and Error Correction (Subset D). Bold indicates the best performance within agent-based frameworks.

Panel of Figure 3 — Figure 3: Efficiency and Reliability Analysis. (a) Multidimensional Capabilities: BioProAgent (red) achieves the most expansive capability envelope across five critical dimensions. Note that identical metrics on different axes represent performance in distinct evaluation subsets. (b) Cost-Efficiency: On long-horizon tasks (Subset C), our system reduces token consumption by 82% compared to AutoGPT while maintaining a 100% success rate. (c) Time-Efficiency: BioProAgent demonstrates superior precision with concise execution time.

Panel of Figure 4 — Figure 4: FSM-Driven Self-Correction Trajectories. (a) Physical Correction. (b) Symbol Grounding Repair.

Constraint	Design Violation ( $s_{design}$ )	FSM Rectification ( $s_{rectify}$ )
Physical ( $\mathcal{K}_{p}$ )	centrifuge(speed=‘25000g’)	centrifuge(speed=‘15000g’)
Trace	$\times$ Exceeds rotor limit (15k)	$\checkmark$ Clamped to safety max
Symbol ( $\Phi$ )	transfer(source=‘buffer_x’)	transfer(source=‘trough_1’)
Trace	$\times$ Grounding Fail (No ID)	$\checkmark$ Remapped to registry
Causal	seal_plate(); transfer(...)	transfer(...); seal_plate()
Trace	$\times$ Blocked op (Sealed)	$\checkmark$ Reordered dependency

Table 5: Representative Error Corrections in Subset D. The FSM interceptor enforces the Rectify phase, guiding the agent from violations to valid execution.

Panel of Figure 5 — Figure 5: Trace Analysis (Subset C). Execution divergence on a long-horizon task. TOP: BioProAgent efficiently transitions to coding via FSM. MIDDLE: ReAct fails early due to context overflow. BOTTOM: AutoGPT wastes resources in a retrieval loop.

	Subset A	Subset B	Subset C	Subset D
Variant	$C_{s}$	$C_{p}$	$S_{code}$	$C_{p}$
Full Model	0.591	0.956	0.668	0.925
w/o FSM	0.262	0.049	0.000	0.782
w/o Verification	0.563	0.902	0.648	0.782
w/o Knowledge	0.561	0.889	0.657	N/A
w/o Clarify	0.403	N/A	0.632	N/A

Table 6: Ablation Study of BioProAgent Components.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition	Description
Problem Formulation
$I$	Natural Language Intent	The initial user query describing the experimental objective.
$\mathcal{E}$	Environmental Context	The set of available lab hardware, reagents, and physical conditions.
$\mathcal{P}$	Protocol Space	The universal set of all possible protocol sequences.
$P^{*}$	Optimal Protocol	The target executable sequence that maximizes scientific validity and physical compliance.
$\mathbb{P}_{\theta}$	Generation Probability	The probability distribution of generating the complete protocol sequence $P$ .

Appendix

Table 7: Summary of Notations and Definitions.

Phase I: Design (Generation & Alignment)
Tool Name	Parameters	Function Description
generate_scientific_draft	query , exp_info , knowledge	Generates a natural language experimental protocol draft based on user intent.
align_draft_to_automation	draft , exp_info	Maps the approved draft to specific hardware operations (Intermediate step).
generate_machine_code	aligned_protocol , suggestion	Synthesizes executable JSON machine code. Accepts suggestions from the Rectify phase.
Phase II: Verify (Hierarchical Validation)
reflect_on_protocol	protocol_text , query	( Scientific Reflector $\mathcal{K}_{s}$ ) Reviews the draft for logic/controls via CoT. Output updates the ternary signal $\sigma_{sci}\in\{-1,1\}$ .

Appendix

Table 8: The complete toolset definition. Tools are clustered by the function in the Design-Verify-Rectify (DVR) workflow. Note that verification tools explicitly implement the

\mathcal{K}_{s}

and

\mathcal{K}_{p}

functions to drive FSM state transitions.

Priority	Signal Condition ( $\sigma_{t}$ )	Target State ( $s_{t+1}$ )	Logic Description (DVR Flow)
1	$(\sigma_{phys}=-1)\land\sigma_{code}$	RECTIFY_CODE	[Rectify] Safety Interlock : If Rule Engine ( $\mathcal{K}_{p}$ ) detects a physical violation (Severity: HALT), immediately force a fix.
2	$(\sigma_{sci}=-1)\land\sigma_{draft}$	RECTIFY_DRAFT	[Rectify] Scientific Review : If Reflector ( $\mathcal{K}_{s}$ ) rejects the protocol logic (e.g., missing controls), require revision.
3	$\neg\sigma_{draft}\land\sigma_{know}$	DESIGN_DRAFT	[Design] Knowledge Ready : If knowledge is retrieved but no draft exists, enter drafting mode.
4	$\sigma_{draft}\land(\sigma_{sci}=1)$	DESIGN_CODE	[Design] Code Generation : Synthesize executable machine code directly from the verified and approved scientific draft.
5	$\sigma_{draft}\land(\sigma_{sci}=0)$	VERIFY_DRAFT	[Verify] Draft Review : Evaluate the newly generated protocol draft against scientific principles (Pending state).
6	$\sigma_{code}\land(\sigma_{phys}=1)$	SUCCESS	[Finish] Completion : Code exists and passes all physical validations ( $\mathcal{K}_{p}$ ). Task complete.

Appendix

Table 9: The Priority Decision Matrix defining the FSM transition logic.

(\sigma_{phys}=-1)

denotes a physical validation failure, while

(\sigma_{sci}=1)

denotes a scientific validation success. Note how the [Rectify] interlocks (Rows 1-2) take strict precedence over [Design] steps.

Device Category	Action Primitive	Parameters (Type: Constraint)	Physical Rules (Examples)
Centrifugation	Centrifuge	$\bullet$ speed_g (int: $500--15000$ ) $\bullet$ time (duration: HH:MM:SS) $\bullet$ temp_C (int: $4--40$ ) $\bullet$ brake (enum: off, slow, fast)	1. Speed must not exceed rotor limit ( $V\leq V_{max}$ ). 2. Temperature requires pre-cooling if $T<25^{\circ}$ C.
Thermal Control	Incubate	$\bullet$ temp_C (int: $4--99$ ) $\bullet$ shake_rpm (int: $0--2000$ ) $\bullet$ duration (time)	1. Shaking prohibited if plate is unsealed. 2. Max temp depends on plate material (PS vs. PP).
Liquid Handling	Transfer	$\bullet$ source (ID string) $\bullet$ dest (ID string) $\bullet$ volume_uL (float: $0.5--1000$ ) $\bullet$ tip_type (enum: p20, p300, p1000)	1. Source volume must be $>$ aspiration volume + dead volume. 2. Tip type must match volume range.
PCR Cycling	Thermocycle	$\bullet$ lid_temp (int: $95--105$ ) $\bullet$ stages (list of [temp, time, cycles])	1. Lid temp must be $>$ reaction temp to prevent condensation.

Appendix

Table 10: Representative Action Schemas from the Hardware Registry (

\Omega

). The Physical Rule Engine (

\mathcal{K}_{p}

) validates generated JSON code against these parameter constraints and physical interlocks.

Model Variant	Gemini	Kimi	GPT-4o
BioProAgent-Full	0.591	0.597	0.515
w/o Knowledge	0.561	0.552	0.497
w/o Clarify	0.403	0.421	0.374
w/o FSM	0.262	0.285	0.105
Correlation ( $r$ ) w/ Gemini	1.00	0.84	0.91

Appendix

Table 11: Multi-Judge Consistency Analysis. Comparison of

C_{s}

scores across three distinct LLMs (

N=640

evaluations). GPT-4o confirms the trends with the highest correlation, validating the robustness of our metric.

Benchmark Subset	Scale & Granularity	Evaluation Focus	Key Metrics
A: Scientific Drafting	160 Protocols	Intent Reasoning	• Intent Alignment
(Cross-domain)	(approx. 3k tokens/doc)	Knowledge retrieval	• Scientific Logic Score
B: Automation Conversion	41 Sub-Experiments	Local Code Mapping	• Parameter Accuracy
(Syn-Bio Focus)	(22 Device APIs)	Schema compliance	• API Validity
C: Long-Horizon Exec.	9 Protocols	Global Orchestration	• Global Success Rate
(Full Pipeline)	(547 Total Steps)	State Persistence (Avg. 60.8 steps)	• ID Consistency

Appendix

Table 12: Overview of the Extended BioProBench .

Category	Comparison Pair (vs. Full Model)	Mean Diff.	P-value
Baselines	ReAct	+0.136	$<10^{-10}$
	AutoGPT	+0.162	$<10^{-11}$
	Reflexion	+0.152	$<10^{-11}$
Ablations	w/o Knowledge	+0.030	$0.025^{*}$
	w/o Clarify	+0.188	$<10^{-13}$
	w/o FSM	+0.329	$<10^{-49}$

Appendix

Table 13: Statistical Significance Analysis (Paired t-test). Comparison of BioProAgent against baselines and ablation variants (

N=160

). All improvements are highly statistically significant.

Task ID	Device Interaction Nodes	Atomic Operation Steps	Involved Consumables
Sample 1	15	55	31
Sample 2	4	24	6
Sample 3	8	16	17
Sample 4	17	35	26
Sample 5	18	55	42
Sample 6	19	27	31

Appendix

Table 14: Complexity Statistics of the 9 Long-Horizon Tasks in Subset C. Unlike standard benchmarks with short trajectories, these tasks involve up to 238 atomic steps and 169 unique consumables. A failure at any step leads to task failure.

Constraint Type	Design Violation ( $s_{design}$ )	Interlock Trigger ( $\neg\mathcal{K}_{p}$ )	FSM Rectification ( $s_{rectify}$ )
Safety Alignment	op.centrifuge( speed=25,000g, duration=15min)	[HALT] Speed 25,000g exceeds rotor limit (Max: 15,000g).	op.centrifuge( speed=15,000g, duration=15min) # Clamped to safe limit
Resource Grounding	op.transfer( source=new_plate, vol=50)	[HALT] Resource ID new_plate not found in Registry $\Omega$ .	op.transfer( source=plate_1, vol=50) # Mapped to valid grounded ID
Causal Rectification	op.seal_plate() op.add_reagent(vol=10)	[HALT] Cannot add_reagent to a sealed container (Physical Collision Risk).	op.add_reagent(vol=10) op.seal_plate() # Sequence causally reordered

Appendix

Table 15: Gallery of Autonomous Corrections. The table compares the raw design violation (left), the specific safety interlock triggered by the Rule Engine (center), and the rectified code after FSM-guided feedback (right). These examples correspond directly to the qualitative traces discussed in Section 4.5.

Error Type	Counts	$C_{p}$	LCS Recall	$Acc_{seq}$
Resource Grounding	10	0.935	1.000	0.999
Safety Alignment	10	0.970	1.000	0.979
Causal Rectification	10	0.870	0.900	0.897

Appendix

Table 16: Granular Performance Breakdown by Error Type (Subset D). While syntactic perfection remains challenging in complex repairs, BioProAgent maintains high

C_{p}

, reinforcing its role as a safety interlock.

		Subset C		Subset D
Method	Backbone	Succ. $\uparrow$	$Acc_{param}$ $\uparrow$	Corr. Succ. $\uparrow$	$C_{p}$ $\uparrow$	Time (s) $\downarrow$
Direct	DeepSeek-V3.2-Think	100.0%	0.414	93.3%	0.948	196.2
Direct	Gemini-3-Pro-Think	100.0%	0.439	73.3%	0.765	81.4
BioProAgent	DeepSeek-V3.2-Think	55.6%	0.366	93.3%	0.955	1313.3
BioProAgent	Gemini-3-Pro-Think	44.4%	0.256	90.0%	0.922	316.1
BioProAgent	Gemini-3-Flash	100.0%	0.718	90.0%	0.925	128.0

Appendix

Table 17: Evaluation of “Thinking” Reasoning Models on Subsets C and D.

Step	BioProAgent (Ours)	AutoGPT (Baseline)
1	[State: CLARIFY_INTENT ] Trigger: $\sigma_{ambiguous}$ Action: clarify_experiment_scope(query=...) Output: "Sample Volume: 50uL, Kit ID: 10038"	[Thought] I need to find the protocol for metabolomics. Action: retrieve_knowledge(query="metabolomics prep")
2	[State: DESIGN_DRAFT ] Trigger: $\sigma_{know}\land\neg\sigma_{draft}$ Action: generate_scientific_draft(...) Grounding ( $\Phi$ ): Mapped "Kit_A" $\to$ ID:10038	[Thought] I found some docs. Now I need specific reagent list. Action: retrieve_knowledge(query="reagent list")
3	[State: VERIFY_DRAFT ] Trigger: $\sigma_{draft}\land(\sigma_{sci}=0)$ Action: reflect_on_protocol(...) Result: Scientific Logic Approved ( $\sigma_{sci}\leftarrow 1$ ).	[Thought] I need to check the temperature settings. Action: retrieve_knowledge(query="temperature settings")
…	(Step 4-7: State transitions through DESIGN_CODE and Physical Rule Verification $\mathcal{K}_{p}$ )	(Repeated retrieval loops for 30+ steps…)
8 / 45	[State: SUCCESS ] Trigger: $\sigma_{code}\land(\sigma_{phys}=1)$ Result: Valid JSON Code committed. Task Success (Total Tokens: 10,329)	[Error] Max Context Exceeded. Result: Timeout / Fail. Task Failure (Total Tokens: 59,278)

Appendix

Table 18: Side-by-Side Execution Log Comparison. BioProAgent utilizes FSM states to progress deterministically, completing the task in 8 steps. In contrast, AutoGPT enters a semantic drift loop, exhausting the context window (Step 45) without producing valid code.

Autonomous Experimentation

Autonomous experimentation uses AI, active learning, and robotics to design, run, and interpret experiments with less human intervention.

See all 140 papers

Jul 25, 2026

Stress-testing large language model agents in a robotic chemistry laboratory

AI is evaluated through knowledge, reasoning and plan generation, yet scientific agency requires reliable physical action and adaptation to evidence. Here, we use a robotic chemistry laboratory as a physical-world testbed to make scientific agency measurable. Its 45 modular workstations exposed as machine-readable skills enabled 4,608 trials. Only 3.3% of trials produced expert-assessed executable workflows under laboratory constraints; even the best system achieved 28.1%. Long-horizon planning remained a challenge: only three executable workflows exceeded 30 operations, although the longest contained 44. Across five rounds, experimental feedback prompted local adjustments but no workflow-level replanning or analytical-method redesign. By making physical executability and evidence-driven replanning measurable, our study provides an evidence-based assessment of deployment readiness and a diagnostic framework to guide closed-loop improvements towards physically grounded autonomous research.

Lulu Guo, Yingkai Sun, Xiaobo Li +13

State Key Laboratory of Precision and Intelligent Chemistry, Hefei National Research Center for Physical Sciences at the Microscale, School of Chemistry and Materials Science, University of Science and Technology of China, Hefei, China · Center for Scientific Intelligence Innovation, Hefei, China · School of Chemistry, School of Computer Science, University of Birmingham, Birmingham, UK

AI Agent Reliability AI Agents for Scientific Discovery AI Agent Evaluation Robotics Laboratory Automation Autonomous Experimentation+1

Ask PDF

Feb 24, 2026

Grounding LLMs in Scientific Discovery via Embodied Actions

Large Language Models (LLMs) have shown significant potential in scientific discovery but struggle to bridge the gap between theoretical reasoning and verifiable physical simulation. Existing solutions operate in a passive "execute-then-response" loop and thus lacks runtime perception, obscuring agents to transient anomalies (e.g., numerical instability or diverging oscillations). To address this limitation, we propose EmbodiedAct, a framework that transforms established scientific software into active embodied agents by grounding LLMs in embodied actions with a tight perception-execution loop. We instantiate EmbodiedAct within MATLAB and evaluate it on complex engineering design and scientific modeling tasks. Extensive experiments show that EmbodiedAct significantly outperforms existing baselines, achieving SOTA performance by ensuring satisfactory reliability and stability in long-horizon simulations and enhanced accuracy in scientific modeling.

Bo Zhang, Jinfeng Zhou, Yuxuan Chen +3

Tsinghua University, Beijing, China.

AI Agents for Scientific Discovery LLM Grounding Large Language Model-Guided Scientific Discovery Embodied Reasoning

Ask PDF

May 1, 2026

RunAgent: Interpreting Natural-Language Plans with Constraint-Guided Execution

Humans solve problems by executing targeted plans, yet large language models (LLMs) remain unreliable for structured workflow execution. We propose RunAgent, a multi-agent plan execution platform that interprets natural-language plans while enforcing stepwise execution through constraints and rubrics. RunAgent bridges the expressiveness of natural language with the determinism of programming via an agentic language with explicit control constructs (e.g., IF, GOTO, FORALL). Beyond verifying syntactic and semantic verification of the step output, which is performed based on the specific instruction of each step, RunAgent autonomously derives and validates constraints based on the description of the task and its instance at each step. RunAgent also dynamically selects among LLM-based reasoning, tool usage, and code generation and execution (e.g., in Python), and incorporates error correction mechanisms to ensure correctness. Finally, RunAgent filters the context history by retaining only relevant information during the execution of each step. Evaluations on Natural-plan and SciBench Datasets demonstrate that RunAgent outperforms baseline LLMs and state-of-the-art PlanGEN methods.

Arunabh Srivastava, Mohammad A., Khojastepour +2

Amir · University of Maryland, College Park, MD · 2NEC Laboratories America, Inc.

Multi-Agent LLM Systems Tool-Using Language Agents LLM Agent Planning

Ask PDF

Abstract

Autonomous Experimentation

Explore similar work

Stress-testing large language model agents in a robotic chemistry laboratory

Grounding LLMs in Scientific Discovery via Embodied Actions

RunAgent: Interpreting Natural-Language Plans with Constraint-Guided Execution