A Neurosymbolic Approach for Constructing Planning Domain Models from Clinical Narratives

# A Neurosymbolic Approach for Constructing Planning Domain Models from Clinical Narratives
Source: [https://arxiv.org/html/2608.21186](https://arxiv.org/html/2608.21186)
Saurabh MathurMichael SkinnerPrasad TadepalliKristian KerstingSriraam Natarajan

###### Abstract

Surgical procedures such as laparoscopic appendectomy are complex, high\-stakes processes, yet formalizing their workflows for decision support remains a significant challenge\. Inducing probabilistic planning domain models in this setting is particularly difficult due to the lack of structured event data and the prevalence of implicit actions in clinical narratives, which neither empirical symbolic methods nor Large Language Models \(LLMs\) can adequately address on their own\. We introduce NSPIN, a neurosymbolic framework for inducing probabilistic planning domain models from unstructured clinical narratives\. Our method extracts and imputes structured event sequences from raw text using a pretrained LLM, then induces a PPDDL model and refines its preconditions with LLM\-proposed revisions, guided by empirical validation\. We evaluate the approach on 2,660 laparoscopic appendectomy notes written by 9 surgeons\. NSPIN yields models that generalize to unseen notes, and expert clinical review indicates its induced knowledge is largely consistent with surgical practice\.

## Introduction

Surgical notes describe clinical findings and how a surgeon handled them in a specific patient, such as a perforated appendix or an unexpected abscess\([19](https://arxiv.org/html/2608.21186#bib.bib19)\)\. However, this knowledge remains confined to the note in which it was recorded and is never used beyond that patient’s chart\. Learning a model of surgical actions and their preconditions and effects generalizes this experience across thousands of individual cases of a given procedure, with direct consequences for patient safety, surgical training, and equitable access to high\-quality care\. Such a model can underpin AI\-in\-the\-loop systems\([20](https://arxiv.org/html/2608.21186#bib.bib20)\)that help improve simulators for training residents by capturing rare complications, enable automated review of operative notes, and inform quality improvement by tracking systematic differences among surgeons and institutions\. With further advances, such models could drive the development of autonomous surgical systems\.

Formalizing surgical procedures requires a representation that is both expressive enough to capture their structure and interpretable enough for clinicians to validate\. Formal planning languages, such as the Planning Domain Definition Language \(PDDL\)\([6](https://arxiv.org/html/2608.21186#bib.bib8)\), offer a powerful framework for this purpose, representing procedures as a set of state predicates and parameterized action schemas\. This declarative symbolic logical representation is highly interpretable, allowing clinicians to validate it against established medical standards\.

NSPINSurgical Notes“Following anesthesia, the patient was prepped and draped off sterilely\. A supraumbilical incision was made\. A 12 mm port was placed … ”Schema CreationCreates the Domain SchemaSequence GenerationExtracts predicate sequences from dataevents temporallyPlanning ModelConstructionSynthesizes formalPDDL operatorsPlanning Model\(:action port\_placement:parameters \(?port\_size ?location\):precondition \(and \(done\_incision\_made\)\):effect \(probabilistic1 \(and \(done\_port\_placement\)\)\)

Figure 1:The NSPIN framework for inducing probabilistic planning models from clinical narratives\.Our neurosymbolic approach extracts structured predicates from unstructured surgical notes using an LLM\. It uses this structured information to synthesize a planning model that generalizes across diverse surgical workflows\. Using an LLM to ground each action and observation in the text allows NSPIN to account for both explicit and implicitly mentioned information\.However, the transition from unstructured clinical narratives to formal symbolic logic represents a significant challenge\. Extracting a formal representation from clinical notes requires handling the linguistic diversity and varied writing styles of different clinicians\([10](https://arxiv.org/html/2608.21186#bib.bib22)\)\. It also requires domain knowledge, since surgeons often omit information that is obvious to them but essential for maintaining a logically consistent state in a formal representation\.

We address this formalization challenge through the specific case of laparoscopic appendectomy\([15](https://arxiv.org/html/2608.21186#bib.bib21)\)in pediatric patients\. The removal of the appendix to manage acute appendicitis is among the most common operative procedures in pediatric surgery\. In a large children’s hospital, pediatric surgeons may perform ten or more such procedures in a single day\. This high frequency and relatively straightforward procedural workflow make the laparoscopic appendectomy an ideal domain for investigation\. It providesa dense, large\-scale dataset of operative notes that are structurally similar enough to allow for comparison, yet linguistically diverse enough to test the limits of automated compilation\. Using this dataset, we build a model of action preconditions and effects that generalizes across different instances of the same procedure\. Beyond serving as an interpretable representation of the procedure itself, this model can also synthesize plans and narratives for downstream uses such as simulator scenario generation and training data augmentation, both of which are important building blocks for deployment\-ready clinical AI systems\.

We make the following key technical contributions toward interpretable, generalizable models of surgical procedures that can support trustworthy, deployable clinical AI: \(1\) We introduce NSPIN, a Neurosymbolic framework that transforms unstructured surgical notes into formal, probabilistic PDDL planning domains by combining the linguistic flexibility of LLMs with the logical rigor of symbolic induction\. \(2\) We exploit the internal world models of pretrained LLMs to fill the gaps in surgical notes, inferring the implicit clinical actions and preconditions essential for logical plan consistency\. \(3\) We evaluate NSPIN on a database of 2,660 real\-world surgical notes by 9 surgeons\. Our results demonstrate that this neurosymbolic approach produces planning models that generalize to unseen surgical workflows more effectively than purely LLM\-based baselines\.

After introducing the necessary technical background, we present our problem and its solution\. We then outline our evaluation before discussing problems for future research\.

## Background

NSPIN is related to three distinct research areas: automated planning, natural language processing, and neurosymbolic learning\. We observe that this type of combination of multiple AI models and systems is necessary when deploying models in high\-stakes domains such as healthcare\.

### Planning Domain Models

Domain models are used to represent planning domains formally\. In this work, we use Probabilistic Planning Domain Definition Language \(PPDDL\)\([30](https://arxiv.org/html/2608.21186#bib.bib7)\)to represent clinical planning domains\. We first introduce classical PDDL\([6](https://arxiv.org/html/2608.21186#bib.bib8);[7](https://arxiv.org/html/2608.21186#bib.bib10)\), which represents deterministic domains, and then discuss PPDDL, which extends it to represent stochasticity\.

Classical PDDL defines planning problem as a tuple⟨𝒫,s0,G⟩\\langle\\mathcal\{P\},s\_\{0\},G\\ranglewhere𝒫=⟨ℱ,O,𝒜⟩\\mathcal\{P\}=\\langle\\mathcal\{F\},O,\\mathcal\{A\}\\rangleis the planning domain\. Here,OOis a finite set of objects,ℱ\\mathcal\{F\}is a finite set of first\-order logic predicate symbols describing properties of objects or the relations between them,s0s\_\{0\}is the initial state,GGis the goal condition, and𝒜\\mathcal\{A\}is a set of parameterized action schemas\. Each statessis a subset of the set of all possible grounded predicatesℱg\\mathcal\{F\}\_\{g\}\([4](https://arxiv.org/html/2608.21186#bib.bib9)\)\. Each action schemaa∈𝒜a\\in\\mathcal\{A\}is defined as the tuple⟨pre​\(a\),add​\(a\),del​\(a\)⟩,\\langle\\text\{pre\}\(a\),\\text\{add\}\(a\),\\text\{del\}\(a\)\\rangle,representing the conditions that must hold for the action to be applicable, the set of facts that become true after executing the action, and the set of facts that become false after executing the action respectively\. Concretely, a grounded actionaga\_\{g\}is applicable in a statessif its preconditions are metpre​\(ag\)⊆s\.\\text\{pre\}\(a\_\{g\}\)\\subseteq s\.Given problems formalized in this notation, classical planning aims to find a plan, defined as a finite sequence of deterministic actionsπ=⟨a\(1\),…,a\(n\)⟩\\pi=\\langle a^\{\(1\)\},\\dots,a^\{\(n\)\}\\ranglethat sequentially transforms the initial states0s\_\{0\}into one satisfying the goal testG\.G\.

In probabilistic domains, classical PDDL must be extended by probabilistic PDDL to model domains in which the outcomes of each action are stochastic\. Instead of an actionaahaving a single, guaranteed outcome defined by one set of add and delete effects, its effects are defined as a probability distribution over multiple possible outcomes\. These stochastic effects are formalized as a set of tuples\{\(p1,add1​\(a\),del1​\(a\)\),…,\(pk,addk​\(a\),delk​\(a\)\)\}\\\{\(p\_\{1\},\\text\{add\}\_\{1\}\(a\),\\text\{del\}\_\{1\}\(a\)\),\\dots,\(p\_\{k\},\\text\{add\}\_\{k\}\(a\),\\text\{del\}\_\{k\}\(a\)\)\\\}wherepip\_\{i\}is the probability ofi−i\-th outcome occurring on executing the actiona,a,and∑i=1kpi=1\.\\sum\_\{i=1\}^\{k\}p\_\{i\}=1\.

The acquisition of these \(P\)PDDL models has traditionally relied on intensive manual engineering by domain experts, which is often brittle and unscalable\. To automate this process, several symbolic inductive learning methods\([13](https://arxiv.org/html/2608.21186#bib.bib31);[16](https://arxiv.org/html/2608.21186#bib.bib32);[22](https://arxiv.org/html/2608.21186#bib.bib33)\)have been proposed to extract action schemas from observed execution\. These methods typically assume access to structured execution trajectories, often represented as state\-action\-state transitions\(s,a,s′\)\(s,a,s^\{\\prime\}\), from which probabilistic action dynamics can be estimated\.

However, clinical narratives are rarely available in such a structured form\. Converting these unstructured text documents to structured traces requires either manual annotation or automatic parsing\. Moreover, human\-written notes often omit critical but implicit intermediary steps\([2](https://arxiv.org/html/2608.21186#bib.bib4)\)\. As a result, even a perfect parser would produce incomplete and logically unsound traces\. These latent dependencies and data sparsity make data\-driven clinical planning domain construction challenging\. Therefore, existing action\-model induction methods cannot be used in this setting, since they assume structured symbolic traces or observed action sequences rather than raw clinical narratives\.

Figure 2:A partial surgical sequence sampled using an action\-transition model and NSPIN\.NSPIN learns action preconditions that constrain the set of possible next actions, along with observational effects each action probabilistically produces\. These emitted observations provide context for a simple data\-driven action\-transition model used for sampling candidate actions\.Bluenodes represent sampled actions,graynodes represent emitted observations,orangearrows indicate candidate precondition relationships, andgrayarrows indicate observation flow\. In this example, port placement is followed by the observation of adhesions, which then provides contributing context for performing an adhesiolysis\.
### Neurosymbolic AI

Our framework, NSPIN, takes a neurosymbolic approach\([5](https://arxiv.org/html/2608.21186#bib.bib11)\)to the problem of formalizing clinical narratives into planning domains\. It exploits the representation learning of neural networks with the structured reasoning of symbolic logic\. Neural networks excel at extracting structured information from unstructured text\([29](https://arxiv.org/html/2608.21186#bib.bib12)\)\. Neural models have also been used as learnable heuristics to efficiently navigate large symbolic search spaces\([3](https://arxiv.org/html/2608.21186#bib.bib13);[18](https://arxiv.org/html/2608.21186#bib.bib30)\)\. However, the use of these methods in medicine is limited by the paucity of training data and the high cost of annotation\([25](https://arxiv.org/html/2608.21186#bib.bib14)\)\.

### Pretrained LLMs

Large language models\([32](https://arxiv.org/html/2608.21186#bib.bib2)\)are a class of neural language models characterized by their large size\. These models are pretrained on large natural language corpora, allowing them to process text and handle linguistic diversity without requiring additional training\. As a result, these models have been used as zero\-shot information extractors to handle non\-standard shorthands and varied terminology, mapping disparate natural language phrases to the same underlying structured form\([11](https://arxiv.org/html/2608.21186#bib.bib6);[23](https://arxiv.org/html/2608.21186#bib.bib5)\)\.

Beyond explicit extraction, their pretraining also makes LLMs effective imputation engines for inferring latent events and implicit affordances that human writers typically omit for brevity\. Moreover, LLMs also embed broad domain knowledge about these clinical narratives\. They have been used to filter out implausible object interactions and state changes to prune the massive symbolic search spaces\([27](https://arxiv.org/html/2608.21186#bib.bib3)\)\. While LLMs excel at linguistic extraction and imputation, they fundamentally struggle with inductive logical reasoning\([9](https://arxiv.org/html/2608.21186#bib.bib1)\), especially in complex domains involving interactions over long horizons\([26](https://arxiv.org/html/2608.21186#bib.bib15)\)\. As a result, they produce overly simplistic models\. Moreover, recent LLM\-based planning formalizers typically assume at least one source of support unavailable when building models from clinical narratives, such as a clean domain description, executable environment feedback, or benchmark tasks with known formal structure\([8](https://arxiv.org/html/2608.21186#bib.bib24);[17](https://arxiv.org/html/2608.21186#bib.bib25);[33](https://arxiv.org/html/2608.21186#bib.bib26)\), making them inapplicable to the task\. Finally, most of these LLM\+PDDL works have focused on deterministic domain generation\([12](https://arxiv.org/html/2608.21186#bib.bib28);[21](https://arxiv.org/html/2608.21186#bib.bib23);[31](https://arxiv.org/html/2608.21186#bib.bib27)\), whereas our setting requires a probabilistic PPDDL domain model to capture stochastic effects across clinical narratives\. Unlike prior single\-instance LLM formalizers, NSPIN uses LLMs for extraction and imputation from raw clinical narratives, while relying on symbolic induction and corpus\-level aggregation to construct and validate probabilistic planning models\.

## The NSPIN Framework

Recall our goal is toinduce a structured model of a clinical procedure from a database of unstructured notes\. We formally define the problem as follows:

Given: A Database𝒟=\{x\(i\)\}i=1N\\mathcal\{D\}=\\\{x^\{\(i\)\}\\\}\_\{i=1\}^\{N\}, where eachx\(i\)x^\{\(i\)\}is an unstructured clinical narrative describing a procedure instance\.To Do: A Probabilistic Planning Domain Model𝒫\\mathcal\{P\}that characterizes the actions, preconditions, and stochastic effects of the procedure\.

This task presents two key challenges\. First, clinical narratives consist of unstructured text, while inducing planning models requires structured action and observation data\. Second, the narratives are highly heterogeneous and sparse, since clinicians have different documenting styles and underlying assumptions, which affect which steps are explicitly recorded and which are treated as implicit knowledge\.

To address these challenges, we introduce NSPIN, a neurosymbolic framework for planning model induction\. NSPIN bridges the representational gap by exploiting a pretrained LLM for extraction and symbolic refinement for logical grounding\. Figure[1](https://arxiv.org/html/2608.21186#Sx1.F1)illustrates the structure of NSPIN, and the subsequent sections describe each stage in detail\.

### Schema Creation

The schema generation stage introduces and defines the formal vocabulary used to represent the planning domain\. In this step, an LLM takes as input a natural language description of the domain, including its rules and constraints, and generates a structured domain schema

𝒫0\\mathcal\{P\}\_\{0\}\. This schema consists of three components: 1\) Typed Object Categories \(O\):A set of semantic classes \(e\.g\.,location,structure,instrument\) that categorize the objects in the surgical environment\. 2\) Observation Predicates \(ℱ\\mathcal\{F\}\):A finite set of first\-order logic predicates that describe the properties of objects and their mutual relations, such as

is\_occluded​\(?​s​t​r​u​c​t​u​r​e\)\\texttt\{is\\\_occluded\}\(?structure\)\. These predicates define the state space of the procedure\.

3\) Action Predicates \(𝒜0\\mathcal\{A\}\_\{0\}\):A set of parameterized action types that name the available surgical steps, for example,

mobilize\_structure​\(?​s​t​r​u​c​t​u​r​e\),\\texttt\{mobilize\\\_structure\}\(?structure\),where each argument is typed by an object category from

OO\.

Note that the action schemas produced at this stage \(𝒜0\\mathcal\{A\}\_\{0\}\) are initialized with empty preconditions and effects,pre​\(a\)=eff​\(a\)=∅\\text\{pre\}\(a\)=\\text\{eff\}\(a\)=\\emptysetfor eacha∈𝒜0\.a\\in\\mathcal\{A\}\_\{0\}\.These are syntactically valid but vacuous placeholders, since each action is always applicable and has no effect on the state\. This stage fixes only the vocabulary and argument structure of each action\. The substantive preconditions and stochastic effects are induced from data in the planning model construction stage\. This constrains the massive symbolic search space to a fixed, clinically relevant vocabulary by exploiting the pretrained LLM’s approximate domain knowledge\.

### Sequence Generation

The sequence generation stage converts the unstructured surgical notes to structured observation and action traces required to induce the planning model\. This process is divided intoinformation extractionandsequence imputationphases\.

##### Information Extraction\.

We formulate information extraction as a constrained grounding problem\. Given an unstructured narrativex∈𝒟x\\in\\mathcal\{D\}and the schema𝒫0=⟨ℱ,O,𝒜0⟩\\mathcal\{P\}\_\{0\}=\\langle\\mathcal\{F\},O,\\mathcal\{A\}\_\{0\}\\rangle, the LLM is prompted to construct a sequence of grounded predicatess=\[u1,…,uk\]s=\[u\_\{1\},\\dots,u\_\{k\}\]where eachu=h⁡\(𝐯\)u=h\(\{\\bf v\}\)combines a predicateh∈ℱ∪𝒜0h\\in\\mathcal\{F\}\\cup\\mathcal\{A\}\_\{0\}with entities𝐯=v1,…,vn\{\\bf v\}=v\_\{1\},\\dots,v\_\{n\}grounded inx,x,as per the schema\. Since LLM outputs are stochastic, we generate K candidate solutions for the same note𝐬=\{s1,…​sK\}\{\\bf s\}=\\\{s\_\{1\},\\dots s\_\{K\}\\\}\(lowercase, one per LLM sample\), and aggregate them to obtain a single consensus sequence, denoted asS0S\_\{0\}\(uppercase\), which we construct in three stages: 1\) Majority Filtering:We retain only those predicate instances supported by a majority of the K candidates\. Concretely, we retain

𝐡=\{\(h,r\):h∈ℱ∪𝒜0,1≤r≤r∗\(h\)\},\{\\bf h\}=\\\{\(h,r\):\\ h\\in\\mathcal\{F\}\\cup\\mathcal\{A\}\_\{0\},\\ 1\\leq r\\leq r^\{\*\}\(h\)\\\},wherer∗​\(h\)r^\{\*\}\(h\)is the maximum number of instances ofhhthat are supported by more than half of theKKcandidates,

r∗\(h\)=max\{r:Count\(h,r,𝐬\)\>K2,r≥0\}r^\{\*\}\(h\)=\\max\\\{r:\\text\{Count\}\(h,r,\{\\bf s\}\)\>\\frac\{K\}\{2\},\\ r\\geq 0\\\}andCount​\(h,r,𝐬\)\\text\{Count\}\(h,r,\{\\bf s\}\)is the number of candidatess∈𝐬s\\in\{\\bf s\}containing at leastrroccurrences ofhh\. 2\) Consensus Grounding:For each retained predicate instance\(h,r\)∈𝐡,\(h,r\)\\in\{\\bf h\},the value of argument slotviv\_\{i\}is chosen by majority vote across the candidates,

vi=arg⁡max𝑣​Counti​\(v,h,r,𝐬\),v\_\{i\}=\\underset\{v\}\{\\arg\\max\}\\ \\text\{Count\}\_\{i\}\(v,h,r,\{\\bf s\}\),whereCounti​\(v,h,r,𝐬\)\\text\{Count\}\_\{i\}\(v,h,r,\{\\bf s\}\)is the number of candidatess∈𝐬s\\in\{\\bf s\}where therr\-th occurrence of predicatehhhasii\-th argument equal tov\.v\. 3\) Temporal Ordering:To order the retained predicate instances𝐡\\mathbf\{h\}, we solve a weighted Kemeny rank aggregation problem\([14](https://arxiv.org/html/2608.21186#bib.bib16)\)\. For\(h,r\),\(h′,r′\)∈𝐡2\(h,r\),\(h^\{\\prime\},r^\{\\prime\}\)\\in\\mathbf\{h\}^\{2\}, we define the pairwise precedence weight as,

w⁡\(\(h,r\),\(h′,r′\)\)=Count​\(\(h,r\)≺\(h′,r′\),𝐬\)w\\big\(\(h,r\),\(h^\{\\prime\},r^\{\\prime\}\)\\big\)=\\text\{Count\}\\big\(\(h,r\)\\prec\(h^\{\\prime\},r^\{\\prime\}\),\\ \{\\bf s\}\\big\)which counts the number of candidatess∈𝐬s\\in\\mathbf\{s\}in which therr\-th occurrence ofhhprecedes ther′r^\{\\prime\}\-th occurrence ofh′h^\{\\prime\}\. The aggregated sequenceS0\{S\}\_\{0\}is the orderingσ\\sigmaof𝐡\\mathbf\{h\}that minimizes the total contradicted precedence weight,

S0=argminσ∑\(h,r\)≺σ\(h′,r′\)w\(\(h′,r′\),\(h,r\)\)\.\{S\}\_\{0\}=\\arg\\min\_\{\\sigma\}\\sum\_\{\(h,r\)\\prec\_\{\\sigma\}\(h^\{\\prime\},r^\{\\prime\}\)\}w\\big\(\(h^\{\\prime\},r^\{\\prime\}\),\(h,r\)\\big\)\.

##### Sequence Imputation\.

While the aggregated sequenceS0S\_\{0\}captures the actions and observations mentioned in the surgical note, surgeons frequently omit obvious intermediary steps or universal surgical protocols\. As a result, the extracted sequenceS0S\_\{0\}lacks the logical continuity required for PPDDL induction\. Some transitions between consecutive instances\(h,r\),\(h′,r′\)∈S0\(h,r\),\(h^\{\\prime\},r^\{\\prime\}\)\\in S\_\{0\}presuppose intermediate actions or observations that are not explicit inx\.x\.Identifying these omitted steps requires domain knowledge of what would have been obvious to the surgeon and therefore not worth recording\.

We address this by exploiting a pretrained LLM’s approximately correct domain knowledge of clinical workflows to infer these omitted steps\. Given the incomplete sequenceS0S\_\{0\}, the LLM is prompted to infer additional grounded instancesu=h⁡\(𝐯\)u=h\(\\mathbf\{v\}\),h∈ℱ∪𝒜0h\\in\\mathcal\{F\}\\cup\\mathcal\{A\}\_\{0\}, that are logically necessary or contextually implied by the surrounding instances inS0S\_\{0\}, with reprompting in case of syntactic or schema violation\.

As in extraction, we generateKKcandidate imputations for the same sequenceS0S\_\{0\}, yieldingKKenriched candidate sequencess1′,…,sK′s\_\{1\}^\{\\prime\},\\dots,s\_\{K\}^\{\\prime\}, each containing the original instances ofS0S\_\{0\}along with the proposed additions\. To ensure robustness, we apply the same three\-stage aggregation procedure used in extraction overs1′,…,sK′s\_\{1\}^\{\\prime\},\\dots,s\_\{K\}^\{\\prime\}to obtain the final enriched sequenceS\.S\.Imputed instances are retained only if supported by a majority of candidates, their arguments are set by consensus vote, and their position relative to all other instances is fixed by Kemeny aggregation\. Since every instance inS0S\_\{0\}is present in allKKenriched candidates by construction, it is trivially retained by majority filtering, soSSstrictly extendsS0S\_\{0\}with additional consensus\-supported instances\. This sequence represents the complete surgical workflow used for planning model construction\.

### Planning Model Construction

This stage transforms the grounded predicate sequences𝐒=\{S\(i\)\}i=1N\{\\bf S\}=\\\{S^\{\(i\)\}\\\}\_\{i=1\}^\{N\}into a planning domain model𝒫\\mathcal\{P\}\. It consists of two steps: a data\-driven induction step that constructs an initial action schema from𝐒\{\\bf S\}, followed by an LLM\-based refinement step that revises the induced preconditions\. The result is a planning model𝒫\\mathcal\{P\}represented in PPDDL111The algorithmic pseudocode for the different components of planning model construction is provided in the supplementary material athttps://github\.com/s\-ranveer/nspin˙ppddl\.

##### Data\-Driven Induction\.

Given the predicate sequences𝐒\{\\bf S\}, we induce an initial PPDDL action schemaαa\\alpha\_\{a\}for each actiona∈𝒜0a\\in\\mathcal\{A\}\_\{0\}, specifying its preconditions and probabilistic effects from empirical co\-occurrence and outcome frequencies\. This induction step is modular, allowing any algorithm that learns action preconditions and effects from structured, action\-segmented sequences to be used\. We consider a greedy, coverage\-based inducer\([28](https://arxiv.org/html/2608.21186#bib.bib29)\)to construct preconditions from observation contexts of sizewwtill atleast a fractionτo​b​s\\tau\_\{obs\}of sequences are satisfied; we estimate effects conditioned on the action arguments via maximum\-likelihood outcome frequencies, filtering out any effects with a probability less thanτp​r​o​b\.\\tau\_\{prob\}\.The resulting schemas are assembled into the domain𝒫1=⟨ℱ,O,𝒜1⟩\\mathcal\{P\}\_\{1\}=\\langle\\mathcal\{F\},O,\\mathcal\{A\}\_\{1\}\\rangle, where𝒜1=\{αa∣a∈𝒜1\}\\mathcal\{A\}\_\{1\}=\\\{\\alpha\_\{a\}\\mid a\\in\\mathcal\{A\}\_\{1\}\\\}\.

#### LLM Refinement\.

While the data\-driven action schema fits the observed data, the learned preconditions may be overly restrictive, including clauses that are merely associational, appearing before an action in the data without being genuinely necessary for that action to occur\. We therefore refine each action’s preconditions using an LLM to obtain constraints that are more semantically grounded in the narrative and less brittle\.

Starting from the domain𝒫1\\mathcal\{P\}\_\{1\}, we split the sequences𝐒\\mathbf\{S\}into training and validation sets and track the best domain found so far\. For each actionaawith failed precondition instances, we extract its satisfied and failed contexts,Ca\+C\_\{a\}^\{\+\}andCa−C\_\{a\}^\{\-\}, from the training set, and prompt the LLM to propose a revised precondition\. We accept a revised precondition only if it does not worsen precondition satisfaction or degrade validation performance, yielding the refined domain𝒫=⟨ℱ,O,𝒜⟩\.\\mathcal\{P\}=\\langle\\mathcal\{F\},O,\\mathcal\{A\}\\rangle\.

## Empirical Evaluation

We now present our empirical evaluations and aim to answer the following questions\. While the first three questions focus on the algorithmic contributions, the last one focuses on the qualitative evaluation for the societal impact\. Q1\)Does LLM\-based imputation of implicit information yield more accurate domain models? Q2\)Does LLM\-refinement help NSPIN induce better planning models? Q3\)Does NSPIN induce better planning domain models from clinical narratives than LLMs? Q4\)How well do NSPIN’s imputed steps, induced preconditions, and generated narratives align with clinical knowledge and practice?

Dataset\.To answer these questions, we obtained a dataset of 2,660 laparoscopic appendectomy surgical notes from 9 different surgeons at Children’s Hospital Medical Center in Dallas, TX222We are not aware of any publicly available equivalent datasets\. Each surgeon differs in the number of notes and average tokens per note, as shown in Figure[3](https://arxiv.org/html/2608.21186#Sx4.F3)\.

![Refer to caption](https://arxiv.org/html/2608.21186v1/Figures/surgeon_distribution.png)Figure 3:Surgical Note lengths and counts across surgeonsMethods\.We evaluate the probabilistic planning domains induced by NSPIN against two LLM\-based baselines:LLM\-onlyandLLM\-Example\. In LLM\-only, the model is provided with the domain description and predicate templates and tasked with generating a PPDDL model zero\-shot\. In LLM\-Example, the model is also provided with an example predicate sequence from the dataset to provide in\-context grounding for the action schemas\. Additionally, we also evaluate a variant of NSPIN without the final LLM\-based precondition refinement\.

To maintain the privacy of the sensitive clinical data, we employ a hybrid LLM strategy: Claude Sonnet 4\.6\([1](https://arxiv.org/html/2608.21186#bib.bib17)\)is used for high\-level schema construction and symbolic refinement, while the local, medically\-tuned Medgemma\-27b\([24](https://arxiv.org/html/2608.21186#bib.bib18)\)is used for the extraction and imputation of procedural traces from the raw narratives, with the extraction performed five times on each surgical note, and each extracted sequence independently imputed\.

All LLM\-based generations and refinement use Claude Sonnet\. We use the Clingo Answer Set Programming \(ASP\) Solver to implement sequence merging333The different prompts, predicate templates, groundings, classification, code for running the different experiments, and the ASP program for the final sequence generation and the different PPDDLs created are provided in the supplementary materialof the five predicate sequences per patient based on the constraints defined before\.

For data\-driven PPDDL construction, we evaluated window sizes\(w∈1,3,5,∞CLOSE\(w\\in\{1,3,5,\\infty\}\), minimum effect\-probability thresholds\(τprob∈0\.005,0\.01,0\.05\)\(\\tau\_\{\\mathrm\{prob\}\}\\in\{0\.005,0\.01,0\.05\}\), and minimum precondition\-support thresholds\(τobs∈0\.7,0\.8,0\.9,1\.0\)\(\\tau\_\{\\mathrm\{obs\}\}\\in\{0\.7,0\.8,0\.9,1\.0\}\), where\(τobs\)\(\\tau\_\{\\mathrm\{obs\}\}\)denotes the minimum fraction of sequences satisfying a candidate precondition\. Grid search on the training set selected w = 3,τprob=0\.01\\tau\_\{\\mathrm\{prob\}\}=0\.01, andτobs=0\.8\\tau\_\{\\mathrm\{obs\}\}=0\.8\.

Metrics\.We evaluate models along two dimensions\.Observation set predictionmeasures how well a model’s learned probabilistic effects predict the set of observations occurring between an action and the next, using a held\-out trace\.Next\-action predictionmeasures how well a learned action\-transition model predicts the subsequent action, while the domain’s preconditions constrain which actions are symbolically applicable\. For both, we report:\(1\) Top\-1/Top\-3 accuracy, whether the ground truth is the top prediction or among the top three;\(2\) average negative log\-likelihood \(Avg NLL\), the probability mass assigned to the ground truth, with lower values indicating better calibration; and, for next\-action prediction,\(3\) true action disallowed rate \(TADR\), the fraction of steps where the learned preconditions incorrectly rule out the ground\-truth action, with lower values indicating better compatibility with observed workflows\. We compute these metrics for each method using 9\-fold cross\-validation, where the PPDDLs are learnt from predicate sequences of all but one surgeon, and evaluated on the held\-out surgeon’s predicate sequences\.

### Results

The generated appendectomy schema comprises 63 predicates in total, including 34 observation predicates, 26 action predicates, and 3 relational predicates, along with 58 typed argument categories and 322 enumerated grounding values\. Tables[1](https://arxiv.org/html/2608.21186#Sx4.T1),[2](https://arxiv.org/html/2608.21186#Sx4.T2), and[3](https://arxiv.org/html/2608.21186#Sx4.T3)show the average evaluation results across 9 folds on the held\-out set, and will be used for answering the questions asked at the beginning of the section\. Statistical significance was assessed using the Wilcoxon signed\-rank test with a significance level ofα=0\.05\\alpha=0\.05; all reported comparisons were statistically significant\.

Table 1:Action set prediction performance comparison across imputed and non\-imputed datasetsTable 2:Observation prediction performance comparison on the imputed dataset across different methodsTable 3:Action prediction performance comparison on the imputed dataset across different methods\(Q1\) LLM\-based Imputation of Sequences\.To evaluate the utility of using LLMs for resolving sparsity in clinical narratives, we compare the action set prediction accuracy of data\-driven models \(NSPIN w/o refinement\) induced from raw extractions versus those enriched via latent sequence imputation\. As shown in Table[1](https://arxiv.org/html/2608.21186#Sx4.T1), we find that imputation improves action set predictions across all metrics\. The model learned from imputed data had a higher Top\-1 accuracy, Top\-3 accuracy, and a lower average negative log likelihood\. This consistent improvement indicates that the imputed sequences recover information that surgeons omitted as clinically obvious, and that this recovered information is essential\. Therefore, we can answer Q1 affirmatively\.

\(Q2\) LLM\-Refinement in NSPIN\.To assess whether LLM\-based refinement of action preconditions can resolve the spurious correlation problem inherent in purely data\-driven induction, we compare the predictive performance of NSPIN with a version without the LLM\-based refinement\. As shown in Table[2](https://arxiv.org/html/2608.21186#Sx4.T2), observation\-set prediction remained stable across both settings, as these metrics are conditioned on action execution and are independent of precondition logic\. However, the refinement phase yielded substantial improvements in all next\-action prediction metrics \(Table[3](https://arxiv.org/html/2608.21186#Sx4.T3)\)\. Notably, the refined model achieved higher Top\-1 and Top\-3 accuracy and a lower Avg NLL, coupled with a significant reduction in the True Action Disallowed Rate \(TADR\)\. This indicates that the refinement step produced less restrictive preconditions that were more compatible with held\-out traces\. For example, for the step of sending the appendix specimen to pathology, the symbolic precondition inducer included descriptive findings such as associated pathology\. The LLM\-based refinement removed these while retaining the core procedural requirement that the specimen must first be retrieved, yielding a precondition that is more semantically faithful to the workflow\. Overall, this improved next\-action prediction and reduced TADR\. So, we can answer Q2 affirmatively\.

\(Q3\) NSPIN vs LLM\-Induced Models\.To evaluate the necessity of neurosymbolic integration, we compare the predictive validity of NSPIN against the purely LLM\-based baselines \(LLM\-OnlyandLLM\-Example\)\. As shown in Table[2](https://arxiv.org/html/2608.21186#Sx4.T2)and Table[3](https://arxiv.org/html/2608.21186#Sx4.T3), NSPIN outperforms the baselines in all metrics, achieving a higher Top\-1 and Top\-3 accuracy, as well as a lower average negative log likelihood\. This indicates that the data\-driven symbolic model provides a stronger explanation of post\-action observations than domains generated directly by the LLM baselines\.

The evaluation of next\-action prediction reveals a more complex trajectory \(Table[3](https://arxiv.org/html/2608.21186#Sx4.T3)\)\. The unrefined model \(NSPIN w/o Refinement\) underperforms the neural baselines, exhibiting a higher TADR\. However, following the one\-step semantic refinement, NSPIN performance improves substantially, ultimately surpassing both neural baselines in predictive accuracy, likelihood, and TADR\. This shift demonstrates that while LLMs possess strong procedural priors, they require symbolic validation to ground those priors in empirical data\. Thus, we can answer Q3 affirmatively\.

\(Q4\) Qualitative evaluation\.To assess whether LLM\-based imputation recovers clinically real but undocumented or missing steps, we considered the top\-10 frequently imputed predicates and sampled 5 notes where they were imputed\. Our domain expert found the imputation clinically correct\. For example, one commonly imputed predicate isfascia\_closure\. In one of the notes, the information extractor failed to extract this despite it being made explicit by the sentence"The umbilical fascia was closed with 0 Vicryl suture\."\. The imputer correctly recovered this information based on its clinical necessity and the partially extracted sequence\. In another example, the predicateappendix\_base\_securedwas imputed by the LLM before transection \(appendix\_transected\)\. Although the note does not explicitly mention securing the appendix, it is typically performed with transection as part of the same mechanical action\. Because transection cannot be completed without securing the appendix, the imputation is justified\.

To assess whether the induced preconditions and their refinement reflect genuine clinical necessity, we selected three actions for expert review:port\_placement, appendix\_base\_secured,andappendix\_removed\. Port placement requires that an incision already be made at the relevant site\. The LLM\-only baseline omitted the precondition completely, allowing ports to be placed anywhere as long as one has not already been placed\. NSPIN, without refinement, listed it as a contributing but non\-required condition\. It was made the necessary condition only after refinement\. Securing the appendix base requires the appendix structure to be mobilized\. This precondition was missing even after refinement; mobilization remained a contributing rather than necessary condition in NSPIN both before and after refinement, while the requirement was completely absent for the LLM baselines\. Finally, for appendix removal, the LLM\-only baselines included stump treatment as a precondition\. While stump treatment typically occurs before removal, it is not a necessary precondition as removal does not logically depend on it\. NSPIN, with and without refinement, correctly identified the full set of relevant preceding actions, but some were rendered only as contributing rather than necessary actions due to the presence of certain observations acting as contributing preconditions as well\.

Our expert found the effects of all 3 actions clinically valid\. In contrast, the LLM\-based baselines had unrealistic effect probabilities\. For example, the probability of injury for port placement was substantially higher than what our expert considered realistic\.

To qualitatively test whether the induced PPDDL model captures sufficient clinical knowledge to generate coherent procedure descriptions, we had the expert review five representative surgical notes synthesized using the model\. We sampled action sequences from the transition model learned during evaluation, using the PPDDL preconditions as guards to determine which actions could occur and emitting the corresponding observations for each sampled action, such as the one in Figure[2](https://arxiv.org/html/2608.21186#Sx2.F2)\. An LLM with the predicate templates, the sampled sequence, and generation rules to generate synthetic notes\. The generated notes generally preserved the high\-level temporal structure of an appendectomy but sometimes missed some nuances that were underrepresented in the training data\. For example, the notes would indicate making a suprapubic incision before placing a camera port, when a different incision would be warranted\. These errors suggest that while the learned model captures the broad procedural workflow, it could benefit from targeted expert knowledge about infrequently observed constraints\.

Taken together, these evaluations indicate that NSPIN recovers clinically real implicit steps, its refined preconditions align with clinical domain knowledge, and its symbolic structure can be used to synthesize plausible surgical narratives\.

## Conclusion

We considered the challenging problem of translating unstructured clinical narratives into a formal representation\. We address two key challenges in the clinical planning domain induction: the linguistic variability inherent in natural language procedure notes and the observational sparsity caused by human reporting bias\. Our work bridges the gap between incomplete, unstructured narratives and the structured representations required for clinical decision support by exploiting the broad priors encoded in pretrained LLMs\. Our empirical evaluation on a corpus of 2,660 surgical notes on laparoscopic appendectomies written by 9 surgeons shows that NSPIN effectively addresses linguistic variability and reporting bias, yielding domain descriptions that generalize to unseen data – a key requirement for many clinical tasks\.

LLM\-based sequence imputation improves downstream predictive performance in a manner consistent with recovering omitted or unmentioned procedural steps that clinicians frequently omit but are logically essential for plan consistency\. Further, we showed that a single step of neural refinement significantly improves the generalizability of domain descriptions by pruning spurious temporal correlations that a purely data\-driven approach would otherwise treat as hard preconditions\. These results suggest that combining LLM\-based extraction and imputation with symbolic induction is a promising approach for constructing useful PPDDL\-like procedural models from clinical narratives\.

There are several directions for future work\. First, evaluation on additional surgical procedures and clinical domains beyond appendectomy is necessary\. Second, the modular design of NSPIN makes it possible to explore alternative methods for inducing PPDDL models from structured predicate sequences\. Finally, an important extension is to move beyond refining only action preconditions and develop methods for refining probabilistic effects as well\. Overall, together with these directions, NSPIN represents a significant step toward intelligent clinical decision support systems capable of reasoning about complex medical procedures\.

## Acknowledgment

The authors gratefully acknowledge support from the NIH award R01NS133142 and the Cluster of Excellence “Reasonable AI" funded by the German Research Foundation \(DFG\) under Germany’s Excellence Strategy, EXC\-3057\.

## References

- Anthropic \(2025\)AnthropicClaude sonnet \(claude 4 family model\)\.Note:https://www\.anthropic\.comAccessed: 2026\-04\-16Cited by:[Empirical Evaluation](https://arxiv.org/html/2608.21186#Sx4.p4.1)\.
- Doppaet al\.\(2011\)J\. R\. Doppa M\. S\. Soroweret al\.Learning rules from incomplete examples via implicit mention models\.InAsian Conference on Machine Learning,pp\. 197–212\.Cited by:[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p5.1)\.
- Ferberet al\.\(2022\)P\. Ferber, F\. Geißer, F\. Trevizan, M\. Helmert, and J\. HoffmannNeural network heuristic functions for classical planning: bootstrapping and comparison to other methods\.Proceedings of the International Conference on Automated Planning and Scheduling32\(1\),pp\. 583–587\.External Links:[Link](https://ojs.aaai.org/index.php/ICAPS/article/view/19845),[Document](https://dx.doi.org/10.1609/icaps.v32i1.19845)Cited by:[Neurosymbolic AI](https://arxiv.org/html/2608.21186#Sx2.SSx2.p1.1)\.
- Fikes and Nilsson \(1971\)R\. E\. Fikes and N\. J\. NilssonSTRIPS: a new approach to the application of theorem proving to problem solving\.Artificial intelligence2\(3\-4\),pp\. 189–208\.Cited by:[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p2.1)\.
- Garcez and Lamb \(2023\)A\. d\. Garcez and L\. C\. LambNeurosymbolic ai: the 3rd wave\.Artif\. Intell\. Rev\.56\(11\),pp\. 12387–12406\.External Links:ISSN 0269\-2821,[Link](https://doi.org/10.1007/s10462-023-10448-w),[Document](https://dx.doi.org/10.1007/s10462-023-10448-w)Cited by:[Neurosymbolic AI](https://arxiv.org/html/2608.21186#Sx2.SSx2.p1.1)\.
- Ghallabet al\.\(1998\)M\. Ghallab, A\. Howe, C\. Knoblock,et al\.PDDL\-the planning domain definition language\.Technical report CVC TR\-98\-003/DCS TR\-1165, Yale center for Computational Vision and Control\.Cited by:[Introduction](https://arxiv.org/html/2608.21186#Sx1.p2.1),[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p1.1)\.
- Ghallabet al\.\(2004\)M\. Ghallab, D\. Nau, and P\. TraversoAutomated planning: theory and practice\.Elsevier\.Cited by:[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p1.1)\.
- Guanet al\.\(2023\)L\. Guanet al\.Leveraging pre\-trained large language models to construct and utilize world models for model\-based task planning\.InAdvances in Neural Information Processing Systems,Vol\.36,pp\. 79081–79094\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/f9f54762cbb4fe4dbffdd4f792c31221-Paper-Conference.pdf)Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Helffet al\.\(2025\)L\. Helff A\. Omaret al\.Slr: automated synthesis for scalable logical reasoning\.arXiv preprint arXiv:2506\.15787\.Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Hieret al\.\(2024\)D\. B\. Hier, M\. D\. Carrithers, T\. S\. Do, and T\. Obafemi\-AjayiEfficient standardization of clinical notes using large language models\.arXiv preprint arXiv:2501\.00644\.Cited by:[Introduction](https://arxiv.org/html/2608.21186#Sx1.p3.1)\.
- Hsu and Roberts \(2025\)E\. Hsu and K\. RobertsLLM\-ie: a python package for biomedical generative information extraction with large language models\.JAMIA open8\(2\),pp\. ooaf012\.Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p1.1)\.
- Huang and Zhang \(2025\)C\. Huang and L\. ZhangOn the limit of language models as planning formalizers\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 4880–4904\.External Links:[Link](https://aclanthology.org/2025.acl-long.242/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.242)Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Juba and Stern \(2022\)B\. Juba and R\. SternLearning probably approximately complete and safe action models for stochastic worlds\.Proceedings of the AAAI Conference on Artificial Intelligence36\(9\),pp\. 9795–9804\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/21215),[Document](https://dx.doi.org/10.1609/aaai.v36i9.21215)Cited by:[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p4.1)\.
- Kemeny \(1959\)J\. G\. KemenyMathematics without numbers\.Daedalus88\(4\),pp\. 577–591\.External Links:ISSN 00115266,[Link](http://www.jstor.org/stable/20026529)Cited by:[Information Extraction\.](https://arxiv.org/html/2608.21186#Sx3.SSx2.SSS0.Px1.p1.4)\.
- Korndorffer Jret al\.\(2010\)J\. R\. Korndorffer Jr, E\. Fellinger, and W\. ReedSAGES guideline for laparoscopic appendectomy\.Surgical endoscopy24\(4\),pp\. 757–761\.Cited by:[Introduction](https://arxiv.org/html/2608.21186#Sx1.p4.1)\.
- Leet al\.\(2024\)H\. S\. Leet al\.Learning safe action models with partial observability\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.38,pp\. 20159–20167\.Cited by:[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p4.1)\.
- Mahdaviet al\.\(2024\)S\. Mahdavi R\. Aokiet al\.Leveraging environment interaction for automated pddl translation and planning with large language models\.InAdvances in Neural Information Processing Systems,Vol\.37,pp\. 38960–39008\.External Links:[Document](https://dx.doi.org/10.52202/079017-1230),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/file/44af065477781e7f8a8589b14a62c489-Paper-Conference.pdf)Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Mangannavaret al\.\(2024\)R\. Mangannavar, S\. Lee, A\. Fern, and P\. TadepalliGraph neural network based action ranking for planning\.arXiv preprint arXiv:2412\.04752\.Cited by:[Neurosymbolic AI](https://arxiv.org/html/2608.21186#Sx2.SSx2.p1.1)\.
- Mathioudakiset al\.\(2016\)A\. Mathioudakis I\. Rousalovaet al\.How to keep good clinical records\.Breathe12\(4\),pp\. 369–373\.Cited by:[Introduction](https://arxiv.org/html/2608.21186#Sx1.p1.1)\.
- Natarajanet al\.\(2025\)S\. Natarajan, S\. Mathur, S\. Sidheekh, W\. Stammer, and K\. KerstingHuman\-in\-the\-loop or ai\-in\-the\-loop? automate or collaborate?\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 28594–28600\.Cited by:[Introduction](https://arxiv.org/html/2608.21186#Sx1.p1.1)\.
- Oswaldet al\.\(2024\)J\. Oswald K\. Srinivaset al\.Large language models as planning domain generators\.InProceedings of the Thirty\-Fourth International Conference on Automated Planning and Scheduling,ICAPS ’24\.External Links:[Link](https://doi.org/10.1609/icaps.v34i1.31502),[Document](https://dx.doi.org/10.1609/icaps.v34i1.31502)Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Pasulaet al\.\(2007\)H\. M\. Pasula, L\. S\. Zettlemoyer, and L\. P\. KaelblingLearning symbolic models of stochastic domains\.Journal of Artificial Intelligence Research29,pp\. 309–352\.Cited by:[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p4.1)\.
- Sadiaet al\.\(2025\)M\. Sadia Z\. Yanget al\.SQUiD: synthesizing relational databases from unstructured text\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 31975–32000\.Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p1.1)\.
- Sellergrenet al\.\(2026\)A\. Sellergren S\. Kazemzadehet al\.MedGemma technical report\.External Links:2507\.05201,[Link](https://arxiv.org/abs/2507.05201)Cited by:[Empirical Evaluation](https://arxiv.org/html/2608.21186#Sx4.p4.1)\.
- Spasicet al\.\(2020\)I\. Spasicet al\.Clinical text data in machine learning: systematic review\.JMIR medical informatics8\(3\),pp\. e17984\.Cited by:[Neurosymbolic AI](https://arxiv.org/html/2608.21186#Sx2.SSx2.p1.1)\.
- Tantakounet al\.\(2025\)M\. Tantakoun, C\. Muise, and X\. ZhuLLMs as planning formalizers: a survey for leveraging large language models to construct automated planning models\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 25167–25188\.Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Valmeekamet al\.\(2023\)K\. Valmeekam, M\. Marquez, S\. Sreedharan, and S\. KambhampatiOn the planning abilities of large language models\-a critical investigation\.Advances in neural information processing systems36,pp\. 75993–76005\.Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Vasuet al\.\(2026\)B\. Vasu, G\. Raffa, and P\. TadepalliLocal\-to\-global logical explanations for deep vision models\.External Links:2601\.13404,[Link](https://arxiv.org/abs/2601.13404)Cited by:[Data\-Driven Induction\.](https://arxiv.org/html/2608.21186#Sx3.SSx3.SSS0.Px1.p1.1)\.
- Yanget al\.\(2022\)Y\. Yang, Z\. Wu, Y\. Yang, S\. Lian, F\. Guo, and Z\. WangA survey of information extraction based on deep learning\.Applied Sciences12\(19\),pp\. 9691\.Cited by:[Neurosymbolic AI](https://arxiv.org/html/2608.21186#Sx2.SSx2.p1.1)\.
- Younes and Littman \(2004\)H\. L\. Younes and M\. L\. LittmanPPDDL1\.0: an extension to pddl for expressing planning domains with probabilistic effects\.Techn\. Rep\. CMU\-CS\-04\-1622\(99\),pp\. 12\.Cited by:[Planning Domain Models](https://arxiv.org/html/2608.21186#Sx2.SSx1.p1.1)\.
- Zhanget al\.\(2024\)T\. Zhang L\. Zhanget al\.PROC2PDDL: open\-domain planning representations from texts\.InProceedings of the 2nd Workshop on Natural Language Reasoning and Structured Explanations \(@ACL 2024\),pp\. 13–24\.External Links:[Link](https://aclanthology.org/2024.nlrse-1.2/)Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.
- Zhaoet al\.\(2023\)W\. X\. Zhao K\. Zhouet al\.A survey of large language models\.arXiv preprint arXiv:2303\.18223\.Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p1.1)\.
- Zhuet al\.\(2025\)W\. B\. Zhuet al\.Language models can infer action semantics for symbolic planners from environment feedback\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),Albuquerque, New Mexico,pp\. 8751–8773\.External Links:[Link](https://aclanthology.org/2025.naacl-long.440/),[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.440)Cited by:[Pretrained LLMs](https://arxiv.org/html/2608.21186#Sx2.SSx3.p2.1)\.