Computational Linguistics Applied to Task Decomposition and Prompt Design

Survey Agent: Computational Linguistics Specialist
Date: 2026-04-06
Context: IAIP Polyphonic Literature Review — Intersection of APE, Computational Linguistics, and Philosophy of AI
Scope: Linguistic theory and analysis as applied to prompt engineering, task decomposition, and the evolution from imperative to interrogative prompting paradigms.

Key Findings

1. Prompts are discourse units with internal rhetorical structure. Rhetorical Structure Theory (Mann & Thompson, 1988) provides a framework for analyzing prompts as hierarchically organized discourse, with nucleus (core instruction) and satellite (context, constraints, examples) relations. Recent work on Enhanced RST (eRST) by Zeldes et al. (2025) extends this to graph-based discourse representations that better model the complex, non-projective relations found in sophisticated multi-component prompts.

2. The linguistic properties of prompts measurably affect LLM output quality, but not in straightforwardly predictable ways. Leidner & Plachouras (2023) showed that neither naturalness nor lower perplexity reliably predict prompt effectiveness; the relationship between linguistic form and output quality is task- and model-dependent. Ma et al. (2024) confirmed this in a large-scale structural analysis of 10,538 real-world prompts, finding that "Capability" and "Demonstration" components outperform simple "Role" specifications.

3. Speech act theory reveals a fundamental asymmetry in human-LLM interaction. Gordon (2024) argues that LLMs are "conversational zombies"—they can produce utterances with perlocutionary effects (persuading, informing) while lacking the intentionality required for genuine illocutionary force under Austin's and Searle's frameworks. This challenges the classical felicity conditions for speech acts and necessitates new theoretical apparatus for human-AI communication.

4. Gricean pragmatics is being systematically operationalized for LLM evaluation. Krause & Vossen (2024) provide the first comprehensive survey of Gricean maxims in NLP, mapping how Quantity, Quality, Relation, and Manner apply to LLM generation and evaluation. Empirical work shows LLMs frequently violate Quantity (over- or under-informing) and Manner (ambiguity), with these violations correlating with user dissatisfaction.

5. Task decomposition in prompting recapitulates compositional semantics. The principle of compositionality—that the meaning of a complex expression derives from its parts and their mode of combination—is structurally mirrored in decomposed prompting (Khot et al., 2023) and self-ask prompting (Press et al., 2023). These methods linguistically decompose complex questions into sub-questions, paralleling how formal semantics builds complex meanings from simpler constituents.

6. Chain-of-thought prompting is linguistically a form of explicit reasoning externalization. Wei et al. (2022) demonstrated that embedding intermediate natural-language reasoning steps in prompts unlocks emergent reasoning abilities in large models. Linguistically, CoT transforms implicit inferential processes into explicit discourse, converting entailment relations into surface-level textual chains.

7. The shift from imperative to interrogative prompting constitutes a change in illocutionary force with measurable semantic consequences. Interrogative prompts invoke question semantics (Hamblin, 1973; Groenendijk & Stokhof, 1984), where the meaning is a set of possible answers, structurally inviting the model to explore an answer space. Imperative prompts invoke command semantics, constraining the model to execute a specified action. This shift from directive to inquisitive framing fundamentally alters the communicative contract between user and model.

8. Instruction tuning reshapes model attention patterns at the linguistic level. Research by Ivison et al. (2024) shows that instruction-tuned models develop increased attention to instruction verbs and semantic structure in lower and middle self-attention layers, with minimal changes to syntactic processing but notable shifts in semantic/instructional interpretation.

9. Prompt design has an information-theoretic foundation. Zhang & Cao (2025) demonstrate at ACL 2025 that prompts function as information selectors, determining which slice of the model's internal representation gets verbalized at each reasoning step. The prompt search space grows combinatorially, and task-specific prompts dramatically outperform generic ones by efficiently routing information extraction.

10. Pragmatic competence remains a frontier challenge for LLMs. Hu et al. (2025) establish through benchmark evaluation (PUB) that LLMs excel at semantic tasks but systematically underperform on pragmatic phenomena including conversational implicature, presupposition accommodation, and deictic reference resolution. This gap has direct implications for the effectiveness of conversational and interrogative prompting strategies.

Theoretical Frameworks Applied to Prompting

1. Rhetorical Structure Theory (RST) — Mann & Thompson (1988); Zeldes et al. (2025)

Application: RST models text as a hierarchy of nucleus-satellite relations. Applied to prompting, the core instruction is the nucleus and contextual elements (role specifications, constraints, examples) are satellites connected by coherence relations (Elaboration, Background, Condition, etc.).

What it reveals: Effective prompts exhibit well-formed rhetorical structure. When satellite information (context, constraints) is properly subordinated to the nucleus (task instruction), coherence is maintained and the model can identify the primary directive. The eRST extension (Zeldes et al., 2025) is particularly relevant for multi-turn prompts where discourse relations cross turn boundaries and may be non-projective.

Key insight for prompt decomposition: Decomposing a complex prompt can be understood as flattening a deep RST tree into a sequence of simpler nucleus-satellite pairs, each expressible as a single coherent instruction.

2. Speech Act Theory — Austin (1962); Searle (1969)

Application: Prompts are analyzed as speech acts with locutionary content (the literal instruction), illocutionary force (the intended communicative function: requesting, commanding, querying), and potential perlocutionary effects (the model's response behavior).

Who applied it: Gordon (2024) in "Speech Acts and Large Language Models" systematically applies Austin-Searle to LLM interaction. Gubelmann (2024) argues from a Kantian-pragmatist perspective that LLMs cannot perform genuine speech acts due to lacking agency. Markl (2025) uses speech act theory to taxonomize representational harms in LLM output as perlocutionary effects without corresponding illocutionary intention.

What it reveals: The imperative-to-interrogative shift in prompting is a shift in illocutionary force from directives (commands that attempt to get the hearer to do something) to questions (requests for information that open a set of acceptable responses). This changes the felicity conditions: directives require the hearer to have the ability and the speaker to have authority; questions require a knowledge asymmetry and the presupposition that an answer exists.

3. Gricean Pragmatics — Grice (1975)

Application: The Cooperative Principle and its four maxims provide a normative framework for evaluating human-LLM interaction.

Who applied it: Krause & Vossen (2024) at INLG provide a comprehensive NLP survey. Chaves & Gerosa (2021) analyze chatbot interactions through a Gricean lens. Participatory design work (2025, arXiv:2503.00858) reinterprets maxims for human-LLM interaction cycles. Rubi et al. (2025) investigate Gricean norms as a basis for effective human-AI collaboration.

What it reveals:

• Quantity: LLMs frequently violate this by over-generating (verbose responses to simple queries) or under-specifying (omitting necessary details). Prompt engineers compensate by explicitly specifying output length and detail level.
• Quality: Hallucination is fundamentally a Quality violation—asserting what is not believed true (though LLMs lack beliefs, the functional equivalent holds).
• Relation: Irrelevant tangents in LLM output violate Relevance. Task-specific prompts improve relevance by narrowing the model's response space.
• Manner: Ambiguous, disorganized, or unnecessarily complex outputs violate Manner. Structured output formatting in prompts (JSON, markdown, numbered lists) is a pragmatic strategy to enforce Manner compliance.

4. Formal Semantics & Compositionality — Frege/Montague tradition

Application: Compositional semantics—the principle that the meaning of a whole derives from the meanings of its parts and their mode of combination—provides the theoretical basis for task decomposition.

What it reveals: Complex prompts can be semantically decomposed along their compositional structure. Each sub-task corresponds to a semantic constituent whose meaning can be independently computed and then combined. Presuppositions in multi-step prompts function as preconditions that must be satisfied (or accommodated) at each stage. Entailment relations between steps ensure logical consistency across the decomposition chain.

5. Question Semantics — Hamblin (1973); Groenendijk & Stokhof (1984)

Application: Formal question semantics treats questions as denoting sets of possible answers (Hamblin) or partitions of logical space (Groenendijk & Stokhof). Applied to interrogative prompting, this framework models the prompt as defining an answer space that the model must navigate.

What it reveals: When a prompt shifts from "Summarize this text" (imperative/directive) to "What are the key points of this text?" (interrogative), the semantic object changes from a command with a single expected execution to a question with a structured set of acceptable answers. This has profound implications for prompt decomposition: interrogative decomposition creates a tree of sub-questions whose answer sets compose into the answer to the parent question—a structure that directly mirrors partition semantics.

6. Discourse Coherence Theory — Hobbs (1979); Asher & Lascarides (2003)

Application: Coherence relations (Cause, Contrast, Elaboration, Narration) explain how discourse segments relate to form coherent text. Applied to prompting, these relations explain why certain prompt orderings and structures are more effective than others.

What it reveals: A prompt that establishes Background before issuing a directive, or that uses Condition relations to specify constraints, leverages the same coherence mechanisms that make natural discourse interpretable. Incoherent prompts—those with unclear or contradictory inter-segment relations—produce degraded model outputs analogously to how incoherent discourse degrades human comprehension.

The Interrogative Turn: Linguistic Analysis

The Phenomenon

Prompt engineering practice has evolved from predominantly imperative constructions ("List X," "Generate Y," "Write Z") toward interrogative and dialogical forms ("What would be the best approach to X?," "How might we think about Y?," "Can you explore Z?"). This is not merely a stylistic preference; it constitutes a linguistically significant shift with theoretical implications across multiple frameworks.

Speech Act Analysis

Under Searle's taxonomy, the shift moves from directives (illocutionary point: getting the addressee to do something; direction of fit: world-to-words) to questions (a sub-type of directives with the illocutionary point of eliciting information; direction of fit: words-to-world). The preparatory conditions change: directives presuppose the hearer can perform the action; questions presuppose the hearer knows the answer and the questioner does not.

For LLM interaction, this shift has functional consequences. Imperative prompts constrain the model to a narrow execution path (low response entropy). Interrogative prompts invite the model to survey a wider response space, often producing more nuanced, hedged, and exploratory outputs. The self-ask method (Press et al., 2023) formalizes this by having the model ask itself follow-up questions—turning the model into both questioner and respondent in an internalized dialogue.

Semantic Analysis

From the perspective of formal question semantics:

• An imperative prompt like "Summarize X" denotes a single action to be performed, evaluated by whether the output satisfies the command.
• An interrogative prompt like "What are the key themes in X?" denotes a set of propositions (Hamblin semantics), each a possible answer. The model's task is to select from this set.

This difference is structurally significant for decomposition. Imperative decomposition yields a sequence of sub-commands (do A, then B, then C). Interrogative decomposition yields a tree of sub-questions whose answers compose hierarchically—a structure directly modeled by partition semantics (Groenendijk & Stokhof). The interrogative mode thus provides a richer semantic framework for structured reasoning.

Pragmatic Analysis

Under Gricean pragmatics, questions trigger different implicature patterns than commands:

• Questions carry the implicature that the questioner does not know the answer (sincerity condition), licensing the respondent to provide information the questioner lacks.
• Commands carry the implicature that the speaker has authority and the hearer has ability (preparatory conditions).

When applied to human-LLM interaction, interrogative prompting implicitly frames the interaction as a knowledge-sharing dialogue rather than a master-servant execution chain. This framing change may explain empirical observations that interrogative prompts elicit more contextual, reasoned, and qualified responses—the model's training on dialogue corpora associates question-answer pairs with explanatory, cooperative communicative behavior.

Dialogical Structure

The interrogative turn also connects to dialogue act theory (Bunt, 2009) and the Question Under Discussion (QUD) framework (Roberts, 2012). In QUD theory, discourse is organized around implicit or explicit questions that interlocutors collaboratively address. Interrogative prompting makes the QUD explicit, providing the model with a clearer discourse structure to navigate.

Multi-turn interrogative prompting creates a discourse tree of QUDs, where each sub-question refines or decomposes the parent question. This is precisely the structure that methods like Tree of Thoughts (Yao et al., 2023) operationalize—each "thought" is a candidate answer to an implicit sub-question, and the tree search navigates the space of possible QUD resolutions.

The Linguistic Significance

The interrogative turn represents a recognition that prompting is communication, not programming. Linguistically, this shift:

1. Moves from monological (one-way instruction) to dialogical (collaborative inquiry) discourse mode.
2. Changes the semantic type of the prompt from an action description to a proposition set.
3. Alters the pragmatic presuppositions from authority/ability to knowledge asymmetry.
4. Creates richer discourse structure amenable to hierarchical decomposition.
5. Better aligns with the training distribution of instruction-tuned models, which are predominantly trained on conversational data.

Key Papers (Annotated)

1. Mann, W.C. & Thompson, S.A. (1988). "Rhetorical Structure Theory: Toward a Functional Theory of Text Organization." Text, 8(3), 243–281.

Key contribution: Foundational theory of discourse organization via nucleus-satellite relations and a taxonomy of rhetorical relations (Elaboration, Background, Evidence, Condition, etc.).
Linguistic framework: Discourse analysis, functional linguistics.
Relevance: Provides the theoretical basis for understanding prompt internal structure. Complex prompts contain RST relations (the instruction is the nucleus; context/constraints/examples are satellites). Decomposing prompts can be modeled as flattening RST trees.

2. Zeldes, A. et al. (2025). "eRST: A Signaled Graph Theory of Discourse Relations and Organization." Computational Linguistics, 51(1), 23–72.

Key contribution: Extends RST to graph structures supporting non-projective, tree-breaking, and concurrent discourse relations with explicit signaling.
Linguistic framework: Enhanced Rhetorical Structure Theory.
Relevance: Directly applicable to multi-component and multi-turn prompts where discourse relations cross segment boundaries, are overlapping, or are non-hierarchical. Provides annotation tools and a 200,000-token annotated corpus.

3. Gordon, J. (2024). "Speech Acts and Large Language Models." PhilArchive.

Key contribution: Applies Austin-Searle speech act theory to LLM interaction, introducing the "conversational zombie" analogy—LLMs simulate speech act form without possessing the intentionality required for genuine illocutionary force.
Linguistic framework: Speech act theory, philosophy of language.
Relevance: Foundational for understanding the philosophical-linguistic status of prompts as speech acts and for analyzing the imperative-to-interrogative shift as a change in illocutionary force type.

4. Gubelmann, R. (2024). "Large Language Models, Agency, and Why Speech Acts are Beyond Them (For Now)." Philosophy & Technology, 37, 45.

Key contribution: Kantian-pragmatist argument that LLMs cannot be genuine language users because speech acts require autonomous agency and self-constitution.
Linguistic framework: Speech act theory, Kantian pragmatics.
Relevance: Establishes theoretical limits on applying speech act theory to LLM-produced outputs, relevant to understanding whether model responses constitute genuine speech acts.

5. Markl, N. (2025). "Taxonomizing Representational Harms using Speech Act Theory." arXiv:2504.00928.

Key contribution: Uses speech act theory to classify representational harms (stereotyping, erasure) in LLM output as perlocutionary effects of machine-generated pseudo-illocutionary acts.
Linguistic framework: Speech act theory, sociolinguistics.
Relevance: Demonstrates that speech act theory applies productively to LLM output analysis, not just input (prompt) analysis.

6. Krause, L. & Vossen, P. (2024). "The Gricean Maxims in NLP — A Survey." Proceedings of INLG 2024, ACL Anthology.

Key contribution: First comprehensive survey mapping Gricean maxims to NLP tasks, benchmarks, and evaluation metrics. Covers Quantity, Quality, Relation, and Manner across data-to-text, dialogue, and generation tasks.
Linguistic framework: Gricean pragmatics, cooperative principle.
Relevance: Provides the definitive current mapping between pragmatic theory and NLP practice, directly applicable to understanding prompt design as cooperative communication.

7. (2025). "Applying the Gricean Maxims to a Human-LLM Interaction Cycle: Design Insights from a Participatory Approach." arXiv:2503.00858.

Key contribution: Participatory design study reinterpreting Gricean maxims specifically for human-LLM interaction, producing actionable design guidelines.
Linguistic framework: Gricean pragmatics, participatory design.
Relevance: Bridges theoretical pragmatics and practical prompt design, offering empirically grounded guidelines for cooperative prompting.

8. Leidner, J.L. & Plachouras, V. (2023). "The Language of Prompting: What Linguistic Properties Make a Prompt Successful?" Findings of EMNLP 2023.

Key contribution: Empirical study correlating linguistic properties of prompts (mood, tense, modality, syntactic complexity, naturalness, perplexity) with LLM output quality across multiple tasks.
Linguistic framework: Corpus linguistics, descriptive linguistics.
Relevance: Foundational empirical work establishing that prompt effectiveness is linguistically analyzable but not reducible to simple surface features.

9. Ma, Y. et al. (2024). "The Death and Life of Great Prompts: Analyzing the Evolution of LLM Prompts from the Structural Perspective." Proceedings of EMNLP 2024.

Key contribution: Large-scale structural analysis of 10,538 real-world prompts, identifying eight structural components and tracking their evolution over time. Finds that "Capability" and "Demonstration" components matter more than "Role" specification.
Linguistic framework: Structural/discourse analysis, corpus linguistics.
Relevance: Provides the most comprehensive empirical picture of prompt structure evolution, framing prompt engineering as an evolving linguistic practice.

10. Wei, J. et al. (2022). "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." Proceedings of NeurIPS 2022.

Key contribution: Demonstrates that including intermediate natural-language reasoning steps in few-shot prompts dramatically improves performance on arithmetic, commonsense, and symbolic reasoning tasks, but only in sufficiently large models (~100B+ parameters).
Linguistic framework: Implicit: discourse structure, reasoning externalization.
Relevance: CoT is linguistically a form of making entailment chains explicit in discourse. The "think step by step" instruction is a meta-pragmatic directive that changes the discourse mode from single-turn response to multi-step reasoning narrative.

11. Yao, S. et al. (2023). "Tree of Thoughts: Deliberate Problem Solving with Large Language Models." Proceedings of NeurIPS 2023.

Key contribution: Generalizes chain-of-thought to tree-structured reasoning, enabling exploration, evaluation, and backtracking across multiple reasoning paths using BFS/DFS.
Linguistic framework: Implicit: question semantics (each "thought" is an answer to an implicit sub-question), discourse tree structure.
Relevance: Operationalizes what question semantics predicts: that interrogative decomposition naturally produces tree structures whose nodes are sub-questions and whose edges are entailment or refinement relations.

12. Khot, T. et al. (2023). "Decomposed Prompting: A Modular Approach for Solving Complex Tasks." Proceedings of ICLR 2023.

Key contribution: Introduces DecomP—a framework where complex tasks are broken into sub-tasks, each handled by specialized prompts or modules that can be independently optimized and reused.
Linguistic framework: Implicit: compositional semantics (complex meaning from composed sub-meanings).
Relevance: DecomP structurally mirrors the principle of compositionality: just as complex expressions are semantically decomposed into sub-expressions, complex tasks are decomposed into sub-tasks whose solutions compose into the final answer.

13. Press, O. et al. (2023). "Measuring and Narrowing the Compositionality Gap in Language Models." Findings of EMNLP 2023.

Key contribution: Identifies the "compositionality gap"—LLMs can answer sub-questions correctly but fail to compose them—and introduces self-ask prompting where the model explicitly generates and answers follow-up sub-questions.
Linguistic framework: Compositional semantics, question semantics.
Relevance: Directly demonstrates that interrogative self-decomposition (asking sub-questions) outperforms imperative linear reasoning. The compositionality gap is a linguistic-computational phenomenon: models process parts but fail at composition, paralleling debates in formal semantics about the nature of compositionality.

14. Zhang, J. & Cao, Y. (2025). "Why Prompt Design Matters and Works: A Complexity Analysis of Prompt Search Space in LLMs." Proceedings of ACL 2025.

Key contribution: Information-theoretic analysis showing that prompts act as information selectors determining which bits of the model's internal representation are verbalized at each reasoning step. The prompt search space grows combinatorially.
Linguistic framework: Information theory, computational complexity.
Relevance: Provides formal grounding for why linguistic choices in prompt design matter—each word shapes an information extraction pathway. Task-specific prompts improve performance by over 50% compared to generic prompts.

15. Ivison, H. et al. (2024). "From Language Modeling to Instruction Following: Understanding the Behavior Shift in LLMs after Instruction Tuning." Proceedings of NAACL 2024.

Key contribution: Mechanistic analysis of how instruction tuning changes model internals—attention heads in lower/middle layers shift to focus on instruction verbs and semantic structure, while syntactic processing remains largely unchanged.
Linguistic framework: Computational psycholinguistics, attention analysis.
Relevance: Demonstrates that instruction tuning creates a linguistically specific change: the model learns to prioritize pragmatic/semantic instruction content over syntactic form, providing a computational basis for why pragmatically well-designed prompts are effective.

16. Hu, J. et al. (2025). "Pragmatics in the Era of Large Language Models." arXiv:2502.12378.

Key contribution: Comprehensive survey of LLM pragmatic capabilities including implicature, presupposition, reference, and deixis. Establishes benchmarks (including PUB) for pragmatic evaluation.
Linguistic framework: Pragmatics, discourse analysis.
Relevance: Maps the frontier of LLM pragmatic competence, identifying precisely where models fail at the pragmatic reasoning that sophisticated prompting strategies demand.

17. (2024). "Combining Discourse Coherence with Large Language Models for More Inclusive and Equitable Responses." Proceedings of LREC-COLING 2024.

Key contribution: Demonstrates that integrating discourse coherence theories (including Hobbs' coherence relations) into LLM pipelines improves intent recognition, robustness, and inclusivity across dialects.
Linguistic framework: Discourse coherence theory.
Relevance: Shows practical benefits of linguistically-informed prompt and pipeline design for equitable AI interaction.

18. Lin, Z. (2024). "Prompt Engineering for Applied Linguistics: Elements, Examples, Techniques, and Strategies." ResearchGate.

Key contribution: Bridges applied linguistics methodology and prompt engineering practice, providing element-based frameworks for prompt construction grounded in linguistic theory.
Linguistic framework: Applied linguistics, pedagogical grammar.
Relevance: Demonstrates growing recognition in the applied linguistics community that prompt engineering is itself a linguistic practice amenable to linguistic analysis.

19. (2025). "A Comprehensive Taxonomy of Prompt Engineering Techniques for Large Language Models." Frontiers of Computer Science, Springer.

Key contribution: Taxonomizes prompt engineering techniques along dimensions including profile/instruction, knowledge, reasoning/planning, and reliability. Covers personalization and cultural/linguistic adaptation.
Linguistic framework: Taxonomic/structural linguistics applied to prompt classification.
Relevance: Provides a systematic classification framework that maps the space of prompt variation, enabling systematic linguistic study of prompt types and their effects.

20. (2024). "Towards an Implementation of Rhetorical Structure Theory in Discourse Coherence Modelling." Proceedings of ALTA 2024.

Key contribution: Integrates RST with Graph Convolutional Networks for coherence modeling, applying RST-based features to distinguish human- from machine-generated text and improve essay grading.
Linguistic framework: RST, neural discourse models.
Relevance: Demonstrates computational operationalization of RST for evaluating the coherence of LLM-generated outputs—directly applicable to assessing whether prompted outputs maintain discourse coherence.

Bridging Concepts

These are the theoretical joints connecting computational linguistics to prompt engineering and task decomposition—the conceptual vocabulary that enables cross-disciplinary analysis.

1. Illocutionary Force (Speech Act Theory → Prompt Design)

The type of speech act a prompt performs (directive, interrogative, assertive) determines the communicative contract between user and model. Shifting illocutionary force from directive to interrogative is the core linguistic mechanism underlying the "interrogative turn" in prompting.

2. Nucleus-Satellite Structure (RST → Prompt Architecture)

The distinction between core instruction (nucleus) and supporting context (satellite) provides a principled vocabulary for prompt component analysis. Effective prompts maintain clear nuclearity; degraded prompts exhibit competing or ambiguous nuclei.

3. Compositionality (Formal Semantics → Task Decomposition)

The Fregean principle that complex meanings derive from parts and their combination is the linguistic analog of task decomposition. DecomP, self-ask, and CoT all operationalize compositionality by decomposing complex semantic objects (instructions, questions) into simpler constituents.

4. Conversational Implicature (Gricean Pragmatics → Prompt Interpretation)

What is implied but not stated in a prompt—and how models handle (or fail to handle) such implicatures—is critical for understanding prompt effectiveness. The gap between what a user means and what they literally say is the domain of conversational implicature.

5. Question Under Discussion (QUD Framework → Discourse-Level Prompting)

The QUD framework (Roberts, 2012) models discourse as organized around implicit questions. Multi-turn prompting creates explicit QUD stacks; interrogative decomposition generates QUD trees. This framework bridges discourse theory and practical prompt decomposition strategy.

6. Felicity Conditions (Speech Act Theory → Prompt Success Conditions)

Austin's felicity conditions—the conditions under which a speech act succeeds—translate to prompt design as the conditions under which a prompt achieves its intended effect. A prompt "misfires" (Austin's term) when preparatory conditions are unmet (e.g., asking the model to perform something beyond its capability) or when essential conditions are violated (e.g., ambiguous intent).

7. Answer Set / Partition Semantics (Formal Question Semantics → Interrogative Prompting)

Hamblin's treatment of questions as denoting answer sets, and Groenendijk & Stokhof's partition semantics, provide the formal machinery for understanding what an interrogative prompt means in a way that imperative semantics cannot capture—the prompt defines a structured space of acceptable responses.

8. Coherence Relations (Discourse Theory → Multi-Step Prompt Design)

Hobbs' coherence relations (Cause, Enablement, Elaboration, Parallel) explain why certain multi-step prompt orderings work better than others. Steps connected by clear causal or enablement relations produce more coherent model reasoning than steps with unclear inter-relations.

9. Presupposition Accommodation (Formal Pragmatics → Prompt Context Management)

Presuppositions in prompts (background assumptions the prompt takes for granted) must be either already in the model's context or accommodatable. Failed presupposition accommodation—where the model lacks the presupposed knowledge—is a major source of prompt failure in multi-step chains.

10. Information Structure (Functional Linguistics → Prompt Focus Management)

The given/new distinction in information structure (Halliday, 1967; Prince, 1981) maps to prompt design: effective prompts establish shared context (given) before introducing new task requirements. This parallels the Topic-Comment structure that governs information flow in natural discourse.

Open Problems & Research Gaps

1. No Systematic Discourse Grammar of Prompts

Despite growing empirical work on prompt structure (Ma et al., 2024), there is no formal grammar or annotation scheme specifically designed for prompt discourse structure. RST and eRST are adaptable but were not designed for the unique characteristics of instructional prompts (e.g., meta-instructions like "think step by step," role assignments, output format specifications). A prompt-specific discourse grammar is needed.

2. Pragmatic Competence Testing Remains Underdeveloped

While Hu et al. (2025) survey the landscape, standardized benchmarks for LLM pragmatic competence—especially for conversational implicature resolution, presupposition accommodation, and indirect speech act interpretation—remain sparse. Current evaluations are largely semantic; pragmatic evaluation is still an emerging frontier.

3. Cross-Linguistic Prompt Linguistics is Nearly Absent

Almost all linguistic analysis of prompting has been conducted on English-language prompts and English-centric models. How prompting strategies interact with typologically diverse languages—languages with different question formation strategies, different honorific systems affecting speech act felicity, different discourse organization principles—is virtually unstudied.

4. The Compositional Semantics of Multi-Agent Prompt Chains is Unformalized

When multiple agents each receive decomposed sub-prompts whose outputs are composed, the compositional semantics of this chain—including how presuppositions propagate, how entailment relations are maintained across agent boundaries, and how coherence is preserved—lacks formal treatment.

5. No Linguistic Theory of Prompt Failure

While prompt engineering practice accumulates empirical knowledge of what fails, there is no systematic linguistic theory of why prompts fail. Speech act theory's "infelicity" and Grice's "maxim violation" provide starting vocabulary, but a comprehensive taxonomy of linguistically-motivated prompt failure modes does not yet exist.

6. The Semantics of Meta-Prompting is Unexplored

Instructions like "think step by step," "you are an expert in X," or "be concise" are meta-level pragmatic operators that modify how subsequent instructions are interpreted. These have no clear analog in standard linguistic theory—they are neither standard speech acts nor standard discourse operators. Their semantics and compositional behavior await formalization.

7. Dialogue Act Theory Has Not Been Systematically Applied to Multi-Turn Prompting

While QUD and dialogue act theory (Bunt, 2009) are relevant, systematic empirical application to multi-turn human-LLM interaction—coding prompt turns for dialogue acts and analyzing how act sequences affect model behavior—remains largely undone.

8. The Relationship Between Syntactic Complexity and Prompt Effectiveness is Non-Linear and Poorly Understood

Leidner & Plachouras (2023) showed that neither simplicity nor complexity reliably predicts effectiveness. The interaction between syntactic structure, model architecture, and task type creates a complex landscape that resists simple generalizations. More nuanced corpus-linguistic work is needed.

9. Linguistic Analysis of the "Prompt Space" is in Its Infancy

Zhang & Cao (2025) provide an information-theoretic account of prompt search space, but this has not been integrated with linguistic description of that space (What types of linguistic variation exist? What are the axes of meaningful variation?). Mapping the linguistic typology of the prompt space is an open challenge.

10. Indigenous and Non-Western Linguistic Frameworks for Prompting Are Unrepresented

Current theoretical frameworks are exclusively drawn from Western linguistic traditions (Anglo-American analytic philosophy, European structural linguistics). Indigenous linguistic frameworks, oral discourse traditions, and non-Western pragmatic theories offer potentially valuable alternative perspectives on instruction, question-asking, and collaborative knowledge construction that have not been explored in the context of prompt design.

Sources

Foundational Linguistic Theory

1. Austin, J.L. (1962). How to Do Things with Words. Oxford University Press.
2. Searle, J.R. (1969). Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.
3. Grice, H.P. (1975). "Logic and Conversation." In Syntax and Semantics 3: Speech Acts, P. Cole & J. Morgan (eds.), Academic Press, 41–58.
4. Mann, W.C. & Thompson, S.A. (1988). "Rhetorical Structure Theory: Toward a Functional Theory of Text Organization." Text, 8(3), 243–281.
5. Hamblin, C.L. (1973). "Questions in Montague English." Foundations of Language, 10(1), 41–53.
6. Groenendijk, J. & Stokhof, M. (1984). Studies on the Semantics of Questions and the Pragmatics of Answers. PhD Dissertation, University of Amsterdam.
7. Hobbs, J.R. (1979). "Coherence and Coreference." Cognitive Science, 3(1), 67–90.
8. Asher, N. & Lascarides, A. (2003). Logics of Conversation. Cambridge University Press.
9. Roberts, C. (2012). "Information Structure in Discourse: Towards an Integrated Formal Theory of Pragmatics." Semantics and Pragmatics, 5(6), 1–69.
10. Bunt, H. (2009). "The DIT++ Taxonomy for Functional Dialogue Acts." In Proceedings of the EDAML 2009 Workshop.
11. Halliday, M.A.K. (1967). "Notes on Transitivity and Theme in English, Part 2." Journal of Linguistics, 3(2), 199–244.

LLM–Linguistics Intersection (2022–2025)

12. Wei, J. et al. (2022). "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." Proceedings of NeurIPS 2022. https://proceedings.neurips.cc/paper_files/paper/2022/hash/9d5609613524ecf4f15af0f7b31abca4-Abstract-Conference.html
13. Yao, S. et al. (2023). "Tree of Thoughts: Deliberate Problem Solving with Large Language Models." Proceedings of NeurIPS 2023. https://proceedings.neurips.cc/paper/2023/hash/271db9922b8d1f4dd7aaef84ed5ac703-Abstract.html
14. Khot, T. et al. (2023). "Decomposed Prompting: A Modular Approach for Solving Complex Tasks." Proceedings of ICLR 2023. https://openreview.net/forum?id=_nGgzQjzaRy
15. Press, O. et al. (2023). "Measuring and Narrowing the Compositionality Gap in Language Models." Findings of EMNLP 2023. https://aclanthology.org/2023.findings-emnlp.378/
16. Leidner, J.L. & Plachouras, V. (2023). "The Language of Prompting: What Linguistic Properties Make a Prompt Successful?" Findings of EMNLP 2023. https://arxiv.org/abs/2311.01967
17. Ye, Q. et al. (2024). "Prompt Engineering a Prompt Engineer." Findings of ACL 2024. https://aclanthology.org/2024.findings-acl.21/
18. Gordon, J. (2024). "Speech Acts and Large Language Models." PhilArchive. https://philarchive.org/archive/GORSAA-12v1
19. Gubelmann, R. (2024). "Large Language Models, Agency, and Why Speech Acts are Beyond Them (For Now)." Philosophy & Technology, 37, 45. https://link.springer.com/article/10.1007/s13347-024-00696-1
20. Krause, L. & Vossen, P. (2024). "The Gricean Maxims in NLP — A Survey." Proceedings of INLG 2024. https://aclanthology.org/2024.inlg-main.39/
21. Ma, Y. et al. (2024). "The Death and Life of Great Prompts: Analyzing the Evolution of LLM Prompts from the Structural Perspective." Proceedings of EMNLP 2024. https://aclanthology.org/2024.emnlp-main.1227/
22. Ivison, H. et al. (2024). "From Language Modeling to Instruction Following: Understanding the Behavior Shift in LLMs after Instruction Tuning." Proceedings of NAACL 2024. https://aclanthology.org/2024.naacl-long.130/
23. Zeldes, A. et al. (2025). "eRST: A Signaled Graph Theory of Discourse Relations and Organization." Computational Linguistics, 51(1), 23–72. https://direct.mit.edu/coli/article/51/1/23/124464/
24. Zhang, J. & Cao, Y. (2025). "Why Prompt Design Matters and Works: A Complexity Analysis of Prompt Search Space in LLMs." Proceedings of ACL 2025. https://aclanthology.org/2025.acl-long.1562/
25. Hu, J. et al. (2025). "Pragmatics in the Era of Large Language Models." arXiv:2502.12378. https://arxiv.org/abs/2502.12378
26. Markl, N. (2025). "Taxonomizing Representational Harms using Speech Act Theory." arXiv:2504.00928. https://arxiv.org/abs/2504.00928
27. (2025). "Applying the Gricean Maxims to a Human-LLM Interaction Cycle: Design Insights from a Participatory Approach." arXiv:2503.00858. https://arxiv.org/abs/2503.00858
28. Chaves, A.P. & Gerosa, M.A. (2021). "How Should My Chatbot Interact? A Survey on Social Characteristics in Human–Chatbot Interaction Design." International Journal of Human–Computer Interaction, 37(8), 729–758.
29. (2024). "Combining Discourse Coherence with Large Language Models for More Inclusive and Equitable Responses." Proceedings of LREC-COLING 2024. https://aclanthology.org/2024.lrec-main.314/
30. (2024). "Towards an Implementation of Rhetorical Structure Theory in Discourse Coherence Modelling." Proceedings of ALTA 2024. https://aclanthology.org/2024.alta-1.1/
31. Lin, Z. (2024). "Prompt Engineering for Applied Linguistics: Elements, Examples, Techniques, and Strategies." ResearchGate.
32. (2025). "A Comprehensive Taxonomy of Prompt Engineering Techniques for Large Language Models." Frontiers of Computer Science, Springer. https://link.springer.com/article/10.1007/s11704-025-50058-z
33. (2024). "An Empirical Categorization of Prompting Techniques for Large Language Models." arXiv:2402.14837. https://arxiv.org/abs/2402.14837
34. Mahowald, K. et al. (2023). "Language Models and Linguistic Theories Beyond Words." Nature Machine Intelligence. https://www.nature.com/articles/s42256-023-00703-8
35. (2025). "Increasing Alignment of Large Language Models with Language Processing in the Human Brain." Nature Computational Science. https://www.nature.com/articles/s43588-025-00863-0
36. Lou, R. et al. (2024). "Large Language Model Instruction Following: A Survey of Progresses and Challenges." Computational Linguistics, 50(3), 1053–1106. https://direct.mit.edu/coli/article/50/3/1053/121669/
37. (2024). "Comparative Analysis of Prompt Strategies for Large Language Models." Electronics, 13(23), 4712. https://www.mdpi.com/2079-9292/13/23/4712
38. (2025). "Large Language Model Prompt Datasets: An In-depth Analysis and Insights." arXiv:2510.09316. https://arxiv.org/abs/2510.09316

This survey was produced as part of the IAIP Polyphonic Literature Review project. It represents the computational linguistics angle of a multi-agent academic survey examining how prompt decomposition engines are evolving from structured instructions to dynamic conversational inquiries. Cross-reference with companion surveys on Automatic Prompt Engineering (technical), Philosophy of AI Mind, and Indigenous Knowledge Systems for the complete literature review.