TELeR Prompting Taxonomy Overview

• TELeR Prompting Taxonomy is a rigorous schema that defines prompts through four dimensions: Turn, Expression, Role, and Level of Detail.
• It enables systematic benchmarking and transparent evaluation by controlling prompt variables to optimize large language model outputs.
• Its practical applications span meta-review synthesis, instructional question generation, and semantic summarization, underpinning reproducible research.

The TELeR Prompting Taxonomy provides a rigorous multidimensional schema for categorizing and engineering prompts for LLMs in complex or ill-defined tasks. It offers both a formal analytic framework and practical guidelines, enabling systematic benchmarking, prompt optimization, and transparency of experimental protocols. TELeR is now widely adopted in LLM behavior evaluation studies across tasks such as peer review meta-synthesis, controlled question generation, and semantic overlap summarization (Santu et al., 2023, Hossain et al., 2024, Knipper et al., 7 Oct 2025, Salvador et al., 2024, Schulhoff et al., 2024).

1. Formal Definition and Four Dimensions

TELeR represents any prompt $P$ as a quadruple:

$$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$

with dimensions:

• Turn: Number of interactive exchanges; $\mathcal{T} = \{$single, multi$\}$
• Expression: Instruction vs. question style; $\mathcal{E} = \{$instruction, question$\}$
• Role: Explicit system/assistant role specification; $\mathcal{R} = \{$defined, undefined$\}$
• Level of Detail ($\mathcal{L}$): Degree of explicitness, constraints, and structured guidance; generally $0 \leq \ell \leq 6$ (Santu et al., 2023, Salvador et al., 2024, Hossain et al., 2024).

This permits construction of $$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$0 distinct prompt categories, allowing controlled experimentation and precise replication.

2. Taxonomy Dimensions: Definitions and Operationalization

Each TELeR dimension is discretized and governs unique aspects of LLM prompting:

Dimension	Discrete Values	Control Mechanism
Turn ($$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$1)	single-turn, multi-turn	Number of exchange rounds
Expression ($$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$2)	instruction, question	Imperative vs. interrogative
Role ($$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$3)	defined, undefined	Persona anchoring
Level of Detail ($$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$4)	0–6 (or 1–4)	Explicitness, constraints, examples

• Turn: Single-turn prompts encapsulate the entire task in one call; multi-turns decompose the task iteratively, enabling decomposition, real-time feedback, and dynamic context injection.
• Expression: Instruction-style (e.g., "Summarize...") yields narrative outputs; question-style (e.g., "What are...") triggers focused, often concise answers.
• Role: Assignment of a meta-role (e.g., "You are a meta-reviewer"; "You are an expert science teacher") modulates LLM formality, field-specific knowledge access, and adherence to desired stylistics (Hossain et al., 2024, Knipper et al., 7 Oct 2025).
• Level of Detail: The most granular axis, controlling scaffolding (e.g., goal statements, sub-tasks, output schemas, rationale requests, retrieval grounding) and defined as a hierarchy, where each higher level either adds constraints, explicit criteria, or externally retrieved evidence (Santu et al., 2023, Knipper et al., 7 Oct 2025, Salvador et al., 2024).

3. Level of Detail: Hierarchical Schemes and Prompt Templates

The Level of Detail dimension ($$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$5) is instantiated via a strictly hierarchical schema:

• Level 0: No directive—raw data only.
• Level 1: Minimal guidance—a single goal statement.
• Level 2: Basic detail—context setup, required format.
• Level 3: Structured considerations—enumerated constraints or sub-task bullets.
• Level 4: Ideal-response specification—output schema or exemplars.
• Level 5: Information retrieval grounding—inject retrieved evidence.
• Level 6: Justification required—mandate rationales for outputs.

Examples from meta-review and science question generation domains:

Level	Prompt Example
1	"As a meta-reviewer, generate a meta-review based on the three reviews below."
2	"Using the three reviews, produce a meta-review addressing: (a) contributions; (b) strengths; (c) weaknesses..."
3	"Draft a meta-review by answering: – What are the core contributions? ... – What suggestions for improvement are mentioned?"
4	"Produce a JSON object: * question: ... * options: ... * answer: ... Ensure the question aligns with the goal of exploring cause-and-effect relationships."

Higher levels systematically add explicit sub-task breakdown, quality constraints, output format specification, and demand justification/explanation for each answer component (Hossain et al., 2024, Knipper et al., 7 Oct 2025, Salvador et al., 2024).

4. Practical Application: Benchmarking and Controlled Comparison

The TELeR taxonomy enables fine-grained benchmarking of LLMs via controlled prompt variants. The standard evaluation pipeline:

1. Select a fixed model $$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$6 and task $$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$7.
2. Define prompt grid $$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$8 by systematic variation of dimensions.
3. For each prompt, generate outputs and evaluate with metrics (e.g., accuracy, ROUGE, SEM-F1, human rating) (Santu et al., 2023, Salvador et al., 2024).
4. Analyze performance as a function of each dimension ($$\text{TELeR}(P) = \left(\text{Turn}(P),\ \text{Expr}(P),\ \text{Role}(P),\ \text{Level}(P)\right) \ \in \ \mathcal{T} \times \mathcal{E} \times \mathcal{R} \times \mathcal{L}$$9).
5. Statistically assess marginal utility of increased detail, choice of role, turn structure, or expression.

This methodology supports apples-to-apples comparison and exposes which axes dominate task accuracy, quality, or faithfulness.

5. Empirical Findings Across Domains

Multiple studies provide converging evidence on the TELeR-driven effects:

• Meta-Review Generation: Fine-grained detail (Level 2–3) yields marked performance gains. Level 2 ("goal plus sub-tasks") is often optimal—higher levels (3–4) sometimes plateau or present diminishing returns. Defined role is essential for style and structure (Hossain et al., 2024).
• Instructional Question Generation: Moderate detail (Levels 2–3) maximizes both format adherence (e.g., JSON parsability up to 0.913) and pedagogical quality (Likert ratings peak at 4.354). Excess schema at Level 4 offers negligible further improvement (Knipper et al., 7 Oct 2025).
• Semantic Overlap Summarization: Minimal instruction (Level 1) is sufficient for tasks with self-explanatory structure; more elaborate scaffolding does not help and may dilute output focus (Salvador et al., 2024).
• Taxonomy-wide Prompt Engineering: In text-based prompting, Task-oriented (T), Example-based (E), Layered Reasoning (L), and Retrieval-Augmented (R) prompting can all be mapped to vectors in TELeR space. Each addresses different complexity axes; for complex reasoning or knowledge-intensive requirements, combining (e.g., L/R) is beneficial (Schulhoff et al., 2024).

6. Composite TELeR/Prompt Type Mappings

Recent work harmonizes traditional prompt classes with TELeR as follows (Schulhoff et al., 2024):

TELeR Family	Principal Variant	Example Techniques
Task-Oriented	⟨single, instruction, role, $\mathcal{T} = \{$0⟩	Zero-shot, Role, Style
Example-Based	⟨single, instruction, (role), $\mathcal{T} = \{$1⟩	Few-shot, In-context exemplars
Layered Reasoning	⟨multi, instruction/question, (role), $\mathcal{T} = \{$2⟩	Chain-of-Thought, Decomposition, Self-critique
Retrieval-Augmented	⟨multi, (instruction), (role), $\mathcal{T} = \{$3⟩	Retrieval-QA, IRCoT, DSP, Toolformer

Instructional prompt engineering typically follows an incremental path: start with Task-Oriented ($\mathcal{T} = \{$4), add exemplars or output criteria (Example-Based, $\mathcal{T} = \{$5), and, for complex tasks, escalate to Layered Reasoning ($\mathcal{T} = \{$6–$\mathcal{T} = \{$7) or Retrieval-Augmented ($\mathcal{T} = \{$8) (Schulhoff et al., 2024, Santu et al., 2023).

7. Prescriptive Guidelines and Best Practices

Consensus guidance derived from comparative studies (Santu et al., 2023, Hossain et al., 2024, Schulhoff et al., 2024, Knipper et al., 7 Oct 2025, Salvador et al., 2024):

• Always report the TELeR quadruple for every prompt (transparent replication).
• For new tasks, begin with minimal or moderate detail (Levels 1–2); only escalate if necessary.
• Assign roles in system prompts for formal or stylistic control.
• Use iterative benchmarking, varying one TELeR axis at a time to identify key contributors.
• For multi-step or highly ill-defined problems, adopt multi-turn (Turn=multi), enumerated sub-tasks (Level $\mathcal{T} = \{$9), request justification (Level 6), and consider retrieval augmentation.
• Monitor for prompt over-specification—excess detail sometimes degrades output quality or confuses the model.
• For factual/structural compliance, prioritize format-specific instructions at Level 2–3; for creative outcomes, allow more flexibility at lower detail levels.
• Employ multiple linguistic prompt variants to mitigate idiosyncratic failure modes, then aggregate or select the best-performing template (Salvador et al., 2024).

Empirical results demonstrate that the optimal TELeR configuration is task-dependent; however, moderate prompt detail and explicit persona anchoring are robustly effective for diverse LLMs and workflows. The TELeR taxonomy forms the methodological backbone for reproducible LLM benchmarking and systematic prompt engineering.