Recursive Meta-Prompting

• Recursive Meta-Prompting (RMP) is a framework that enables LLMs to iteratively generate, refine, and optimize prompts using recursive self-improvement cycles.
• It leverages formal monadic structures and semantic gradient descent, ensuring stable convergence and composable prompt refinements across different tasks.
• RMP has demonstrated significant improvements in mathematical reasoning, video summarization, and long-context inference, showcasing its versatility in multi-modal applications.

Recursive Meta-Prompting (RMP) is a formal paradigm in which LLMs are orchestrated to not only generate outputs to task prompts, but to iteratively generate, refine, and evaluate their own prompts through a recursive self-improvement process. Unlike traditional prompt-engineering, which depends on static, expert-designed templates or ad-hoc feedback, RMP enables fully automated, compositional, and principled prompt optimization through higher-order meta-instructions—enabling LLMs to bootstrap high-quality task performance, discover modular prompt hierarchies, and scale to tasks infeasible for manual engineering (Zhang et al., 2023, Zhang et al., 31 Dec 2025, Fu, 17 Dec 2025, Hu et al., 22 Apr 2025).

1. Formal Foundations: Monad and Semantic Computation

Recursive Meta-Prompting is mathematically structured by modeling prompt generation and refinement as a monadic self-improvement loop. Given a category $\mathcal{P}$ of prompts, RMP defines a monad $(\mathcal{M}_P,\,\eta,\,\mu)$ such that:

• The endofunctor $\mathcal{M}_P:\mathcal{P}\to\mathcal{P}$ implements a single refinement step by applying a meta-meta-prompt to an initial prompt.
• The unit $\eta$ injects an unstructured raw task description $T$ into the space of structured prompts:

$\eta_T: T \mapsto P_0 = \text{InitialPrompt}(T).$

• The multiplication $\mu$ collapses nested layers of refinement, guaranteeing stability and associativity:

$$\mu_{P}:\;\mathcal{M}_P\bigl(\mathcal{M}_P(P)\bigr)\to \mathcal{M}_P(P).$$

The monad laws—left identity, right identity, and associativity—guarantee that recursive refinement and composition of prompt improvements yields consistent results regardless of grouping, underpinning both theoretical soundness and compositionality of the RMP loop (Zhang et al., 2023, Fu, 17 Dec 2025).

In the Meta-Prompting Protocol, natural language prompts are treated as variables in a semantic computation graph, supporting "textual gradients"—text-based surrogate gradients derived from auditor critiques to perform prompt updates via feedback loops. The update step at iteration $t$ is expressed as

$I_{t+1} = I_t - \eta\,\nabla_I L_t,$

where the loss $(\mathcal{M}_P,\,\eta,\,\mu)$0 is semantic, computed as $(\mathcal{M}_P,\,\eta,\,\mu)$1, with $(\mathcal{M}_P,\,\eta,\,\mu)$2 a normalized auditor score, and $(\mathcal{M}_P,\,\eta,\,\mu)$3 is constructed via parsing structured critiques, enabling prompt optimization as gradient descent in prompt-space (Fu, 17 Dec 2025).

2. Algorithmic Structure and Implementation

Recursive meta-prompting proceeds by iteratively conducting a generate–evaluate–refine cycle:

1. Prompt Generation: The current meta-prompt (or prompt scaffold) is used by the LLM to generate task outputs.
2. Evaluation: The output is assessed, either by a dedicated evaluator LLM, an auxiliary scoring function, or a task-specific auditor. Evaluation can be based on semantic relevance, coherence, coverage, brevity, or external rule sets (Hu et al., 22 Apr 2025).
3. Prompt Refinement: Based on critiques and evaluation, a prompt optimizer LLM (or an explicit meta-meta-prompt) rewrites the prompt, leading to a new meta-prompt for the next iteration.

Concrete algorithmic pseudocode (e.g., (Zhang et al., 2023)) captures the refinement loop: $(\mathcal{M}_P,\,\eta,\,\mu)$6 This pattern generalizes to settings with adversarial feedback (Generator–Auditor–Optimizer trinity), multi-agent pipelines (as in ViSMaP's three-LLM scheme), or REPL-based recursive LLMs (Zhang et al., 31 Dec 2025, Fu, 17 Dec 2025, Hu et al., 22 Apr 2025).

3. Key Instantiations Across Application Domains

Mathematical and Reasoning Benchmarks

In "Meta Prompting for AI Systems" (Zhang et al., 2023), RMP-driven Qwen-72B outperforms few-shot Chain-of-Thought (CoT) prompting on challenging mathematics (MATH: +11.1 pp, GSM8K: +4.6 pp, Game of 24: 100% success). The process requires only a zero-shot, content-agnostic meta-prompt and does not depend on human-curated demonstrations. Performance gains are attributed to the stable convergence properties of the monadic refinement process and substantial token-efficiency, as the prompts remain compact and modular.

Video Summary Generation

ViSMaP (Hu et al., 22 Apr 2025) applies RMP in multi-modal settings, using a tri-LLM pipeline (generator, evaluator, optimizer) to iteratively improve prompts for hour-long video summarization. Formalization employs a meta-optimization loop:

$(\mathcal{M}_P,\,\eta,\,\mu)$4

where $(\mathcal{M}_P,\,\eta,\,\mu)$5 is an evaluator score over summaries. Rapid convergence is observed (typically 4–6 cycles), and empirical evaluations demonstrate that fully unsupervised ViSMaP achieves near-parity with state-of-the-art supervised methods on benchmarks (e.g., Ego4D-HCap, ROUGE-L within 2.7 points of the best supervised system).

Long-Context Inference

Recursive LLMs (RLM) (Zhang et al., 31 Dec 2025) instantiate RMP by allowing a base LLM to recursively decompose and process input prompts far exceeding its context window. RLMs operate by symbolically manipulating prompt "memory" as an external environment, spawning sub-queries over relevant slices, and recursively aggregating results. RLMs dramatically improve performance and cost-efficiency on tasks such as CodeQA and BrowseComp-Plus, attaining, for example, 91.3% accuracy (BrowseComp+) at <$1 per query, outperforming both base LLMs and summary-agent baselines by large margins.

Software Engineering via Semantic Feedback Loops

The Meta-Prompting Protocol (Fu, 17 Dec 2025) operationalizes RMP in mission-critical LLM orchestration by treating prompts as differentiable nodes in a semantic computation graph. The Adversarial Trinity structure (Generator, Auditor, Optimizer) enables robust prompt self-optimization, leveraging DSPy (Declarative Self-Improving Pipelines) for modular composition, and TextGrad for automatic textual differentiation. Audited feedback functions as surrogate gradients, systematically improving prompt scaffolds to mitigate hallucinations and deliver deterministic performance guarantees.

4. Theoretical Properties and Guarantees

• Stable Convergence: Monad laws ensure that nesting of prompt refinements is associative, enabling stable convergence of multi-step optimize-evaluate cycles. Theoretical results guarantee that the sequence of prompt improvements is robust to variation in association order (Zhang et al., 2023).
• Compositionality: The functorial structure of the monad enables prompt modules for subtasks to be composed, refined independently, and integrated seamlessly into larger workflows.
• Convergence Criteria: Termination relies on explicit convergence predicates—either by prompt edit distance, stochastic stability in evaluator scores, or functional correctness. Providing formal, model-agnostic convergence bounds is an open research problem.
• Semantic Gradient Descent: By modeling prompt updates as "gradient descent" in prompt-space using textual critiques, RMP establishes a correspondence between natural language optimization and classic supervised learning—bridging black-box LLM control with differentiable programming (Fu, 17 Dec 2025).

5. Limitations and Open Questions

• Convergence and Optimality: While monad laws guarantee compositional stability, they do not ensure optimality or termination within finite iterations. Oscillations or plateauing may occur, especially with poorly-specified initial prompts.
• Quality of Initialization: The effectiveness of RMP is sensitive to the choice of initial prompt scaffold. If the base prompt is excessively vague, the refinement process may converge to suboptimal solutions.
• Error Propagation: Mistakes in earlier refined prompts can propagate or become entrenched unless corrected via self-consistency or external evaluators.
• Automation Challenges: Certain instances, such as error identification in recursive chain-of-feedback frameworks, may require manual intervention. Automating these remains a subject of ongoing investigation (Ahn et al., 2024).
• Generality Across Modalities: While RMP generalizes to multi-modal tasks (e.g., video, code, QA), adaptation to settings beyond reconstructive text generation is an open area.

6. Empirical Performance and Comparative Analysis

Method / Domain	Metric	RMP Score	Baseline	Delta
MATH (Qwen-72B)	PASS@1	46.3%	CoT: 35.2%	+11.1 pp
GSM8K (Qwen-72B)	Accuracy	83.5%	CoT: 78.9%	+4.6 pp
Game of 24 (Qwen-72B)	Success	100%	ToT: 74%	+26 pp
Ego4D-HCap (ViSMaP)	ROUGE-L	29.9	Supervised: 32.6	–2.7
BrowseComp+ (RLM, GPT-5)	Accuracy	91.3%	Base: 0%; Summary: 70.5%	+20.8/+91.3

Token and cost efficiencies are achieved in all cases, as RMP typically uses compact, content-agnostic meta-prompts, avoiding the combinatorial scaling of few-shot templates.

7. Relationship to Other Recursive Frameworks

RMP subsumes or generalizes several lines of work in prompt self-improvement for LLMs:

• Recursive Chain-of-Feedback (R-CoF): A divide-and-conquer, recursion-inspired protocol that iteratively rewrites only incorrect steps in an LLM’s chain-of-thought, yielding strong gains on MATH problems (0/50→31/50 accuracy with a single R-CoF invocation) (Ahn et al., 2024).
• Recursive LLMs (RLMs): RMP viewed through the lens of inference-time decomposition and recursive sub-querying, scaling LLM reasoning beyond fixed context windows (Zhang et al., 31 Dec 2025).
• Meta-Prompting Protocol: An explicit adversarial feedback unrolling with modular, differentiable prompt scaffolds, mapping LLM orchestration to software engineering workflows (Fu, 17 Dec 2025). This suggests that RMP is a unifying abstraction for recursive prompt engineering, capable of adapting to virtually any LLM-augmented reasoning domain characterized by self-refinement, modularity, and iterative critique.

---

Recursive Meta-Prompting is thus a foundational method for recursive, self-optimizing prompt engineering. Its monadic formalism, empirical efficacy, and principled handling of compositionality and feedback mark it as a central contribution to the discipline of advanced LLM control (Zhang et al., 2023, Zhang et al., 31 Dec 2025, Fu, 17 Dec 2025, Hu et al., 22 Apr 2025, Ahn et al., 2024).