LLM Rewriting Strategies

• LLM Rewriting Strategies are algorithmic frameworks that transform inputs to improve retrieval, factual consistency, and output quality in downstream tasks.
• They integrate methods like prompt rewriting, data curation, and reinforcement learning to optimize performance across diverse applications.
• Empirical results show enhanced metrics such as higher retrieval hit rates, improved text quality, and boosted code generation accuracy.

LLM rewriting strategies are algorithmic and procedural frameworks designed to transform input data, prompts, queries, or retrieved knowledge to enhance the performance, robustness, efficiency, or factuality of downstream LLM-based systems. These strategies are motivated by the observation that the initial formulation of natural language inputs—whether user queries for retrieval-augmented generation, style or toxicity in writing, under-optimized prompts, or raw pre-training data—often exhibits mismatches or inefficiencies that can be systematically rectified by automated rewriting. The resultant improvements manifest in areas such as retrieval precision, semantic robustness, data efficiency, domain adaptation, intent disambiguation, and the overall quality or appropriateness of generated outputs.

1. Taxonomy and Objectives of LLM Rewriting Strategies

LLM rewriting strategies span a broad range of objectives and scenarios, including:

• Query and Prompt Rewriting: Transforming raw user inputs or prompts to close the gap between user expression and what is optimal for retrieval or model response, as in retrieval-augmented generation and black-box LLM scenarios (Ma et al., 2023, Kong et al., 16 Jan 2024, Sarkar et al., 21 Mar 2025).
• Knowledge Curation and Summarization: Refining or synthesizing retrieved knowledge to maximize relevance and supportiveness for downstream generation (Qiao et al., 12 Jun 2024).
• Data and Training Corpus Rewriting: Systematically upgrading the quality, consistency, and utility of pre-training datasets—e.g., by enforcing style, self-containment, or algorithmic quality (Fujii et al., 5 May 2025).
• Content Moderation and Appropriateness: Rewriting user-generated content to mitigate toxicity or inappropriateness while preserving core content (Ziegenbein et al., 5 Jun 2024, Zhuo et al., 21 Apr 2025).
• Task-Selective and Adaptive Rewriting: Employing diverse rewriting strategies per query or input, adaptively selecting among them based on the scenario or downstream needs (Li et al., 20 Nov 2024, Li et al., 2023).
• Robustness via Multi-Perspective Rewriting: Generating multiple rewrites to cover demographic, linguistic, or intent diversity for increased resilience in ranking and retrieval (Li et al., 2023).
• Generic and Multi-Objective Text Rewriting: Employing decoupled reward models and reinforcement learning to optimize for competing rewriting objectives—such as factuality, style, minimal editing, and coherence—within a single unified model (Li et al., 9 Mar 2025).

The common thread is that the rewriting function is itself typically realized by an LLM—either in a frozen, prompted mode, fine-tuned via supervised learning, or further optimized via reinforcement learning using explicit or implicitly simulated feedback.

2. Methodological Foundations

2.1 Pipeline Integration

Rewriting strategies are deployed at various stages across systems:

• Rewrite–Retrieve–Read (RRR) Pipeline: Inserts a rewriting step before retrieval to close the gap between user query phrasing and the optimal retrieval vocabulary (Ma et al., 2023).
• End-to-End Data Rewriting Pipelines: Upgrade pre-training corpora via multi-stage filtering and LLM-based transformation to boost LLM downstream capabilities (e.g., SwallowCode’s four-stage process spanning syntax validation, lint, style-guided rewriting, and self-contained optimization (Fujii et al., 5 May 2025)).
• Prompt/Query Rewriting with Reinforcement Learning: Both prompt and query rewrites can be optimized end-to-end using RL, with the reward tightly coupled to the performance on the downstream task or retrieval hit ratio (Kong et al., 16 Jan 2024, Ma et al., 2023).

2.2 Rewriting Mechanisms and Learning Paradigms

Mechanisms can be divided as follows:

Mechanism	Training or Control Method	Key Features
Few-shot Prompting	Manual prompt design	Rapid deployment, limited adaptivity
Supervised Fine-Tuning	Clean pseudo pairs, instruction data	Scalable with curated or synthetic data
Reinforcement Learning	Task-driven or classifier-based	Direct downstream optimization, RLHF, PPO
Direct Preference Optimization (DPO)	Pairwise preference alignment	Learned ranking per rewriting objective
Heuristic Rewarding	Rule- or metric-based multi-signal	Reduces label cost, enables task-specific RL

A notable methodological theme is the explicit modeling of the rewriting policy as a Markov Decision Process (MDP), where the generation of the rewrite proceeds token by token, with rewards shaped by downstream task performance, semantic similarity, supportiveness, or appropriateness (e.g., (Ma et al., 2023, Qiao et al., 12 Jun 2024, Ziegenbein et al., 5 Jun 2024)).

3. Evaluation Metrics and Empirical Outcomes

LLM rewriting strategies are validated using a variety of metrics tailored to the downstream application:

• Retrieval and QA: Exact Match (EM), F1-score, retrieval hit rate (Ma et al., 2023, Li et al., 20 Nov 2024).
• Text Rewriting Benchmarks: Natural language inference (NLI) scores, edit distances, SARI, GLEU, Updated-ROUGE (Shu et al., 2023, Zhu et al., 2023).
• Code and Math Synthesis: pass@k metrics for code (e.g., HumanEval), accuracy on GSM8K or MATH (Fujii et al., 5 May 2025).
• Toxicity and Appropriateness: Semantic similarity (e.g., BERTScore), appropriateness classifier accuracy, content preservation checks (Ziegenbein et al., 5 Jun 2024, Zhuo et al., 21 Apr 2025).
• Prompt Optimization: Task performance metrics (EM, F1, perplexity), relative gains over manual prompts (Kong et al., 16 Jan 2024).
• Robustness: Variance in ranking (VNDCG, VNAP) across multi-perspective rewrites (Li et al., 2023).

Empirical results consistently indicate that inserting a rewriting stage (e.g., in retrieval-augmented QA, prompt optimization, or pre-training corpus construction) outperforms both direct and standard baseline systems. For example, in code generation, rewriting low-quality samples yields substantial pass@1 increases (+17) over state-of-the-art filtered datasets (Fujii et al., 5 May 2025). In retrieval-augmented QA, trainable rewriters achieve higher EM and F1, with a clear increase in retrieval hit ratios (Ma et al., 2023). In prompt engineering, reinforcement-learning-optimized rewriting achieves up to 80% performance gains depending on the base task (Kong et al., 16 Jan 2024).

4. Trade-Offs, Challenges, and Adaptive Strategies

4.1 Efficiency vs. Effectiveness

One trade-off is between the computational cost of LLM-based rewriting (especially at inference time) and real-world practicality. Several works address this by confining expensive rewriting to the training phase (Anand et al., 2023), or by distilling LLM knowledge into lighter models for online serving (e.g., MiniELM in e-commerce (Nguyen et al., 29 Jan 2025), on-device text rewriting agents (Zhu et al., 2023)).

4.2 Semantic and Syntactic Fidelity

LLM-generated rewrites can introduce hallucinations or semantic/syntactic errors. Counterexample-guided iterative refinement (Liu et al., 14 Mar 2024), logic- and sample-based semantic checks (Dharwada et al., 18 Feb 2025), or classifier-based semantic similarity constraints (Ziegenbein et al., 5 Jun 2024) are employed to address these issues.

4.3 Objective Decoupling

Rewriting often involves competing or multi-objective constraints—such as instruction following, coherence, conciseness, and factual consistency. Decoupled reward modeling, using separate reward models per objective and an aggregated, dynamically weighted reward in RL, is shown to produce high-quality, adaptive rewrites across diverse tasks (Li et al., 9 Mar 2025).

4.4 Adaptive and Multi-Strategy Selection

Rather than relying on a single rewriting strategy, recent frameworks (such as DMQR-RAG) employ multiple rewriting engines—clarification, keyword extraction, pseudo-answer enrichment, and content distillation—selected adaptively using LLM-based gating to minimize unnecessary rewrites and maximize coverage (Li et al., 20 Nov 2024).

5. Real-World Implications and Applications

LLM rewriting strategies are increasingly embedded in practical systems across domains:

• Search and QA: Rewrite–Retrieve–Read frameworks, context-aware query rewriting, and multi-agent demographic-centric rewrites to boost robustness and recall (Ma et al., 2023, Anand et al., 2023, Li et al., 2023).
• Code Synthesis and Math: Quality-driven pre-training corpus rewriting demonstrably elevates LLM mathematical reasoning and code-writing capabilities (Fujii et al., 5 May 2025).
• MT and Translation: Input simplification via LLMs—often aided by quality estimation—improves translation quality across diverse language pairs (Ki et al., 23 Feb 2025).
• Content Moderation: RL-based and classifier-informed rewriting of inappropriate or toxic argumentation enables scalable, automation-enhanced, preemptive moderation (Ziegenbein et al., 5 Jun 2024, Zhuo et al., 21 Apr 2025).
• Prompt Engineering: Reinforcement-learned prompt rewriting automates and improves a process traditionally reliant on human trial and error (Kong et al., 16 Jan 2024, Sarkar et al., 21 Mar 2025).
• Mobile and Real-Time Systems: Knowledge distillation, heuristic RL, and cascading of on-device and cloud LLMs construct privacy-aware, low-latency agents (Zhu et al., 2023).

6. Formulations, Algorithms, and Representative Equations

Key mathematical constructs from various works include:

• MDP-based RL for Rewriting:

$$R(s_t, a_t) = R_{LM}(\tilde{z}, y) - \beta \cdot \mathrm{KL}(\pi_\theta || \pi_0)$$

(Ma et al., 2023)

• Aggregated Decoupled Reward in RL:

$$r_{\varphi'}(x, y) = \sum_{o=1}^{O} w_{o}^{t} \cdot r_{\varphi_o}(x, y)$$

(Li et al., 9 Mar 2025)

• Supportiveness (Perplexity Ratio) for Knowledge Rewriting:

$ss(q, c) = \frac{P_{raw}}{P_{retrieval}}$

(Qiao et al., 12 Jun 2024)

• Upper Confidence Bound in Monte Carlo Tree Search for Token Selection:

$$\text{UCB}(n, a) = V(n') + \beta(n) \cdot P_{\text{LLM}}(a | n.\text{state}) \cdot \sqrt{\frac{\log(\text{visits}[n])}{1+\text{visits}[n']}}$$

(Dharwada et al., 18 Feb 2025)

• LTCS Reward for Plan Generation:

$$R(P_g, P_r) = \begin{cases} 1, & \text{if } P_g \text{ valid} \ \frac{|LCCS(P_g, P_r)|}{|P_r|}, & \text{otherwise} \end{cases}$$

(Huang et al., 14 Dec 2024)

These formulations govern the optimization or selection of rewritten outputs in line with the downstream system’s requirements.

7. Future Directions and Open Challenges

Ongoing and open research questions include:

• Automated Adaptation: How to generalize rewriting strategies to previously unseen tasks or domains (domain adaptation) and new LLM architectures with minimal manual tuning.
• Explainability and Transparency: Developing approaches that can surface the factors driving rewriting decisions, especially in high-stakes applications (Zhuo et al., 21 Apr 2025).
• Multi-Objective, Multi-Signal RL: Advancing decoupled and compositional reward models for ever more complex rewriting objectives (Li et al., 9 Mar 2025).
• Robustness and Distribution Shift: Addressing the brittleness of rewriting strategies under distributional, length, or domain shift in both the input and output space (Huang et al., 14 Dec 2024).
• Data Efficiency and Labeling Cost: Further reducing annotation or preference requirements by leveraging heuristic or simulated reward signals at scale (Zhu et al., 2023, Nguyen et al., 29 Jan 2025).
• Human-in-the-loop Systems: Integrating user and domain expert feedback dynamically to refine LLM rewriting in operational settings.

A plausible implication is that as LLMs and their deployment contexts grow increasingly diverse, rewriting will become both a core research topic and an engineering necessity for bridging gaps between user intent, knowledge sources, and emergent model behavior, with fine-grained, adaptive control over all stages of the language processing pipeline.