Robustness via Multi-Perspective Rewriting

• The paper demonstrates that multi-perspective rewriting improves retrieval accuracy and reduces model hallucinations, with metrics like MRR and NDCG showing significant gains.
• Methodologies such as prefix-driven rewriting, strategy pooling, and multi-agent simulation actively optimize input diversity for enhanced robustness in tasks like conversational search and SQL query optimization.
• Empirical evaluations reveal that aggregating diverse rewrites dramatically enhances robustness, evidenced by improvements in H@5, P@5, and out-of-distribution accuracy across numerous applications.

Robustness via Multi-Perspective Rewriting encompasses algorithmic and architectural strategies that deploy multiple, diverse rewritings or reformulations of an input—whether a query, utterance, or command—to enhance both performance and robustness across retrieval, ranking, and generation tasks. These techniques systematically introduce linguistic and semantic diversity into the rewriting pipeline or evaluation process, mitigating model overfitting to particular phrasing, ameliorating the impact of distributional shifts, and providing guardrails against hallucinations and other model failures. Multi-perspective rewriting is most prominently instantiated in LLM driven retrieval-augmented generation (RAG), conversational search, SQL query optimization, e-commerce relevance, semantic ranking, and black-box adaptation for out-of-distribution robustness.

1. Core Principles and Motivation

Multi-perspective rewriting is predicated on the observation that single-form rewritings or static query formulations are insufficiently robust: models may become brittle under paraphrasing, demographic variation, query ambiguity, or adversarial surface forms. By diversifying rewrites across multiple axes—semantic, syntactic, demographic, or information-theoretic—systems can ensure that retrieved or generated outputs are less sensitive to input idiosyncrasies. Robustness is further amplified by fusing the signals from these perspectives, whether at the query, passage, or response level.

The foundational motivations for this approach include:

• Simulating real-world input variation (demographic language, syntactic diversity, paraphrase, style shift) to avoid overfitting.
• Increasing recall in retrieval and ranking by mining complementary document sets.
• Mitigating hallucinations in generation by exposing models to richer and more varied context.
• Enabling black-box robustness when model parameters cannot be updated.

2. Multi-Perspective Rewriting Methodologies

Multiple paradigms instantiate this principle, each aligned to the requirements of their domain.

a) Prefix-Driven Multi-Faceted Rewriting

In "Multi-Faceted Self-Consistent Preference Alignment for Query Rewriting in Conversational Search" (MSPA-CQR) (Cao et al., 8 Apr 2026), rewriting is decomposed along three axes, each governed by a prefix ([REWRITE], [RETRIEVAL], [RESPONSE]) that conditions the LLM to generate rewrites optimized for paraphrase clarity, retrieval efficacy, and response informativeness, respectively. Each prefix is applied to the same base prompt, and the resulting rewrites are concatenated for downstream passage retrieval.

b) Strategy Pool and Adaptive Selection in RAG

DMQR-RAG (Li et al., 2024) formalizes multi-query rewriting by defining a suite of four strategies:

• General Query Rewriting (GQR): denoised, intent-preserving reformulation.
• Keyword Rewriting (KWR): extraction of core keywords.
• Pseudo-Answer Integration (PAR): enrichment with predicted answer content.
• Core Content Extraction (CCE): distillation to the root question. Retrieval is then performed across all rewrites, followed by adaptive selection to prune unnecessary perspectives for computational efficiency.

c) Multi-Agent Simulation and Genetic Search

OptAgent (Handa et al., 4 Oct 2025) uses a committee of LLM-based agents—each implemented as the same base model sampled at different temperatures—to simulate diverse customer perspectives in e-commerce search. These agentic judgments inform a fitness function for an evolutionary algorithm, which iteratively generates, crosses over, and mutates natural-language query rewrites, guided by ensemble agent feedback.

d) Demographic Persona Rewriting

In Agent4Ranking (Li et al., 2023), the LLM is prompted via Chain-of-Thought (CoT) to extract the information need and then re-express the query in the linguistically typical style of different demographic groups (middle-aged man, middle-aged woman, student, elder). These persona-generated rewrites feed through a Mixture-of-Experts ranking model governed by a loss combining accuracy and output-distribution alignment.

e) Black-Box Test-Time Augmentation

LLM-TTA (O'Brien et al., 2024) applies paraphrasing and in-context style transfer to generate multiple semantic variants of each test-time input, aggregating predictions across rewrites to improve classification robustness when retraining is infeasible.

f) SQL Rewriting via Diverse Prompts and Rule Injection

LITHE (Dharwada et al., 18 Feb 2025) addresses SQL query optimization by invoking an ensemble of prompt templates, domain-specific rule-based prompt injections (e.g., redundancy removal), and token-probability-guided exploration, each producing alternative rewrites that are later filtered for semantic equivalence and execution cost.

3. Algorithmic Formulations and Architecture

The technical implementations follow modular or compositional paradigms, integrating multi-perspective rewriting at different junctures:

• Sampling and Selection: Candidate rewrites are generated either in parallel (prefixes, strategies, agents) or iteratively (genetic algorithms), followed by scoring, reranking, or adaptive subset selection.
• Preference Scoring: MSPA-CQR computes self-consistency scores for each candidate along the three perspectives using neural-encoder-based NLI and retrieval overlap metrics. DMQR-RAG employs information-complementary strategy pools.
• Fusion and Aggregation: Rewrites are typically concatenated (MSPA-CQR), their retrieved documents pooled and reranked (DMQR-RAG), or their predictions ensembled (LLM-TTA).
• Robust Optimization: Direct Preference Optimization (DPO) and its prefix-guided, multi-faceted variants align model outputs with multi-perspective-preferred candidates.
• Evaluation and Fitness: For subjective tasks (OptAgent), agentic fitness is computed by averaging agent-graded relevance scores over products, combined with purchase-based economic signals.

A representative example from MSPA-CQR is the MDPO loss: $$\mathcal{L}_{\mathrm{MDPO}}(\theta) = - \mathbb{E}_{(pr, x, rq_+, rq_-) \sim \mathcal D} \left[\log \sigma \left(\beta \left[\log \frac{\pi_\theta(rq_+ \mid pr, x)}{\pi_\mathrm{ref}(rq_+ \mid pr, x)} - \log \frac{\pi_\theta(rq_- \mid pr, x)}{\pi_\mathrm{ref}(rq_- \mid pr, x)}\right]\right)\right]$$ where $\pi_\theta$ is the fine-tuned LLM and $\pi_\mathrm{ref}$ a frozen reference model (Cao et al., 8 Apr 2026).

4. Empirical Impact and Evaluation

Studies consistently report both aggregate effectiveness and substantial improvements in robustness metrics when deploying multi-perspective rewriting.

• MSPA-CQR (Cao et al., 8 Apr 2026):
  • TopiOCQA (BM25): MRR increases from 30.6 to 41.4; NDCG from 29.5 to 39.5; R@100 from 75.2 to 77.4.
  • Outperforms prior feedback-informed baselines and remains robust to domain shift (TREC CAsT 19–21).
  • Ablations indicate all perspectives are required for maximal recall.
• DMQR-RAG (Li et al., 2024):
  • On AmbigNQ, RAG-Fusion baseline H@5=86.33% increases to 88.08%; P@5 from 53.62% to 62.43%.
  • Industrial deployment exhibits a 2.0% absolute gain in H@5 and 10.0% in P@5.
  • Removal of any single rewriting strategy reduces P@5, with PAR most critical for diversity.
• OptAgent (Handa et al., 4 Oct 2025):
  • Achieves a mean fitness gain of +3.36% over a Best-of-N baseline and +21.98% over original user queries.
  • On niche queries (tail, multilingual), gains are larger (e.g., +4.50% on tail, +32.36% for Italian).
• Agent4Ranking (Li et al., 2023):
  • Reduces variance in NDCG@10 by 26% versus CharacterBert on Robust04.
  • Improves ranking accuracy by a statistically significant, though small, margin.
• LLM-TTA (O'Brien et al., 2024):
  • Average OOD accuracy gain of +4.48 pp for BERT; +2.81 pp for T5.
  • Entropy-based selective augmentation reduces LLM calls by ≈57% at negligible accuracy cost.
• LITHE (Dharwada et al., 18 Feb 2025):
  • TPC-DS: geometric mean speedup 30.6× vs. prior SOTA 2.1× on feasible productive rewrites.

5. Component Analysis and Ablation Findings

Component-level analysis in these works indicates that robustness gains are always multi-factorial:

• Use of all perspectives or strategies yields the largest gains; any reduction (e.g., omitting PAR in DMQR-RAG or a facet in MSPA-CQR) degrades performance.
• Multi-agent or multi-demographic approaches outperform single-agent or single-style baselines in both mean effectiveness and variance reduction.
• Methods that employ only a vanilla ensembling or consensus step without explicit multi-perspective structuring (e.g., DPO without prefixes) show meaningful drops in performance.
• The distributional alignment (e.g., Jensen-Shannon penalty in Agent4Ranking) is necessary to make the ensemble predictions coherent and stable.

A plausible implication is that the complementarity among perspectives is crucial: information-preserving rewrites boost precision, information-expanding variants improve recall, and reductionist forms catch broad or under-specified information needs. The fusion of outputs ameliorates under- and over-specification errors exhibited by single-perspective approaches.

6. Limitations and Open Challenges

While multi-perspective rewriting delivers tangible robustness improvements, multiple challenges persist:

• Computational overhead due to LLM inference across several rewrites can be substantial; adaptive strategy selection and selective augmentation partially mitigate this.
• Cost and latency remain high for settings requiring real-time feedback or user interactivity, especially for multi-agent simulations or multiple demographic rewrites.
• Coverage of perspectives is still generally hand-crafted or limited (e.g., Agent4Ranking uses only four demographics); richer, learned latent perspective spaces may afford further robustness.
• Full semantic equivalence checking is nontrivial outside narrow domains (see LITHE’s combined use of theorem proving and sampling for SQL).

7. Cross-Domain Extensions and Future Directions

The foundational architectural and algorithmic motifs of multi-perspective rewriting translate naturally to domains beyond text retrieval and ranking.

• Graph query optimization, code refactoring, and natural language generation/reformulation can benefit from similar ensembles of prompts, schema- or profile-injected rules, and token-probability-driven exploration (Dharwada et al., 18 Feb 2025).
• Research into learnt or automatically discovered perspectives, rather than hand-coded or stylistic ones, is ongoing.
• Efficient methods for aligning, aggregating, or fusing perspective-driven outputs—especially under resource constraints—remain a critical area of investigation.

Systematic, multi-perspective rewriting represents a general-purpose, empirically validated mechanism for imparting robustness to LLM-driven systems under distributional shift, demographic heterogeneity, and input ambiguity. The unifying evidence demonstrates that complementary views, fused via principled optimization or aggregation, outperform single-form or single-model approaches in both effectiveness and stability.