PaperlayerView exact revision on arXiv

Source-linked AI summary

MAPS: Macro-Adaptive Parallel Reasoning for Regulating Search Momentum in Retrieval-Augmented Generation

Xinze Li, Dingwen Zhang, Zhenghao Liu, Yukun Yan, Shuo Wang, Yu Gu, Ge Yu, Maosong Sun

arXiv:2610.29792v1cs.CL

TL;DR

Iterative RAG methods can generate sub-queries that drift from the original information need, and how reasoning shapes that drift remains unclear. MAPS regulates accumulated search momentum by disentangling historical trajectories and exploring complementary directions in parallel. It consistently outperforms representative baselines across tasks and backbones.

  • Problem

    In iterative RAG, subsequent sub-queries may become misaligned with the original information need, and the influence of evolving reasoning on search direction remains unclear.

    Source: §1 · p. 2, §3 · p. 4

  • Method

    MAPS reduces historical trajectory influence by retaining refined knowledge and reasoning state, then generates parallel sub-queries for complementary retrieval directions and consolidates their evidence.

    Source: §1 · p. 2, §3 · p. 4

  • Results

    MAPS consistently outperforms baselines across tasks, with improvements of 6.8% on Qwen3-32B and 5.8% on Qwen3.6-35B-A3B.

    Source: §1 · p. 2, §5.1 · p. 6

  • Takeaways & Limitations

    The analysis finds that sub-queries tend to follow accumulated retrieval trajectories, while MAPS alleviates query-intent drift by disentangling history and introducing parallel reasoning.

    Source: §7 · p. 10

Abstract

from arXiv · show

Retrieval-Augmented Generation (RAG) improves the performance of Large Language Models (LLMs) by retrieving external knowledge. Recently, some RAG methods interleave reasoning and retrieval, enabling LLMs to progressively generate sub-queries and retrieve external knowledge. However, these methods may suffer from redundant or ineffective sub-queries and insufficiently diverse query exploration. Moreover, due to the tightly interleaved nature of reasoning and retrieval, it remains unclear how the evolving reasoning process influences query intent. To characterize the evolution of query intent during reasoning, we introduce search momentum, which captures the accumulated retrieval tendency induced by historical changes in generated sub-queries along the reasoning trajectory. Through momentum-based analysis, we find that the query generation process tends to continuously evolve along existing retrieval trajectories. Based on this observation, we propose MAPS, a Macro-Adaptive Parallel Reasoning framework for regulating search momentum and mitigating query-intent drift. MAPS first retains the retrieved knowledge and a refined reasoning state, thereby reducing the influence of historical reasoning trajectories on subsequent sub-query generation. Subsequently, MAPS extends the conventional sequential reasoning process into a parallel reasoning paradigm. At each iteration, MAPS identifies complementary retrieval directions, generates corresponding sub-queries for parallel retrieval, and consolidates the retrieved evidence to update the knowledge and reasoning states. Experiments on six question-answering benchmarks demonstrate that MAPS consistently outperforms representative RAG methods across different backbone models. All code and resources are available at https://github.com/NEUIR/MAPS.

1 INTRODUCTION

Iterative RAG can still generate ineffective or insufficiently diverse sub-queries, while the effect of evolving reasoning on search direction remains unclear. MAPS regulates accumulated search momentum through disentanglement and parallel exploration, and the paper reports gains across datasets and backbones.

1 INTRODUCTION

  • Multi-hop questions often require evidence across interdependent documents, making single-step retrieval insufficient for comprehensive answers.RAG gives models access to external knowledge during inference to improve response factuality and reliability.Source: §1 · p. 1
  • Iterative reasoning-and-retrieval methods can produce redundant or ineffective sub-queries, long retrieval chains, and insufficiently diverse exploration.Because reasoning and retrieval are tightly interleaved, how evolving reasoning influences search direction remains unclear, especially when evidence lies beyond the current trajectory.Source: §1 · p. 2
  • Historical changes in sub-queries accumulate as search momentum, and subsequent queries increasingly align with that trajectory, potentially amplifying early deviations from the original information need.This motivates regulating momentum to preserve alignment with the original query intent.Source: §1 · p. 2
  • MAPS retains refined retrieval knowledge and reasoning state, then identifies complementary directions for parallel sub-query generation and retrieval.It consolidates the resulting evidence to update the knowledge and reasoning states for the next iteration.Source: §1 · p. 2
  • MAPS significantly improves over all baseline methods across multiple datasets and backbone models.The reported analyses also find reduced query-intent drift, stronger relevance between sub-queries and evidence, earlier acquisition of prerequisite knowledge, and fewer unnecessary retrieval iterations.Source: §1 · p. 2

2 ANALYZING QUERY INTENT DRIFT DURING LLM REASONING

This section formalizes search momentum as accumulated movement in retrieval intent and tests whether it predicts subsequent query drift. Its analysis finds that deeper reasoning is associated with greater retention of past search dynamics and stronger alignment between momentum and the next query’s deviation.

2 ANALYZING QUERY INTENT DRIFT DURING LLM REASONING

  • The section frames query drift as a consequence of iterative retrieval trajectories that can influence the intent of subsequent sub-queries and downstream reasoning.It investigates whether accumulated retrieval-intent change predicts that subsequent drift.Source: §2 · p. 2
  • The reasoning trajectory records each reasoning step, generated sub-query, and retrieved-evidence observation, which supplies context for generating the next step and query.The generated sub-query retrieves documents from which the model derives the corresponding observation.Source: §2 · p. 3
  • Search momentum m_t is an exponentially weighted aggregation of historical query movements, with its magnitude representing accumulated shift and its direction representing the predominant intent evolution.The latest semantic displacement is combined with retained prior dynamics, where β_t is constrained to [0, 1).Source: §2 · p. 3
  • The analysis measures whether momentum aligns with the next sub-query’s deviation from the original query; stronger alignment indicates predictive correspondence with future drift.The reported table presents fitted momentum coefficients and held-out mean directional alignment scores across dataset settings.Source: §2 · pp. 3–4, Table 1 · p. 3
  • Both fitted β_t and held-out directional alignment S_t increase as reasoning progresses, indicating that later queries increasingly follow accumulated search dynamics.The authors conclude that subsequent reasoning does not naturally eliminate accumulated momentum.Source: §2 · p. 4

3 MAPS: REGULATING INTENT DRIFT THROUGH PARALLEL REASONING

MAPS reduces the influence of historical retrieval trajectories by retaining accumulated knowledge and a refined reasoning state, then explores complementary retrieval directions in parallel. Retrieved evidence is consolidated to update these states for subsequent reasoning.

Reasoning Disentanglement for Search Momentum Regulation

  • Historical sub-queries can reinforce retrieval tendencies and lead to query-intent drift, so MAPS retains accumulated knowledge and a compact refined reasoning state instead of the full trajectory.The model conditions subsequent reasoning on the original query, accumulated knowledge observations, and refined state, avoiding explicit exposure to historical sub-queries and reasoning traces.Source: §3 · pp. 4–5

Parallel Reasoning for Search Direction Calibration

  • Reasoning stops when accumulated knowledge satisfies current information needs or the maximum allowed iterations are reached, after which the model answers from the query, accumulated observations, and refined state.T denotes the termination iteration or maximum number of allowed reasoning iterations.Source: §3 · p. 5
  • MAPS identifies multiple complementary retrieval directions before generating independent sub-queries, expanding the search space early rather than conditioning each query on the preceding retrieval trajectory.This parallel approach is intended to reduce the risk that search momentum becomes dominated by one deviated direction.Source: §3 · p. 5
  • Each sub-query independently retrieves top-K documents, which are aggregated into a unified evidence set for refinement.The refinement agent consolidates evidence from different directions into a knowledge observation and updates the refined reasoning state for subsequent iterations.Source: §3 · p. 5

4 EXPERIMENTAL METHODOLOGY

The evaluation covers single-hop, multi-hop, and reasoning question answering, comparing MAPS with representative RAG baselines. Experiments use two Qwen backbones, a Wikipedia retrieval corpus, and Cover-EM as the primary metric.

Datasets

  • Evaluation spans single-hop NQ and PopQA, multi-hop 2WikiQA, HotpotQA, and MuSiQue, and the Diamond subset of GPQA for reasoning QA.The GPQA Diamond subset contains 198 graduate-level science questions; 1,000 questions are randomly sampled from each other dataset.Source: §4 · pp. 5–6

Baselines

  • Baselines include Vanilla LLM and Vanilla RAG, adaptive Self-RAG, reasoning-driven IRCoT and Search-o1, and sub-query methods S3 and ParallelSearch.The passages describe these methods as spanning no-retrieval prompting, retrieval, adaptive retrieval, interleaved reasoning and retrieval, and sub-query optimization or parallelization.Source: §4 · p. 6

Implementation Details

  • Experiments use Qwen3-32B and Qwen3.6-35B-A3B, English Wikipedia’s December 2018 dump, E5-base-v2 with FAISS, top-5 passages per step, and a maximum of 10 iterations.Cover-EM is the primary evaluation metric.Source: §4 · p. 6

5 EVALUATION RESULTS

MAPS outperforms the evaluated baselines, and the analyses report that it stabilizes search momentum and maintains query alignment during iterative reasoning. Ablations and behavioral comparisons examine the roles of parallel retrieval, complementary directions, and disentanglement.

5.1 OVERALL PERFORMANCE

  • MAPS consistently outperforms all baselines across tasks, improving by 6.8% and 5.8% on Qwen3-32B and Qwen3.6-35B-A3B, respectively.MAPS also improves over Search-o1 by over 9% and consistently outperforms ParallelSearch across datasets.Source: §5.1 · pp. 6–7

5.2 ABLATION STUDY

  • MAPS outperforms all ablated variants, with over a 4% improvement over MAPS w/o Retrieval Direction.The comparisons associate complementary retrieval directions with diverse information needs, parallel reasoning with early evidence exploration, and disentanglement with reduced dependence on historical retrieval tendencies.Source: Table 3 · p. 7, §5.2 · p. 7

5.3 EVOLUTION OF SEARCH MOMENTUM DURING ITERATIVE REASONING

  • MAPS shows less search-momentum deviation from the original query than Search-o1 and stronger alignment than MAPS w/o Disentanglement.Its search momentum becomes more stable as reasoning progresses.Source: Figure 1 · p. 8, §5.3 · p. 8

5.4 ANALYSIS OF MAPS RETRIEVAL BEHAVIOR DURING ITERATIVE REASONING.

  • After early declines, MAPS sub-queries recover and stabilize their similarity to golden evidence, while retaining higher similarity to the original query than Search-o1.MAPS has more samples terminating early and more requiring at least five retrieval rounds, while maintaining strong answer performance across termination steps.Source: Figure 2 · p. 8, §5.4 · pp. 8–9

6 RELATED WORK

Related work covers retrieval-augmented generation, reasoning-driven iterative retrieval, and efforts to improve sub-query generation, while noting that historical retrieval trajectories can still influence later queries.

6 RELATED WORK

  • Early RAG systems retrieve documents from the original query and use them as context, but complex questions may require evidence distributed across multiple documents.Some work also studies adaptive decisions about when to retrieve.Source: §6 · p. 9
  • Reasoning-driven RAG methods progressively generate sub-queries and incorporate retrieved evidence into later reasoning, but depend heavily on sub-query quality.Irrelevant or erroneous sub-queries may shift subsequent retrieval trajectories.Source: §6 · p. 9
  • S3 trains a lightweight sub-query generator and ParallelSearch retrieves multiple sub-queries in parallel, but these methods do not eliminate historical-trajectory influence.ParallelSearch is described as acquiring more comprehensive evidence through multiple sub-queries.Source: §6 · pp. 9–10

7 CONCLUSION

The conclusion introduces search momentum to characterize accumulated retrieval tendencies and reports that sub-queries tend to follow those trajectories. MAPS combines reasoning disentanglement with parallel reasoning and is reported to alleviate query-intent drift.

7 CONCLUSION

  • Search momentum captures the accumulated retrieval tendency induced by historical changes in generated sub-queries along the reasoning trajectory.Source: §7 · p. 10
  • The authors report that generated sub-queries tend to evolve continuously along the accumulated retrieval trajectory, leading to deviations in query intent.Source: §7 · p. 10
  • MAPS combines disentangled historical reasoning trajectories with parallel reasoning to alleviate query-intent drift.Source: §7 · p. 10

AI USE STATEMENT

The paper states that generative AI tools were used only for grammar checking and language polishing; the authors independently developed and verified the research work.

AI USE STATEMENT

  • Generative AI tools were used to check grammatical errors.Source: AI USE STATEMENT · p. 10
  • The tools also assisted with language polishing to improve manuscript clarity.Source: AI USE STATEMENT · p. 10
  • The authors state that they independently developed and verified the research ideas, experimental design, implementation, analysis, and conclusions without AI tools.Source: AI USE STATEMENT · p. 10

A.1 FITTING THE SEARCH MOMENTUM COEFFICIENTS

This section describes how MAPS’s search-momentum coefficients are fitted and evaluated. The procedure uses query trajectories from three multi-hop datasets and held-out directional-alignment scores.

Data Construction for Momentum Fitting

  • The fitting data contain 1,000 questions each from 2WikiQA, MuSiQue, and HotpotQA, with trajectories shorter than three retrieval iterations discarded.Separate non-overlapping sets of 1,000 questions per dataset are used for held-out evaluation, and coefficients are fixed after fitting.Source: §A.1 · p. 14

Fitting Objective

  • The fitting objective equally averages the first four stepwise alignment scores, and β is estimated by minimizing it with L-BFGS-B for 500 iterations.Later steps are excluded because fewer trajectories reach them, potentially making coefficient estimates unreliable.Source: §A.1 · p. 14

Evaluation and Results across Methods

  • The held-out evaluation reports fitted coefficients and directional-alignment scores for MAPS and MAPS without disentanglement across three datasets and pooled MHQA.Table 4 covers MAPS; Table 5 covers MAPS without disentanglement.Source: Table 4 · p. 14, §A.1 · p. 14, Table 5 · p. 15

A.2 EXPERIMENTAL DETAILS FOR ANALYZING THE EVOLUTION OF SEARCH MOMENTUM

The appendix describes how search-momentum visualizations are constructed, presents a case study contrasting Search-o1 with MAPS, and explains the prompts used in MAPS. The case study illustrates how MAPS’s parallel exploration and evidence-guided refinement can avoid a misleading retrieval trajectory.

A.2 EXPERIMENTAL DETAILS FOR ANALYZING THE EVOLUTION OF SEARCH MOMENTUM

  • The momentum visualizations jointly project query embeddings with t-SNE, center the original questions and iteration-wise momentum, and plot iterations on the z-axis.Closed contours show the point-cloud range at each iteration; Figure 4 distinguishes Search-o1’s sequential queries from MAPS’s branched parallel queries.Source: §A.2 · p. 15, Figure 4 · p. 16

A.3 CASE STUDY

  • The case study shows MAPS identifying both relevant directors in its first iteration and answering correctly, while Search-o1 follows an unsupported director interpretation into off-topic evidence.MAPS later retrieves Simón’s birth date and infers Borderie was born earlier from his 1953 directorial debut, without retrieving Borderie’s exact birth date.Source: §A.3 · p. 15, Table 6 · p. 17

A.4 PROMPT TEMPLATES

  • MAPS’s prompts assess whether accumulated knowledge suffices, generate complementary retrieval directions and parallel sub-queries, and refine reasoning from retrieved evidence.Figures 5–7 show the templates for directions, parallel sub-queries, and refined reasoning, respectively.Source: §A.4 · p. 15, Figure 5 · p. 18, Figure 6 · p. 19

Related papers

Loading 2610.29792v1…