acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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 datasets will be released on GitHub.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Retrieval-Augmented Generation, Parallel Reasoning, Query Intent Drift

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RUnqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper studies query intent drift in reasoning-based RAG, where later sub-queries become increasingly influenced by previous retrieval trajectories, causing early search errors to accumulate over multiple reasoning steps. To characterize this behavior, the authors introduce search momentum, which captures the accumulated semantic direction of sub-query changes, and show that later queries tend to follow this momentum rather than naturally correcting earlier deviations. Based on this observation, they propose MAPS, which combines reasoning disentanglement—keeping only retrieved knowledge and a refined reasoning state instead of the full history of reasoning traces and sub-queries—with parallel reasoning, where multiple complementary retrieval directions are proposed and searched in parallel at each iteration. Experiments on NQ, PopQA, 2WikiQA, HotpotQA, MuSiQue, and GPQA using Qwen3-32B and Qwen3.6-35B-A3B show that MAPS consistently outperforms baselines such as Vanilla RAG, Self-RAG, IRCoT, Search-o1, S3, and ParallelSearch. Ablation and analysis further suggest that both parallel reasoning and disentanglement help reduce search drift and keep retrieval better aligned with the original question and supporting evidence.

Critical strengths
  1. The paper propose search momentum to measure the deviation of multi-hop qa searches and use it to discover problems in the trajectory.
  2. The paper propose reasoning disentanglement and parallel reasoning to alleviate the query drift and the experiments are good.
Critical weaknesses
  1. MAPS uses parallel queries while equal-budget comparison and efficiency analysis is needed to see whether the improvement comes from larger queries and computational budget. Only comparison with ParallelSearch is not enought while budget should be reported. Also another easy baseline pass@k should be reported.
  2. There seems no relation between search momentum and MAPS. Since MAPS does not explicitly estimate or regulate the momentum variable during inference, it is unclear whether search momentum is an essential mechanism underlying the method or mainly an explanatory diagnostic.
Questions for authors

The proposed notion of search momentum is interesting but its empirical validation is somewhat post-hoc. The momentum coefficients are explicitly fitted to maximize alignment with future query deviation, making the observed correlation partly induced by the construction itself. Can we use the momentum to guide the future and then make the momentum more relative to the MAPS method?

2 comments
AuthorsOpenReview ↗

Reply to RUnq 2

Reply to Weakness 1: We thank the reviewer for this comment. We address retrieval-budget matching in first point of our General Response. Specifically, we set the number of passages retrieved per iteration by Naive-RAG and Search-o1 to match the average number retrieved per iteration by MAPS. Under the matched passage retrieval budget, MAPS still achieves higher average performance, indicating that its advantage does not simply stem from retrieving more passages.

Furthermore, we compute pass@3 for both Search-o1 and MAPS. For each question, we independently sample three answers from each method under the same experimental settings. An instance is considered successful if at least one of the three answers is correct. As shown in the table below, MAPS consistently outperforms Search-o1.

MethodNQPopQA2WikiQAMuSiQueHotpotQAGPQAAvg.
Qwen3.6-35B-A3B
MAPS72.274.177.956.466.688.872.7
search-o164.368.166.147.555.883.564.2
Qwen3-32B
MAPS70.870.568.948.659.470.864.8
search-o160.359.258.331.745.36553.3

Reply to Weakness 2: We thank the reviewer for this comment. In our paper, search momentum not only provides key motivation for the design of MAPS, but also helps assess whether the model becomes trapped in a particular local search drift during reasoning. Although deviations in the retrieval direction can be valuable during reasoning, as the process progresses, the accumulated context may cause the model to increasingly rely on the existing retrieval trajectory, thereby reducing its exploration of potentially useful directions. The search momentum experiments in Section 2 of our paper reveal this phenomenon and motivate the key mechanisms of MAPS, which calibrate the retrieval trajectory in the early stages of reasoning and help prevent subsequent reasoning from becoming trapped in a locally biased retrieval direction.

Reply to Question 1: We thank the reviewer for raising this question. Search momentum is computed as a weighted sum of historical changes in query intent, characterizing the accumulated retrieval tendency arising from historical sub-query movements along the reasoning trajectory. The coefficient β\beta is fitted only on the training trajectories and then held fixed. The correlation between search momentum and subsequent sub-query deviations is evaluated on independent test trajectories. Therefore, fitting and validation are conducted on different trajectories, and the observed correlation is independently verified on the test trajectories rather than induced by the construction of the metric itself.

The connection between search momentum and MAPS is directly reflected in the method design. During reasoning, local deviations in the early retrieval stages reflect the model’s exploration of diverse and potentially useful directions. However, as reasoning progresses, the growing context causes the model to rely increasingly on the existing trajectory, gradually reducing its exploration of other valuable directions. We observe this phenomenon in the search-momentum experiments in Section 2 of our paper, which directly motivates the key mechanisms of MAPS. Thus, search momentum not only provides motivation for MAPS, but also helps assess whether the reasoning process has become trapped in a particular local retrieval direction. Furthermore, we believe that search momentum can be further used to assess retrieval quality and determine appropriate stopping points, providing a promising direction for future research.

AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer,

Thank you again for your time and thoughtful feedback. As the discussion period is approaching its conclusion, we would like to briefly follow up on our rebuttal. We have carefully addressed each of your comments and sincerely hope that our responses have clarified the concerns you raised.

If you have any remaining questions or concerns, we would be happy to provide further clarification during the remaining discussion period.

Best regards,
The Authors

Reviewer_24USOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how earlier retrieval decisions can steer iterative RAG toward redundant or unproductive searches. It defines search momentum as a weighted accumulation of query-embedding displacements and analyzes its alignment with later query deviations. MAPS then combines a refined knowledge state, complementary search-direction proposals, and parallel retrieval to reduce dependence on the full search history.

Critical strengths

Pros

  1. The architecture targets a clear retrieval failure. Refined knowledge states may reduce reinforcement of early errors.

  2. The evaluation covers diverse QA tasks and relevant baselines. It includes single-hop, multi-hop, and scientific reasoning questions.

  3. The analysis examines more than final accuracy. Ablations and analyses of query similarity, evidence similarity, and stopping rounds reveal intermediate behavior.

Critical weaknesses

Cons

  1. Momentum alignment does not establish harmful drift, because useful searches can also move away from the original query.

  2. The comparisons do not match total inference and retrieval budgets.

Questions for authors

Questions

  1. Can momentum predict retrieval failure and guide online decisions?

  2. Which components remain useful under a fixed total budget?

4 comments
AuthorsOpenReview ↗

Reply to 24US

Reply to Weakness 1: We thank the reviewer for this comment. You are correct that sub-query deviations partly reflect the model’s exploration of different retrieval directions. However, as reasoning proceeds, later steps increasingly rely on previous reasoning states and retrieved evidence, making the model more likely to continue along the existing trajectory and reducing its exploration of other potentially valuable directions. At each step, MAPS explores multiple retrieval directions in parallel to recalibrate the retrieval drift accumulated in earlier steps. Through this continued exploration of diverse directions, MAPS identifies more accurate retrieval directions, stabilizes search momentum, and provides more valuable retrieval results for subsequent reasoning steps.

Reply to Weakness 2: We thank the reviewer for this comment. We present detailed experiments and analyses on retrieval-budget matching in the first point of our General Response. We conduct a comparison with matched numbers of retrieved passages. MAPS still achieves higher average performance than Search-o1 under the matched retrieval budget. This suggests that MAPS’s gains do not simply stem from retrieving more passages, but primarily result from its parallel retrieval strategy and reasoning disentanglement mechanism.

Reply to Question 1: Thank you for raising this question. As you pointed out, deviations in the search direction can sometimes be valuable. However, as reasoning progresses, the accumulated context and historical retrieval results may cause the model to increasingly rely on the existing retrieval trajectory, thereby reducing its exploration of potentially useful directions. We observe this phenomenon in the search-momentum experiments in Section 2 of our paper, and this observation supports the design of MAPS’s key mechanisms, which calibrate the retrieval drift accumulated in earlier steps and provide more valuable evidence for subsequent reasoning. Furthermore, we believe that search momentum could be used to assess retrieval quality and determine appropriate stopping points, making this a promising direction for future work.

Reply to Question 2: We thank the reviewer for the question. To analyze the role of each component in MAPS under a fixed retrieval budget, we fix the total retrieval budget per round at 10 documents and separately ablate the two mechanisms of parallel reasoning and reasoning-trajectory disentanglement. Specifically, when performing parallel retrieval, MAPS (N=2) generates two sub-queries per round, with each sub-query retrieving 5 documents. MAPS w/o Parallel Reasoning removes parallel reasoning and generates only one sub-query, but each sub-query retrieves 10 documents, thereby ensuring that the total number of retrieved documents remains the same. MAPS w/o Disentanglement retains the same parallel retrieval budget but removes the reasoning disentanglement mechanism. As shown in the table below, MAPS achieves higher average performance than MAPS w/o Disentanglement and MAPS w/o Parallel Reasoning under the same per-iteration retrieval budget. These results demonstrate that both parallel reasoning and reasoning disentanglement remain effective under a fixed retrieval budget, further indicating that MAPS’s advantage does not simply stem from retrieving more passages. We will include this experiment in the revised version of the paper.

Method2WikiMuSiQueHotpotQAGPQAAvg.
w/o Parallel Reasoning58.53447.838.444.7
w/o Disentanglement58.936.647.844.146.8
MAPS (N=2)5937.44944.447.5
AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer,

Thank you again for your time and thoughtful feedback. As the discussion period is approaching its conclusion, we would like to briefly follow up on our rebuttal. We have carefully addressed each of your comments and sincerely hope that our responses have clarified the concerns you raised.

If you have any remaining questions or concerns, we would be happy to provide further clarification during the remaining discussion period.

Best regards,
The Authors

Thanks for your response, which has addressed my concern. I prefer to keep my score with the weak acceptance.

Thank you for confirming that our response has addressed your concern. We sincerely appreciate your support and the time and care you have devoted to reviewing our work. If you have any further questions, we would be happy to discuss them and provide additional clarification during the remaining discussion period.

Reviewer_ZyCdOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies query-intent drift in iterative retrieval-augmented generation. It introduces “search momentum,” an exponentially weighted aggregation of semantic shifts between consecutive sub-queries, and shows that later sub-queries tend to align increasingly with this accumulated momentum. Based on this observation, the authors propose MAPS, which (i) disentangles historical reasoning traces and sub-queries from the accumulated knowledge and refined reasoning state, and (ii) replaces sequential sub-query generation with parallel exploration of multiple complementary retrieval directions. Experiments on NQ, PopQA, 2WikiQA, HotpotQA, MuSiQue, and GPQA with Qwen3-32B and Qwen3.6-35B-A3B show consistent improvements over Vanilla RAG, Self-RAG, IRCoT, Search-o1, S3, and ParallelSearch, with ablation studies and qualitative case analyses supporting the design.

Critical strengths

A clear problem formulation and a useful diagnostic analysis of query-intent drift. The paper identifies a real failure mode in iterative RAG: historical sub-queries can reinforce an increasingly deviated retrieval trajectory. The search-momentum formulation is simple and interpretable, and the empirical observation that later sub-queries become more aligned with accumulated momentum is a useful diagnostic. This analysis motivates the proposed method and gives the paper a coherent narrative beyond a purely architectural contribution.

Consistent empirical gains across datasets and backbones, with informative ablations. MAPS is evaluated on six QA benchmarks spanning single-hop, multi-hop, and reasoning QA, and on two Qwen backbones. The gains over strong iterative and parallel-search baselines are consistent, and the ablations isolate the contributions of retrieval-direction proposal, parallel reasoning, and disentanglement. The case study in Appendix A.3 is also helpful for illustrating how sequential search can drift while parallel exploration recovers relevant evidence. These results support the practical effectiveness of the proposed framework.

Critical weaknesses

The comparison does not control for inference cost, so the source of the gains is ambiguous. MAPS generates multiple retrieval directions per iteration, converts each direction into a sub-query, retrieves top-5 documents for each sub-query, and then calls a refinement agent to consolidate evidence and update the reasoning state. This is substantially more LLM inference and retrieval than single-query iterative baselines. The paper reports no comparison of LLM calls, generated tokens, retrieved passages, wall-clock latency, or monetary cost. As a result, it is difficult to tell whether the gains come from the proposed disentanglement and parallel-reasoning mechanism, or simply from exploring more queries and retrieving more evidence. This is consequential because the paper’s central claim is that MAPS regulates search momentum, not merely that more compute improves QA. At minimum, the authors should report a budget-matched comparison, e.g., matching the number of retrieved passages or LLM calls between MAPS and ParallelSearch/Search-o1, or provide a cost-quality curve.

The mechanism evidence is suggestive but not a clean causal validation, and the evaluation lacks statistical robustness checks. The search-momentum analysis is interesting, but the coefficients are fitted separately for each method, and the comparison of momentum trajectories across methods is descriptive. It is not established that the observed momentum stabilization is caused by disentanglement and parallel reasoning, rather than by differences in query generation, evidence pooling, or the fitted coefficients themselves. In addition, the main results appear to use sampled evaluation sets (1,000 questions per dataset, 198 for GPQA) and do not report repeated runs, confidence intervals, or significance tests. Several improvements are large, but some dataset-level comparisons are closer, and the paper does not clarify which conclusions are robust. This weakens the strength of the mechanism claim and the reliability of fine-grained rankings. A stronger validation would fix momentum coefficients across methods, ablate the number of parallel directions under a fixed retrieval budget, and report variance across seeds or bootstrap confidence intervals.

Taken together, these weaknesses do not invalidate the empirical gains, but they leave the central mechanism claim and the practical recommendation insufficiently supported in the current form. I lean toward borderline rejection, but would be willing to reconsider if the authors provide budget-matched comparisons, clearer cost accounting, and more rigorous statistical and mechanistic validation.

Questions for authors
  1. Inference-cost control. Can the authors report LLM-call counts, generated-token counts, number of retrieved passages, and latency for MAPS and the main baselines? Can they provide a budget-matched comparison, e.g., matching the number of retrieved passages or LLM calls between MAPS and ParallelSearch/Search-o1? If MAPS still outperforms under matched budgets, this would substantially strengthen the claim that the gains come from the proposed mechanism rather than from additional inference compute.

  2. Mechanism validation. The search-momentum coefficients are fitted separately per method. Can the authors clarify whether the cross-method momentum comparison uses the same fitted coefficients, or whether each method’s coefficients are re-fitted? Can they provide an ablation that fixes the momentum coefficients and varies only the number of parallel retrieval directions, or otherwise isolates the effect of disentanglement from the effect of exploring more queries?

  3. Statistical robustness. The main results appear to use sampled evaluation sets and a single run per configuration. Can the authors report repeated runs, bootstrap confidence intervals, or significance tests for the main comparisons, at least for the average score and the closest dataset-level comparisons? If not, can they clarify which conclusions should be treated as robust and which are merely suggestive?

4 comments
AuthorsOpenReview ↗

Reply to ZyCd (1/2)

Reply to Weakness 1: We thank the reviewer for this comment. Regarding the concern about controlling inference costs, we provide a unified set of experiments and responses in the first point of our General Response. Specifically, we conduct additional experiments with matched numbers of retrieved passages. Compared with the baseline models under the matched retrieval budget, MAPS still achieves higher average performance. This indicates that MAPS’s performance gains do not simply stem from retrieving more documents, but primarily result from its parallel reasoning strategy and reasoning disentanglement mechanism.

Reply to Weakness 2: We thank the reviewer for these comments. We address the concerns and additional experiments below:

  1. First, regarding whether momentum stabilization results from reasoning disentanglement and parallel reasoning, we discuss this question in the experiments in Section 5.3 of our paper. Specifically, we use t-SNE to jointly project the original query embeddings and the corresponding search momentum vectors of MAPS, MAPS w/o Disentanglement, and Search-o1 at different reasoning steps into a two-dimensional space. MAPS w/o Disentanglement retains the parallel reasoning strategy but does not employ reasoning disentanglement. Compared with MAPS, both Search-o1 and MAPS w/o Disentanglement exhibit larger directional deviations of search momentum from the original query. These results indicate that MAPS’s parallel reasoning mechanism mitigates the tendency of search momentum to become dominated by a single retrieval direction, while reasoning disentanglement reduces the influence of historical retrieval tendencies on the evolution of search momentum. Together, these findings also support the contribution of both MAPS mechanisms to momentum stabilization.

  2. Furthermore, to further assess the robustness of our experimental results, we conduct three independent runs for both MAPS and Search-o1 in the second point of our General Response. As shown in Table 3 of the second point of our General Response, both methods maintain stable performance across runs, with MAPS consistently outperforming Search-o1. These results indicate that MAPS’s advantage is not attributable to the randomness of a single run and demonstrate its robustness.

  3. In addition, we conduct ablation experiments by varying the number of parallel retrieval directions while keeping the retrieval budget fixed. Specifically, we fix the number of retrieval directions and sub-queries that MAPS generates at each iteration to N=1,2,3N=1,2,3, respectively, to examine whether its performance gains simply stem from retrieving more passages. The results in the table below show that increasing the number of parallel sub-queries improves performance to some extent. Nevertheless, MAPS achieves higher average performance than all fixed-budget ablation variants while generating an average of two sub-queries per iteration. This indicates that MAPS’s advantage does not simply stem from generating more sub-queries, but from adaptively selecting retrieval directions according to the current reasoning state and combining this strategy with reasoning disentanglement to support effective and robust reasoning.

MethodNQPopQA2WikiQAMuSiQueHotpotQAGPQAAvg.
Qwen3.6-35B-A3B
MAPS(N=1)56.366.862.839.549.863.856.5
MAPS(N=2)63.469.365.044.752.676.661.3
MAPS(N=3)62.768.967.245.754.876.862.7
MAPS64.169.365.845.255.179.263.1
Qwen3-32B
MAPS(N=1)60.161.254.133.145.244.049.6
MAPS(N=2)60.864.059.037.449.044.452.4
MAPS(N=3)63.764.759.137.347.545.352.9
MAPS63.464.859.437.648.862.156.0
AuthorsOpenReview ↗

Reply to ZyCd (2/2)

Reply to Question 1: We thank the reviewer for this question. Regarding the retrieval budget matching experiment, we provide a unified experimental evaluation and response in the first point of our General Response. We conduct a comparison with matched numbers of retrieved passages. MAPS still achieves higher average performance than Search-o1 under the matched retrieval budget. This suggests that MAPS’s gains do not simply stem from retrieving more passages, but primarily result from its parallel retrieval strategy and reasoning disentanglement mechanism. Furthermore, the table below reports the number of retrieved passages and the inference latency of MAPS and the main baseline methods. We will include these additional experimental results in the revised version of the manuscript.

Table. Runtime Latency on Each Dataset.

MethodNQPopQA2WikiQAHotpotQAMuSiQueGPQAAvg. (s)
Search-o1454.6419.1802.8585.6528.7450.3540.2
ParallelSearch687.0632.51792.02381.02102.72027.01603.7
MAPS1272.02288.12073.41451.91694.8866.41607.8

Table. Average Number of Passages Retrieved per Iteration for Different Methods.

MethodNQPopQA2WikiQAHotpotQAMuSiQueGPQAAvg.
Search-o11111111
ParallelSearch1.0661.0151.3961.3351.1941.4151.237
MAPS2.112.2922.4322.3472.2362.7152.355

Reply to Question 2: We thank the reviewer for this question. In our paper, we fit the search-momentum coefficients separately for each method using its trajectories and then hold these coefficients fixed for evaluation on independent test trajectories. Search momentum is an analytical metric derived from model-generated reasoning trajectories, which provides support and motivation for the subsequent design of MAPS. It is neither a model parameter nor an inference-time control variable and therefore cannot be held fixed.

To distinguish the effect of reasoning disentanglement from that of exploring more queries, we design two ablation variants: MAPS w/o Parallel Reasoning and MAPS w/o Disentanglement. MAPS w/o Parallel Reasoning replaces multi-directional exploration with sequential retrieval using a single query at each iteration. In contrast, MAPS w/o Disentanglement retains parallel retrieval but exposes the complete historical reasoning trajectory to subsequent query generation. The results in the table below show that removing either component reduces performance, supporting the respective contributions of parallel reasoning and reasoning disentanglement to MAPS. These experimental results are also presented in the ablation study in Section 5.2 of the paper.

Model / MethodNQPopQA2WikiQAHotpotQAMuSiQueGPQAAvg.
Qwen3-32B
MAPS63.464.859.448.837.662.356.1
w/o Parallel Reasoning60.161.254.145.233.145.349.8
w/o Disentanglement61.058.853.148.334.345.150.1
Qwen3.6-35B-A3B
MAPS64.169.365.855.145.279.263.1
w/o Parallel Reasoning56.366.862.849.839.563.856.5
w/o Disentanglement54.064.964.452.942.678.159.5

Reply to Question 3: We conduct three independent runs for both MAPS and Search-o1 in the second point of our General Response. As shown in Table 3 of the second point of our General Response, both methods maintain stable performance across runs, with MAPS consistently outperforming Search-o1. These results indicate that MAPS’s advantage is not attributable to the randomness of a single run and demonstrate its robustness.

AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer,

Thank you again for your time and thoughtful feedback. As the discussion period is approaching its conclusion, we would like to briefly follow up on our rebuttal. We have carefully addressed each of your comments and sincerely hope that our responses have clarified the concerns you raised.

If you have any remaining questions or concerns, we would be happy to provide further clarification during the remaining discussion period.

Best regards,
The Authors

AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer,

We hope this message finds you well.

As the author–reviewer discussion period is coming to an end, we would like to follow up once more and sincerely thank you for the time and effort you have devoted to reviewing our paper.

We have provided detailed, point-by-point responses to all of your comments and sincerely hope that our responses have addressed the concerns you raised. We would greatly appreciate it if you could let us know whether our responses have resolved your concerns. If any questions or concerns remain, we would be very happy to provide further clarification while the discussion is still open.

Thank you again for your thoughtful feedback and careful consideration of our work.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their efforts and valuable suggestions. Here, we try to address the common concerns.

1. Matched Retrieval Budget

Several reviewers raise a common concern about whether MAPS’s performance gains partly stem from retrieving more passages per iteration. To address this concern, we conduct additional experiments with matched fixed retrieval budgets. At each iteration, MAPS adaptively generates multiple parallel sub-queries, each retrieving the top-5 passages. As shown in Table 1, we compute the average number of passages retrieved per iteration by MAPS on each dataset, denoted by BMAPSB_{\text{MAPS}}. We then set the number of passages retrieved by Naive-RAG and Search-o1 to the corresponding BMAPSB_{\text{MAPS}} per Iteration, thereby ensuring an identical passage budget per retrieval iteration across all methods.

Table 1. Average Number of Retrieved Passages.

Model2WikiQAGPQAHotpotQAMuSiQueNQPopQA
Qwen3.6-35B-A3B1016101087
Qwen3-32B131812121113

As shown in Table 2, MAPS maintains higher average performance across both backbone models and outperforms the baselines on multiple datasets under the matched per-iteration retrieval budget. For baseline methods, simply increasing the number of retrieved passages expands the evidence pool but does not explicitly encourage the acquisition of complementary knowledge. Moreover, the additional passages may contain redundant or irrelevant information, thereby diluting useful evidence and introducing noise. In contrast, MAPS explores multiple complementary retrieval directions adaptively and reduces interference from historical retrieval trajectories through reasoning disentanglement, enabling more relevant evidence acquisition. These results suggest that MAPS’s gains are not simply attributable to retrieving more passages per iteration, but support the effectiveness of its adaptive parallel retrieval strategy and reasoning disentanglement mechanism. We will include this experiment in the revised version of the manuscript.

Table 2. Performance of Different Models under a Fixed Budget.

ModelMethod2WikiQAGPQAHotpotQAMuSiQueNQPopQAAvg
Qwen3.6-35B-A3BMAPS0.6580.7920.5510.4520.6410.6930.631
Qwen3.6-35B-A3BNaive-RAG0.5320.48440.4770.2780.5810.6170.4949
Qwen3.6-35B-A3BSearch-o10.4720.77230.4470.3400.4790.5960.5177
Qwen3-32BMAPS0.5940.6210.4880.3760.6340.6480.560
Qwen3-32BNaive-RAG0.3110.44190.4120.1970.5620.5990.4205
Qwen3-32BSearch-o10.4250.45760.4250.2430.5480.5810.4466

2. Repeated Runs of the Experiments

We conduct three independent runs for both MAPS and Search-o1 while keeping the evaluation sets and experimental settings identical. We report the results of all three runs along with their corresponding averages to characterize performance variability. As shown in the table below, both MAPS and Search-o1 exhibit stable performance across repeated runs, with MAPS consistently outperforming Search-o1. These results indicate that MAPS's performance advantage is not attributable to the randomness of a single run, further demonstrating its robustness.

Table 3. Repeated Runs of the Experiments.

MethodNQPopQA2WikiQAHotpotQAMuSiQueGPQAAvg.
Qwen3.6-35B-A3B
MAPS (run1)64.169.365.855.145.279.263.1
MAPS (run2)63.668.165.754.745.578.462.7
MAPS (run3)64.268.465.054.945.978.162.8
MAPS (avg.)64.068.665.554.945.578.662.9
Search-o1 (run1)53.062.056.344.034.176.354.3
Search-o1 (run2)52.862.355.544.433.875.254.0
Search-o1 (run3)53.163.455.144.233.675.754.2
Search-o1 (avg.)53.062.655.644.233.875.754.2
Qwen3-32B
MAPS (run1)63.464.859.448.837.662.156.0
MAPS (run2)63.764.159.649.536.263.456.1
MAPS (run3)62.465.961.148.937.262.356.3
MAPS (avg.)63.264.960.049.137.062.656.1
Search-o1 (run1)53.156.543.740.023.746.443.9
Search-o1 (run2)54.156.542.940.923.345.343.8
Search-o1 (run3)54.857.143.140.123.645.544.0
Search-o1 (avg.)54.056.743.240.323.545.743.9

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.