acceptodds
Under review as a conference paper at ICLR 2027

Recursive Dense Retrieval

Abstract

*Recursive retrieval* comes in a few flavors: *single-hop* retrieval iteratively expands context around a single concept, whereas *multi-hop* retrieval focuses on *depth*, where it iteratively continues an evidence chain, and/or *width*, searching for complementary pieces of evidence. Although all scenarios require, at each retrieval step, decisions on (i) what to search for next and (ii) how to organize the evidence gathered along the way, both encoder training and inference strategies largely remain scenario-specific. In this work, we introduce **RDR**, a recursive dense retrieval framework that unifies single- and multi-hop retrieval. At the core of RDR is an instruction-conditioned encoder trained to map a retrieval goal and observed artifacts into a dense next-step query. Repeated encoding and index lookups build a retrieval tree, and we produce readouts (i.e., evidence rankings) by aggregating evidence scores across the tree's nodes. Because the instruction is part of the embedded state, the same encoder can both refine single-hop queries and follow multi-hop evidence chains. We train an RDR-8B encoder on **Recursive Retrieval Worlds** (RRW), a 4.75 million-sample dataset compiled from existing retrieval data, contextual transformations, constructed worlds, and recovery supervision. RDR performs strongly on a wide range of single- and multi-hop evaluations. In multi-hop retrieval, with a four-level retrieval tree with a branching factor of three, RDR reaches 80.14 and 25.28 Recall@5 on MuSiQue and BrowseComp+, respectively, outperforming state-of-the-art encoders and multi-hop retrievers on all four multi-hop benchmarks. In single-hop retrieval, RDR's tree-based readout improves the encoder's five-benchmark mean by 2.5 points, most on conversational retrieval (70.79 nDCG@10) and on multi-aspect reasoning queries, where it nearly closes the gap to the strongest reasoning retriever. On average, RDR outperforms the strongest evaluated baselines in both task families, substantially on multi-hop retrieval (by 5.8 Recall@5) and by 3.6 nDCG@10 on single-hop retrieval. We will publicly release RDR-8B, RRW data, and the evaluation code.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
dense retrieval, recursive retrieval, multi-hop retrieval, reasoning-intensive retrieval, instruction-conditioned embeddings, test-time scaling, synthetic training data

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YquPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Recursive Dense Retrieval (RDR). The authors fine-tune Qwen3-Embedding-8B on 4.75 million training data from Recursive Retrieval Worlds (RRW). During inference, RDR constructs a retrieval tree and combines document scores across search states using a level-balanced global readout. The method is evaluated on a range of single-hop and multi-hop retrieval benchmarks.

Critical strengths
  • RDR outperforms GRITHopper's native retrieval procedure on all four multi-hop benchmarks. It also achieves the highest five-suite single-hop mean among the models evaluated in Table 2. These results support the practical value of a unified framework across different retrieval tasks.

  • The paper compares the initialization and fine-tuned encoder under the same search procedure. It also compares readout methods on fixed traversals. These controls help identify the sources of the performance gains and strengthen the empirical findings.

Critical weaknesses
  • The proposed method combines existing ideas in context-conditioned recursive retrieval and tree-based score aggregation. RDR's main difference is that it aggregates document scores across learned, sequentially conditioned search states rather than only along selected retrieval paths. The experiments show that this combination is effective, but the methodological contribution appears incremental.

  • RRW is presented as a major contribution, but the effect of its specific data construction is not sufficiently isolated. The current comparisons do not distinguish this effect from that of additional fine-tuning on conventional retrieval data.

Questions for authors
  • The paper compares RDR-8B with the original Qwen3-Embedding-8B, but not with a model fine-tuned on the same amount of conventional retrieval data. How can the authors distinguish the contribution of RRW's data construction from the effect of additional fine-tuning?
  • How does the proposed readout, which assigns equal weight to each search level, compare with assigning equal weight to every search node? Using the same retrieval tree and per-node scores would help isolate the effect of level balancing.
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/3)

Dear Reviewer YquP,

Thank you for your review. We are glad that you see the practical value of a unified framework across retrieval tasks and that the matched-search and fixed-traversal controls help identify the sources of the gains.

Line numbers below refer to the submitted paper. New material is identified separately and marked in the attached revised-paper diff.

Comment: The proposed method combines existing ideas in context-conditioned recursive retrieval and tree-based score aggregation. [...] the methodological contribution appears incremental.

RDR combines a learned state operator with global readout. MDR (Xiong et al., 2021), BeamDR (Zhao et al., 2021), and GRITHopper (Erker et al., 2026) score selected retrieval paths; COR (Park et al., 2025) aggregates a tree of document-derived query views rather than learned state transitions (Section 2, lines 141–157; Appendix K, lines 1749–1755).

Both components matter: replacing learned states with document-as-query states lowers global-readout performance from 55.15 to 40.41 mean Recall@5 (Section 5.4, Table 3, lines 449–463). The same tree with Qwen3-Embedding-8B (Zhang et al., 2025) before RRW fine-tuning gains little (Section 5.1, Table 1, lines 331–358). The new recovery ablation further demonstrates the importance of explicitly training the recovery instruction (Table R1 in our next response below, also reported to ttJR; revised Appendix E.2, Table 22 in the attached diff).

Revised Text — Section 3; existing comparison in Section 2 / Appendix K

For both task families, the encoder weights, tree configuration, and level-balanced readout remain fixed; only the task input (instruction and query) changes.

We continue with the Rebuttal in the next comment.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/3)

Comment: RRW is presented as a major contribution, but the effect of its specific data construction is not sufficiently isolated. [...] How can the authors distinguish the contribution of RRW's data construction from the effect of additional fine-tuning?

Thank you for this question. Full training runs require thousands of A100 GPU-hours, so a further matched-size conventional-data run is not feasible within the rebuttal period and our available compute budget (revised Appendix A). We view Qwen3-Embedding-8B (Zhang et al., 2025), the checkpoint we train from (Section 4.4, lines 303–310), as a relevant conventional-retrieval baseline: it has already been trained on large-scale retrieval data, including synthetic supervision. The targeted recovery ablation below is the closest controlled training comparison we can provide within this budget. Started before and completed during the rebuttal period, it uses the same recipe for one epoch on RRW without recovery samples.

Table R1. New recovery-data ablation, Recall@5, RDR(4,3). Continuation is the primary readout; recovery re-encodes the same traversal. Full RDR rows: submitted Tables 1 and 20. Complete comparison: revised Appendix E.2, Table 22.

TrainingInstructionMuSiQueBC+FanOutQAFRAMESMean
Qwen3-Embedding-8B (no RRW)continuation62.6816.9127.6568.2543.87
recovery59.7515.6129.1767.1842.93
RRW without recovery samplescontinuation77.7721.5839.4776.8653.92
recovery71.3820.5436.7871.6750.09
RRW (RDR-8B)continuation80.1425.2837.7877.4155.15
recovery80.4724.7340.3176.2255.43

Removing recovery samples reduces mean Recall@5 by 1.23 under continuation and 5.34 under recovery. Recovery then falls 3.83 points below the ablation’s own continuation ranking. This supports recovery-specific supervision, one component of RRW’s construction. Unlike GRITHopper (Erker et al., 2026), which has no explicitly trained recovery instruction, RDR selects this learned mode with unchanged encoder weights (Section 5.1, lines 359–365; Appendix K). We add these results to the revised Appendix E.2, Table 22 in the attached diff.

Revised Text — Appendix E.2, Table 22

We evaluate a run started before and completed during the rebuttal period, trained with RDR-8B’s recipe for one epoch on RRW without its 1,043,474 recovery samples (3,706,250 samples, 1,809 steps). Evaluation follows Table 21. Full RRW training adds 1.23 mean Recall@5 under continuation and 5.34 under recovery (Table 22). Without recovery supervision, recovery scores below continuation (50.09 vs. 53.92); with full RRW, both instructions perform similarly (55.43 vs. 55.15). This supports recovery-specific supervision. GRITHopper (Erker et al., 2026) has no explicitly trained recovery instruction; RDR invokes this learned mode with unchanged encoder weights.

Revised Text — Section 6

A recovery-data ablation supports instruction-conditioned recovery behavior (Appendix E.2).

Revised Text — Appendix A — training cost

A full training run requires thousands of A100 GPU-hours. We report one full RRW run and one ablation without recovery supervision (Appendix E.2); multi-seed estimates and further data ablations remain future work.

We continue with the rebuttal in the next comment.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 3/3)

Comment: How does the proposed readout, which assigns equal weight to each search level, compare with assigning equal weight to every search node? Using the same retrieval tree and per-node scores would help isolate the effect of level balancing.

Thank you for this suggestion; we add this new comparison to revised Appendix F (Table 23), marked in the attached diff. On identical traversals and per-state scores, level balancing performs better on average: 55.15 vs. 52.96 mean Recall@5 on multi-hop, and 59.34 vs. 58.90 mean nDCG@10 on single-hop (difference 0.44, 95% CI [0.26, 0.62]).

Table R2. New node-versus-level comparison on identical traversals; paired bootstrap, 2,000 resamples. Differences are node minus level.

SuiteMetricEqual weight per levelEqual weight per nodeDifference [95% CI]
MuSiQueRecall@580.1477.17−2.97 [−3.82, −2.12]
BrowseComp+Recall@525.2824.94−0.34 [−1.00, 0.33]
FanOutQARecall@537.7838.62+0.84 [−0.59, 2.25]
FRAMESRecall@577.4171.12−6.29 [−7.57, −5.10]
Multi-hop meanRecall@555.1552.96−2.19
NanoBEIRnDCG@1068.1367.86−0.26 [−0.75, 0.21]
BRIGHTnDCG@1030.8131.05+0.24 [−0.06, 0.55]
TopiOCQAnDCG@1070.7969.72−1.07 [−1.52, −0.60]
ToolRetnDCG@1048.1147.89−0.22 [−0.64, 0.17]
CoIR (sample)nDCG@1078.8777.98−0.89 [−1.20, −0.57]
Single-hop meannDCG@1059.3458.90−0.44 [−0.62, −0.26]

Multi-hop retrieval needs the complete evidence chain: finding only hops three and four still leaves hops one and two missing. With 1, 3, 9, and 27 states per level, node weighting gives the deepest level 67.5% of the mass and the root 2.5%; level balancing gives each depth 25%. Of the gold documents lost from the top 5 under node weighting, 62% on MuSiQue and 92% on FRAMES were in the root’s top 5 (new analysis in revised Appendix F).

Single-hop differences are smaller: node weighting slightly improves the reasoning-intensive BRIGHT benchmark (Su et al., 2025) (+0.24 nDCG@10), while the other four suites decrease by 0.22–1.07 points. Level balancing remains better on average.

Revised Text — Appendix F, Table 23 — node weighting and interpretation

A multi-hop question needs the complete evidence chain. Equal weight per node shifts 67.5% of the mass to the deepest level and 2.5% to the root, versus 25% per level. On MuSiQue and FRAMES, 62% and 92% of gold documents lost from the top 5 were in the root’s top 5. Single-hop differences are smaller: node weighting slightly improves BRIGHT (Su et al., 2025) (+0.24), while the other four suites decrease by 0.22–1.07 nDCG@10; level balancing remains better on average (59.34 vs. 58.90).

Thank you again; the two analyses you suggested strengthen the paper. Please also see the uploaded paper diff, which highlights the additions in blue and shows the revisions proposed for the camera-ready version.

AuthorsOpenReview ↗

Follow-up on the training and level-balancing ablations

Dear Reviewer YquP,

thank you again for your careful review and initial weak-accept recommendation. We appreciate your recognition of the practical value of unified retrieval and the evidence provided by our matched-search and fixed-traversal controls.

Our responses clarify the methodological distinction, present the recovery-data ablation, and add your suggested node-versus-level comparison on identical traversals and scores, together with an explanation of why level balancing helps retain the complete evidence chain. The corresponding revisions are highlighted in the uploaded paper diff.

We hope our responses address your concerns. With less than 48 hours remaining in the discussion period, we would be grateful if you could let us know whether any questions or concerns remain, so that we have time to address them before the deadline.

Thank you again for your time and consideration.

Best,

The authors

Reviewer_WYmEOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper aims to improve complex single/mutli-hop retrieval mechanism by training a tree-based retriever model with global aggregation capability. The LLM-assisted curated training dataset is balanced mixture of single and multi-hop retrieval samples with detailed quality filtering. The proposed method improves over the existing methods, including domain-specific retrievers, across variety of task families, such as coding, tool-use, conversation and reasoning. The paper also analyzes particular cases how the retriever can be best configured depending on quality requirements, latency budgets and add-ons like re-ranking.

Critical strengths

Strong Generalization: The paper demonstrates, given strong supervision signal, the proposed method can achieve SoTA performance over existing methods across diverse tasks. Many prior approaches have been domain-specific and/or are limited in their method, whereas this work try to generalize the tree-based retrieval in a unified framework.

Detailed Ablations: The work also visualizes the rationale behind their approach and how they affect the downstream performance. For example, how the width and depth affect the performance (e.g., RDR(4,3) exceeds RDR(2,39) at 40 states) and what are the cost-quality trade-offs across configurations (e.g., RDR(3,3) with 0.36 s/query vs RDR(4,3)'s 1.02 s/query).

Critical weaknesses

LLM Disclosure and AI Use statement: It is unclear from the paper that which specific LLMs are used to curate the dataset and the judge. It would beneficial to include these summaries clearly in paper appendix with corresponding LLM versions (with citations where applicable) and/or AI Use Statement, rather than putting in external repos which are complex to navigate. The cost of those LLM usage is also unclear.

Questions for authors

Distillation vs actual methodological gain: Is the main contribution of the paper distilling strong supervisory signals from large generative models to retrieval methods? Does the compared prior baselines utilize these kinds of signals? What is the total monetary and token cost of these LLM usage, specially for the ones used during trainings dataset curation?

Edit: The authors have clarified their methodological contribution and LLM usage in their pipeline. Overall, I increased the Soundness from 2 -> 3, keeping original Rating of 3.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/2)

Dear Reviewer WYmE,

Thank you for your positive assessment. We are glad that you see RDR’s generalization across task families as its main strength and found the depth, width, and cost analyses useful.

Line numbers below refer to the submitted paper. New material is identified separately and marked in the attached revised-paper diff.

Comment: Distillation vs actual methodological gain: Is the main contribution of the paper distilling strong supervisory signals from large generative models to retrieval methods? Does the compared prior baselines utilize these kinds of signals?

We report a new ablation without recovery samples to examine whether the gains come only from distilling LLM-generated supervision. Recovery samples insert a hard negative, including mined distractors, or an out-of-context document while retaining the original positive target (Section 4.2, lines 272–278 of the submitted paper). They add explicit supervision for recovering from misleading context. The run used the same recipe for one epoch on RRW without its 1,043,474 recovery samples; it was started before and completed during the rebuttal period.

Table R1. Recovery-data ablation, Recall@5, RDR(4,3). Continuation is the primary readout; recovery re-encodes the same traversal. Full RDR rows: submitted Tables 1 and 20. Ablation: revised Appendix E.2, Table 22.

TrainingInstructionMuSiQueBC+FanOutQAFRAMESMean
Qwen3-Embedding-8B (no RRW)continuation62.6816.9127.6568.2543.87
recovery59.7515.6129.1767.1842.93
RRW without recovery samplescontinuation77.7721.5839.4776.8653.92
recovery71.3820.5436.7871.6750.09
RRW (RDR-8B)continuation80.1425.2837.7877.4155.15
recovery80.4724.7340.3176.2255.43

Including recovery supervision improves mean Recall@5 by 5.34 points under the recovery instruction and by 1.23 under continuation. This supports recovery-specific supervision as an additional source of gains beyond synthetic query generation alone.

We do use LLM-generated data, especially for context lifting. Synthetic query or instruction supervision is also used by Qwen3-Embedding (Zhang et al., 2025), ReasonIR (Shao et al., 2025), and Promptriever (Weller et al., 2025). RRW additionally supervises transitions: context lifting renders the document or dialogue observed by a state, while the target specifies what to retrieve next given that context (Section 4.2, lines 265–269; Appendix C.1).

Our methodological contribution also includes global readout: it scores candidate documents across all learned states, rather than only along individual retrieval paths as in MDR (Xiong et al., 2021) and GRITHopper (Erker et al., 2026) (Section 2, lines 141–157; Appendix K). With the encoder and traversal fixed, replacing path-product scoring with global readout improves mean Recall@5 from 52.69 to 55.15 (+2.46) and mean single-hop nDCG@10 from 57.65 to 59.34 (+1.69), without additional training or LLM calls (Section 5.4, Table 3, lines 433–460 of the submitted paper).

Revised Text — Appendix C.1 — construction families

LLMs render the observed document or dialogue for context lifting, while the target specifies what to retrieve next given that context. For constructed worlds, we first generate a latent structure (entities, relations, and evidence layout) and then render its documents, so each transition’s target follows from that structure (models in Table 11).

Revised Text — Appendix E.2, Table 22

We evaluate a run started before and completed during the rebuttal period, trained with RDR-8B’s recipe for one epoch on RRW without its 1,043,474 recovery samples (3,706,250 samples, 1,809 steps). Evaluation follows Table 21. Full RRW training adds 1.23 mean Recall@5 under continuation and 5.34 under recovery (Table 22). Without recovery supervision, recovery scores below continuation (50.09 vs. 53.92); with full RRW, both instructions perform similarly (55.43 vs. 55.15). This supports recovery-specific supervision. GRITHopper (Erker et al., 2026) has no explicitly trained recovery instruction; RDR invokes this learned mode with unchanged encoder weights.

We continue with the rebuttal in the next comment.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/2)

Comment: LLM Disclosure and AI Use statement: It is unclear from the paper that which specific LLMs are used to curate the dataset and the judge. [...] The cost of those LLM usage is also unclear

Thank you; we agree that this belongs in the paper. The revision adds a new Table 11 to Appendix C.1 and refers to it from the AI use statement:

• Context lifting (Chat-ReasonIR, Doc-ReasonIR, Doc-Promptriever): GPT-4o-mini (gpt-4o-mini-2024-07-18) via the OpenAI Batch API (OpenAI, 2024); source supervision: (Shao et al., 2025) and (Weller et al., 2025).

• Constructed company and web worlds, the high-recall pipelines, their negatives, and quality filtering: Qwen3.5-122B-A10B, served locally with vLLM (filtering at temperature 0) (Qwen Team, 2026a); vLLM: (Kwon et al., 2023).

• Listwise reranking judge (Appendix J): Qwen3.6-27B, served locally (Qwen Team, 2026b).

Context lifting issued 1,079,247 GPT-4o-mini batch requests with 2.44 billion input and 0.39 billion output tokens, about 300 USD at batch prices. All other generation and filtering ran locally on our own GPUs.

Revised Text — AI use statement

LLMs generated and transformed synthetic training examples; Table 11 lists the model and version used in each construction stage and for the reranking judge.

Revised Text — Appendix C.1, Table 11

Table 11 lists the language models used in each construction stage. Context lifting issued 1,079,247 GPT-4o-mini requests through the OpenAI Batch API, with 2.44 billion input and 0.39 billion output tokens (about 300 USD at batch prices). All other generation and filtering ran locally on our own GPUs.

Thank you again for your review and the concrete suggestions, which improve the paper’s transparency. Please also see the uploaded paper diff, which highlights the additions in blue and shows the revisions proposed for the camera-ready version.

References not yet part of the paper:

Kwon, Woosuk, Zhuohan Li, Siyuan Zhuang, et al. 2023. “Efficient Memory Management for Large Language Model Serving with PagedAttention.” In Proceedings of the 29th Symposium on Operating Systems Principles, 611–626. Association for Computing Machinery. https://doi.org/10.1145/3600006.3613165.

OpenAI. 2024. “GPT-4o mini: Advancing Cost-Efficient Intelligence.” July 18, 2024. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/.

AuthorsOpenReview ↗

Follow-up on recovery supervision and LLM disclosure

Dear Reviewer WYmE,

thank you again for your constructive review and initial weak-accept recommendation. We appreciate your positive assessment of RDR’s generalization across task families and its depth, width, and cost analyses.

Our responses add the recovery-data ablation and explain the fixed-encoder readout controls to clarify the methodological contributions beyond synthetic supervision. We also specify the LLMs and versions used, report the context-lifting token and API costs, and incorporate these disclosures into the revised paper.

We hope our responses address your concerns. With less than 48 hours remaining in the discussion period, we would be grateful if you could let us know whether any questions or concerns remain, so that we have time to address them before the deadline.

Thank you again for your time and consideration.

Best,

The authors

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

RDR (Recursive Dense Retrieval) is a dense retriever built on the idea that one learned encoder can decide what to search for next from what it has already retrieved. It argues that refining a single query, following an evidence chain, and collecting complementary evidence all need this step, yet existing retrievers treat each scenario separately or hand the step to an LLM agent. RDR fine-tunes Qwen3-Embedding-8B on author-proposed Recursive Retrieval Worlds (RRW). RRW has 4.75M samples, each mapping an instruction, a query, and already-seen documents to the next target document, drawn from existing retrieval and multi-hop data. At inference, the encoder is applied recursively under a "continuation" (multi-hop) or "recovery" (single-hop) instruction to build a retrieval tree. A global readout then ranks every candidate by averaging the scores all states give it, rather than scoring it only along the path that found it. On four multi-hop benchmarks (MuSiQue, BrowseComp+, FanOutQA, FRAMES) and six single-hop benchmarks (BRIGHT, BRIGHT-Pro, NanoBEIR, ToolRet, CoIR, TopiOCQA), RDR reports higher Recall@5 than the multi-hop specialist GRITHopper-7B on every multi-hop suite. It also reports the best average single-hop nDCG@10 among the evaluated embedders.

Critical strengths
  1. Training the retriever itself for recursive retrieval is an interesting idea, and the dataset built to do it is solid. Instead of handing the search update to an LLM, RDR teaches one embedding model to map a goal plus already-seen documents to the next query, with instructions that switch between refining a query and continuing an evidence chain. The RRW dataset turns this into 4.75M uniform training samples. The construction is documented, with quality checks, de-duplication, and a public data viewer. This matters because a standard embedder does not benefit from recursion.
  2. The experimental results are promising. RDR beats the multi-hop GRITHopper-7B on all four multi-hop suites, most clearly on BrowseComp+ (Table 1). It also has the best average single-hop nDCG@10 among the evaluated embedders (Table 2). With no LLM in the loop, this makes RDR a practical option for multi-step search.
Critical weaknesses
  1. The experiments show that RDR works but not why, so the paper's two main claims ("one learned operator" and "the readout drives the gains") are not yet supported. How to understand the role of data against the proposed methodology? There is no ablation of the RRW training data. On MuSiQue, where both models train on the same five datasets, GRITHopper's encoder inside RDR's tree ties RDR (80.06 vs 80.14; Table 13). The large margins appear on BrowseComp+ and FanOutQA, which resemble RRW's synthetic long-chain worlds.
Questions for authors
  1. As the large margins appear on BrowseComp+ and FanOutQA. Has the author done data audit between the training set against these test set to rule out information leakage?
  2. Do the authors have even a coarse RRW ablation, e.g., removing the synthetic web/company/KG worlds or removing recovery samples, on BrowseComp+ and FanOutQA? This helps us understand whether the multi-hop gains should be credited to the recursive framework or to targeted training data.
  3. Why does recursion beat a single lookup, and how is self-retrieval (since the query now has the document) avoided?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1)

Dear Reviewer ttJR,

Thank you for your careful review. We are glad that you find training the retriever itself for recursive retrieval an interesting idea, consider RRW solid and well documented, and see RDR, with no LLM in the loop, as a practical option for multi-step search. Below we connect the existing controlled comparisons, add the recovery-data ablation you suggested, and point to the corresponding clarifications in the revised PDF. Line numbers below refer to the submitted paper.

Comment: The experiments show that RDR works but not why, so the paper’s two main claims (“one learned operator” and “the readout drives the gains”) are not yet supported.

“One learned operator” means the same encoder weights, tree configuration, and level-balanced readout serve single-hop and multi-hop retrieval. Only the input (instruction and query) changes (Section 3, lines 184–215; Appendix B.5, lines 925–930). The instruction selects recovery or continuation against a fixed document index.

RDR benefits from recursion, whereas its initialization does not consistently benefit. The comparison below summarizes existing results for Qwen3-Embedding-8B (Zhang et al., 2025) and RDR:

EvaluationMetricQwen single lookupQwen treeRDR single lookupRDR tree
Multi-hop mean (4 suites)Recall@543.1143.8742.7055.15
Single-hop mean (5 suites)nDCG@1055.7954.4456.8359.34

Sources: multi-hop means calculated from Table 1 (lines 325–337); single-hop means from Appendix B.2, Table 5 (lines 879–894).

Global Readout: Table 3 and Section 5.4 already address the readout claim (lines 433–460): with global readout, chain, beam, and tree search perform similarly. Changing the readout has a much larger effect:

SearchReadoutMulti-hop mean R@5Single-hop mean nDCG@10
ChainTerminal state → global32.19 → 53.9154.31 → 57.95
Beam (W=13)Path product → global42.77 → 53.8856.68 → 59.60
TreePath product → global52.69 → 55.1557.65 → 59.34

Global readout outperforms terminal-state or path scoring for each search strategy. Transfer to GRITHopper (Erker et al., 2026) is discussed below (Appendix C.3, Table 13, lines 1076–1090).

Revised Text — Section 3; existing readout controls in Section 5.4, Table 3

For both task families, the encoder weights, tree configuration, and level-balanced readout remain fixed; only the task input (instruction and query) changes.

Comment: On MuSiQue, where both models train on the same five datasets, GRITHopper's encoder inside RDR's tree ties RDR (80.06 vs 80.14; Table 13). The large margins appear on BrowseComp+ and FanOutQA, which resemble RRW's synthetic long-chain worlds.

GRITHopper (Erker et al., 2026) essentially ties RDR on MuSiQue (80.06 vs. 80.14), but trails by 11.01, 9.15, and 6.07 Recall@5 on BrowseComp+, FanOutQA, and FRAMES under the same tree and readout (Appendix C.3, Table 13, lines 1076–1090). Our claim concerns generalization across task families, as stated in the Introduction (lines 82–110) and the evaluation setup (lines 303–312).

We also report GRITHopper’s single-hop results, absent from the submitted paper; they are now included as a new row in Table 2 of the revised version and marked in the attached paper diff. With the same evaluation harness, it is below both RDR and Qwen3-Embedding-8B (Zhang et al., 2025) on all four evaluated suites:

Single-hop nDCG@10. GRITHopper: new evaluation; RDR and Qwen3: submitted in Table 2 (lines 341–350).

ModelBRIGHTNanoBEIRToolRetCoIR
GRITHopper-7B11.5661.3635.5654.73
Qwen3-Embedding-8B23.1771.7444.7279.93
RDR-8B30.8468.1348.0478.88

These results support broader generalization; the resemblance to RRW’s constructed worlds is addressed by the audit below.

Revised Text — Table 2; Appendix C.3

Table 2 includes GRITHopper-7B’s (Erker et al., 2026) single-hop results under the same evaluation harness: 11.56, 61.36, 35.56, and 54.73 nDCG@10 on BRIGHT, NanoBEIR, ToolRet, and CoIR. It is below both RDR-8B and Qwen3-Embedding-8B (Zhang et al., 2025) on all four suites (Table 2).

We continue with the rebuttal in the next comment.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2)

Comment: There is no ablation of the RRW training data. [...] Do the authors have even a coarse RRW ablation, e.g., removing the synthetic web/company/KG worlds or removing recovery samples, on BrowseComp+ and FanOutQA?

Thank you for this suggestion. We report a run started before and completed during the rebuttal period: the same training recipe, one epoch on RRW without its 1,043,474 recovery samples. Evaluation follows Section 4.4 and Appendix B; the full model’s results are taken from Tables 1 and 20 of the submitted paper. The ablation is new and is added in revised Appendix E.2, Table 22.

Table R1. New recovery-data ablation, Recall@5, RDR(4,3). Continuation is the primary readout; recovery re-encodes the same traversal. Full RDR rows: submitted Tables 1 and 20. Complete comparison: revised Appendix E.2, Table 22.

TrainingInstructionMuSiQueBC+FanOutQAFRAMESMean
Qwen3-Embedding-8B (no RRW)continuation62.6816.9127.6568.2543.87
recovery59.7515.6129.1767.1842.93
RRW without recovery samplescontinuation77.7721.5839.4776.8653.92
recovery71.3820.5436.7871.6750.09
RRW (RDR-8B)continuation80.1425.2837.7877.4155.15
recovery80.4724.7340.3176.2255.43

Recovery supervision adds 1.23 mean Recall@5 under continuation and 5.34 under recovery. Without it, recovery falls 3.83 points below the ablation’s own continuation ranking. This supports learned recovery behavior: unlike GRITHopper (Erker et al., 2026), which has no explicitly trained recovery instruction, RDR selects this mode with the same encoder weights (Section 5.1, lines 359–365; Appendix K).

Unfortunately, full training runs require thousands of A100 GPU-hours, so further full-scale training ablations are not feasible within the rebuttal period and our available compute budget (revised Appendix A).

Revised Text — Appendix E.2, Table 22

We evaluate a run started before and completed during the rebuttal period, trained with RDR-8B’s recipe for one epoch on RRW without its 1,043,474 recovery samples (3,706,250 samples, 1,809 steps). Evaluation follows Table 21. Full RRW training adds 1.23 mean Recall@5 under continuation and 5.34 under recovery (Table 22). Without recovery supervision, recovery scores below continuation (50.09 vs. 53.92); with full RRW, both instructions perform similarly (55.43 vs. 55.15). This supports recovery-specific supervision. GRITHopper (Erker et al., 2026) has no explicitly trained recovery instruction; RDR invokes this learned mode with unchanged encoder weights.

Revised Text — Section 6

A recovery-data ablation supports instruction-conditioned recovery behavior (Appendix E.2).

Revised Text — Appendix A — training cost

A full training run requires thousands of A100 GPU-hours. We report one full RRW run and one ablation without recovery supervision (Appendix E.2); multi-seed estimates and further data ablations remain future work.

Comment: Why does recursion beat a single lookup, and how is self-retrieval (since the query now has the document) avoided?

Two mechanisms address the concern about embedding collapse and self-retrieval: training distinguishes observed context from the next target, while global readout combines evidence across states.

Training: to discourage simply retrieving the observed document again, 653,515 of 4,749,724 RRW samples (about 14%) use a document already in their context as the hard negative, contrasted with the positive next target (new quantitative breakdown in revised Appendix C.1). This trains the state update to seek missing information.

Inference: recursion produces several state-conditioned rankings instead of one query vector. Global readout aggregates support across these states, so an earlier document can still contribute when retrieved again. At every root-score level, documents retrieved by more states are more often relevant (Figure 5; Section 5.4, lines 458–463). The full top 200 per state contributes to readout (Section 3, lines 207–215); repeated evidence is not simply discarded.

Revised Text — Appendix C.1 — observed documents as negatives; Figure 5 for the readout analysis

About 14% of all RRW samples (653,515 of 4,749,724) use a document that is already in their context as the hard negative, contrasting it with the positive target. This explicitly trains the encoder to distinguish what has already been observed from what is still missing.

We continue with the rebuttal in the next comment

AuthorsOpenReview ↗

Rebuttal by Authors (Part 3)

Comment: As the large margins appear on BrowseComp+ and FanOutQA. Has the author done a data audit between the training set an*d these test set's to rule out information leakage?

We audited the multi-hop benchmarks you questioned, BrowseComp+ (Chen et al., 2026) and FanOutQA (Zhu et al., 2024), together with MuSiQue (Trivedi et al., 2022) and FRAMES (Krishna et al., 2025). Exact matching found no test question in any of the 4,749,724 normalized RRW queries, including the question part of context-bearing states.

Beyond exact matching, we ran an n-gram audit following GPT-3’s overlap analysis (Brown et al., 2020) and PaLM’s criterion (Chowdhery et al., 2023): at least 70% of an item’s word 8-grams must occur in training. We compared these multi-hop test queries and gold documents against all RRW training text and the public training data of GRITHopper (Erker et al., 2026) and Llama-Embed-Nemotron-8B (NVIDIA, dataset release).

Table R3. Multi-hop test items meeting the n-gram criterion (%); new audit in revised Appendix C.1, Table 11b. FanOutQA and FRAMES gold documents include all passages of gold pages.

SuiteItem typeRRWLlama-Embed-Nemotron-8BGRITHopper
MuSiQueGold documents29.2758.5348.56
BrowseComp+Gold documents0.040.040.00
FanOutQAGold documents0.030.020.01
FRAMESGold documents0.930.570.60
All fourQueries0.000.000.00

No test query meets the criterion in any corpus. Document overlap is below 1% on BrowseComp+, FanOutQA, and FRAMES, and RRW has the lowest rate on MuSiQue. The higher document-level overlap on MuSiQue is consistent with the reuse of Wikipedia passages across retrieval datasets. Neither BrowseComp+ nor FanOutQA supplies training questions: the company and web worlds are constructed from generated entity/persona seeds and latent evidence structures (Section 4.2, lines 265–269; Appendix C.1).

Revised Text — Appendix C.1 — multi-hop excerpt of the audit; Table 11b

We find no normalized exact match between RRW queries and the evaluated multi-hop questions. We additionally apply PaLM’s n-gram criterion (Chowdhery et al., 2023), following GPT-3’s overlap analysis (Brown et al., 2020): at least 70% of an item’s word 8-grams occur in training. Table 11b compares test queries and gold documents against all RRW text and the public training data of GRITHopper (Erker et al., 2026) and Llama-Embed-Nemotron-8B (NVIDIA, dataset release). No test query meets the criterion in any corpus. Document-level rates are below 1% on BrowseComp+, FanOutQA, and FRAMES; RRW has the lowest rate on MuSiQue (29.27% versus 58.53% and 48.56%).

Thank you again for the constructive questions. We hope that the ablation and the clarifications address your concerns. Please also see the uploaded paper diff, which highlights the additions in blue and shows the revisions proposed for the camera-ready version.

References not yet part of the paper:

Brown, Tom B., Benjamin Mann, Nick Ryder, et al. 2020. “Language Models Are Few-Shot Learners.” In Advances in Neural Information Processing Systems 33, edited by H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, 1877–1901. Curran Associates. https://proceedings.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html.

Chowdhery, Aakanksha, Sharan Narang, Jacob Devlin, et al. 2023. “PaLM: Scaling Language Modeling with Pathways.” Journal of Machine Learning Research 24 (240): 1–113. https://jmlr.org/papers/v24/22-1144.html.

NVIDIA. 2025. embed-nemotron-dataset-v1. Public training data of Llama-Embed-Nemotron-8B. Hugging Face dataset. https://huggingface.co/datasets/nvidia/embed-nemotron-dataset-v1.

AuthorsOpenReview ↗

Follow-up on the recovery ablation and data audit

Dear Reviewer ttJR,

thank you again for your thoughtful review and initial weak-accept recommendation. We appreciate your recognition of the value of training the retriever for recursive retrieval, RRW’s construction, and RDR’s practical benefits without an LLM at inference.

Our responses bring together the operator and readout controls, add the requested recovery-data ablation and multi-hop overlap audit, and clarify generalization and self-retrieval.

These additions are also reflected in the uploaded paper diff. We hope our responses address your concerns. With less than 48 hours remaining in the discussion period, we would be grateful if you could let us know whether any questions or concerns remain, so that we have time to address them before the deadline.

Thank you for your time and consideration.

Best, The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Rebuttal Summary and Discussion Status

Dear Area Chair,

Dear Program Chairs,

Thank you for overseeing the review process, and we thank all three reviewers for their thoughtful assessments. All three recommended weak acceptance (3, 3, 3) based on the submitted version, before the additional experiments and clarifications presented in our rebuttal.

We particularly appreciate Reviewer WYmE’s engagement during the discussion. Following our clarifications of the methodological contribution and LLM usage, they raised Soundness from 2 to 3 while maintaining their positive overall recommendation.

Our responses provide additional evidence on the main questions:

  • Training and methodological contributions: a new recovery-data ablation supports the contribution of recovery-specific supervision, alongside the existing fixed-encoder comparisons isolating the benefit of global readout.
  • Generalization and data overlap: additional GRITHopper single-hop results and an exact-match and n-gram audit of the multi-hop benchmarks address the questions about generalization and training–test overlap.
  • Readout and transparency: the requested node-versus-level comparison uses identical traversals and scores and shows that level balancing performs better on average (55.15 vs. 52.96 mean Recall@5 on multi-hop retrieval). The revision also documents the LLMs used and the context-lifting token and API costs.

Reviewers ttJR and YquP have not yet posted follow-up responses. Their currently posted scores remain those assigned before our rebuttal. We believe the additional experiments and clarifications substantially address their main concerns and further strengthen the empirical support for our claims. We would be grateful if this evidence and the uploaded paper diff, which highlights the additions in blue, could be considered alongside the initial reviews in the final assessment. We remain available for any further questions.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.