acceptodds
Under review as a conference paper at ICLR 2027

Continual Graph Retrieval-Augmented Generation under Corpus Expansion via Path Memory

35 views

Abstract

GraphRAG supports multi-hop reasoning for complex tasks, but becomes less reliable as both documents and queries evolve over time. Corpus expansion can weaken previously useful evidence paths while introducing new evidence that remains undiscovered, and evolving queries further shift graph exploration. These dynamics challenge both retention and uptake of supporting evidence, characterized by evidence path dilution. We propose EviWeaver, a non-parametric continual framework that maintains and refreshes path memory from prior graph exploration. EviWeaver preserves useful evidence paths and incorporates newly relevant support entirely at retrieval time without retraining. Experiments across three multi-hop question answering benchmarks show that EviWeaver improves retrieval robustness and answer quality under continual corpus and query evolution.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Graph, Retrieval, Multi-hop Question Answering, Expanding Corpora, Continual Retrieval

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MjRmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies GraphRAG as the underlying triple store keeps growing. It identifies a phenomenon it calls path dilution: when new triples and edges are inserted, the Personalized-PageRank mass that previously flowed from query seeds to supporting documents is split among the newly added neighbors, so the same query at a later step no longer surfaces the document it used to. The proposed framework, EviWeaver, (i) decomposes the query into atomic sub-questions and builds bridge-aware seed distributions over both query-mentioned and intermediate bridge entities; (ii) after retrieval, runs a backward PPR from the retrieved documents to score each traversed edge by its contribution, retaining the top edges as a per-query path memory; and (iii) at later steps either strengthens the stored edges at retrieval time (a query-specific weight bonus) or refreshes the memory when a staleness statistic s_new crosses a threshold. All adaptation happens at retrieval time — no retraining, no structural graph edit — and the method is evaluated on MuSiQue, 2WikiMultiHopQA, and HotpotQA under fixed-, query-expanding, and delayed-evidence protocols.

Critical strengths
  1. The observation that prior incremental / continual GraphRAG work updates the graph or the retriever without explicitly preserving which path made an old query succeed is a reasonable one, and the "path dilution" vocabulary is useful.
  2. Backward attribution of PPR flow to individual edges (rather than to nodes or documents) is a clean and implementable idea.
Critical weaknesses
  1. The central problem is scoped to a special case, not characterized as such. The paper treats "corpus expansion dilutes previously correct evidence" as a general law of continual GraphRAG. It is not. It holds only under an implicit assumption that the triples/documents added during expansion are irrelevant distractors. When the newly added content is correct, on-topic, and supplementary, a retriever that changes its evidence set is not failing — it is behaving correctly. The paper never separates three qualitatively different expansion regimes: (a) noise/distractor insertion, (b) complementary correct evidence, (c) corrections or superseding facts. The phenomenon EviWeaver defends against is regime (a). As written, the abstract and introduction overclaim beyond that.

  2. The "good evidence" and "memory is stale" signals are not independently validated. Refresh is triggered by s_new, the share of query–triple similarity mass assigned to triples outside the stored memory. The paper treats this as evidence that new evidence is better. But: (i) there is no ground-truth comparison at refresh time of new-vs-old quality — the system cannot actually know whether the new triples are better; (ii) in the worst case the system must pay for both a memory-augmented pass over the old rails and an unweighted exploration of the new content, and the paper does not quantify this dual-search cost; (iii) "supporting document" is operationalized as the benchmark gold passage, which does not transfer to open-ended settings where relevance is multi-faceted. The decision boundary that drives the whole method is therefore heuristic and under-justified.

  3. Edge re-weighting is the wrong primitive for a sizable class of queries. The method systematically biases importance toward previously traversed edges. This is defensible when the task is "find the one gold passage that makes this multi-hop fact true" — exactly the HotpotQA / MuSiQue / 2WikiMultiHopQA family. It breaks for aggregative queries — "how many documents mention X?", "which topics co-occur with X?", coverage/trend questions. On such queries every newly added document should count, and there is no principled notion of edges worth preserving versus edges worth diluting; the method would, by construction, under-count new evidence. The paper does not discuss this boundary at all.

  4. Backward attribution is noisy and can over-commit to incidental paths. STPPR assigns nonzero credit to essentially every edge on the flow paths; selecting a top-k and then up-weighting them (with an explicit bias toward low-degree entities) encodes a strong prior that the first successful traversal was the right one. Under a noisy seed, a spurious bridge, or a one-off good hit, this incident is baked into memory and then strengthened on every future query. The ablation in Table 1 removes modules, but does not measure the false-memory rate — how often the stored path was, on a later oracle check, wrong, and how much strengthening amplifies that error.

Questions for authors

Q1. Can you report a regime where newly added documents are correct and supplementary (e.g., a new edition of the same topic) rather than distractors? Does EviWeaver still help, or does path memory now fight against legitimate new evidence?

Q2. When s_new triggers a refresh, how often does the system actually rebuild, and what is the retrieval-time cost of running both the memory-augmented PPR and the fresh bridge-aware construction?

7 comments
AuthorsOpenReview ↗

Response to Reviewer MjRm (Part 1/6)

We thank you for your constructive review and detailed questions about our evaluation scope and memory maintenance mechanism. We appreciate your recognition of the motivation for preserving paths that supported earlier retrieval and your assessment that the “path dilution” vocabulary is “useful.” We also thank you for describing backward attribution of PPR flow to individual edges as “a clean and implementable idea.”

Your distinction among the three expansion regimes helped us clarify the evaluation scope. Your questions also motivated additional analyses of complementary evidence uptake, refresh costs, and the effects of strengthening memory constructed from imperfect retrieval.

As a brief summary, we first provide our main responses and the corresponding revisions to the manuscript:

  • Complementary correct evidence. Our delayed-evidence setting directly evaluates regime (b). Adaptive refresh supports new evidence uptake and improves the joint retrieval of existing and newly available evidence (Sections 3 and 5.2, and Appendix D.3).
  • Refresh signal and cost. The signal measures retrieval shift rather than evidence quality. Only one retrieval branch is executed, and refresh is selected in 20.8% of the evaluated decisions after initial construction (Section 4.4 and Appendix E.1, Table 19).
  • Query scope. We distinguish multi-hop evidence retrieval under a top-k budget from open-ended and aggregative queries that require broader evidence coverage (Section 6).
  • Memory errors and strengthening. We directly measure gold-document loss and recovery under strengthening and analyze how its effects depend on initial retrieval success (Section 5.5 and Appendix D.6, Tables 15–16).

We provide detailed responses below. If you have any additional questions or comments, we would be happy to discuss further.

AuthorsOpenReview ↗

Response to Reviewer MjRm (Part 2/6)

W1 & Q1. Expansion with Complementary Correct Evidence

The central problem is scoped to a special case, not characterized as such. The paper treats "corpus expansion dilutes previously correct evidence" as a general law of continual GraphRAG. It is not. It holds only under an implicit assumption that the triples/documents added during expansion are irrelevant distractors. When the newly added content is correct, on-topic, and supplementary, a retriever that changes its evidence set is not failing — it is behaving correctly. The paper never separates three qualitatively different expansion regimes: (a) noise/distractor insertion, (b) complementary correct evidence, (c) corrections or superseding facts. The phenomenon EviWeaver defends against is regime (a). As written, the abstract and introduction overclaim beyond that.

Can you report a regime where newly added documents are correct and supplementary (e.g., a new edition of the same topic) rather than distractors? Does EviWeaver still help, or does path memory now fight against legitimate new evidence?

Response to W1 and Q1.

Your distinction among these three expansion regimes helped us clarify the scope of our evaluation. Our evaluation covers both regimes (a) and (b), including complementary evidence arriving later in the delayed-evidence setting.

  • (a) Noise/distractor insertion. The fixed-query protocol of HippoRAG2 and our query-expanding setting evaluate retrieval as additional documents arrive.
  • (b) Complementary correct evidence. Our delayed-evidence setting releases some supporting documents after step 250, removing the assumption that all required evidence is available from the beginning.

The structural mechanism of path dilution does not depend on the added content being incorrect or irrelevant. When a new document connects to an entity on an existing path, that entity gains neighbors. PPR importance is then distributed across more transitions, weakening the contribution of the existing path and potentially pushing previously retrieved gold documents out of the top-k. For the multi-hop queries in our setting, newly arriving gold documents complement existing gold documents rather than replace them. Both are needed. Losing still-required evidence is therefore a retrieval failure, even when the newly added evidence is correct.

We have clarified in Section 3 that existing supporting evidence remains valid and that new documents may provide complementary evidence. Extensions to regime (c), including fact corrections and contradictory information, are identified as future work in Section 6.

Does path memory obstruct new evidence, and does EviWeaver still help?

We compare three methods in the delayed-evidence setting at the highest delayed-query ratio (100%) used in our evaluation.

  • HippoRAG2 performs PPR-based retrieval without path memory.
  • w/o Refresh continues strengthening the initial memory without rebuilding it.
  • EviWeaver uses the full method, including adaptive refresh.

New gold recall measures retrieval of supporting documents at their first arrival, averaged across queries. It directly examines whether memory obstructs newly available evidence. The all-gold joint retrieval rate measures the proportion of queries for which all supporting documents appear together in the top-5 at step 1000.

All values below are percentages.

DatasetMethodNew gold recall, step 500New gold recall, step 750All-gold joint retrieval, step 1000
MuSiQueHippoRAG286.295.247.4
w/o Refresh86.290.546.6
EviWeaver89.0100.049.1
2WikiMultiHopQAHippoRAG288.888.562.4
w/o Refresh89.286.956.3
EviWeaver87.690.266.0
  • EviWeaver achieves the highest result in five of the six comparisons. In particular, it achieves the highest all-gold joint retrieval rate on both datasets. On 2WikiMultiHopQA, this rate reaches 66.0%, compared with 56.3% for w/o Refresh. This shows that EviWeaver incorporates new evidence while retaining access to existing evidence.

  • Memory without refresh can hinder retrieval when new evidence first arrives. At step 750, w/o Refresh has lower new gold recall than HippoRAG2 on both MuSiQue (90.5% vs. 95.2%) and 2WikiMultiHopQA (86.9% vs. 88.5%). With adaptive refresh, EviWeaver achieves higher new gold recall than both HippoRAG2 and w/o Refresh on both datasets.

AuthorsOpenReview ↗

Response to Reviewer MjRm (Part 3/6)

W2 & Q2. Meaning of the Refresh Signal, Effectiveness, and Cost

The "good evidence" and "memory is stale" signals are not independently validated. Refresh is triggered by s_new, the share of query–triple similarity mass assigned to triples outside the stored memory. The paper treats this as evidence that new evidence is better. But: (i) there is no ground-truth comparison at refresh time of new-vs-old quality — the system cannot actually know whether the new triples are better; (ii) in the worst case the system must pay for both a memory-augmented pass over the old rails and an unweighted exploration of the new content, and the paper does not quantify this dual-search cost; (iii) "supporting document" is operationalized as the benchmark gold passage, which does not transfer to open-ended settings where relevance is multi-faceted. The decision boundary that drives the whole method is therefore heuristic and under-justified.

When s_new triggers a refresh, how often does the system actually rebuild, and what is the retrieval-time cost of running both the memory-augmented PPR and the fresh bridge-aware construction?

Response to W2 and Q2.

The refresh signal snews_{\text{new}} measures how retrieval has shifted toward information added since memory construction. It does not judge whether newly added evidence is better. We have clarified this distinction in Section 4.4.

1. Meaning of the refresh signal

snews_{\text{new}} is the share of similarity mass in the current top-KK retrieved triples assigned to triples added to the triple store after the latest memory construction time tmt_m. A triple is considered new based on its absence from the triple store at tmt_m, not its absence from the stored memory edges.

A larger share means that triples unavailable when memory was constructed have become prominent retrieval candidates for the current query. These triples can affect seed selection and subsequent PPR exploration. The signal therefore informs whether to rebuild memory for the current graph. EviWeaver aggregates query-level snews_{\text{new}} values using RMS to choose between strengthening and refresh.

When refresh is selected, EviWeaver performs bridge-aware seed construction over the current triple store, including both existing and newly added triples. It retrieves documents on the unmodified current graph G(t)G^{(t)} and reselects the transitions that contributed to retrieval. Newly added triples receive no special priority, and transitions leading to existing evidence can remain in the rebuilt memory. Refresh therefore does not assume that new evidence is better.

2. Effectiveness and sensitivity of the decision boundary

The delayed-evidence comparison in our response to W1 and Q1 shows that strengthening memory without refresh can hinder new evidence retrieval. EviWeaver mitigates this effect and improves the joint retrieval of existing and newly available evidence.

The refresh criterion is a heuristic based on retrieval shift, but its gains do not depend on a narrowly tuned threshold. In Appendix D.4, Table 11, average Recall@5 ranges from 81.00% to 81.29% for τc∈[0.20,0.40]\tau_c \in [0.20, 0.40], a variation of only 0.29 percentage points. Every evaluated threshold outperforms HippoRAG2's 79.55%. A lower threshold triggers refresh more readily and incurs additional QD cost, while retrieval quality remains stable across the evaluated range.

Reviewer_f4WCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies evidence path dilution in continual GraphRAG, where corpus expansion can weaken previously useful retrieval paths and make newly available evidence difficult to discover. The authors propose EviWeaver, a retraining-free framework that constructs bridge-aware retrieval seeds, stores query-conditioned edge memories through backward tracing, and adaptively strengthens or refreshes these memories as the graph evolves. The paper also introduces query-expanding and delayed-evidence evaluation protocols. Experiments on three multi-hop QA benchmarks show improved retrieval robustness and answer quality over existing baselines, particularly under substantial corpus expansion.

Critical strengths

1. Clear problem formulation and a well-motivated core idea. The paper identifies evidence path dilution as a meaningful challenge in continual GraphRAG and distinguishes between evidence retention and uptake. Maintaining query-conditioned path memory is an intuitive and substantive idea, with a coherent connection between the proposed components and the underlying problem.

2. Convincing evidence for improved retrieval robustness. The longitudinal results demonstrate substantially reduced performance degradation under corpus expansion. Notably, on 2WikiMultiHopQA, EviWeaver maintains stronger retrieval and answer quality despite starting below HippoRAG2. The ablations further show that path memory contributes meaningfully to later-stage performance, supporting the central mechanism rather than merely reporting an overall improvement.

Critical weaknesses

1. Inference budget is not fully controlled. EviWeaver uses substantially more LLM calls and tokens than HippoRAG2 (2.55 vs. 1.04 calls per query). Although the ablations support the contribution of path memory, it remains unclear how much of the overall improvement could be reproduced by allocating comparable inference compute to stronger query decomposition or repeated graph exploration. This limits the attribution of gains specifically to persistent path memory.

2. Evidence for uptake is less direct than for retention. The delayed-evidence experiments show encouraging improvements, but aggregate Recall@5 does not isolate retrieval of newly arrived supporting documents. It is therefore difficult to determine how much of the uptake improvement comes from adaptive memory refresh rather than ordinary re-retrieval on the expanded graph.

Questions for authors
  1. Compute-matched comparison: How would EviWeaver compare against a baseline that uses a similar LLM inference budget for repeated query decomposition or graph exploration, without maintaining persistent path memory?

  2. New evidence uptake: Can the authors report retrieval recall specifically for supporting documents when they first become available, ideally separating the contributions of memory refresh and repeated retrieval?

6 comments
AuthorsOpenReview ↗

Response to Reviewer f4WC (Part 1/3)

We sincerely thank you for your careful review and concrete suggestions. We appreciate your recognition of our “Clear problem formulation and a well-motivated core idea” and your description of query-conditioned path memory as “an intuitive and substantive idea.”

We also appreciate your assessment of the “Convincing evidence for improved retrieval robustness,” particularly the stronger retrieval and answer quality on 2WikiMultiHopQA despite starting below HippoRAG2. Your observation that “path memory contributes meaningfully to later-stage performance” directly recognizes the mechanism our ablations examine.

Following your suggestions, we have conducted additional experiments to separate the contribution of persistent path memory from additional inference computation and to directly evaluate new evidence uptake.

  • Comparison under the same expected token budget. EviWeaver outperforms a memory-free query decomposition (QD) baseline on both datasets, supporting the contribution of path memory beyond additional QD computation (Section 5.3 and Appendix D.7, Table 17).
  • Direct evaluation of uptake and refresh. We measure new evidence recall at its first arrival and compare EviWeaver with w/o Refresh. Adaptive refresh improves the final joint retrieval of existing and newly available evidence on both datasets (Section 5.2 and Appendix D.3, Table 10).

Please find our detailed responses below and let us know if you have any additional questions or comments.

AuthorsOpenReview ↗

Response to Reviewer f4WC (Part 2/3)

Q1. Comparison under the Same Expected Token Budget

Inference budget is not fully controlled. EviWeaver uses substantially more LLM calls and tokens than HippoRAG2 (2.55 vs. 1.04 calls per query). Although the ablations support the contribution of path memory, it remains unclear how much of the overall improvement could be reproduced by allocating comparable inference compute to stronger query decomposition or repeated graph exploration. This limits the attribution of gains specifically to persistent path memory.

Compute-matched comparison: How would EviWeaver compare against a baseline that uses a similar LLM inference budget for repeated query decomposition or graph exploration, without maintaining persistent path memory?

Response to Q1.

Following your suggestion, we constructed a baseline without persistent path memory that mixes QD-based bridge-aware retrieval and base retrieval. Base retrieval constructs the seed distribution from the original query without query decomposition. Both branches run PPR on the unmodified current graph. We compute the baseline's expected Recall@5 under the same expected LLM token budget as EviWeaver.

The table reports average Recall@5 across delayed-query ratios of 0–100% and steps 500, 750, and 1000. The comparison conditions and results for each ratio are provided in Appendix D.7, Table 17.

MethodMuSiQue Recall@5 (%)2WikiMultiHopQA Recall@5 (%)
HippoRAG272.3486.76
Same-budget QD without memory72.9187.35
EviWeaver73.9888.50

Allocating additional computation to QD reproduces part of EviWeaver's improvement over HippoRAG2. However, EviWeaver achieves higher Recall@5 on both datasets under the same expected token budget. These results support the additional contribution of persistent path memory beyond allocating the same budget to memory-free retrieval. We now discuss this comparison in Section 5.3, with details in Appendix D.7, Table 17.

AuthorsOpenReview ↗

Response to Reviewer f4WC (Part 3/3)

Q2. New Evidence Retrieval at First Arrival and the Contribution of Refresh

Evidence for uptake is less direct than for retention. The delayed-evidence experiments show encouraging improvements, but aggregate Recall@5 does not isolate retrieval of newly arrived supporting documents. It is therefore difficult to determine how much of the uptake improvement comes from adaptive memory refresh rather than ordinary re-retrieval on the expanded graph.

New evidence uptake: Can the authors report retrieval recall specifically for supporting documents when they first become available, ideally separating the contributions of memory refresh and repeated retrieval?

Response to Q2.

Following your suggestion, we measure recall for newly arrived gold documents at the step when they first enter the corpus. These documents are absent from the graph before that step, making this their first opportunity to be retrieved. We also report recall for existing gold documents at the same step and the proportion of queries for which all gold documents are jointly retrieved in the top-5 at the final step.

To isolate the contribution of refresh, we compare EviWeaver with w/o Refresh, which continues strengthening the existing memory without rebuilding it. Both methods rerun retrieval on the expanded graph at every step. EviWeaver rebuilds memory when refresh is selected, whereas w/o Refresh retains the previous memory.

New and existing gold recall are averages of per-query recall. The table averages results over delayed-query ratios of 25%, 50%, 75%, and 100%.

DatasetMetric (%)w/o RefreshEviWeaver
MuSiQueNew gold recall, step 50085.786.8
MuSiQueExisting gold recall, step 50053.153.7
MuSiQueNew gold recall, step 75091.194.9
MuSiQueExisting gold recall, step 75068.770.4
MuSiQueAll-gold joint retrieval rate, step 100043.045.1
2WikiMultiHopQANew gold recall, step 50090.289.4
2WikiMultiHopQAExisting gold recall, step 50086.584.5
2WikiMultiHopQANew gold recall, step 75086.091.1
2WikiMultiHopQAExisting gold recall, step 75075.777.9
2WikiMultiHopQAAll-gold joint retrieval rate, step 100054.863.7

The benefit of refresh does not appear at every step. On 2WikiMultiHopQA at step 500, both new and existing gold recall are slightly lower. At step 750, however, EviWeaver improves both new and existing gold recall on both datasets. It also achieves a higher final all-gold joint retrieval rate on both datasets.

These results show that retaining the same memory can miss newly available evidence even when retrieval is repeated on the expanded graph, and adaptive refresh mitigates this problem. EviWeaver reuses useful existing paths while updating memory as new evidence arrives, supporting the joint retrieval of existing and newly available evidence.

We have included this direct evaluation of uptake in Section 5.2 and reported the new-gold recall and final all-gold joint retrieval results in Appendix D.3, Table 10.

Reviewer_S1bxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EviWeaver, a continual GraphRAG framework that maintains query-specific path memory to preserve previously retrieved evidence as the corpus expands. The method uses query decomposition, backward PPR tracing, and adaptive edge reweighting to improve retrieval robustness.

Critical strengths
  1. Maintaining retrieval performance under corpus expansion is an interesting problem.

  2. The paper provides ablation studies and analyzes how graph expansion affects retrieval performance.

Critical weaknesses
  1. Questionable problem setting. The framework assumes that previously answered queries are repeatedly processed whenever the corpus expands. However, most real-world RAG queries are one-time requests. It is unclear why the system should continuously maintain and refresh query-specific path memories without new user requests. The authors should justify this assumption and evaluate more realistic query workloads.

  2. Unrealistic continual evaluation. The experiments use three static multi-hop QA datasets, MuSiQue, 2WikiMultiHopQA, and HotpotQA. Corpus expansion and delayed evidence are artificially simulated by splitting existing benchmark documents into batches. This does not adequately reflect real continual settings involving evolving facts, outdated information, and changing query distributions.

  3. Limited generalizability beyond PPR-based GraphRAG. The proposed path dilution problem is closely related to the probability propagation mechanism of PPR. However, many recent GraphRAG methods use different retrieval strategies, including LLM-guided graph traversal and agentic retrieval. The authors do not compare against these methods, making it unclear whether the proposed approach addresses a general continual GraphRAG problem or primarily a limitation of PPR-based retrieval.

  4. Limited novelty and efficiency concerns. The method mainly combines query decomposition, PPR-based path scoring, and adaptive edge reweighting. The technical novelty is relatively limited. Moreover, Table 14 shows that EviWeaver requires substantially more LLM calls and tokens than HippoRAG 2, while the improvements are relatively modest. The additional cost is difficult to justify, especially given the assumption that historical queries are repeatedly processed.

  5. Lack of support for evolving facts. The proposed framework assumes append-only corpus expansion, where existing nodes and edges are preserved. However, real-world continual GraphRAG systems must also handle updated, outdated, or contradictory facts. For example, when an existing fact changes, the proposed path memory may continue reinforcing obsolete evidence. The paper does not evaluate fact updates or knowledge deletion, which further limits the realism and practical applicability of the proposed continual setting.

Questions for authors

please refer to the weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer S1bx (Part 1/6)

We sincerely thank you for your careful review and constructive feedback. We appreciate your recognition that “Maintaining retrieval performance under corpus expansion is an interesting problem.” We also thank you for highlighting our empirical analysis: “The paper provides ablation studies and analyzes how graph expansion affects retrieval performance.”

Our main responses to your questions and corresponding revisions are summarized as:

  • Persistent information needs and repeated evaluation. We clarify why the same information need remains relevant as new documents arrive and how the query-expanding setting evaluates both newly introduced and earlier queries (Sections 1 and 3).
  • Evaluation scope. Our protocols evaluate retention and uptake under corpus expansion where existing supporting evidence remains valid. We explicitly distinguish this setting from fact corrections, deletions, and contradictory information (Sections 3 and 6).
  • Comparison beyond PPR-based retrieval. We add HopRAG results on all three datasets to examine retrieval robustness under LLM-guided graph traversal (Appendix A.2 and Tables 4-5).
  • Technical contributions and inference cost. We explain the challenges addressed by path memory and show that EviWeaver outperforms a memory-free query decomposition (QD) baseline under the same expected token budget (Sections 1 and 5.3, Appendix D.7, Table 17, and Appendix E.1, Tables 18–19).

Please refer to the detailed responses below. We would be glad to discuss any further questions you may have during the rebuttal phase.

AuthorsOpenReview ↗

Response to Reviewer S1bx (Part 2/6)

W1. Persistent Information Needs and Repeated Evaluation

Questionable problem setting. The framework assumes that previously answered queries are repeatedly processed whenever the corpus expands. However, most real-world RAG queries are one-time requests. It is unclear why the system should continuously maintain and refresh query-specific path memories without new user requests. The authors should justify this assumption and evaluate more realistic query workloads.

Response to W1.

Our study targets information needs that persist as the corpus expands. A previously processed query can remain relevant when an investigation is ongoing or the required evidence is still incomplete. When subsequent documents arrive, the same information need must be evaluated against the expanded corpus. In particular, registered queries require updated results even when users do not submit the same question again.

This setting has a substantial research history. Continuous queries [1] and NiagaraCQ [2] maintain registered queries and provide results as matching data arrive. For document streams, PDSum [3] studies continuous summarization of evolving document sets, CREAM [4] studies continual retrieval over dynamic query and document streams, and StreamingQA [5] studies QA over knowledge arriving over time. Although these works address different tasks, they share the need to maintain and update information access as data continue to arrive. In multi-hop QA with GraphRAG, HippoRAG2 [6] introduced the fixed-query corpus-expansion protocol on which our evaluation builds. We extend this protocol to query arrivals and the arrival of complementary evidence.

In practical workflows, the multi-hop evidence required for a persistent information need can arrive at different times. Real-world applications under this continual multi-hop QA setting may include:

  • Ongoing investigations. The FBI's N-DEx provides notifications when new records relevant to an ongoing investigation are added [7]. As an illustrative example, an existing record may connect incident A to vehicle C, while a subsequent record connects vehicle C to incident B. The information need persists, but the connecting evidence becomes complete only later. Retrieving both records is therefore necessary.
  • Living systematic reviews. Cochrane living systematic reviews continuously monitor new evidence for an ongoing research question [8]. They also require identifying and linking multiple reports of the same study. Later reports may contain study details or results absent from earlier reports [9]. For example, connecting the study population and design in an earlier report with additional results in a later report requires access to both documents.

Our work addresses the retrieval stage of these workflows, where access to existing evidence must be preserved while newly available complementary evidence is retrieved. We do not assume that every historical one-time request must be continuously re-executed. Reusing the same query at each expansion step measures whether retrieval quality is maintained for the same information need as the corpus expands. The query-expanding setting additionally introduces new queries over time, allowing us to evaluate retrieval for newly introduced queries together with retention for earlier queries.

Your comment helped us make the motivation for revisiting queries more explicit. We revised the Introduction to explain persistent information needs through living systematic reviews and clarified the evaluation target in Section 3.

References

[1] Continuous Queries over Append-Only Databases (SIGMOD 1992)

[2] NiagaraCQ: A Scalable Continuous Query System for Internet Databases (SIGMOD 2000)

[3] PDSum: Prototype-driven Continuous Summarization of Evolving Multi-document Sets Stream (WWW 2023)

[4] CREAM: Continual Retrieval on Dynamic Streaming Corpora with Adaptive Soft Memory (KDD 2026)

[5] StreamingQA: A Benchmark for Adaptation to New Knowledge over Time in Question Answering Models (ICML 2022)

[6] From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025)

[7] National Data Exchange (N-DEx) (FBI)

[8] Chapter 21: Living Systematic Reviews (Cochrane Handbook)

[9] Chapter 5: Collecting Data (Cochrane Handbook)

AuthorsOpenReview ↗

Response to Reviewer S1bx (Part 3/6)

W2 & W5. Purpose and Scope of Continual Evaluation

W2. Unrealistic continual evaluation. The experiments use three static multi-hop QA datasets, MuSiQue, 2WikiMultiHopQA, and HotpotQA. Corpus expansion and delayed evidence are artificially simulated by splitting existing benchmark documents into batches. This does not adequately reflect real continual settings involving evolving facts, outdated information, and changing query distributions.

W5. Lack of support for evolving facts. The proposed framework assumes append-only corpus expansion, where existing nodes and edges are preserved. However, real-world continual GraphRAG systems must also handle updated, outdated, or contradictory facts. For example, when an existing fact changes, the proposed path memory may continue reinforcing obsolete evidence. The paper does not evaluate fact updates or knowledge deletion, which further limits the realism and practical applicability of the proposed continual setting.

Response to W2 and W5.

The purpose of our evaluation is to measure the effects of corpus expansion on retention and uptake under varying, realistic scenarios. Diversifying when each query's supporting documents become available allows us to distinguish the loss of existing evidence from the uptake of newly available evidence.

Our starting point is the corpus-expansion protocol of HippoRAG2 [6], which adds document batches while the supporting evidence for the evaluation queries is available from the beginning. We further extend this evaluation to more realistic and practical scenarios with query-expanding and delayed-evidence settings to cover the arrival of new queries and complementary evidence. We have added this positioning to the related work in Section 2.

SettingChange evaluated
Fixed-query (Default)Retrieval of existing evidence as the corpus expands while queries remain fixed
Query-expanding (Our extension)Retrieval for new queries and retention for earlier queries as queries and documents arrive
Delayed-evidence (Our extension)Joint retrieval of existing and newly available supporting evidence for earlier queries

In the delayed-evidence setting, some gold documents arrive only at later steps. Preserving previously retrieved evidence alone is therefore insufficient. Documents are selected for delayed arrival independently of their position in the reasoning chain, so even first-hop evidence can arrive later. This setting evaluates situations in which the required evidence becomes progressively available as documents arrive.

Our evaluation assumes that existing supporting evidence remains valid as documents and complementary evidence are added. It does not directly evaluate changes in the validity of existing facts.

Related work addresses validity changes directly. TimelyRAG [10] retrieves document versions appropriate to the query time, while ERASE [11] edits or deletes existing knowledge entries when new documents arrive. EviWeaver addresses the loss of retrieval access to valid evidence that remains in the corpus. A system that also handles fact updates needs both evidence validity management and the preservation of retrieval access.

We have explicitly stated the validity assumption in Section 3.

We study retrieval robustness under corpus expansion, where existing documents remain available and existing supporting evidence remains valid. New documents may provide complementary evidence that was unavailable when a query was first processed.

We have also identified extensions to fact corrections, deletions, and contradictory information as future work in Section 6.

References

[6] From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025)

[10] TimelyRAG: Semantic-Temporal Hybrid Retrieval for Time-Critical Question Answering in Overlapping-Evolving Documents (Findings of EMNLP 2026)

[11] Language Modeling with Editable External Knowledge (Findings of NAACL 2025)

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

Dear Reviewers and Area Chair,

We sincerely thank the reviewers for their thoughtful feedback and constructive suggestions. Below, we summarize the strengths acknowledged by the reviewers and our responses with the updates in the manuscript to address their questions and concerns.

Key Strengths Highlighted by Reviewers

  1. Novelty and motivation
    • Importance of retrieval robustness under corpus expansion (S1bx, f4WC).
    • Clear problem formulation and distinction between retention and uptake (f4WC).
  2. Methodology
    • Intuitive and substantive path-memory design with coherent components (f4WC).
    • Identification of a gap in preserving previously successful retrieval paths and a useful path-dilution perspective (MjRm).
    • Clean and implementable backward attribution of PPR flow to individual edges (MjRm).
  3. Performance
    • Convincing evidence of reduced degradation and stronger later-stage retrieval and answer quality (f4WC).
    • Ablation studies and graph-expansion analyses supporting path memory’s contribution to later-stage performance (S1bx, f4WC).
  4. Clear and easy-to-follow presentation (S1bx, f4WC, MjRm).

Key Revisions to the Manuscript

To address the reviewers’ concerns, we have made the following updates.

  1. Problem setting and scope. Clarified persistent information needs and the coverage of distractor insertion and complementary evidence. Expanded the discussion of corrections, deletions, contradictory information, and open-ended or aggregative queries (Sections 1, 3, and 6).
  2. Comparison with LLM-guided graph traversal. Added HopRAG on all three datasets using the same queries and cumulative corpus at each expansion step (Section 2, Appendix A.2, and Tables 4–5).
  3. Comparison under the same expected token budget. Added a comparison against query decomposition without persistent path memory under the same expected LLM token budget (Appendix D.7, Table 17).
  4. Direct evaluation of new evidence uptake and refresh. Added measurements of new supporting-document recall at first arrival and final joint retrieval of all supporting documents, including comparisons with w/o Refresh (Appendix D.3, Table 10).
  5. Refresh signal and inference costs. Clarified the meaning of the refresh signal and branch selection. Reported refresh frequency and per-query costs for strengthening and refresh (Section 4.4 and Appendix E.1, Table 19).
  6. Retrieval harm and recovery from path memory. Added controlled comparisons of gold documents lost and recovered through strengthening, along with analysis by initial retrieval success, including harmful memory cases (Appendix D.6, Tables 15–16).

We hope these revisions and additional experiments address the reviewers’ concerns and further substantiate our contributions. We sincerely thank the reviewers and Area Chair for their 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.