acceptodds
Under review as a conference paper at ICLR 2027

CAN RETRIEVAL FAILURES DRIVE GRAPH EVOLUTION? SELF-IMPROVING GRAPHS FOR EVIDENCE RECOVERY IN MULTI-HOP QA

Abstract

Frontier models are increasingly expected to be self-improving, refining their reasoning, tools, and environments rather than remaining static after training. However, optimization typically targets only the policy, while the environment remains fixed. Multi-hop question answering makes this gap concrete: answering requires retrieving, connecting, and verifying evidence across documents, yet existing retrieval-augmented and graph-based systems operate over fixed retrieval environments, and reinforcement learning (RL) approaches typically rely on terminal answer rewards. Such rewards provide limited guidance on which retrieval decisions contribute to success or failure, and policy optimization alone cannot repair deficiencies in how the environment represents and connects evidence. We present SIG-R1, a self-improving graph method that combines retrieval-step rewards with failure-driven adaptation of a searchable hypergraph. SIG-R1 assigns process-aware credit to individual retrieval steps, diagnoses recurring failures, and admits only corpus-supported edits. Experiments show that, with Qwen3.5-9B, SIG-R1 improves macro-F1 by 6.4 points over Graph-R1, including a 12.0-point gain on HotpotQA. Three-seed ablations show higher mean F1 with graph evolution and dataset-dependent additional gains from process rewards. These results support coupling retrieval-policy learning with evidence-constrained environment adaptation. Our code is available at https://anonymous.4open.science/status/SIG-R1-DC07.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Multi-hop Question Answering,Retrieval-Augmented Generation,Hypergraph,Reinforcement Learning,Graph Evolution,Failure-driven Adaptation,Evidence Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3dNpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SIG-R1, a framework that jointly optimizes retrieval policies and graph-based retrieval environments for multi-hop question answering. It introduces step-level process rewards and failure-driven hypergraph evolution to improve evidence retrieval. Experiments on three QA benchmarks demonstrate improvements over Graph-R1, with ablation studies examining the contributions of the two components.

Critical strengths
  1. Clear presentation. The paper is generally well-organized and easy to follow. The figures and tables are clearly designed and effectively illustrate the motivation, proposed framework, and experimental results.
  2. Evaluation across multiple backbones. The paper evaluates the proposed method on multiple backbone models, including Qwen3.5-2B, Qwen3.5-4B, Qwen3.5-9B, and Llama-3.1-8B.
Critical weaknesses
  1. Insufficient baseline comparisons and benchmark coverage. Table 1 includes only three retrieval-based RL baselines: Search-R1, R1-Searcher, and Graph-R1. This limited comparison is insufficient to convincingly establish the effectiveness of the proposed method. Moreover, Graph-R1 reports results on six benchmarks (2WikiMultiHopQA, HotpotQA, MuSiQue, NQ, PopQA, and TriviaQA), while Table 1 only includes three of them.
  2. Insufficient efficiency analysis. The proposed method introduces periodic failure diagnosis, graph evolution, and retrieval index updates, which may incur substantial computational overhead. However, the paper lacks a detailed analysis of training time, computational cost, and retrieval latency compared to Graph-R1.
Questions for authors
  1. Can SIG-R1 be extended to more challenging open-web search benchmarks, such as BrowseComp and WideSearch? How would the proposed graph evolution mechanism scale to large, dynamically changing retrieval environments?
7 comments
AuthorsOpenReview ↗

Reply to Reviewer 3dNp

Thank you for the constructive suggestions. We added two recent RL controls, the three missing in-domain benchmarks, a BrowseComp-Plus-Lite fixed-web extension, a Bamboogle transfer diagnostic, and detailed efficiency measurements. We respond to each weakness and the open-web question separately.

Evaluation scope. Our controlled comparisons use Qwen3.5-4B, where SIG-R1 already improves over Graph-R1 on all three datasets. The submitted 9B results show a larger margin (+6.4 macro F1 and +12.0 HotpotQA), consistent with favorable scaling. Due to time and compute constraints, we could not complete the full 9B controlled suite during the rebuttal; we will add it in the camera-ready version.

W1. Insufficient baseline comparisons and benchmark coverage

The submitted Table 1 already includes four RL methods: R1, Search-R1, R1-Searcher, and Graph-R1 . MuSiQue, PopQA, and TriviaQA were also evaluated as zero-shot targets with fixed graphs in §4.6/Table 6. To better address your concern about stronger baselines and complete coverage, we added two recent RL controls, HyperGraphPro and PathRouter, and in-domain training results on the three missing datasets. This lets us test whether SIG-R1's advantage persists under stronger baselines and across all six datasets.

Table N1. Controlled RL comparison; Qwen3.5-4B, 40 updates, 128 test questions per dataset; per-dataset test F1 (%).

MethodNQ F12Wiki F1HotpotQA F1
Graph-R145.9268.3564.85
HyperGraphPro47.4465.3366.36
PathRouter48.5070.0254.64
SIG-R150.0571.9966.10

Table N2. Per-dataset F1 contrasts.

ContrastNQ Δ F12Wiki Δ F1HotpotQA Δ F1
SIG-R1 − Graph-R1+4.13+3.64+1.25
SIG-R1 − HyperGraphPro+2.61+6.66−0.26
SIG-R1 − PathRouter+1.55+1.97+11.46

SIG-R1 improves over Graph-R1 on all three datasets (+4.13 NQ, +3.64 2Wiki, +1.25 HotpotQA), and is best on NQ and 2Wiki; HyperGraphPro is 0.26 F1 higher on HotpotQA. Against PathRouter, SIG-R1 is higher on all three datasets. These are controlled comparisons under identical model, data, retrieval, and update budgets.

We next added the three missing in-domain training evaluations, using one final protocol: Qwen3.5-4B, update 40, 6,144 response tokens, and identical answer-closing rules.

Table N3. Full six-dataset coverage: newly added in-domain results.

DatasetGraph-R1 F1SIG-R1 F1ΔF1ΔEM
MuSiQue49.0149.40+0.39+2.34
PopQA45.7348.65+2.92+3.13
TriviaQA69.0370.62+1.58+0.78
Mean54.5956.22+1.63+2.08

PopQA and TriviaQA show clear gains; MuSiQue is essentially neutral on F1. The mean improvement is +1.63 F1 and +2.08 EM. Together with NQ/2Wiki/HotpotQA, this provides the requested six-dataset coverage; we present it as additional coverage rather than uniform superiority.

W2. Insufficient training-cost and retrieval-efficiency analysis

To better address your concern about efficiency, we measured matched training on two H200 GPUs with Qwen3.5-4B: 40 updates for the in-domain runs and 10 updates for BrowseComp-Plus-Lite.

Table N4. Training wall-clock.

DatasetGraph-R1SIG-R1
MuSiQue5.54 h6.01 h
PopQA4.98 h5.12 h
TriviaQA5.05 h5.20 h
HotpotQA2.20 h2.87 h
BrowseComp-Plus-Lite, 10 updates2.88 h3.22 h

Table N5. Graph growth and inference-side cost.

MeasurementResult
Probe tokens in the accounted HotpotQA run4.86M
HotpotQA graph growth100/215,455 hyperedges (0.046%)
Test-time graph evolutionNone
Mean same-query CPU retrieval latency, initial → evolved graph0.136 s → 0.121 s

Table N4 reports the measured training wall-clock; Table N5 reports graph growth and inference-side cost. Evolution is entirely training-side: there is no test-time graph evolution, and inference uses the fixed graph. The retrieval latency is a retrieval-only diagnostic, not an end-to-end generation-speedup claim.

Q1. Can SIG-R1 extend to BrowseComp/WideSearch and large dynamic retrieval environments?

Yes. SIG-R1 is corpus-agnostic: it requires only retrievable evidence-bearing units, persistent training failures, and provenance-checked source text. The same failure-attribution/edit loop and process-reward interface apply directly to open-web and multi-row environments.

End-to-end fixed-web extension. To better address your question about open-web extension, we ran BrowseComp-Plus-Lite: 128 training, 64 development, and the unchanged 128-question test split; 5,000 fixed web documents; Qwen3.5-4B; and 10 updates for both methods. The graph contains deterministic title, metadata, and body entities with cross-document entity sharing; only training evidence is eligible for edits. Diagnosis runs every two updates with the two-round persistence gate, and both arms use identical data, initialization, graph, and optimization budget. At test time the graph is fixed.

Table N6. End-to-end fixed-web extension on BrowseComp-Plus-Lite; Qwen3.5-4B, 128 test questions.

MeasurementGraph-R1SIG-R1
Test F117.9121.41
Test EM15.6218.75
Training wall-clock2.88 h3.22 h

The F1 difference is +3.50; EM increases by 3.13 percentage points. SIG-R1 committed 18 corpus-grounded H1 edits across four productive rounds, with the first edits at update 4; 244 proposals failed the retrievability gate.

Fixed-corpus policy transfer. To isolate policy transfer without target-corpus adaptation, we also evaluated Bamboogle, a web-search-focused benchmark of 125 two-hop questions. Both source-trained policies are evaluated zero-shot on the same fixed initial HotpotQA graph, with no Bamboogle training or graph edits.

Table N7. Zero-shot transfer to Bamboogle; Qwen3.5-4B, 125 questions.

MethodF1EMFormat
Graph-R120.8516.0099.60
SIG-R127.7521.60100.00
SIG-R1 − Graph-R1+6.89 F1+5.60 EM+0.40 pp

The paired-bootstrap 95% CI for the F1 improvement is [+1.00, +12.97]. This shows that the learned policy transfers to new web-search questions without carrying over source-graph edits. Because the graph is fixed, this is a question-level transfer test rather than an open-web evaluation; the graph-evolution mechanism itself is edit-driven and can operate on web-built, dynamically updated graphs.

For an evolution round with mrm_r accepted edits, its work can be expressed as

m_r\left(C_{\mathrm{quote}}+C_{\mathrm{ownership}}+C_{\mathrm{embedding}}+C_{\mathrm{index\ append}}\right),$$ rather than regenerating the full corpus graph. New hyperedge vectors are appended and edited entity rows are re-encoded. The measured HotpotQA run illustrates this sparsity: **only 0.046% hyperedge growth**. | Environment | Status | Additional requirement | |:--|:--|:--| | BrowseComp-Plus-Lite | **Evaluated in Table N6** | — | | WideSearch multi-row completion | Directly compatible; not yet evaluated | Structured table-output actions, item-level rewards, and row-level evaluation | | Dynamically changing web corpus | Mechanism supports sparse gated updates | Timestamped provenance, freshness revalidation, deletion propagation, stale-edit rollback | **The key scaling point is that adaptation is edit-driven and sparse, not corpus-regeneration-driven.** For genuinely live web deployment, the freshness and rollback mechanisms above are necessary systems extensions; WideSearch additionally requires its structured output/reward interface.
Reviewer_ZYczOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Building on Graph-R1, this paper proposes SIG-R1 to co-evolve a retrieval policy and a hypergraph index during RL training for multi-hop QA. It adds step-level retrieval rewards and a periodic graph repair step that uses a 512-question training probe to insert hyperedges and entity texts built from gold supporting sentences. On NQ, 2WikiMultiHopQA, and HotpotQA, SIG-R1 improves average F1 over Graph-R1 across four models.

Critical strengths
  1. Practical problem setup and careful edit design. Separating search policy errors from missing links in the graph is well motivated, every edit must pass quote and ownership gates before commit (L686-689), and the zero-shot transfer to fixed target graphs tests the policy without carrying source edits (Table 6).

  2. Consistent F1 gains with an informative ablation. Average F1 improves for all four policy models, and the largest gain at 9B on HotpotQA comes with a similar EM gain (59.64 to 71.66 F1, 52.34 to 64.84 EM, Table 1). With the outcome reward unchanged, graph evolution alone adds 3.92 F1 on NQ and 1.43 on 2Wiki (Table 2, L376).

Critical weaknesses
  1. The gains may come from adding training answers to the graph rather than from the evolution process. As I read Section 3.4, each edit copies the annotated gold sentence of a training question the policy keeps failing on (L080, L254), and the largest batch already arrives at update 10 of 40 (L407-409). The paper needs a control that adds the same sentences once before training, otherwise the failure-driven schedule is not shown to matter.

  2. The evidence does not show that the edits fix missing links in the graph. In the first case study, the answer sentence was already in the top-1 retrieved chunk and was only cut off by the response-length limit (L959-967), and both methods retrieve evidence for the same 70.1% of test answers (Table 3). I suspect that longer excerpts or simple reranking would recover much of this gain, and the paper should test this before crediting graph repair.

Questions for authors
  1. What happens if the same training gold sentences are added to the graph once before training and plain Graph-R1 is trained on it? Please also report how many test questions actually retrieve committed edits.
  2. What does the final SIG-R1 policy score on the initial graph, and the Graph-R1 policy on the evolved graph? This cross evaluation, together with a component ablation on HotpotQA where the gain is largest, would separate a better policy from a better graph.
  3. Why does the case study run dev-split questions (L916) on the evolved graph, whose committed edges already contain their answers (Tables 10 and 11)?
8 comments
AuthorsOpenReview ↗

Reply to Reviewer ZYcz

Thank you for identifying the central confound between policy learning, answer-bearing edits, and retrieval visibility. Your questions prompted three-seed controls that substantially sharpen our claims. We first respond to each weakness and then to each question.

W1.The gains may come from adding training answers rather than evolution

The policy-by-graph cross evaluation directly separates the trained policy from the evolved graph. Section 3.4 and Appendix B.7 state that the edits use training-split gold evidence; the causal question is whether test performance depends on retrieving those edits at test time. To test this, we added the requested cross evaluation and a committed-edit exposure audit.

Table Z1. HotpotQA policy × graph; Qwen3.5-4B, three seeds, 40 updates; test F1.

Trained policyInitial graphEvolved graphΔ graph (evolved − initial)
Graph-R161.4361.34−0.09
SIG-R165.6665.92+0.26

Table Z1b. Policy contrasts.

ContrastΔ F195% CI
SIG-R1 − Graph-R1, initial graph+4.23[+0.77, +7.75]
SIG-R1 − Graph-R1, evolved graph+4.58—

Table Z2. Direct exposure to committed edits on held-out tests.

DatasetTest questions retrieving a committed editCorrect among exposed cases
2Wiki0/1280
HotpotQA1/1281
NQ2/1280

The policy contrast is +4.23 F1 on the initial graph (95% CI [+0.77, +7.75]) and +4.58 on the evolved graph, whereas the graph contrast is only −0.09 for Graph-R1 and +0.26 for SIG-R1. This is the separation the reviewer requested: at test time, the trained policy carries the gain, while the evolved graph alone does not. Direct test-time lookup of inserted training answers therefore cannot explain the aggregate gain.

The graph's value is training-time, not test-time. A policy trained on the initial graph cannot exploit the evolved graph (−0.09), but training alongside evolution adds +3.65 F1 (evolution-only 65.08 vs. Graph-R1 61.43, Table Z6 in Q2). This is the co-evolution effect: the edits must be experienced during training to change retrieval behavior. The exposure audit (0–2 of 128 test questions) confirms that the gain is not test-time lookup; it is the policy learning to use the edited environment.

The evidence therefore supports a policy effect rather than answer lookup. We will use the precise term train-supervised, failure-driven environment adaptation.

AuthorsOpenReview ↗

Reply to Reviewer ZYcz

W2.The edits may not fix missing links; longer excerpts or reranking may explain the gain

Appendix C.1 records that the answer is at character 2,463, outside the visible prefix, and that an existing answer-bearing hyperedge does not enter the demonstrated top-10 retrieval path. We agree that Case 1 should not be described as a completely absent answer-bearing edge, and we will replace “the edge is absent” with retrieval-visibility/path failure.

The submitted evidence also distinguishes evidence access from evidence use:

Table Z3. Submitted 4B evidence-conditioned profile (Table 3).

MetricGraph-R1SIG-R1Difference
Answers with evidence present70.1%70.1%0.0 pp
Correct and evidence present47.4%53.7%+6.3 pp
Incorrect and evidence present22.7%16.4%−6.3 pp

To directly address the visibility explanation, we increased the tool-observation budget without changing the policy or graph:

Table Z4. Longer-observation control on HotpotQA, seed 1.

Method and observation budgetTest F1
Graph-R1, 1,024 tokens55.26
Graph-R1, 4,096 tokens54.66
SIG-R1, original 1,024-token budget67.39

A fourfold observation budget does not recover the gain in this single-seed control. The benefit is therefore not explained simply by truncating longer raw chunks; it also reflects how the learned policy uses available evidence. The observation-budget control shows that exposing more raw text does not recover the gain, but it does not rule out a learned reranker.

We agree with the reviewer's assessment. A dedicated reranker control and a three-seed version of the observation-budget study are not yet included in this response; we will add both, together with the full four-way evidence profile and the corrected case-study wording, in the camera-ready version.

Q1.What if the same gold sentences are added before training? How many test questions retrieve the edits?

We ran exactly this control, together with a text-only arm containing identical evidence. Table Z2 above reports the requested exposure counts.

Table Z5. Same-evidence controls on 2Wiki; outcome-reward-only, three seeds, 40 updates; test F1 mean ± SD.

SettingWhat is insertedWhenTest F1
Online evolutionAccepted graph editsDuring training66.37 ± 4.89
Static-full (oracle timing)Same graph editsOnce before training68.99 ± 2.10
Text-onlySame evidence, plain textOnce before training65.06 ± 2.81

The table separates two questions: structure (static-full vs. text-only) and timing (online vs. static-full).

Table Z5b. Key contrasts.

ContrastΔ F195% CI
Static-full − Text-only (structure)+3.93[+0.36, +7.53]
Online − Static-full (timing)−2.62[−6.16, +0.95]

Static-full is an oracle-timing ablation, not a competing method: it inserts the final accepted edit set, which is produced by the online failure-discovery loop. The final edit set is not available in advance: the online loop performs the failure attribution that determines which persistent failures are graph defects and therefore which edits are worth committing. Static-full reuses that output; it does not show that the online process is unnecessary. Because static-full has a fixed edit set, its lower variance is expected; the online arm must discover edits during training.

What the comparison isolates is the timing of insertion. The structure contrast is positive, so structured graph insertion outperforms plain-text insertion of the same evidence. The online loop's contribution is failure attribution and edit selection, not the insertion schedule. In the main Table 2 protocol, which includes process rewards, full SIG-R1 reaches 70.48 on 2Wiki; the three arms above are outcome-only controls used to isolate timing.

Q2. What are the SIG-R1 policy on the initial graph and Graph-R1 policy on the evolved graph? Please ablate HotpotQA.

Table Z1 gives the requested cross evaluation: 65.66 for SIG-R1 on the initial graph and 61.34 for Graph-R1 on the evolved graph. We also added the requested three-seed HotpotQA component ablation:

Table Z6. Component ablation across three datasets; Qwen3.5-4B; test F1 (three-seed mean).

DatasetGraph-R1Process onlyEvolution onlyFull SIG-R1
NQ46.8548.4050.7751.05
2Wiki67.0667.3568.4970.48
HotpotQA61.4367.3565.0865.92

The components are dataset-dependent. On NQ and 2Wiki, the full model is best (51.05 and 70.48) and evolution is the larger standalone contributor. On HotpotQA, full SIG-R1 improves over Graph-R1 by +4.49 (65.92 vs. 61.43); within the component arms, process-only is strongest (67.35), consistent with the small edit budget on HotpotQA (100 edits, 0.046% growth). The full model is therefore the best configuration on two of three datasets and improves over Graph-R1 on all three.

We report this dependence explicitly rather than claim uniform component synergy. Together with Tables Z1–Z5, the ablation supports two concrete sources of value: a policy advantage that persists on the initial graph and a representation advantage for structured evidence insertion.

Reviewer_j5xROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SIG-R1, a framework that jointly optimizes an RL-based retrieval policy and environment for multi-hop question answering. The method introduces two main components, process-aware rewards that provide feedback for retrieval actions, and failure-driven graph evolution that identifies failures and repairs the graph. Experiments on NQ, 2WikiMultiHopQA, and HotpotQA demonstrate improvements over Graph-R.

Critical strengths

The paper proposes to jointly optimize the retrieval policy as well as the enviroment, which helps resolve poorly represented evidence problem.

The method consistently improves average F1 over Graph-R1 across four backbone models

Critical weaknesses
  1. The paper should be desk rejected the first place since it is format is not using correct ICLR template.
  2. I highly doubt the paper is AI generated, at least highly rely on AI for paper writing. The paper lists nothing but 11 observations in the experimental sections. There are multiple explanations of what each analysis does and does not establish. The paper uses many structural bullet point listings.
  3. Process-level reward shaping, failure attribution, and adaptive graph construction, are not original contribution but existing research. The paper does not sufficiently distinguish its technical contributions from previous approaches.
  4. Graph evolution uses ground-truth answers and supporting evidence from training examples, so it is more similar to supervised instead of the claimed 'self-evolving'.
  5. The paper does not provide analysis of training cost and inference efficiency compared with fixed-graph baselines.
Questions for authors

see weakness

6 comments
AuthorsOpenReview ↗

Reply to Reviewer j5xR

Thank you for the direct criticism. We respond separately to formatting and presentation, technical novelty, supervision, and efficiency. We also conducted additional controls that allow us to narrow the claims to what the evidence establishes.

W1. ICLR template compliance

The manuscript used the ICLR template; we additionally compressed float, section, and paragraph spacing to improve presentation. We will restore the default ICLR 2027 template settings and correct the conference-year header. We respectfully ask that this formatting issue be assessed separately from the scientific content: it does not affect the method, experiments, or claims.

W2. AI-assisted writing and the eleven Observations

The submitted experiments are organized around five research questions ; the eleven Observations summarize results from Tables 1–3 and 6 and Figure 5. Numbered observations and structured lists are conventional in published papers. For example, AlphaEdit (ICLR 2025 Best Paper) uses Obs 1–7, and LLMs in Learning on Graphs (SIGKDD Explorations) uses Observation 1–19. The AI Use Statement discloses grammar, phrasing, and clarity assistance only; the authors remain responsible for the ideas, code, experiments, results, and claims.

The current presentation is repetitive, and we will revise it: consolidate the eleven Observations into the five RQ narratives, keep one explicit limitation per analysis, move general caveats to the discussion, and replace structural lists with table-led prose. Writing style alone is not evidence of AI authorship, and we ask that the technical content be evaluated on the revised manuscript.

W3. Process rewards, failure attribution, and adaptive graphs are existing ideas

SIG-R1's contribution is the closed loop that couples policy optimization with environment adaptation. The individual primitives have precedents; we do not claim to have invented them. Graph-R1 optimizes a retrieval policy on a fixed graph. TIPS uses teacher-likelihood step rewards; HyperGraphPro uses target-hyperedge coverage; PathRouter uses answer correctness combined with evidence-path overlap. EvoGraph-R1 edits multimodal hypergraphs; SAGE and Agentic-KGR study graph memory or KG construction.

Prior settingWhat it optimizesRemaining issue addressed by SIG-R1
Graph-R1: outcome RL on a fixed graphRetrieval policy onlyA policy cannot learn a path that is absent or inaccessible in the environment
TIPS / HyperGraphPro / PathRouter: process-oriented supervisionRetrieval-step creditBetter actions alone do not modify defective evidence representation
SAGE / EvoGraph-R1 / Agentic-KGR: memory or graph adaptationEnvironment or memory in distinct settingsAn edit does not by itself teach the policy to exploit the changed environment

Every five policy updates, SIG-R1 probes 512 training questions; only graph-attributed persistent failures trigger deterministic H1 hyperedge additions or H2 entity-text edits. Each edit passes verbatim-quote, ownership, and retrievability gates. Accepted edits update the graph and index, and the policy continues training in the changed environment; the graph remains fixed at test time (§§3.3–3.5). The contribution is therefore persistent-failure routing + constrained editing + subsequent policy learning as one closed training loop.

Two new controls support this interpretation:

Additional experimentControlled comparisonResultSupported interpretation
HotpotQA policy × graph, three seedsSIG-R1 vs. Graph-R1 policy on the same initial graph+4.23 F1; 95% CI [+0.77,+7.75]The learned policy retains an advantage without carrying over evolved-graph edits
2Wiki identical-evidence insertion, three seedsStructured graph insertion vs. text-only insertion+3.93 F1; 95% CI [+0.36,+7.53]Structured representation adds value beyond placing the same text in the index

These controls isolate the policy and representation channels; the closed loop is what produces the edit set and teaches the policy to use it. That interaction is the contribution we will emphasize in the revision.

W4. Gold training evidence makes this supervised rather than “self-evolving”

The paper's title, abstract, method name (SIG-R1), and central claim use self-improving, not self-evolving. “Self-evolution” appears as the label of §3.4 and Algorithm 1 for the graph-update procedure; it is not a claim of annotation-free or unsupervised learning. In our context, “self-improving” means that both the policy and its retrieval environment improve during training, which is fully compatible with the disclosed use of training-split gold evidence.

The supervision boundary is

\qquad \mathcal D_{\mathrm{dev/test}}^{\mathrm{questions,\ answers,\ evidence}} \cap \text{probe/edit inputs}=\varnothing .$$ The reviewer's conclusion that the method is “more similar to supervised” therefore evaluates a claim the paper does not make. We do not claim that the graph evolves without supervision; we claim that the policy and environment improve together during training, within the boundary above. To remove ambiguity, we will rename the §3.4/Algorithm 1 label to **train-supervised, failure-driven environment adaptation**, while retaining “self-improving” in the title and abstract. **W5. Missing training-cost and inference-efficiency analysis** **Table J1. Measured training wall-clock (two H200 GPUs, Qwen3.5-4B, 40 updates).** | Dataset | Graph-R1 | SIG-R1 | |:--|--:|--:| | MuSiQue | 5.54 h | 6.01 h | | PopQA | 4.98 h | 5.12 h | | TriviaQA | 5.05 h | 5.20 h | | HotpotQA | 2.20 h | 2.87 h | **Table J2. Graph and inference-side measurements.** | Measurement | Result | |:--|--:| | Probe tokens in the accounted HotpotQA run | 4.86M | | HotpotQA graph growth | 100/215,455 hyperedges (0.046%) | | Test-time graph evolution | None | | Mean same-query CPU retrieval latency, initial → evolved graph | 0.136 s → 0.121 s | Table J1 reports the measured wall-clock; Table J2 reports graph growth and inference-side measurements. Evolution is entirely training-side: there is no test-time graph evolution, and inference uses the fixed graph. The retrieval latency is a retrieval-only diagnostic, not an end-to-end generation-speedup claim.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.