acceptodds
Under review as a conference paper at ICLR 2027

Graph-Guided Diffusion Language Models for Efficient Multi-Hop Reasoning

Abstract

Graph Retrieval-Augmented Generation (GraphRAG) has remarkably enhanced large language models on complex reasoning by leveraging structured entity topologies. However, existing frameworks heavily rely on standard autoregressive language models where the nature of inherent sequential generation severely hinders overall inference efficiency. Inspired by Diffusion Language Models (DLMs) that offer massive parallelism via continuous refine-in-parallel decoding, we aim to accelerate GraphRAG in the discrete space. However, it remains non-trivial for two challenges. First, partially denoised drafts are highly dynamic and uncertain, making dynamic graph grounding non-trivial. Second, raw denoising states are inherently noisy and unstable, making synchronous graph retrieval and multi-hop aggregation computationally prohibitive. To this end, we present Ladder, a novel framework that bridges diffusion language modeling with GraphRAG through graph-guided parallel decoding. Specifically, we propose an event-driven self-clocking retrieval, inspired by our key insight that of target entities emerge early in the partially denoised state, leading final commitment by an average of – steps. This mechanism dynamically triggers graph retrieval only when the set of graph-linkable entities expands, yielding an asynchronous self-clocking policy that bypasses learned gates or heuristic thresholds. An incomplete-query graph propagation module is designed to process the newly emerging entity queries using a specialized graph foundation model, continuously aggregating multi-hop evidence to sharpen parallel predictions and accelerate overall decoding convergence. Extensive experiments on three challenging multi-hop QA benchmarks show that Ladder raises average exact match from to while achieving a latency reduction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Graph Retrieval-Augmented Generation; Diffusion Language Model

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_k5HsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LADDER, which integrates graph retrieval into the denoising process of a diffusion language model for multi-hop QA. Motivated by the observation that about 88% of final entity spans appear in the intermediate draft of Dream-7B before they are committed, with mean lead times of 5.7 to 9.6 steps, LADDER triggers retrieval only when a new graph entity emerges in the draft, and fine-tunes GFM-RAG on these entity events to retrieve both missing and previously found evidence. On HotpotQA, 2WikiMultihopQA and MuSiQue, with MuSiQue held out from retriever training, it reports an average EM increase from 39.6% to 45.2% and a 4.1× latency reduction over a multi-round Dream-7B baseline, and also compares against six autoregressive retrieval pipelines. Ablations cover refresh schedules, training objectives and the effect of evidence quality on denoising steps.

Critical strengths
  1. Diffusion-specific retrieval timing. Entities in the final response usually appear in the intermediate draft before they are committed, and the paper turns this into a retrieval trigger without a learned gate.

  2. Effective refresh policy. Under a fixed backbone and retriever, entity-triggered refresh outperforms retrieving once, fixed-interval and confidence-based schedules at similar latency.

  3. Retrieval as a source of decoding speed. Better evidence in the prompt reduces the number of denoising steps, a property of diffusion decoding that standard autoregressive decoding lacks.

  4. Favorable quality-latency trade-off. LADDER obtains the best EM on all three benchmarks with lower latency than all iterative baselines.

Critical weaknesses
  1. Source of the speedup. The 4.1× gain is measured against a Dream-7B baseline that runs multiple IRCoT rounds, while retrieving once within a single trajectory is already slightly faster than LADDER. The efficiency gain therefore comes mainly from single-pass decoding rather than from entity-triggered retrieval.

  2. Source of the accuracy gains. The main comparison changes the decoding loop, the trigger and the retriever training at once, and no experiment isolates the effect of retriever fine-tuning. On held-out MuSiQue, LADDER is roughly on par with the strongest baseline, and no significance tests are reported.

  3. Evidence for the acceleration claim. Random passages already remove a large share of denoising steps, which may partly reflect the generator's fine-tuning format rather than evidence quality. The step-count experiment also uses fixed evidence, so it does not show that refreshing retrieval during denoising speeds up convergence.

Questions for authors
  1. Could the authors report LADDER with the pretrained GFM-RAG retriever, and the IRCoT baseline with the fine-tuned retriever? This would separate the contribution of retriever fine-tuning from that of the decoding scheme.

  2. Could the authors report the number of denoising steps for each refresh schedule, and for a generator not fine-tuned exclusively with documents in the prompt? This would show whether entity-triggered refresh itself speeds up convergence, and why random passages already reduce the step count.

  3. Could the authors add an autoregressive baseline with a single retrieval round, and describe the inference setup used for the autoregressive models? This would clarify how LADDER's latency compares with an efficient autoregressive pipeline.

  4. The paper could discuss retrieval-based speculative decoding, which also uses external text to accelerate autoregressive generation, to better position its efficiency argument.

4 comments
AuthorsOpenReview ↗

Response (Part 1/2)

Thank you for the questions about efficiency attribution, retriever adaptation, and convergence. Manuscript table numbers refer to the submitted PDF. Times are seconds per question, and averages give equal weight to HotpotQA, 2WikiMultihopQA, and MuSiQue.

[W1,Q3] Source of the speedup and a single-retrieval AR baseline

The 4.1× speedup in manuscript Table 2 compares multi-round Dream-7B IRCoT with pretrained GFM-RAG (11.30 s/question) against LADDER (2.79 s/question). Both systems commit tokens in parallel within each denoising step. With the pretrained retriever held fixed, switching to a single trajectory with entity-triggered refresh reduces latency from 11.30 to 2.93 s/question. Following your suggestion, we have named the 4.1× comparator in the abstract and TL;DR.

Within one trajectory and with the event-tuned retriever fixed, manuscript Table 3 provides the schedule comparison. LADDER improves EM from 37.9 to 45.2 over retrieve-once for an additional 0.24 s/question (2.55 to 2.79). Compared with per-step refresh, it reduces latency by 36.3% (4.38 to 2.79 s/question) and improves EM from 44.5 to 45.2.

Table K1. Single-retrieval AR and DLM comparison. LADDER results are from manuscript Tables 1–2. The two HippoRAG 2 configurations use the same top-7 passages retrieved from the question alone.

ConfigurationEM ↑F1 ↑R@7 ↑Time ↓
Qwen2.5-7B + HippoRAG 2, retrieve once37.446.059.81.64
Dream-7B + HippoRAG 2, retrieve once38.947.959.82.68
LADDER45.254.667.12.79

The single-retrieval Qwen2.5-7B baseline takes 1.64 s/question versus 2.79 for LADDER. LADDER scores 7.8 points higher in EM and 8.6 points higher in F1. Appendix B describes the matched generator supervision. Manuscript timings use a single NVIDIA H20 and a warmed graph index. The two retrieve-once runs use batch size 1, with identical warm-up and timing boundaries covering online retrieval and generation.

[W2,Q1] Retriever tuning and MuSiQue

We agree that a checkpoint swap is the direct control. Manuscript Table 4 compares training objectives, with 41.9 EM for retrieve-only training and 45.2 for the full objective. Both of those checkpoints are tuned. Table K2 instead crosses execution mode with pretrained versus event-tuned GFM-RAG. Within each mode, the generator, controller, query and seed construction rules, evidence budget, and evaluator are fixed.

Table K2. Execution mode and GFM-RAG checkpoint on Dream-7B. The pretrained IRCoT and event-tuned entity-refresh results are from manuscript Tables 1–2.

ModeGFM-RAG checkpointEM ↑F1 ↑R@7 ↑Time ↓
IRCoTPretrained39.647.458.111.30
IRCoTEvent-tuned40.948.960.711.19
Entity refreshPretrained42.451.863.82.93
Entity refreshEvent-tuned45.254.667.12.79

With the execution mode fixed, event-conditioned tuning improves EM by 1.3 points under IRCoT and 2.8 points under entity refresh. With the checkpoint fixed, switching to single-trajectory entity refresh improves EM by 2.8 points with pretrained weights and 4.3 points with tuned weights. Within entity refresh, tuning also raises R@7 from 63.8 to 67.1 and reduces latency from 2.93 to 2.79 s/question.

MuSiQue was held out from retriever optimization and model selection. In manuscript Table 2, LADDER takes 3.95 s/question versus 14.24 for Youtu-GraphRAG on Qwen2.5-7B. The accuracy point estimates differ by +0.6 EM points (18.8 versus 18.2) and −0.4 F1 points (30.7 versus 31.1) for LADDER.

AuthorsOpenReview ↗

Response (Part 2/2)

[W3,Q2] Dynamic convergence and the generator's training format

Convergence under dynamic refresh. Manuscript Table 11 measures convergence with the evidence fixed at the start. Table K3 reports denoising steps and GNN forwards under four refresh policies within a single trajectory. We hold the Dream-7B and event-tuned GFM-RAG checkpoints, confidence threshold (τc = 0.85), and output budget fixed.

Table K3. Denoising, retrieval, and accuracy by refresh policy. EM and total time are macro-averages from manuscript Table 8. LADDER’s GNN-forward count is the macro-average of manuscript Table 6.

ScheduleEM ↑Mean denoising stepsGNN forwards/questions/question ↓
Retrieve once37.924.51.002.55
Every step44.521.921.374.38
Confidence drop42.922.82.972.76
LADDER45.221.83.202.79

LADDER uses 2.7 fewer denoising steps than retrieve-once (21.8 versus 24.5). It uses almost the same number of steps as per-step refresh (21.8 versus 21.9), with 3.20 versus 21.37 GNN forwards per question and 2.79 versus 4.38 s/question. Compared with confidence-drop refresh, LADDER gains 2.3 EM points at an additional 0.03 s/question.

Generator training format. The no-evidence result in manuscript Table 11 may reflect both the generator’s training format and Appendix H’s instruction to answer only from supplied facts. On 2Wiki, we compare the current checkpoint with one trained using mixed-context SFT. Both use the same base model, training questions, targets, and number of updates. Mixed-context SFT leaves the evidence slot empty in 30% of training examples. We evaluate both on the same questions with a shared prompt that allows answers when the evidence slot is empty. The random-passage and graph-evidence prompts have matched token lengths.

Table K4. Matched training-format control on 2Wiki. Values are mean denoising steps / EM.

EvidenceCurrent checkpoint, matched promptMixed-context SFT, matched prompt
No evidence39.2 / 32.030.8 / 34.2
Random passages24.4 / 32.926.1 / 34.1
Fixed t=0 GFM-RAG passages15.5 / 57.316.3 / 56.8

With the prompt held fixed, mixed-context SFT reduces the no-evidence step count from 39.2 to 30.8 and narrows the gap between no evidence and random passages from 14.8 to 4.7 steps. This suggests that the training format contributes to the slowdown with empty evidence. With the mixed-context checkpoint, random passages yield nearly the same EM as no evidence (34.1 versus 34.2). Replacing them with graph-retrieved passages reduces the mean step count from 26.1 to 16.3 and raises EM from 34.1 to 56.8.

[Q4] Retrieval-based speculative decoding

Following your suggestion, we have added a discussion of REST to the related work in the final version. REST retrieves candidate continuations for verification by an autoregressive model. LADDER retrieves supporting passages that condition subsequent parallel denoising in multi-hop QA. The revised text ties the efficiency claims to the systems and inference settings in each comparison. REST, NAACL 2024

Table K5. End-to-end comparison with REST. Both IRCoT variants pair Qwen2.5-7B with HippoRAG 2. LADDER uses Dream-7B. The IRCoT and LADDER rows use results from manuscript Tables 1–2.

MethodAverage EM ↑Average F1 ↑R@7 ↑s/question ↓
IRCoT41.349.262.611.65
IRCoT + REST41.349.262.69.05
LADDER45.254.667.12.79

With REST, the Qwen2.5-7B IRCoT system takes 9.05 s/question versus 11.65 for standard IRCoT, with the same EM, F1, and R@7. Compared with IRCoT + REST, LADDER achieves a 3.24× end-to-end speedup and scores 3.9 points higher in EM and 5.4 points higher in F1.

Thank you for the authors’ response and the additional experiments. I will maintain my current decision for weak acceptance.

Thank you for your great support and for taking the time to consider our responses and additional experiments.

Reviewer_BLSCOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

The paper proposes LADDER, a graph-guided dynamic RAG framework for diffusion language models. Instead of running graph retrieval at every denoising step, the authors introduce a scheduling mechanism that only refreshes retrieval when new entities emerges. They fine-tune a Graph Foundation Model to make it robust to the noisy drafts that arise at intermediate diffusion states. On three multi-hop QA datasets, LADDER dominates the quality-speed trade-off against prior GraphRAG approaches, achieving a 4x speedup over the strongest baseline.

Critical strengths
  1. The use of graph-guided retrieval is well-motivated and a natural fit for dynamic diffusion retrieval: anchoring retrieval in a structured entity graph might be more robust against partial queries (arising in intermediate diffusion states) than dense retrieval.

  2. The speedup over prior GraphRAG baselines is significant (~4x) and consistent across all three datasets, including the held-out dataset "MuSiQue".

Critical weaknesses
  1. No non-graph baseline is tested. While LADDER dominates prior GraphRAG approaches in the quality-speed tradeoff, it is unclear whether the performance gains stem from the graph-guided retrieval or from the retrieval timing schedule alone.

  2. The empirical analysis showing that key entities emerge very early in the denoising trajectory is convincing. However, the core insight that intermediate denoising states carry retrieval signal originates from earlier dynamic retrieval work (e.g. SARDI), which should be referenced there.

Several issues need to be fixed in the presentation: E_t and Eq. 3 are used in Section 3 but only defined 2 pages later. Additionally, there is a significant amount of citation errors (some examples below):

  • Zhu et al., 2022 (Neural Bellman-Ford Networks): Published at NeurIPS 2021, not 2022.
  • Mavromatis & Karypis, 2024 (GNN-RAG): Published at ACL Findings 2025, not 2024. Both year and venue are wrong.
  • Gutierrez et al., 2025a (HippoRAG): Published at NeurIPS 2024, not 2025. Author name: "Jimenez Gutierrez" -> "Jiménez Gutiérrez".
  • Jnger et al., 2026 (SARDI): Published at ICML 2026. Author name misspelled: "Paul Jnger" should be "Paul Jünger".
  • Lou et al., 2024; Galkin et al., 2024; Gutierrez et al., 2025b; Luo et al., 2025; Nie et al., 2025; Yao et al., 2023: Preprint version cited even though published versions exists at ICML 2024, ICLR 2024, ICML 2025, NeurIPS 2025, NeurIPS 2025, and ICLR 2023.
Questions for authors
  1. How do non-graph based retrieval methods compare against LADDER? Is it possible to isolate the contribution of the graph-structured retrieval from the retrieval timing mechanism?
3 comments

Thank you for the comments on causal attribution and presentation. Times are seconds per question, and averages give equal weight to HotpotQA, 2WikiMultihopQA, and MuSiQue.

[W1,Q1] Graph retrieval and refresh timing

The graph-free E5-base-v2 IRCoT baseline scores 40.7 average EM at 11.06 s/question. Table L1 compares flat and graph retrieval backends under the same entity-triggered policy.

Table L1. Retrieval backends under the same entity-trigger rule. All runs use the same Dream-7B checkpoint, corpus, graph-derived entity matcher, query construction rules, and K = 7.

Retrieval backendAverage EM ↑Average F1 ↑s/question ↓
BM2540.449.02.89
E5-base-v241.450.62.85
Pretrained GFM-RAG42.451.82.93

Pretrained GFM-RAG gains 1.0 EM and 1.2 F1 points on average over E5-base-v2, with an additional 0.08 s/question. Its MuSiQue EM is 17.2 versus 17.5 for E5-base-v2. This comparison measures the effect of changing the entire retrieval backend before event-conditioned tuning. The retrieved passages can change later drafts and trigger events.

Table L2. Refresh policies with the same retriever. All values are from manuscript Table 3. The Dream-7B checkpoint, event-tuned GFM-RAG checkpoint, single-trajectory execution, and K = 7 are fixed.

Refresh policyAverage EM ↑Average R@7 ↑s/question ↓
Random refresh (15%)40.459.72.81
Every 5 steps42.662.92.98
Every step44.565.84.38
Entity-triggered refresh (LADDER)45.267.12.79

With the retriever held fixed, entity-triggered refresh gains 4.8 EM and 7.4 R@7 points over random refresh at similar latency. It gains 2.6 EM points over refreshing every five steps and reduces latency from 2.98 to 2.79 s/question. Compared with per-step refresh, it gains 0.7 EM points and reduces latency by 36.3%. The policies share the same passage budget per call and can make different numbers of calls.

Full LADDER combines graph retrieval with event-conditioned tuning. Compared with E5-base-v2 under the same entity trigger, it gains 3.8 EM points (45.2 versus 41.4) at comparable latency (2.79 versus 2.85 s/question).

[W2] Attribution to SARDI

Following your suggestion, we have cited SARDI where Section 3 first discusses early entity signals and made the attribution explicit in the motivation and efficiency discussion. SARDI establishes both retrieval lookahead from intermediate drafts and the connection between evidence grounding and parallel decoding. Our contribution builds on those observations by using a newly matched graph entity to trigger retrieval and seed graph propagation, together with retriever adaptation to the incomplete queries encountered at those events. In our span-level analysis, about 88% of final graph-linkable entity spans emerge before commitment, with mean lead times of 5.7–9.6 denoising steps across datasets (manuscript Table 5). SARDI, ICML 2026

[W3] Notation and references

Following your constructive suggestions, we have revised Section 3 of the final version to define the proxy draft rₜ, the cumulative set of matched entities Eₜ, and the set of newly matched entities ΔEₜ at first use. Before Eq. (3), we explain that rₜ retains committed tokens and fills masked positions with their current highest-probability predictions.

We have corrected the publication details for NBFNet (NeurIPS 2021), GNN-RAG (Findings of ACL 2025), HippoRAG (NeurIPS 2024), and SARDI (ICML 2026), including the spellings of “Jiménez Gutiérrez” and “Paul Jünger.” We now cite the proceedings versions of SEDD, ULTRA, HippoRAG 2, GFM-RAG, LLaDA, and ReAct.

Thank you for your detailed response and additional experiments, isolating the gains of both the graph-guided retrieval and retrieval-refresh-timing contributions.

I highly suggest adding the new flat-retrieval rows from Table L1 to the paper (+a version with retrieval at every step), which you could then call "SARDI" to have at least one diffusion-language-model dynamic retrieval baseline in the paper.

I think the analysis showing that key entities emerge very early in the denoising trajectory is interesting, and the use of graph-based retrieval to exploit this, while being robust against noisy diffusion states, is convincing. In light of the new analyses, I have raised my score.

Following your suggestion, we have added the flat-retrieval results from Table L1 and a per-step flat-retrieval baseline to the revised paper, with appropriate attribution to SARDI. Thank you for your great support and constructive suggestions. We appreciate your careful consideration of our response and additional analyses.

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

This paper proposes LADDER, a novel framework that bridges diffusion language modeling with GraphRAG to achieve efficient multi-hop reasoning. LADDER leverages parallel refine-in-parallel decoding via Diffusion Language Models. Specifically, it introduces an event-driven self-clocking retrieval policy that dynamically triggers graph propagation when graph-linkable entities emerge in the partially denoised draft. Additionally, it uses an incomplete-query graph propagation module powered by a specialized Graph Foundation Model trained with discovery, retention, and anchor objectives. Experimental results across three multi-hop QA benchmarks show that LADDER significantly improves exact match accuracy while reducing the latency compared to existing baselines.

Critical strengths
  • It couples diffusion language modeling with structured graph retrieval, replacing slow sequential thought-level retrieval rounds with sparse, parallel-guided updates.

  • It demonstrates performance gains while reducing end-to-end latency

Critical weaknesses

1. Insufficient evidence that CA-guided evolution itself provides the claimed benefit.

The main experiments compare CARVE against raw Wikipedia and only two randomly selected reformulation strategies. However, CARVE selects the best candidate through a multi-round search, while the random baselines are not similarly selected for quality. This selection asymmetry makes it difficult to attribute the improvements specifically to CA-guided evolution rather than simply finding a stronger reformulation. Moreover, the paper lacks comparisons with competitive existing reformulation methods or a best-of-budget unguided search. Consequently, the central methodological claim remains insufficiently supported.

2. The improvements may primarily arise from factual redundancy rather than the proposed CA-based explanation.

The final selected strategy, Dual-Track Factual Redundancy, explicitly repeats each fact using different syntactic realizations. Such repetition and paraphrasing may independently improve factual learning, regardless of CA. The paper does not include controlled comparisons against simple factual repetition or paraphrasing strategies, making it unclear whether the improvements reflect the proposed CA-based principle or established benefits of repeated and diverse factual exposure. This alternative explanation directly weakens the interpretation of the main results.

Questions for authors
  1. Does CA-guided evolution outperform a matched-budget unguided search? In particular, how does CARVE compare against the lowest-CA, fidelity-valid strategy selected from an equally sized pool of candidates generated without evolutionary feedback? This comparison would help establish whether iterative CA-guided evolution contributes beyond candidate selection.

  2. Can the authors disentangle the effects of CA from factual redundancy and paraphrasing? How does the selected Dual-Track strategy compare against simpler repetition or paraphrasing baselines under matched token budgets and comparable factual exposure? Reporting both CA and downstream factual recall would help determine whether the proposed explanation accounts for the observed improvements.

1 comment

Thank you for your review and for summarizing LADDER. We noticed that the “Critical Weaknesses” and “Questions for Authors” refer to CARVE, CA-guided Wikipedia reformulation, and Dual-Track Factual Redundancy. Our submission studies entity-triggered graph retrieval during diffusion decoding for multi-hop question answering. Could you please check whether those two sections were intended for our submission and update them with any concerns or questions you had about LADDER? As part of that check, we would also be grateful if you could confirm whether the current overall rating of 2 reflects your assessment of our paper. We would appreciate the opportunity to respond to your feedback during the discussion.

Reviewer_BhY3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper moves graph retrieval into the denoising trajectory of a diffusion language model, removing retrieval from the autoregressive critical path. LADDER triggers a graph refresh when new graph-linkable entities appear in the partially denoised draft (no learned gate, no threshold) and fine-tunes GFM-RAG on the resulting event traces. On three multi-hop QA benchmarks with Dream-7B it reports 45.2 average EM versus 39.6 for a diffusion-side IRCoT bridge, at 2.79 s versus 11.30 s per question.

Critical strengths

Graph retrieval inside DLM denoising is unexplored; RAG for DLMs has been studied only with flat retrieval, so the combination is new. Six established GraphRAG baselines are run under one controller, with matched fine-tuning on both generators. Using an emerging entity as both refresh trigger and propagation seed is a clean, gate-free design. Weaknesses 1.Backbone confound. Five baselines run on Qwen2.5-7B while the method runs on Dream-7B, and the sole diffusion-side system is GFM-RAG — the retriever LADDER itself fine-tunes. The gains are therefore not cleanly attributable to the proposed mechanism.

  1. Necessity of the graph is not established. [1] already reports accuracy gains from RAG on DLMs (Dream, LLaDA) without any graph structure, and no BM25 or dense retriever baseline appears in the paper.

  2. The efficiency comparison spans different operating points. HippoRAG 2 is natively single-step, so its five-round IRCoT latency reflects a different execution unit from LADDER's single trajectory. [1] Unlocking the Potentials of Retrieval-Augmented Generation for Diffusion Language Models: A Semantic Drift Perspective.

Critical weaknesses

Could the authors add HippoRAG 2 and Youtu-GraphRAG on Dream-7B under the same controller (K = 7, identical prompt)? Could the authors add BM25 and a dense retriever to isolate what the graph contributes beyond retrieval in general? Could the authors report single-pass HippoRAG 2 and Youtu-GraphRAG on Dream-7B as the efficiency reference, since a single-pass run is the like-for-like comparison for a single-trajectory method?

Questions for authors

Could the authors add HippoRAG 2 and Youtu-GraphRAG on Dream-7B under the same controller (K = 7, identical prompt)? Could the authors add BM25 and a dense retriever to isolate what the graph contributes beyond retrieval in general? Could the authors report single-pass HippoRAG 2 and Youtu-GraphRAG on Dream-7B as the efficiency reference, since a single-pass run is the like-for-like comparison for a single-trajectory method?

3 comments

Thank you for highlighting the need to control the generator, retrieval backend, and inference procedure. Manuscript table numbers refer to the submitted PDF. Times are seconds per question, and averages give equal weight to HotpotQA, 2WikiMultihopQA, and MuSiQue.

[W1,Q1] Backbone control

To control for the generator, we evaluate HippoRAG 2 and Youtu-GraphRAG using the same Dream-7B checkpoint, corpus, reader prompt, and K = 7, with at most five IRCoT rounds. LADDER uses the same Dream-7B checkpoint and per-retrieval passage budget with its within-trajectory controller.

Table B1. Same-backbone comparison. LADDER results are from manuscript Tables 1–2.

Dream-7B agentEM ↑F1 ↑R@7 ↑Time ↓
HippoRAG 2(IRCoT)41.850.463.012.51
Youtu-GraphRAG(IRCoT)43.152.265.013.70
LADDER45.254.667.12.79

With Dream-7B held fixed, LADDER gains 3.4 average EM points over HippoRAG 2 and 2.1 EM and 2.4 F1 points over Youtu-GraphRAG. The systems retain their respective retrieval backends and inference procedures.

[W2,Q2] BM25 and dense retrieval

We agree that prior RAG results on Dream and LLaDA make flat retrieval an essential control. Table B2 reports BM25 and E5-base-v2 under both IRCoT and entity-triggered refresh. The entity-refresh configurations share the graph-derived entity matcher, trigger rule, query format, evidence formatting, and K = 7. BM25 and E5-base-v2 directly rank passages. Their IRCoT configurations provide fully non-graph reference pipelines. Actual refresh times can differ because retrieved evidence changes subsequent drafts. Prior work on RAG for DLMs

Table B2. Retrieval backend and execution mode on Dream-7B. Each cell reports EM / F1 / s/question. LADDER results are from manuscript Table 2.

MethodModeHotpotQA2WikiMultihopQAMuSiQueAverage
Diffusion agent (Dream-7B)
BM25IRCoT40.6 / 48.3 / 10.2757.8 / 62.4 / 10.0215.4 / 27.0 / 11.6237.9 / 45.9 / 10.64
Dense E5-base-v2IRCoT44.0 / 52.6 / 10.9160.5 / 66.0 / 10.2917.6 / 29.5 / 11.9740.7 / 49.4 / 11.06
BM25Entity refresh43.6 / 53.1 / 2.6161.8 / 67.0 / 2.0515.9 / 26.9 / 4.0140.4 / 49.0 / 2.89
Dense E5-base-v2Entity refresh45.2 / 55.1 / 2.5861.4 / 67.8 / 2.0017.5 / 29.0 / 3.9741.4 / 50.6 / 2.85
Pretrained GFM-RAGEntity refresh46.8 / 56.4 / 2.7063.3 / 69.8 / 2.0717.2 / 29.1 / 4.0342.4 / 51.8 / 2.93
LADDEREntity refresh49.4 / 59.6 / 2.5567.4 / 73.5 / 1.8818.8 / 30.7 / 3.9545.2 / 54.6 / 2.79

Switching from IRCoT to entity refresh raises average EM from 37.9 to 40.4 for BM25 and from 40.7 to 41.4 for E5-base-v2. This switch changes both the generation loop and the refresh policy. Under the shared entity trigger, pretrained GFM-RAG reaches 42.4 EM, a 1.0-point gain over E5-base-v2. With event-conditioned tuning, LADDER scores 3.8 EM and 4.0 F1 points above E5-base-v2 at comparable latency (2.79 versus 2.85 s/question).

[W3,Q3] Single-pass efficiency references

Each retrieve-once baseline in Table B3 makes one question-only retrieval request at the controller level and keeps the top seven passages fixed throughout one Dream-7B trajectory. End-to-end timing includes online query processing, any model calls made by the retriever, and generation.

Table B3. Single-trajectory efficiency on Dream-7B. LADDER results are from manuscript Table 2.

Dream-7B configurationEM ↑F1 ↑Time ↓
HippoRAG 2, retrieve once38.947.92.68
Youtu-GraphRAG, retrieve once40.149.53.14
LADDER, entity refresh45.254.62.79

LADDER gains 6.3 EM points over single-retrieval HippoRAG 2 at an additional 0.11 s/question. It gains 5.1 EM points over single-retrieval Youtu-GraphRAG while reducing latency by 0.35 s/question.

Reviewer_BhY3OpenReview ↗

Reply 1

Noted. Could the authors please provide the source code for the study, particularly the code used to generate Figure 3? I did not sufficiently consider the reproducibility of the work during my initial review. If the code cannot be provided, I will need to reconsider my assessment of the overall quality of the manuscript.

Thank you for asking for the code. The anonymous repository includes the LADDER implementation and scripts to regenerate the per-question results and plot Figure~3.

https://anonymous.4open.science/r/LADDER-C264/

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.