acceptodds
Under review as a conference paper at ICLR 2027

Conditioning-Aware Adaptive Noising for Factual and Edit-Sensitive Graph-to-Sequence Generation

Abstract

Graph-to-sequence (G2S) generation is commonly evaluated by similarity to reference text, but high reference similarity does not imply reliable generation, which additionally requires factual grounding in the input graph and edit sensitivity to controlled graph edits. Diffusion language models are well suited to these properties because they iteratively revise the output, while the noise schedule controls how much target information is retained during denoising. However, adaptive noising is typically guided by denoising-side signals rather than by support from the conditioning graph. Across WikiOFGraph, GenWiki, and TekGEN, we find that denoising difficulty is only weakly and inconsistently associated with token-level graph support. We therefore introduce DLM4G, which separates adaptive noising into two roles: a global adaptive schedule uses token-wise denoising difficulty to identify positions that benefit from retaining more signal, while dynamic conditioning modulation uses decoder-to-graph cross-attention and conditioning sensitivity-the change in prediction when graph conditioning is removed, to control how strongly that schedule is applied for the current graph during both training and inference. We also introduce Factual Grounding (FGT) and Edit Sensitivity Rate (ESR) to measure the two reliability properties directly, and validate both against human and semantic assessments. The 63M-parameter DLM4G improves BLEU, chrF++, and METEOR over diffusion baselines on all three benchmarks while remaining competitive with 770M and 7-8B models. Within DLM4G, conditioning-aware modulation over difficulty-only adaptation yields absolute gains of 4-7 points in [email protected], 6-9 points in ESR@2, 5-7 points in relation-edit ESR, and approximately 3 BLEU points across the three benchmarks.

Primary area:
generative models
Keywords:
Graph-to-Sequence Generation, Diffusion Language Models, Adaptive Noising, Graph Conditioning, Factual Grounding, Edit Sensitivity, Reliable Generation.

Reviews

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

The paper introduces DLM4G, which combines a global adaptive schedule based on token-wise denoising difficulty with dynamic conditioning modulation using graph attention and conditioning sensitivity. It also introduces Factual Grounding (FGT) and Edit Sensitivity Rate (ESR) to evaluate reliable graph-to-sequence generation. Across WikiOFGraph, GenWiki, and TekGEN, conditioning-aware modulation improves grounding, edit sensitivity, and generation quality over difficulty-only adaptation.

Critical strengths
  1. The paper studies three complementary properties of reliable G2S generation: generation quality, factual grounding, and edit sensitivity.
  2. DLM4G combines a global adaptive schedule with dynamic conditioning modulation using graph attention and conditioning sensitivity.
  3. Across three benchmarks, conditioning-aware modulation improves grounding, edit sensitivity, and generation quality over difficulty-only adaptation.
Critical weaknesses
  1. It is unclear why positions that depend more on the graph should receive less noise. Adding more noise might instead encourage the model to rely more on the graph. Comparing these two choices would help justify the design.

  2. The ablations show that graph attention and conditioning sensitivity help, but do not show whether the global adaptive schedule is necessary. Could these conditioning signals achieve similar results on their own? An additional comparison would help answer this.

Questions for authors

see weakness. My current assessment is borderline. I would maintain a positive recommendation if the authors adequately address my concerns.

2 comments
AuthorsOpenReview ↗

Response to W1

We thank the reviewer for the positive assessment and constructive questions. We address both concerns below with additional experiments.

W1: "It is unclear why positions that depend more on the graph should receive less noise. Adding more noise might instead encourage the model to rely more on the graph. Comparing these two choices would help justify the design."


We agree that stronger corruption may encourage greater reliance on graph conditioning. However, greater reliance does not necessarily ensure accurate factual reconstruction. Consider a graph containing (Company C, founded_by, Person A), (Company C, led_by, Person B), and (Company C, founded_in, 2013), with the target sentence:

“Company C was founded by Person A in 2013 and is led by Person B.”

Consider two illustrative intermediate recovery states (conceptual representations, not literal Gaussian-corrupted outputs):

  • Case A — Moderate corruption: “Company C was founded by Per?? A in 2013 and is led by Per?? B.” The sequence retains contextual clues about the relationships and their associated entities.
  • Case B — Strong corruption: “Company C was ???? by ?????? ? in 2013 and is ??? by ?????? ?.” The graph still contains the correct facts, but the sequence provides weaker clues about their local realization.

In Case B, the denoiser must reconstruct not only the entities but also their correct associations with relation-bearing positions. An imperfect denoiser might generate:

“Company C was founded by Person B in 2013 and is led by Person A.”

Both entities are present in the graph and output, but their relationships are reversed. The problem is not missing graph information, but whether the denoiser can correctly retrieve and align that information under strong corruption. More noise may increase dependence on the graph while simultaneously removing information from the noisy sequence ztz_t that helps determine which graph fact belongs at each output position.

Our attention analysis is consistent with this intuition (Figure 3a). On WikiOFGraph, graph-attention AUROC increases from 0.61 at high noise to 0.83 as denoising progresses, indicating that graph-supported positions are localized more reliably as the sequence representation is refined. This motivates retaining sufficient signal at difficult, graph-dependent positions. Difficulty-aware noise allocation is also supported by prior work (SeqDiffuSeq, Yuan et al., 2024; Meta-DiffuB, Chuang et al., 2024).

Inverse-noising ablation. To test the alternative, we trained an inverse-noising variant that reverses the adaptive schedule shift in Eq. 1 while retaining graph-conditioning modulation (Eq. 2), so positions receiving stronger adaptation undergo increased corruption.

Variant (WikiOFGraph)BLEU[email protected]ESR@2Relation ESR
Base schedule0.600.770.530.52
Inverse (more noise)0.550.620.410.47
DLM4G (less noise)0.650.830.600.58

The inverse-noising variant underperforms even the base schedule. Compared with DLM4G, factual grounding decreases from 0.83 to 0.62, ESR@2 from 0.60 to 0.41, and relation-edit ESR from 0.58 to 0.47. These results are consistent with more difficult factual and relational recovery under stronger corruption, as illustrated above.

Thus, stronger corruption at graph-dependent positions does not translate into better factual reconstruction in our setting. Instead, retaining additional signal at difficult, graph-dependent positions achieves substantially better grounding and edit sensitivity. We will clarify this motivation and incorporate the inverse-noising comparison into the revised manuscript.

AuthorsOpenReview ↗

Response to W2

W2: "The ablations show that graph attention and conditioning sensitivity help, but do not show whether the global adaptive schedule is necessary. Could these conditioning signals achieve similar results on their own? An additional comparison would help answer this."


We thank the reviewer for identifying this missing comparison. We agree that the original component ablations (Table 4) evaluate graph-conditioning signals alongside the global difficulty schedule, without isolating their independent effectiveness.

We therefore introduce a conditioning-only variant that removes the global difficulty-based schedule while retaining graph attention and conditioning sensitivity. In this variant, noise adaptation is determined by graph-conditioning signals without using token-wise reconstruction difficulty. This directly tests whether conditioning information alone can achieve performance comparable to the full framework.

Variant (WikiOFGraph)BLEU[email protected]ESR@1ESR@2Relation ESR
Base schedule0.600.770.640.530.52
Difficulty only (LL)0.620.790.650.540.53
Conditioning only (A+ΔA+\Delta)0.630.790.660.570.54
Full (L+A+ΔL+A+\Delta)0.650.830.680.600.58

The conditioning-only variant improves over the base schedule and performs comparably to or better than difficulty-only adaptation. However, neither component alone matches the full framework. Adding difficulty-aware adaptation to conditioning-only improves BLEU from 0.63 to 0.65, [email protected] from 0.79 to 0.83, ESR@2 from 0.57 to 0.60, and relation-edit ESR from 0.54 to 0.58.

Our conditioning-only ablation therefore provides empirical evidence that the global difficulty-based schedule contributes beyond graph conditioning alone, with the strongest performance achieved by combining both sources of information.

The new conditioning-only and inverse-noising ablations are preliminary single-run evaluations. We will include three-seed statistics, precise schedule configurations, and reproducibility details in the revised manuscript.

We sincerely thank the reviewer for their constructive feedback, which has helped strengthen the paper. We hope our responses and the new experiments address the concerns raised, and, if so, we would greatly appreciate it if the reviewer would consider increasing their score. We remain happy to address any further questions during the discussion period.

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

This paper proposes DLM4G, a encoder-decoder transformer based on diffusion language model (DLM) for generating text from serialized knowledge graphs. Motivated by the observation that adaptive noising is typically guided by denoising-side signals rather than by support from the conditioning graph, the paper introduce a global adaptive schedule that retains more signal at positions that are harder to recover, measured by denoising loss. During both training and inference, this schedule is modulated by two conditioning signals: decoder-to-graph cross-attention and conditioning sensitivity, measuring the change in the denoising prediction when graph conditioning is removed. Two metrics (FGT and ESR) are introduced for measuring factual grounding and edit sensitivity. Experiments on three datasets assess generation quality, factual grounding and responses to controlled graph edits against fine-tuned PLMs, zero-shot LLMs and diffusion baselines, and ablations attribute gains in FGT and ESR to the conditioning signals.

Critical strengths
  1. This paper addresses a relevant problem by considering factual grounding and responsiveness to graph edits alongside conventional generation quality.
  2. Exploring graph-conditioning signals for adaptive noising in DLMs is a potentially useful research direction.
Critical weaknesses
  1. The paper is imprecise and unpolished. For example, the central concept "graph-supported position" is neither explicitly defined nor clearly illustrated. Yet both the motivation (Fig. 1) and the diagnostic experiments (Fig. 3) rely on it, making the paper hard to evaluate.

  2. The main new idea is to modulate a difficulty-based schedule with conditioning signals, but its components are largely existing. The paper also does not make clear how this design relates to graph-factual support. Both signals measure aggregate dependence on the input rather than graph inductive bias. How do the authors expect these aggregate signals to capture fact-level graph support, and what aspect of the design is specific to graph inputs?

  3. The comparison with larger models does not isolate the contribution of the method. As stated in App. D.6, the fine-tuned GPT-2 and T5 baselines are trained on 100K pairs, while DLM4G uses the full training sets, and the 7-8B LLMs are evaluated zero-shot, a different supervision setting altogether. The matched 100K-pair comparison (Table 3) includes only T5-Small and SeqDiffuSeq, and under this budget DLM4G-2.o drops from 0.654 to 0.433 BLEU on WikiOFGraph, indicating that training data has a large effect. As a result, the claim of remaining competitive with 770M and 7-8B models (Abstract, Sec. 4.4) conflates differences in method, model size, and task-specific training data. Matched-data comparisons with T5-Large and a small fine-tuned LLM (e.g., Qwen3-1.7B) would help; otherwise the claim should be qualified accordingly.

Questions for authors
  1. How exactly is an output position labeled as graph-supported or unsupported in Figs. 1 and 3? What threshold separates high- and low-loss positions, and why was it chosen?

  2. Table 4 and App. G.4 show that using both AA and Δ^\hat{\Delta} helps, but the specific form ω=A+(1−A)Δ^\omega=A+(1-A)\hat{\Delta} is not motivated. How should ω\omega be interpreted when AA is high but Δ^\hat{\Delta} is close to zero, i.e., the model attends to the graph but removing it barely changes the prediction? This seems relevant given that attention is noted to be weakly discriminative at high noise levels (Limitations). Why is this form chosen over alternatives such as a product or a learned combination, are AA and Δ^\hat{\Delta} on comparable scales, and how is ω\omega distributed across positions and timesteps in practice?

11 comments
AuthorsOpenReview ↗

Response to W1, Q1

W1 & Q1: Definition of graph-supported positions and the high/low-loss threshold


We agree that the definition and labeling procedure should have been stated explicitly, and we will revise the manuscript for clarity more broadly. We clarify both below and will incorporate them into Section 4.2.

(1) Graph-supported positions. For each graph–reference pair (G,S)(\mathbf{G}, \mathbf{S}), we extract and canonicalize entities, attribute values and relation labels using the fixed Qwen2.5-7B-Instruct pipeline described in Appendix E.1. Extracted units are aligned to their corresponding spans in the reference sequence and mapped to output token positions.

A position is labeled graph-supported if it belongs to the token span of a canonicalized unit present in both the conditioning graph and the reference sequence. All remaining non-padding positions are labeled unsupported. Unsupported does not imply factually incorrect; it covers linguistic tokens that do not directly realize an extracted graph fact.

(2) Token-level illustration (Appendix D.2). Consider a graph containing:

  • ⟨USA, hosted, 1994_FIFA_World_Cup⟩
  • ⟨USA, capital, Washington_D.C.⟩
  • ⟨1994_FIFA_World_Cup, top_scorer, Hristo_Stoichkov⟩

The corresponding reference sequence is:

The United States hosted the 1994 FIFA World Cup; its capital is Washington, D.C., and the tournament's top scorer was Hristo Stoichkov.

The bold spans correspond to graph-supported entities and relations. Their constituent output tokens are labeled graph-supported, whereas tokens such as The, its, is, and and was are labeled unsupported. The same support-labeling procedure is used in Figures 1 and 3.

(3) High/low-loss threshold. Denoising difficulty and graph support are computed at different granularities:

  • Denoising difficulty: The running loss ℓˉi\bar{\ell}_i in Eq. 1 is accumulated across non-padding occurrences of output position ii during training. For Figure 1, high- and low-loss positions are separated using the median of ℓˉi\bar{\ell}_i across valid output-position indices.
  • Graph support: Determined independently for each graph–reference pair using the token-span alignment described above.

We use the median because it gives a fixed, parameter-free threshold that requires no dataset-specific selection. Figure 1 aggregates these labels across all non-padding tokens in the full training split. Because examples have variable lengths, splitting position indices at the median does not necessarily produce equal proportions of token occurrences in the high- and low-loss groups.

Figure 3 uses the same graph-support labels to evaluate attention and conditioning sensitivity, without requiring the high/low-loss partition.

We will add these definitions, the labeling procedure and the token-level illustration to the revised manuscript.

AuthorsOpenReview ↗

Response to W2

W2: "The main new idea is .....?"


Our contribution is a graph-conditioning-aware adaptive noising framework that integrates global reconstruction difficulty with input-specific graph dependence to control signal retention at individual output positions. Rather than treating reconstruction difficulty as a sufficient indicator of factual relevance, DLM4G dynamically modulates the adaptive noise schedule using graph attention and conditioning sensitivity for the current graph and noisy sequence.

(1) Methodological distinction. Prior adaptive diffusion approaches have explored token-position reconstruction difficulty (SeqDiffuSeq; Yuan et al., 2024), contextualized scheduling (Meta-DiffuB; Chuang et al., 2024) and learned adaptive noise processes (MuLAN; Sahoo et al., 2024b).

DLM4G addresses a complementary problem: how should a globally learned, difficulty-based schedule be adapted for the specific graph and noisy sequence being processed?

As shown in Fig 1, denoising difficulty and graph support are only weakly and inconsistently associated across datasets. A position that is generally difficult to reconstruct need not contain graph-supported information in a particular example.

DLM4G therefore combines two levels of information: (1) the global schedule captures position-wise reconstruction difficulty accumulated during training, while (2) dynamic graph-conditioning signals determine how strongly this adaptation is applied for the current input. Unlike prediction-level guidance or loss reweighting, this interaction directly modulates the diffusion noise schedule during training and generation.

(2) How do these signals relate to factual graph support? The two signals measure different aspects of graph dependence:

  • Graph attention (AA): For each output position, At,iA_{t,i} takes the peak attention to any single graph token rather than the total attention mass (Section 3.2). Higher values indicate attention concentrated on a specific part of the graph, such as a particular entity or relation token, rather than diffusely spread across the input.
  • Conditioning sensitivity (Δ^\hat{\Delta}): Measures, per output position, how much the denoising prediction changes when graph conditioning is removed, quantifying the graph's influence on that specific prediction.

Both signals are therefore position-level rather than sequence-level aggregates. We do not assume that either signal identifies the correct supporting triple. Instead, we evaluate their relationship to graph-supported positions empirically using the labels clarified in W1/Q1.

Fig 3(a) shows that graph attention increasingly distinguishes supported from unsupported positions, reaching AUROC values of 0.83 / 0.79 / 0.73 on WikiOFGraph / GenWiki / TekGEN. Because support labels use aligned graph-unit spans (W1/Q1), AUROC measures how well attention distinguishes graph-supported from unsupported positions. Fig 3(b) further shows that conditioning sensitivity remains consistently higher for graph-supported positions. These findings support their use as complementary, position-specific indicators of graph dependence, rather than exact factual-attribution mechanisms.

(3) Does their integration improve factual reliability? Our conditioning-only ablation tests whether graph-conditioning signals are sufficient without the global difficulty schedule.

Variant (WikiOFGraph)BLEU[email protected]ESR@2Relation ESR
Base schedule0.600.770.530.52
Difficulty only (LL)0.620.790.540.53
Conditioning only (A+ΔA+\Delta)0.630.790.570.54
Full (L+A+ΔL+A+\Delta)0.650.830.600.58

Neither component alone matches the full framework. Furthermore, the same-checkpoint interventions in App G.6 show that rank-reversing the conditioning weights reduces [email protected] from 0.83 to 0.72, indicating that their position-wise assignment matters rather than merely their aggregate magnitude. The conditioning-only results are preliminary single-run evaluations; we will report three-seed statistics in the revised manuscript.

(4) What is specific to graph inputs? We agree that the underlying scheduling formulation is conditioning-generic and does not introduce an explicit graph-topological inductive bias, such as message passing (Appendix A.3). Its graph-specific application addresses the requirement that generated text correctly express entities, attributes and relationships from an input graph, and respond appropriately when those facts are edited. DLM4G uses the serialized relational graph as its conditioning source to dynamically regulate signal retention without requiring explicit triple-to-output alignment during generation.

This distinction is evaluated through factual grounding and controlled graph edits, including relation-edit ESR, where DLM4G achieves 0.58 / 0.48 / 0.46 across the three benchmarks.

AuthorsOpenReview ↗

Response to W3

W3: "The comparison with larger models does not isolate the contribution of the method. Matched-data comparisons with T5-Large and a small fine-tuned LLM (e.g., Qwen3-1.7B) would help; otherwise the claim should be qualified accordingly."


We thank the reviewer for highlighting the differences in training data, initialization, model size and supervision across the original comparisons. We agree that these differences prevent the larger-model comparisons from independently isolating the contribution of our method. We address this through additional matched-data experiments and explicit qualification of our larger-model claims.

(1) Additional matched-data comparison. We extend the 100K-pair experiment to include GPT-2 Base and T5-Large alongside T5-Small, SeqDiffuSeq and DLM4G-2.o. All models are trained from scratch on identical 100K graph–text pairs, using batch size 128, a maximum of 200K optimizer steps, checkpoint selection based on validation-set BLEU, and three independent seeds. Only tokenizer vocabularies are retained; no pretrained model weights are used.

BLEU (mean ± standard deviation, 3 seeds):

ModelWikiOFGraphGenWikiTekGEN
GPT-2 Base0.233 ± 0.0030.102 ± 0.0040.113 ± 0.003
T5-Large0.315 ± 0.0040.182 ± 0.0050.146 ± 0.003
T5-Small0.345 ± 0.0050.223 ± 0.0030.162 ± 0.005
SeqDiffuSeq0.405 ± 0.0010.247 ± 0.0040.185 ± 0.003
DLM4G-2.o0.433 ± 0.0040.293 ± 0.0030.191 ± 0.003

DLM4G-2.o achieves the highest BLEU across all three datasets under this common from-scratch protocol, including against T5-Large. T5-Large performs below T5-Small in this setting; its validation-set BLEU peaks at approximately 120K–130K steps, and we report the best validation-selected checkpoint. Matching task-specific data, initialization regime and optimizer steps does not equate model capacity or computational cost.

We also acknowledge that DLM4G's BLEU decreases from 0.654 under full-data training to 0.433 under the 100K-pair setting, demonstrating that training-data scale substantially affects performance. This reinforces the importance of separating full-data results from matched-supervision comparisons.

(2) Qualifying comparisons with larger pretrained and zero-shot models. The new T5-Large experiment is a from-scratch comparison, not a replacement for the original pretrained T5-Large baseline. Similarly, the T5-Small result above is obtained from scratch and replaces the pretrained, fine-tuned T5-Small row in the revised Table 3.

Pretrained T5-Large fine-tuned on the same 100K pairs reaches 0.658 BLEU on WikiOFGraph versus 0.433 for DLM4G trained from scratch, reflecting the advantage of pretraining.

As a reference point, DLM4G trained on the full split reaches 0.654 BLEU against 0.658 for pretrained T5-Large, with approximately 12 times fewer parameters, although this comparison is not matched in pretraining or training data.

Likewise, the 7–8B LLM baselines are evaluated zero-shot, representing a different supervision regime from task-trained DLM4G. We therefore present the pretrained and zero-shot comparisons as reference points rather than evidence of a controlled advantage.

Regarding the small fine-tuned LLM comparison (e.g., Qwen3-1.7B), we would appreciate the reviewer's guidance on whether this experiment is important for their assessment. If so, we will prioritize it and report the results during the discussion period. Otherwise, following the reviewer's suggestion, we will qualify the claim accordingly, as described above.

(3) Isolating the methodological contribution. The matched from-scratch comparisons establish comparative performance under a common task-specific supervision protocol. More importantly, the within-DLM4G ablations (Table 4) directly isolate the contribution of conditioning-aware adaptive noising by holding the architecture and training conditions fixed.

We will revise the Abstract, Section 4.4 and Conclusion to explicitly qualify comparisons with larger pretrained and zero-shot models as reference comparisons, with the main claims resting on the matched-data from-scratch results and the within-architecture ablations.

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

The paper proposes Diffusion Language Model for Graphs (DLM4G), which can generate text from knowledge graph. The framework can separate adaptive noise into two parts. A global adaptive schedule identifies positions that benefit from retaining more signal. Dynamic conditioning modulation uses decoder-to-graph cross-attention and conditioning sensitivity-the change in prediction when graph conditioning is removed. The paper also proposes Factual Grounding (FGT) and Edit Sensitivity Rate (ESR) to measure the two reliability properties directly.

Critical strengths
  1. The paper is well motivated. Figure 1 shows that the association between denoising difficulty and graph support is weak. It then proposes adaptive noising based on this finding. Table 4 shows that adding conditional information further improves the model compared to the difficulty-only baseline. The paper also shows that shuffling or reversal will reduce performance.
  2. The edit sensitivity rate is very useful. The validation of 500 samples in the Appendix shows a strong correlation between it and human annotation.
Critical weaknesses
  1. The factuality metrics cannot detect some factual errors. FGT compares entity sets. An output can mention all correct entities with an incorrect relationship. ESR removes edit direction. Replacing A with B and incorrectly replacing B with A can produce the same change set.
  2. Some baseline comparisons have problems. DLM4G uses more training data compared to GPT2 and T5 baselines. The paper provides a matched-data experiment. However, it only compares T5-small. T5-large also needs to be included.
Questions for authors
  1. What exactly is matched in the 100K-pair comparison? The paper includes a valuable comparison using the same 100K training pairs. However, Table 9 reports three training epochs for T5. The diffusion configurations use much larger step budgets. Did you reuse those settings in the matched-data experiment?
4 comments
AuthorsOpenReview ↗

Response to W1

We thank the reviewer for recognizing the motivation behind DLM4G and the usefulness of ESR. We address the concerns below with additional evaluations and controlled experiments.


W1(a): FGT cannot detect correct entities with an incorrect relation.


We agree that FGT measures entity-level grounding, as acknowledged in our Limitations. To evaluate factual reliability beyond entity matching, we provide three complementary assessments:

  1. AlignScore (Zha et al., 2023) evaluates graph–text semantic consistency. DLM4G-2.o achieves 0.86 / 0.77 / 0.75 on WikiOFGraph / GenWiki / TekGEN, compared with 0.84 / 0.74 / 0.71 for the strongest baselines (Tables 1, 16).
  2. Human grounding judgments on 500 examples assess whether the information expressed in the output is supported by the graph, rather than merely checking entity overlap. These correlate with [email protected] at ρ=0.78\rho=0.78 (Table 14).
  3. Relation-edit evaluation modifies the graph predicate while keeping the subject and object fixed, testing relational changes that entity-level FGT cannot distinguish. DLM4G-2.o achieves relation-edit ESR of 0.58 / 0.48 / 0.46 across the three benchmarks (Tables 15–16). Additionally, target-edit realization verifies whether the correct edited relation is expressed (0.65 on WikiOFGraph; see W1(b)).

Thus, FGT captures entity grounding, while AlignScore, human judgments, and predicate-level edit experiments provide complementary evidence of semantic consistency and relational correctness. We will clarify these distinct evaluation roles in Section 4.3.


W1(b): ESR is direction-insensitive.


We agree that ESR's symmetric-difference formulation does not explicitly encode edit direction. Consider an original graph containing ⟨France, capital, Paris⟩, which is edited to ⟨France, capital, Lyon⟩. The corresponding correct output sequences would be:

  • Original graph: "The capital of France is Paris."
  • Edited graph: "The capital of France is Lyon."

However, if the generated sequences are reversed—producing "The capital of France is Lyon" for the original graph and "The capital of France is Paris" for the edited graph—ESR still observes the same symmetric difference {(Paris,Lyon)}\{(\text{Paris}, \text{Lyon})\}. Thus, ESR can register responsiveness to the edit even when both generated sequences contradict their respective conditioning graphs.

To explicitly verify directional correctness, we report target-edit realization (Appendix E.7), measuring whether the intended edited graph fact is correctly expressed in the output.

We use the fixed Qwen2.5-7B-Instruct extraction and canonicalization pipeline (Appendix E.1), applied identically across models. For relation edits, the pipeline extracts facts as (subject, relation, object) triples and matches the complete extracted triple against the edited graph fact. Since the subject and object remain fixed, successful realization requires the new predicate to be correctly associated with that subject–object pair, rather than merely appearing elsewhere in the text. Consequently, reversed relations are not counted as successful realizations.

This follows our controlled graph-intervention protocol, where only the edited graph unit changes and paired generations use identical sampling randomness where applicable.

Target-edit realization on WikiOFGraph:

ModelValue edit, k=1k=1Value edit, k=2k=2Relation edit
T5-Small0.710.630.60
T5-Large0.740.680.63
DiffuSeq0.730.640.61
SeqDiffuSeq0.730.670.63
DLM4G-2.o0.760.700.65

DLM4G-2.o leads across all three settings, preserving its ranking under a direction-aware evaluation. We also examined the specific reversal failure mode: exact reversed substitutions occur in only 0.4% of evaluated edit pairs for DLM4G-2.o, versus 0.4–0.8% for the baselines.

Thus, while ESR measures edit responsiveness without explicitly encoding direction, target-edit realization verifies correct directional realization, and the identified reversal failure mode is rare in practice. We will clarify the triple-matching procedure in Appendix E.7.

AuthorsOpenReview ↗

Response to W2 and Q1

W2 + Q1: Training-data fairness, T5-Large, and matched settings.


We agree that the original 100K-pair comparison matched training examples but not initialization or optimization budgets. T5-Small used pretrained weights and three fine-tuning epochs (approximately 9.4K steps, batch size 32; Table 9), whereas DLM4G and SeqDiffuSeq were trained from scratch for up to 200K steps (batch size 128).

We now distinguish three experimental settings.

(i) Within-framework ablation. Holding the architecture and training protocol fixed, we isolate the contribution of conditioning-aware noising (Table 4). Compared with difficulty-only (LL), the full L+A+ΔL+A+\Delta schedule improves:

  • [email protected]: 0.79 → 0.83 (WikiOFGraph), 0.65 → 0.71 (GenWiki), 0.62 → 0.69 (TekGEN).
  • ESR@2: 0.54 → 0.60, 0.47 → 0.55, 0.43 → 0.52.
  • Relation-edit ESR: 0.53 → 0.58, 0.42 → 0.48, 0.39 → 0.46.

These comparisons isolate the effect of conditioning-aware adaptation within the same architecture.

(ii) Pretrained T5-Large comparison. T5-Large (770M parameters), fine-tuned on 100K task pairs, is already included in Tables 1–2. We acknowledge its advantage over DLM4G under the original 100K protocol. However, DLM4G-2.o (63M parameters), trained from scratch on the full task split, achieves comparable or higher BLEU (0.654 / 0.469 / 0.253 versus 0.658 / 0.361 / 0.199) and higher [email protected] (0.83 / 0.71 / 0.69 versus 0.79 / 0.61 / 0.59). This demonstrates competitive performance with approximately 12 times fewer parameters, not a data-matched comparison.

(iii) Controlled 100K-pair from-scratch comparison. To address the reviewer's request directly, we additionally trained T5-Large and the other baselines under a common protocol:

  • Identical 100K training pairs and held-out evaluation inputs.
  • Randomly initialized model weights, retaining only tokenizer vocabularies.
  • Batch size 128, maximum 200K optimizer steps, and best checkpoint selected by best validation-set BLEU score.
  • Three independent seeds per model and dataset.

BLEU (mean ± std, 3 seeds):

ModelWikiOFGraphGenWikiTekGEN
GPT-2 Base0.233±0.0030.102±0.0040.113±0.003
T5-Large0.315±0.0040.182±0.0050.146±0.003
T5-Small0.345±0.0050.223±0.0030.162±0.005
SeqDiffuSeq0.405±0.0010.247±0.0040.185±0.003
DLM4G-2.o0.433±0.0040.293±0.0030.191±0.003

DLM4G-2.o achieves the highest BLEU on all three datasets, including against T5-Large, under this controlled from-scratch setting. Relative to T5-Small, it also improves [email protected], ESR@2, and relation-edit ESR across all three benchmarks.

T5-Large's validation set BLEU score peaks around 120K–130K steps before declining; we report its best checkpoint. While task-specific data and optimizer-step budgets are matched, computational cost and model capacity differ across architectures.

This protocol removes pretraining for all models. When pretrained T5-Large is fine-tuned on the same 100K pairs, it reaches 0.658 BLEU on WikiOFGraph versus 0.433 for DLM4G-2.o trained from scratch, reflecting the substantial advantage of pretraining. We therefore treat the from-scratch results as a controlled comparison under common task-specific supervision, not as evidence that DLM4G-2.o outperforms pretrained models under matched data.

We will distinguish these protocols explicitly in Section 4.4 and Appendix F.3, replacing the existing Table 3 with the controlled comparison.

We sincerely thank the reviewer for their constructive feedback, which has helped strengthen the paper. We hope our responses and the new experiments address the concerns raised, and, if so, we would greatly appreciate it if the reviewer would consider increasing their score. We remain happy to address any further questions during the discussion period.

AuthorsOpenReview ↗

Gentle Reminder

We thank the reviewer for the constructive feedback. As the discussion deadline is approaching, we would appreciate your consideration of our updated response. We have now uploaded a revised manuscript incorporating the additional experiments and clarifications discussed during the rebuttal.

In particular, we have addressed your concerns regarding:

  1. FGT and ESR: Clarified the entity-level scope of FGT and added direction-aware edit realization and content-preservation analyses (§4.3; Appendix E.7).
  2. Matched-data comparisons: Added GPT-2 Base, T5-Small, T5-Large, and DLM4G-2.o under the same 100K-pair from-scratch setting. These results are now explicitly reported in Table 2 (p. 9), with further details in Appendix F.3. DLM4G-2.o achieves the best BLEU, chrF++, and METEOR among these baselines across all three datasets.
  3. Training protocol: Clarified the matched settings, including batch size 128, up to 200K optimization steps, checkpoint selection, and three independent seeds (§4.4; Appendix F.3).

We have also revised our experimental claims to clearly distinguish comparisons across different training and pretraining settings.

Your concerns largely overlap with those raised by Reviewer ZzKC, which we have addressed through the manuscript revisions and additional experiments. A detailed summary of these changes, with their exact locations, is available in our latest response to Reviewer ZzKC, titled "Summary of Manuscript Revisions and Additional Response to W3."

We hope the revised manuscript and additional results adequately address your concerns. If you find them satisfactory, we would be grateful if you could reconsider increasing your score before the discussion deadline. We would also appreciate knowing if any substantive concerns remain.

Thank you again for your time and consideration.

Note: Reviewer ZzKC has reassessed the revised manuscript and improved their presentation and soundness scores. Reviewer fbmH has also acknowledged our rebuttal and maintained their positive recommendation. We hope the revisions and additional experiments similarly address your concerns.

Thanks for your detailed response. My concerns have been addressed. I have updated my score to 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Consolidated Rebuttal Summary and Manuscript Revisions

Dear Area Chair,

We thank you and the reviewers for the constructive feedback. During the rebuttal, we ran additional experiments, strengthened the methodological justification, clarified the evaluation protocols, and uploaded a revised manuscript. The main concerns and revisions are summarized below.

Summary of Rebuttal and Additional Evidence


Reviewer concernClarifications and additional evidenceLocation
1. Definition of graph-supported positions (ZzKC)Defined via canonicalized graph-unit extraction and token-span alignment; clarified the high/low loss partition with a worked example.Intro; §4.2; App. D.2
2. Novelty and contribution of conditioning (ZzKC)Clarified the two-level mechanism combining global reconstruction difficulty with input-specific conditioning. Over difficulty-only adaptation on WikiOFGraph: [email protected] 0.79 → 0.83, ESR@2 0.54 → 0.60, relation ESR 0.53 → 0.58, with consistent gains on all three datasets.§3.2; §4.4, Table 3; App. G.3–G.4
3. Attention, sensitivity, and aggregation (ZzKC)Clarified both signals as complementary measures of graph dependence. Soft-OR achieves FGT 0.83 / ESR 0.60 vs. 0.79 / 0.49 for a learned gate; attention AUROC 0.83 / 0.79 / 0.73.§3.2; §4.2; App. C.2
4. Importance of position-specific scheduling (ZzKC)Same-checkpoint interventions: shuffled weights reduce FGT 0.83 → 0.76, rank reversal to 0.72; classifier-free guidance remains below the full scheduler (ESR@2 0.56 vs. 0.60).§4.4; App. G.6–G.7
5. Reduced vs. increased corruption (fbmH)Inverse noising yields BLEU 0.55 / FGT 0.62 / ESR@2 0.41 vs. 0.65 / 0.83 / 0.60 for our schedule.Response to fbmH, W1
6. Necessity of the global schedule (fbmH)Conditioning-only: 0.63 / 0.79 / 0.57; full framework: 0.65 / 0.83 / 0.60, showing both components contribute.§4.4, Table 3
7. FGT relation limits and ESR direction (acPv)Scoped FGT to entity level; added AlignScore, relation-edit ESR (0.58 / 0.48 / 0.46), direction-aware realization (0.76 / 0.70 / 0.65 on WikiOFGraph), and preservation analysis.§4.3; App. E.2–E.7
8. Baseline fairness (acPv, ZzKC)Controlled 100K-pair from-scratch comparison with GPT-2 Base, T5-Small, T5-Large, SeqDiffuSeq, and DLM4G (batch 128, ≤200K steps, three seeds). DLM4G achieves the highest BLEU: 0.433 / 0.293 / 0.191.§4.4, Table 2; App. F.3
9. Clarity and positioning (ZzKC)Revised title, abstract, method, experiments, and limitations; distinguished conditioning dependence from factual correctness and positioned conditioning-aware noising as the core contribution.Title; Abstract; §3.2; §4.4; §5

Reviewer Reassessment

ReviewerInitialCurrentOutcome
acPv23Raised overall recommendation after rebuttal
ZzKC22Increased soundness (1→3), presentation (1→2); acknowledged that the claims are "largely well scoped" and agreed the revised title fits better.
fbmH33Maintained positive assessment after our responses

Average rating increased from 2.33 to 2.67.

Closing Remarks


The rebuttal strengthened the paper in three ways:

  • Methodological validation: ablations and controlled interventions isolate the contribution of conditioning-aware scheduling.
  • Evaluation rigor: direction-aware edits, relation-level analysis, human validation, and matched-training comparisons address metric and fairness concerns.
  • Scientific clarity: the revision defines the conditioning signals, scopes the claims, and acknowledges the limits of cross-regime comparisons.

We would be grateful if the Area Chair would consider the revised manuscript and reviewer reassessments and recommend acceptance of our submission.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.