acceptodds
Under review as a conference paper at ICLR 2027

Learning Probabilistic Graphs from Natural Language

25 views

Abstract

Natural-language descriptions contain knowledge about uncertain systems, but using that knowledge in probabilistic models requires an explicit account of variables, states, dependencies, and probabilities. Recovering these elements together is difficult because errors in one component can change the meaning of the others. We introduce DUAL-BN, a two-stage framework for learning this translation through bidirectional supervision and structured feedback. A shared language model first learns to extract probabilistic graph components from text and reconstruct descriptions from graphs. Group Relative Policy Optimization (GRPO) then refines jointly generated components using reference-graph fidelity, semantic consistency of the text–graph–text cycle, and a penalty for component counts exceeding the reference. The probability targets are parent-specific pairwise conditional tables. On PRISM-BN backbones unseen during training, DUAL-BN increases Node F1 from 0.77 to 0.91 and mapped Edge F1 from 0.68 to 0.84 over supervised fine-tuning, while reducing forward KL divergence on matched pairwise tables from 0.87 to 0.66. The trained policy is evaluated through staged component queries with reference-assisted filtering. The gains accompany a decrease in conditional State F1. A separate study of composition from supplied partial graphs examines how such representations can support larger probabilistic backbones. On ten descriptions rated by three evaluators, mean overall reconstruction quality rises from 3.4/5 for SFT to 4.1/5 for DUAL-BN. The results support combining token-level supervision with feedback on the structure and recoverable information of graph predictions as a step toward building probabilistic models from language.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
probabilistic graphical models, semantic parsing, structured prediction, text-to-graph generation, probabilistic knowledge extraction

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

This paper proposes DUAL-BN, a framework for extracting probabilistic graphs from natural-language descriptions. The method combines bidirectional supervised fine-tuning (SFT) with Group Relative Policy Optimization (GRPO). The reinforcement learning reward integrates reference-graph fidelity, text–graph–text reconstruction consistency, and a complexity penalty.

Critical strengths
  1. Meaningful research problem. Extracting structured probabilistic representations from natural language is a challenging and potentially useful task.

  2. Promising empirical improvements. The proposed method improves Node F1 from 0.7745 to 0.9106 and mapped Edge F1 from 0.6762 to 0.8415 over SFT.

Critical weaknesses
  1. The discrepancy between training and evaluation. During GRPO, the model generates all graph components jointly, whereas evaluation relies on a staged extraction procedure with reference-assisted filtering. Consequently, the reported scores do not directly measure the quality of complete graphs generated from text alone.

  2. The individual contributions of the proposed components are insufficiently isolated. The experiments do not adequately establish which components are responsible for the observed improvements. GRPO without the complexity penalty yields only modest improvements over SFT, especially for Edge F1 (0.6762 to 0.6855). Most of the reported structural gains appear after introducing the complexity penalty.

  3. An overly simplified task formulation. DUAL-BN is built directly upon PRISM-BN [5], adopting much of its evaluation protocol. PRISM-BN identifies accurate conditional probability estimation as one of the central challenges of the task. However, rather than directly addressing this challenge, DUAL-BN simplifies the original full multi-parent conditional probability tables into parent-specific pairwise tables, thereby reducing the complexity of probability reconstruction. Under this simplified setting, although DUAL-BN achieves lower KL divergence than SFT, it is difficult to determine whether the proposed method truly addresses the key challenges of the original benchmark.

  4. Unclear substantive novelty and contributions over prior work. DUAL-BN primarily combines existing techniques, including bidirectional learning, cycle-consistency feedback, GRPO, and complexity regularization. Bidirectional text–graph transformation and cycle-consistency learning have already been explored in CycleGT [1] and Semantic Parsing with Dual Learning [2], while GRPO [3] and complexity regularization in Bayesian network learning [4] are also established techniques. Therefore, the paper should clarify what substantive methodological novelty or new mechanism is introduced beyond applying these existing techniques to the PRISM-BN task.

  5. Although the authors disclose the use of generative AI to assist with manuscript writing, the paper has still not undergone sufficient human review and revision. First, the research motivation in the Introduction lacks clear logical connections. The second paragraph emphasizes the challenges of maintaining consistency among the components of probabilistic graphs, while the third paragraph directly shifts to the limitations of SFT's token-level loss without explaining the relationship between the two. Although the fourth paragraph introduces methods such as bidirectional supervision, cycle-consistency feedback, and GRPO, it does not clearly explain how these designs correspond to and address the specific challenges identified earlier. Second, the Introduction contains considerable defensive language. For example, before adequately establishing the research motivation, the authors emphasize that local pairwise conditional probability tables cannot constitute a complete Bayesian network. Furthermore, the Introduction reports detailed changes in multiple experimental metrics, which would be more appropriate in the experimental section.

[1] Guo, Q., Jin, Z., Qiu, X., Zhang, W., Wipf, D., & Zhang, Z. (2020). CycleGT: Unsupervised Graph-to-Text and Text-to-Graph Generation via Cycle Training. Proceedings of the 3rd International Workshop on Natural Language Generation from the Semantic Web (WebNLG+), 77–88.

[2] Cao, R., Zhu, S., Liu, C., Li, J., & Yu, K. (2019). Semantic Parsing with Dual Learning. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 51–64.

[3] Shao, Z., et al. (2024). DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv preprint arXiv:2402.03300.

[4] Lam, W., & Bacchus, F. (1994). Learning Bayesian Belief Networks: An Approach Based on the MDL Principle. Computational Intelligence, 10(3), 269–293.

[5] Bhattacharya, A., Singh, N., Pokhriyal, N., & Vosoughi, S. (2026). PRISM-BN: A Controlled Corpus and Benchmark for Text-to-Parameterized Bayesian Network Extraction. arXiv preprint arXiv:2609.21673.

Questions for authors
  1. Can the authors provide an end-to-end evaluation without reference-assisted filtering to verify whether the reported improvements still hold?

  2. Why does DUAL-BN simplify full multi-parent conditional probability tables into pairwise tables, and how does this address the challenges identified by PRISM-BN [5]?

  3. What is the substantive novelty of DUAL-BN over prior work, and why are the original PRISM-BN [5] extraction baselines and simpler structural regularization methods not included?

  4. Can the authors provide ablations to demonstrate the independent contributions of bidirectional supervision, cycle-consistency feedback, and complexity regularization?

3 comments

Thank you for the feedback. We added a matched SFT–DUAL-BN evaluation without reference filtering and controlled ablations of bidirectional supervision, cycle feedback, and complexity regularization.

1. Graph reconstruction without reference filtering

We evaluated SFT and DUAL-BN on the same 50 descriptions. Each stage passed its own predictions directly to the next. We used reference graphs only after generation for matching and scoring.

ModelNode F1 ↑Mapped Edge F1 ↑Forward Pairwise KL ↓Bidir. KL ↓
SFT0.61200.70020.68000.4692
DUAL-BN0.68410.71880.43570.2937

Without reference filtering, DUAL-BN improved Node F1 by 0.0721 and mapped Edge F1 by 0.0186, while lowering both KL scores. KL measures probability agreement on recovered edges with comparable tables.

Joint generation. We also compared SFT and DUAL-BN when both models generated all graph components together in a single pass from the input description.

ModelNode F1 ↑State F1 ↑Mapped Edge F1 ↑Forward Pairwise KL ↓Bidir. KL ↓
SFT0.58110.63320.62841.04600.9238
DUAL-BN0.63440.61120.63680.91200.7200

Under joint generation, DUAL-BN improved Node F1 by 0.0533 and reduced both KL scores relative to SFT. State F1 decreased by 0.0220. Thus, the node recovery and probability agreement gains persisted in one-pass generation, with a trade-off in state recovery.

Mapped Edge F1 is the Edge F1 metric reported in the paper.

We align predicted nodes to reference nodes, then score directed edges after mapping their endpoints. Predicted edges with an unaligned endpoint are excluded, and reversed edges do not count as matches. In this evaluation, alignment is used only for scoring after generation.

2. Why DUAL-BN uses pairwise tables

DUAL-BN trains on the parent-specific pairwise tables provided by PRISM-BN [5, §3.2]. For a child with four binary parents, a full conditional probability table requires 16 distributions, one per joint parent assignment. The pairwise targets require eight, two per parent. This reduces the number of distributions to predict and lets us evaluate recovered edges when the model identifies only part of a node’s parent set. Our target captures parent-specific probabilities. Full Bayesian-network parameterization would also require modeling joint parent effects.

3. Novelty and baseline comparisons

CycleGT [1] uses back-translation to learn from unpaired text and graphs. Cao et al. [2] use validity and reconstruction rewards for semantic parsing. DUAL-BN builds on these ideas by training a model to recover graph structure and pairwise probabilities together. Each GRPO candidate jointly predicts variables, states, directed edges, and probability tables. The same candidate receives three scores:

  • Reference fidelity scores structural matches and numerical agreement. It distinguishes graphs that express the same qualitative relationship but assign different probabilities.
  • Cycle feedback asks the shared model to reconstruct the description from the predicted graph alone, then scores how much of the source meaning is preserved.
  • Excess complexity penalizes surplus counts relative to the reference, including represented probability parameters. Its edge count includes edges incident to unmatched nodes, which mapped Edge F1 excludes.

The cycle ablation shows the value of the reconstruction signal. With the starting checkpoint, fidelity weight, and excess penalty fixed, adding cycle feedback raised mapped Edge F1 from 0.7819 to 0.8415 and lowered forward pairwise KL from 0.7332 to 0.6594.

We prioritized a comparison against the same extractor’s SFT starting checkpoint to measure what DUAL-BN training adds. The gains persisted when the extractor had to pass its own predictions between stages. PRISM-BN’s prompted results provide context, and its full-CPD KL scores measure a different target from our pairwise KL. For regularization, we used a simple excess-count penalty and switched it on and off while holding the other rewards fixed. A size cap at decoding or pruning after generation would test a different intervention: controlling graph size at inference. To summarize Prior work. CycleGT uses cycle training to exploit unpaired data through back-translation. Cao et al. apply reconstruction rewards to logical forms. What is new here. DUAL-BN uses text reconstruction as a reward on a jointly generated probabilistic graph that contains continuous parameters. That reward is scored together with structural and numerical fidelity on the same GRPO candidate. Why it matters. Token-level SFT does not penalize a small edit that reverses a dependency or shifts a probability, while a graph-level reward does. Evidence. The cycle ablation isolates this signal's contribution beyond fidelity.

4. Component ablations

We evaluated all configurations on the same 73 descriptions from five held-out PRISM-BN backbones using the original staged evaluation. The table includes the penalty-off run from Table 3.

SFTGRPO rewardsExcess penaltyNode F1 ↑Conditional State F1 ↑Mapped Edge F1 ↑Forward Pairwise KL ↓Bidir. KL ↓Coverage ↑
Forward only——0.73220.81260.65780.87120.744167.20%
Bidirectional——0.77450.87380.67620.87070.616066.18%
Forward onlyFidelity0.050.73500.85620.76910.72290.598974.55%
BidirectionalFidelity0.050.86220.83710.78190.73320.544174.67%
BidirectionalFidelity + cycle00.80820.88370.68550.71890.530274.22%
Bidirectional (full DUAL-BN)Fidelity + cycle0.050.91060.84860.84150.65940.535782.30%

Coverage is the percentage of reference probability columns with comparable predictions.

  • Bidirectional supervision. With the GRPO rewards and penalty fixed, Node F1 rose from 0.7350 to 0.8622 and mapped Edge F1 from 0.7691 to 0.7819, with trade-offs in conditional State F1 and forward KL.
  • Cycle feedback. With the SFT checkpoint, fidelity reward, and penalty fixed, adding cycle feedback improved all six metrics. Coverage increased from 74.67% to 82.30%.
  • Complexity regularization. With fidelity and cycle feedback fixed, adding the penalty improved mapped Edge F1 (0.6855 → 0.8415) and forward KL (0.7189 → 0.6594), while conditional State F1 decreased and bidirectional KL rose slightly.

We will restructure the introduction so that each challenge maps to a design choice:

  • Cross-component consistency → a joint graph-level reward
  • A token-level loss that is blind to graph errors → GRPO over whole graphs
  • Information loss → cycle reconstruction

The pairwise-table discussion will move to the task definition, and detailed metrics will move to Results.

We hope these results address your concerns.

Reviewer_D2RhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Authors proposed DUAL-BN, a two-stage framework for reconstructing probabilistic graph components from natural language. The approach combines bidirectional supervised fine-tuning with GRPO using rewards for reference fidelity, text-graph consistency, and excess complexity. On held-out PRISM-BN backbones, the proposed method substantially improves Node and mapped Edge F1 over SFT, while improving forward pairwise KL but decreasing conditional State F1.

Critical strengths

The problem is interesting and the proposed combination of structural and semantic feedback is reasonable. Proposed DUAL-BN provides a coherent framework combining bidirectional text–graph supervision with GRPO-based structural and semantic feedback. The method shows substantial improvements over SFT, particularly for node and mapped-edge reconstruction, with improved forward pairwise KL on held-out backbones.

Critical weaknesses

The evidence currently comes from one extraction model, one benchmark family, a small held-out test set, and a very small human evaluation. Broader benchmarking, evaluation across multiple LLM backbones, stronger end-to-end evaluation, and a better-powered and more clearly documented human study would be needed to establish the generality of the proposed approach.

Reference matching determines downstream evaluation inputs, and probabilities are evaluated only for successfully recovered edges. The experiments do not separately isolate the effects of bidirectional supervision, cycle feedback, GRPO optimization, and the individual reward components, limiting attribution of the reported gains. Abalation study is missing.

Questions for authors
  1. How well does DUAL-BN generalize? Can the authors provide results on additional benchmark families and/or LLM backbones?

  2. Given the relatively small held-out test set, can the authors provide uncertainty estimates to establish the robustness of the reported improvements?

  3. Can the authors clarify and strengthen the human evaluation, including evaluator expertise, blinding/randomization, inter-rater agreement, and statistical uncertainty? Would a larger human study support the same conclusions?

  4. Since reference matching determines downstream evaluation inputs and probability metrics are computed only on successfully recovered edges, can the authors provide a fully end-to-end evaluation without reference-assisted filtering?

  5. Can the authors provide an ablation study isolating the contributions of bidirectional supervision, cycle-consistency feedback, GRPO optimization, reference-fidelity rewards, and complexity regularization?

2 comments
AuthorsOpenReview ↗

DUAL-BN Ablations, End-to-End Evaluation, and Robustness

Thank you for the constructive feedback. Our additional experiments show that GRPO improves on matched SFT without reference-assisted filtering and that cycle feedback improves reconstruction when the fidelity reward and excess penalty are held fixed. We also report an additional prompted-model evaluation, a reference sensitivity analysis, and details of the human study.

1. Generalization across descriptions and LLMs

We generated an additional set of descriptions using GPT-6.2 Astra and evaluated three prompted LLMs on the same inputs.

ModelNode F1 ↑State F1 ↑Edge F1 ↑
GPT-5.60.92500.87660.85542
GPT-6 Astra0.95200.91200.9060
Claude Opus 50.86520.88070.8893

All three models achieved Node F1 above 0.86 and Edge F1 above 0.85 on these descriptions. This experiment shows that multiple prompted LLMs can perform the reconstruction task. Our original held-out evaluation provides separate evidence that DUAL-BN transfers to source backbones excluded from training. Transfer of the DUAL-BN training gains to another base LLM or benchmark family remains untested.

2. Robustness and uncertainty

We evaluated DUAL-BN on the same descriptions and source backbones using the original PRISM-BN reference subgraphs and alternative subgraphs generated by GPT-6.2 Astra. With the alternative references, Node F1 increased by 0.09, mapped Edge F1 by 0.12, and conditional State F1 by 0.10. Forward pairwise KL decreased by 0.0003, and bidirectional KL increased by 0.0007.

The choice of reference affected structural F1 more than the aggregate KL scores on comparable tables. This experiment measures sensitivity to reference construction. To estimate uncertainty in the gains over SFT, we paired predictions on the same 73 descriptions and bootstrapped the five held-out source backbones. DUAL-BN improved Node F1 by 0.1361 (exploratory 95% bootstrap interval: [0.0622, 0.2190]) and mapped Edge F1 by 0.1653 ([0.0983, 0.2291]). Both intervals were above zero. We interpret them as exploratory given the small number of test backbones.

3. Human evaluation

The same three evaluators completed both tasks (Appendices C.7 and F), and all three were well versed in probabilistic graph reconstruction. Each task included ten cases, two from each of five domains. All evaluators rated every displayed output using the same 1–5 rubric, with no missing ratings. We randomized the order of model outputs in the subgraph evaluation.

Observed inter-rater agreement was 83% for nodes, 86% for states, and 86% for edges. Regularized DUAL-BN outscored SFT on all five subgraph criteria, including overall quality (4.10 vs. 3.41). It received the highest mean node, state, and edge ratings among the four models, while GPT-6 Astra received the highest overall and probability-table ratings.

Each model’s mean for a given criterion combines 30 ratings from three evaluators on the same ten cases. We interpret these differences descriptively because observations are repeated across cases and raters. We plan to include an expanded study in the final version, with more cases and independent evaluators, to test whether the differences persist and estimate their uncertainty.

4. Evaluation without reference-assisted filtering

We evaluated SFT and GRPO on the same 50 descriptions. Each stage received the preceding stages’ predictions directly. Reference graphs were used only after generation for matching and scoring.

ModelNode F1 ↑Mapped Edge F1 ↑Forward Pairwise KL ↓Bidirectional KL ↓
SFT0.61200.70020.68000.4692
GRPO0.68410.71880.43570.2937

The gains over SFT persisted without reference filtering: Node F1 increased by 0.0721, mapped Edge F1 increased by 0.0186, and both KL scores decreased. Mapped Edge F1 uses aligned endpoints, and KL measures probability agreement on recovered edges with comparable tables.

5. Ablation of supervision, GRPO refinement, and reward components

We evaluated all configurations on the same 73 descriptions from five held-out PRISM-BN backbones. The table combines the new ablations with the penalty-off comparison from Table 3. GRPO runs used seed 42 and 2,000 update steps.

SFTGRPO rewardsExcess penaltyNode F1 ↑Conditional State F1 ↑Mapped Edge F1 ↑Forward Pairwise KL ↓Bidir. KL ↓Coverage ↑
Forward only——0.73220.81260.65780.87120.744167.20%
Bidirectional——0.77450.87380.67620.87070.616066.18%
Forward onlyFidelity0.050.73500.85620.76910.72290.598974.55%
BidirectionalFidelity0.050.86220.83710.78190.73320.544174.67%
BidirectionalFidelity + cycle00.80820.88370.68550.71890.530274.22%
Bidirectional (full DUAL-BN)Fidelity + cycle0.050.91060.84860.84150.65940.535782.30%

Coverage is the percentage of reference probability columns with comparable predictions. Mapped Edge F1 is the Edge F1 metric reported in the paper. We align predicted nodes to reference nodes, then score directed edges after mapping their endpoints. Predicted edges with an unaligned endpoint are excluded, and reversed edges do not count as matches. This ablation uses the original reference-assisted staged protocol. In the separate 50-description evaluation in Section 4, reference alignment is used only after generation for scoring.

  • Bidirectional supervision. Adding graph-to-text examples improved Node F1 and mapped Edge F1 before GRPO. With the fidelity reward and excess penalty fixed during GRPO, it also raised Node F1 from 0.7350 to 0.8622 and mapped Edge F1 from 0.7691 to 0.7819, with trade-offs in conditional State F1 and forward KL.
  • GRPO refinement. Starting from the same bidirectional SFT checkpoint, adding fidelity-based GRPO with the excess penalty raised mapped Edge F1 from 0.6762 to 0.7819 and lowered forward KL from 0.8707 to 0.7332. This estimates the gain from the refinement stage as a whole. The independent effect of the fidelity reward was not separately estimated.
  • Cycle feedback. Holding the SFT checkpoint, fidelity weight, and excess penalty fixed, adding cycle feedback improved all six metrics. Mapped Edge F1 increased from 0.7819 to 0.8415, forward KL decreased from 0.7332 to 0.6594, and coverage increased from 74.67% to 82.30%.
  • Complexity regularization. With fidelity and cycle feedback fixed, adding the excess penalty improved mapped Edge F1 from 0.6855 to 0.8415 and forward KL from 0.7189 to 0.6594. Coverage increased by 8.08%, while conditional State F1 and bidirectional KL worsened.
Reviewer_ty7eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DUAL-BN, a two-stage framework for extracting probabilistic graphs from natural-language descriptions. The method first uses bidirectional supervised fine-tuning to learn text-to-graph extraction and graph-to-text reconstruction. It then applies Group Relative Policy Optimization (GRPO) to refine the predicted graph components, including variables, states, directed edges, and pairwise conditional probability tables. The reward function combines reference-graph fidelity, text reconstruction consistency, and a penalty for excessive graph components.Experiments on the PRISM-BN benchmark show that DUAL-BN improves node and edge F1 scores over the supervised fine-tuning baseline, while reducing the KL divergence of matched pairwise probability tables. However, the full model shows a decrease in conditional State F1. The paper also explores the integration of partial graphs into larger probabilistic structures and includes human evaluations of reconstruction quality.

Critical strengths
  1. The paper combines bidirectional text–graph learning with GRPO to improve the extraction of probabilistic graph components. The use of structural fidelity, semantic reconstruction, and complexity control in the reward function is a reasonable way to encourage more consistent graph predictions.
  2. The experimental results show clear improvements over the SFT baseline, particularly in Node F1 and mapped Edge F1. The comparison between GRPO with and without the complexity penalty is also useful, as it helps demonstrate the contribution of this component to the reported performance gains.
Critical weaknesses
  1. The paper uses reference matching to guide the prediction of subsequent graph components. Although this provides useful component-level results, it differs from the joint graph generation used during GRPO training. In addition, the edge and probability metrics are evaluated only on selected matched components, making it difficult to assess the overall quality of the recovered graphs. I would like to see an end-to-end evaluation without reference-assisted filtering to better understand the method's performance.
  2. The reported improvements are encouraging, but the contributions of some key components, particularly bidirectional supervision and semantic reconstruction feedback, are not evaluated separately. The main comparisons are also limited to SFT and GRPO variants. Moreover, the test set includes only 73 generated descriptions from five held-out backbones. Additional ablation studies and experiments on more diverse or real-world text would provide stronger evidence of the method's effectiveness and generalizability.
Questions for authors
  1. Can the authors demonstrate that graph-to-text supervision and cycle-consistency feedback provide improvements beyond forward-only supervision and graph-fidelity rewards?
  2. Given that natural language often provides only vague probabilistic information, how reliable are the specific probability values predicted by the model?
  3. Could the authors provide or discuss the performance of the model when generating complete graphs directly from text, without reference-assisted filtering? This would help assess whether the improvements observed under the staged evaluation protocol also hold in an end-to-end setting.
  4. How sensitive are the training rewards and evaluation results to the LLM-based semantic judgments? Could the authors include additional experiments to assess the reliability of these judgments?
1 comment
AuthorsOpenReview ↗

Evaluation Sensitivity to the Choice of Semantic Judge

Thank you for the constructive feedback. We added a controlled ablation to isolate graph-to-text supervision and cycle feedback, a pilot study of probability estimation, and the evaluation without reference-assisted filtering.

1. Graph-to-text supervision and cycle feedback

We evaluated five configurations on the same 73 descriptions from five held-out backbones using the original staged protocol. All runs used seed 42, both SFT configurations trained for three epochs, and each GRPO run used 2,000 steps and an excess penalty of λ=0.05\lambda=0.05. The GRPO comparisons held the fidelity coefficient fixed.

ConfigurationNode F1 ↑State F1 ↑Edge F1 ↑Forward KL ↓Bidir. KL ↓Coverage ↑
Forward-only SFT0.73220.81260.65780.87120.744167.20%
Bidirectional SFT0.77450.87380.67620.87070.616066.18%
Forward SFT + fidelity GRPO0.73500.85620.76910.72290.598974.55%
Bidirectional SFT + fidelity GRPO0.86220.83710.78190.73320.544174.67%
Full DUAL-BN (+ cycle feedback)0.91060.84860.84150.65940.535782.30%

State F1 is conditional on matched nodes, Edge F1 uses mapped endpoints, and coverage is the share of reference probability columns with comparable predictions.

Graph-to-text supervision improved node and edge recovery both before and after GRPO. With the same fidelity-based GRPO reward, it raised Node F1 from 0.7350 to 0.8622 and mapped Edge F1 from 0.7691 to 0.7819, with trade-offs in conditional State F1 and forward KL.

Cycle feedback improved all six metrics over bidirectional SFT with fidelity-based GRPO. Mapped Edge F1 rose from 0.7819 to 0.8415, forward KL fell from 0.7332 to 0.6594, and coverage increased by 7.63 percentage points. The full model improved mapped Edge F1 by 0.0724 over forward-only SFT with fidelity-based GRPO. These comparisons isolate the gains on generated descriptions from held-out backbones. Evaluation on naturally occurring text would extend this generalization test.

2. Reliability of predicted probabilities

We held the reference edge and ordered states fixed and varied what the description said about two target probabilities: explicit values, qualitative ordering, or no comparison cue. The table reports the same three cases valid in all conditions.

Text conditionTarget-entry MAE ↓Forward KL ↓Ordering agreement
Explicit values0.03920.37443/3
Qualitative ordering0.06140.21043/3
No comparison cue0.04770.36243/3

On the three shared cases, explicit wording yielded the lowest MAE on the two target probabilities, while qualitative wording yielded the lowest forward KL across the full distributions. All three conditions recovered the reference ordering, including when the cue was withheld. Qualitative wording can constrain the ordering while leaving the exact probabilities underspecified.

3. Graph generation without reference-assisted filtering

We evaluated the GRPO model on 50 descriptions using the four-stage prediction pipeline. Each stage passed its predictions directly to the next. Reference graphs were used only after generation for matching and scoring.

ModelNode F1 ↑Edge F1 ↑Pairwise KL ↓Bidirectional KL ↓
SFT0.61200.70020.68000.4692
GRPO0.68410.71880.43570.2937

The GRPO model achieved a Node F1 of 0.6841 and a mapped Edge F1 of 0.7188 without reference filtering during generation. The KL scores measure probability agreement on recovered edges with comparable tables.

4. Sensitivity to semantic judgments

We rescored the same predictions using GPT-6.1 Sol as an alternative semantic judge. Relative to the original judge, Node F1 decreased by 0.006 and mapped Edge F1 decreased by 0.002. These small changes quantify sensitivity to the judge used for evaluation. The rescoring did not vary the judge used during GRPO training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.