acceptodds
Under review as a conference paper at ICLR 2027

Overjet-1: A Multi-modal Generalist Vision-Language Model for Grounded Dental Image Understanding

Abstract

A complete dental examination is an inherently multi-modal, multi-view diagnostic task: no single imaging source provides complete insight into underlying pathologies. Clinicians aggregate evidence across sources such as panoramic and intraoral radiographs, color photographs, cephalograms, periodontal charts, and 3D CBCT scans. With the advent of deep learning and vision-language models (VLMs), significant progress has been made in dental image analysis. However, current VLMs and task-specific models process single images in isolation, missing crucial cross-modal context and allowing errors to compound across findings. We present Overjet-1, a generalist dental VLM built on Gemma 4 26B-A4B that natively processes multi-image, multi-modal inputs. Overjet-1 is post-trained on 1,631,594 supervised records spanning 122 dental tasks. These records are organized through the Dental Graph, a case-level representation that unifies tooth segmentations, finding regions, clinical measurements, and 3D geometry across all available multi-modal sources. The graph supplies the spatial coordinates for each training target, links observations of the same tooth across views, and constructs chain-of-thought (CoT) reasoning paths through graph traversal. We introduce Dental Consistency Reinforcement Learning (DC-RL), a novel framework extending Group Relative Policy Optimization (GRPO). DC-RL computes verifiable, deterministic reward signals directly from grounded facts in the Dental Graph, simultaneously optimizing diagnostic correctness and spatial localization without relying on extensive manual annotation or CoT distillation. Evaluated on MMOral-OPG-Bench, Overjet-1 achieves a new state of the art of 71.5%, compared to the best baseline of 61.3%. We also introduce Overjet-Bench, a more comprehensive benchmark covering 9 task families across single-image, multi-image, and multi-modal setups. Overjet-1 achieves a macro score of 78.8% on Overjet-Bench, outperforming top proprietary and open-source baselines (33.5%) significantly on real-world tasks. To our knowledge, Overjet-1 is the first model natively trained and evaluated on both multi-image and multi-modal dental tasks with grounded supervision. Code is available at https://anonymous.4open.science/r/Overjet-1-code-D205; Overjet-Bench will be released upon acceptance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language Models, Dental Image Analysis, Multi-Image Reasoning, Visual Grounding, Reinforcement Learning, Multimodal Learning

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_9K7JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Overjet-1, a dental vision-language model that processes multiple images and modalities within a patient examination. The authors construct a Dental Graph linking tooth identities, clinical findings, and geometry across views, and use it to generate 1.63 million supervised records covering 122 tasks, including annotation-derived reasoning traces. The model is trained through supervised fine-tuning and Dental Consistency Reinforcement Learning (DC-RL), which combines task correctness with grounding, evidence, format, and tooth-layout rewards. Evaluation includes the external MMOral-OPG benchmark, the patient-disjoint internal Overjet-Bench, external COde anomaly classification, specialist comparisons, and ablations on two backbones. The authors report 71.5% closed-answer accuracy and a 75.1% open-answer judge score on MMOral, and a 78.8 macro score on Overjet-Bench. Additional modality-removal and shuffling experiments assess input utilization, while non-dental experiments examine the broader usefulness of consistency rewards.

Critical strengths

S1: A useful case-level formulation and substantial grounded supervision. The Dental Graph explicitly links tooth identities, clinical findings, and geometry across views and modalities, allowing the same model to address localization, numbering, question answering, reporting, and cross-input tasks. This targets a practically important limitation of evaluating dental images independently, and the construction of 122 task types is a substantial contribution.

S2: A meaningful empirical package with relevant controls. The paper includes external MMOral and COde evaluations, patient-disjoint internal evaluation, SFT and DC-RL ablations, a second backbone, and case-level confidence intervals. Table 11 additionally evaluates modality removal and patient-mismatched inputs on two 200-case slices. These results provide substantive evidence of post-training gains and of dependence on the supplied modalities, beyond the qualitative examples.

Critical weaknesses

W1: The clinical validity of the main internal benchmark is not sufficiently established. Much training supervision comes from production systems, and Appendix A acknowledges clinically plausible annotation errors. Overjet-Bench uses the same clinical database, but Appendix F.1 does not make clear whether its reference boxes and clinical facts were independently adjudicated rather than inherited from source systems. Dentist verification is explicitly described for the displayed case studies (Appendix J), which does not establish benchmark-wide reliability. Moreover, Appendix G.3 assigns zero to parsing failures, and reports are assessed with ROUGE-L/token F1; no cross-model parsing audit or blinded agreement with clinical judgments is reported. Thus, the 45.3-point macro advantage may combine clinical improvement with agreement to annotation and output conventions. MMOral and COde support external transfer, but do not directly resolve these concerns for cross-view grounding and reporting. A bounded audit of existing benchmark references and predictions would make this central result more convincing.

W2: The reward specification makes the claimed isolated effect of spatial consistency difficult to interpret. For at least two predicted tooth boxes, Equation (6) defines the tooth-box term as 0.9 i + 0.1 s, whereas Appendix D.3's no-spatial control uses i for the corresponding term. The contrast therefore changes the absolute-grounding coefficient as well as removing consistency, despite statements that absolute grounding remains unchanged (Appendix G.4). Table 3's 3.4-point gain supports the full reward design, but not unambiguously the addition of s at fixed weights. Also, Appendix D.3 describes a sum of geometry components, while Algorithm 3 selects them through mutually exclusive branches. The handling of a target containing both tooth and lesion geometry, including how the grounding score remains bounded, needs clarification. Please identify the actual implementation and reconcile the formulae and the interpretation of the ablation.

Questions for authors

Q1: How were Overjet-Bench reference boxes and clinical facts obtained, and which were independently annotated or adjudicated by dentists rather than derived from the source systems? Please distinguish benchmark-wide validation from the verified examples in Appendix J. Can existing audit results, or a small blinded subset of the current test records and model predictions, quantify reference errors and agreement between the automatic scorer and clinical correctness? A per-model parsing-failure breakdown would also help assess whether the ranking is sensitive to output conventions.

Q2: What exact geometry aggregation and no-spatial reward were used in the reported runs? In particular, does the control retain the 0.9 coefficient on absolute tooth-box grounding, and are tooth boxes, finding boxes, and keypoints scored jointly or through exclusive branches? If the implementation differs from the written definition, please correct it. If the coefficient changes, please provide any existing weight-matched evidence or qualify the conclusion as a comparison between reward designs rather than an isolated addition of spatial consistency.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9K7J — Part 1/2

Thank you very much for the detailed review. We greatly appreciate that you recognized the case-level formulation of the Dental Graph, the scale of the grounded supervision across 122 task types, and the empirical package of external evaluations, patient-disjoint evaluation, ablations on two backbones and case-level confidence intervals. Your questions concern the reliability of the benchmark and the exact reward specification, and we have addressed both in the revision. See below:

Q1, Part 1. Benchmark annotation

How were Overjet-Bench reference boxes and clinical facts obtained, and which were independently annotated or adjudicated by dentists rather than derived from the source systems?

We first note that we have revised and substantially extended Overjet-Bench in the revised manuscript to thoroughly evaluate the efficacy of the model on multi-image and multimodal tasks. The benchmark now includes 3,000 QA pairs across single-image, multi-image and multimodal questions.

The ground truth, including reference boxes and clinical facts for teeth, anatomy, restorations, lesions and periodontal status, was constructed by a team of 31 expert annotators: 17 specialists and 14 general dentists. The test set was labeled from scratch, with no pre-filling from the source systems. Each case was labeled by one annotator and then checked by one reviewer, so every case was reviewed by at least two experts.

From the labeled clinical facts, QA pairs were constructed using the same generation step as the training QA pairs. The validity of the QA pairs was checked manually by a team of 31 annotators. The final 3,000 QA pairs were selected from the validated pairs.

Q1, Part 2. Benchmark validation and illustrative examples

Please distinguish benchmark-wide validation from the verified examples in Appendix J.

The overjet-bench annotation/adjudication is separate from the verification of the illustrative examples in Appendix J. Each QA pair correctness was validated by the team of expert annotators, whereas Appendix J lists examples of model predictions.

Q1, Part 3. Scoring agreement and clinical correctness

Can existing audit results, or a small blinded subset of the current test records and model predictions, quantify reference errors and agreement between the automatic scorer and clinical correctness?

Agreement between scoring metrics and clinical correctness

We distinguish evaluation criteria by question type. Multiple-choice answers are evaluated by matching the selected option to the reference. Grounding is evaluated through task-specific box matching at IoU ≥ 0.5, with tooth-identity and source-view requirements where applicable. These metrics quantify agreement with the annotated targets (Appendix F.1).

For whole-image and per-tooth reports, we use ROUGE-L; open clinical synthesis uses token F1. These metrics measure content overlap with the dentist-annotated and checked references. Because the outputs mostly describe structured dental information, overlap provides a useful signal of agreement with the reference content. The clinical correctness is more directly evaluated by fact based questions (multiple-choice, grounding, and keyword extraction, etc.). The report-generation family also only accounts for 135 of the 3,000 benchmark examples (4.5%), including 90 evaluated with ROUGE-L (3.0%) and 45 with token F1 (1.5%) (Table 12).

AuthorsOpenReview ↗

Response to Reviewer 9K7J — Part 2/2

Q1, Part 4. Parsing failures and ranking sensitivity

A per-model parsing-failure breakdown would also help assess whether the ranking is sensitive to output conventions.

A per-model parsing-failure breakdown

The system prompt used in our evaluation specifies a structured output format; however, models may occasionally fail to comply despite using a standardized answer schema and parser. We therefore quantified, for each model, how often outputs failed to parse, separately from parseable but incorrect answers (Table 16). Across the 12 evaluated models, the average first-pass parsing-failure rate was 1.18%. MedVLM-R1 had the highest rate at 3.17% (95/3,000), compared with 0.13% (4/3,000) for Overjet-1. To assess the effect on the results, we reran failed outputs with up to three additional generations using the same input, system prompt, and parser for each model. After retries, every model’s Overall score changed by at most 1.03 points, and the top four models retained their rankings. Overjet-1’s parsing failures decreased to zero, its score changed from 73.53 to 73.66, and it remained first. Thus, recovering parsing failures leaves the main conclusion unchanged.

Q2. Geometry aggregation and the no-spatial control

What exact geometry aggregation and no-spatial reward were used in the reported runs? In particular, does the control retain the 0.9 coefficient on absolute tooth-box grounding, and are tooth boxes, finding boxes, and keypoints scored jointly or through exclusive branches? If the implementation differs from the written definition, please correct it. If the coefficient changes, please provide any existing weight-matched evidence or qualify the conclusion as a comparison between reward designs rather than an isolated addition of spatial consistency.

We thank the reviewer for the close reading of the reward definition and apologize for the inconsistent descriptions. In our experiments, the no-spatial control retained the 0.9 coefficient on absolute tooth-box grounding: for predictions with at least two valid tooth boxes, the full reward uses 0.9i + 0.1s, while the control uses 0.9i. The single-box behavior and the other reward terms remain unchanged, and the topology-only contribution for positive targets without reference geometry is disabled in the control. We have corrected the no-spatial expression in Appendix D.3 and aligned it with Appendix G.4 so that the fixed-coefficient comparison is explicit.

Your comments have helped us clarify the benchmark validation and reward specification. We would be happy to address any remaining questions. If our responses and additional results resolve your concerns, we would be grateful if you could consider a higher score.

Reviewer_9K7JOpenReview ↗

Post-Rebuttal Assessment

Thank you for the detailed response and additional analyses. The clarification regarding independent expert annotation substantially addresses my concern about reference-label provenance. The parsing-failure analysis also alleviates my concern about output-format effects on the ranking. I regard the fixed-coefficient no-spatial control as clarified and withdraw the corresponding concern.

My remaining concern is the validation of clinical text-output evaluation. Expert verification of reference annotations and QA validity strengthens the benchmark, but does not establish how well ROUGE-L and token F1 reflect the clinical correctness of model-generated reports and syntheses. The response explains these metrics but does not provide the requested agreement analysis or a blinded clinical audit of model predictions.

The sample proportion of text-output questions also does not directly establish their contribution to Overall. Under the original nine-family macro average, the reporting family contributes one ninth regardless of its sample count. Please clarify the revised aggregation weights. A small clinical-scoring audit would help establish the practical impact of this remaining concern.

The response has improved my assessment and substantially narrowed my concerns. I nevertheless retain my Weak Reject recommendation at present because the validation supporting the clinical text-quality claims remains incomplete.

Reviewer_o6hQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper develops a dental vision-language model for interpreting multiple images and modalities from the same examination. Dental Graph links tooth identities, findings, and geometry to generate grounded answers and reasoning traces; DC-RL rewards answer correctness and spatial and evidence consistency. Overjet-Bench covers 29 tasks in nine families. The main claim is consistent case-level interpretation with explicit image evidence. Validation covers Overjet-Bench and external dental benchmarks, with ablations of grounding supervision, reasoning traces, and spatial consistency.

Critical strengths

Overjet-1 addresses a convincing clinical need and combines substantial data construction with a benchmark for grounded interpretation across dental images and modalities. The 71.5% closed-question accuracy on MMOral and 76.63% zero-shot accuracy on COde provide useful external evidence of the system's effectiveness.

The matched spatial-consistency ablation in Table 3 improves the Overjet-Bench macro score by 3.4 points, with a paired 95% interval of [1.2, 5.6]. This supports the utility of the dental spatial constraints.

Critical weaknesses

The distinctive benefit of cross-view supervision is not sufficiently isolated. DentVLM already uses rationales and anatomical-region supervision, although this differs from cross-view coordinate grounding. Table 11 supports multi-source utilization, but changing test inputs does not isolate the training contribution of Dental Graph's identity links. Table 3 removes coordinates or reasoning traces and omits cross-input scores. DC-RL's spatial ablation tests within-view constraints, leaving the separate contribution of evidence reward e unclear. More targeted comparisons are needed to distinguish better single-image predictions from better cross-view association.

The reliability of generated supervision needs stronger quantitative support. Gemini 3.5 Flash-Lite usually rewrites traces without seeing source images. The checks in C.2 are useful, but remaining semantic errors and inherited annotation errors are not quantified. Audited tasks and dentist-verified examples do not give representative error estimates for generated traces and benchmark references. Tables 12–13 and F.1 specify localization recall at IoU 0.5, which does not penalize unsupported extra regions. Precision and F1, with tooth-identity matching rules made explicit, would help assess returned evidence. G.3's broader metric description also needs alignment with the reported score.

Questions for authors

Please first report cross-input scores for full SFT, DC-RL without spatial consistency, and full DC-RL. If feasible, compare independent per-view targets with joint cross-view targets, keeping images, tooth labels, coordinates and training budget matched. Shared tooth labels would still provide association cues; the comparison should test the added benefit of joint supervision.

Could the authors ablate evidence reward e with other rewards and training conditions fixed, and report wrong-tooth and wrong-view attribution rates on cross-input cases? D.3 retains the source view for reference geometry but defines clinical claims as tooth–finding–polarity tuples. How does the evidence parser or reward catch a correct finding assigned to the wrong image?

What fraction of Gemini-generated targets was rejected or corrected, and what errors remain in a representative clinical audit of generated traces and benchmark references? Please distinguish rewriting errors from inherited annotation errors and clarify how the checks detect changes in clinical meaning.

F.1 specifies within-view recall at IoU 0.5. Does tooth localization additionally require correct tooth identity? Please report precision and F1 alongside the 90.8 recall score using the same matching rules, and align G.3's description with the metrics actually reported.

8 comments
AuthorsOpenReview ↗

Response to Reviewer o6hQ — Part 1/3

Thank you very much for the thoughtful review. We greatly appreciate that you recognized the clinical motivation of the work, the external evaluations on MMOral and COde, and the value of the spatial-consistency results, and we have used your critical feedback on experimental setups to improve the paper. In summary:

  • W1 / Q1 (joint cross-view supervision): cross-input scores for full SFT and both DC-RL variants, and a matched comparison of per-view and joint supervision (Section H.4, Tables 3 and 21).

  • W1 / Q2 (evidence reward): a matched ablation of e, wrong-tooth and wrong-view attribution rates, and a clarification of how the evidence reward handles source views (Appendix D.3).

  • W2 / Q3 (reliability of supervision): an audit of Gemini-rewritten targets and how the benchmark references were annotated (Appendix C.2, Appendix F.1).

  • W2 / Q4 (localization metrics): the tooth-identity matching rule is stated explicitly in Appendix F.1; supplementary precision and F1 alongside recall are reported in our response below. We address each question in detail below.

Q1. Joint cross-view supervision

Please first report cross-input scores for full SFT, DC-RL without spatial consistency, and full DC-RL. If feasible, compare independent per-view targets with joint cross-view targets, keeping images, tooth labels, coordinates and training budget matched. Shared tooth labels would still provide association cues; the comparison should test the added benefit of joint supervision.

We acknowledge that the distinct benefit of cross-view supervision was not sufficiently demonstrated or isolated in the submission. We agree that a controlled comparison of joint and per-view supervision is a valuable analysis, and we thank the reviewer for the precise suggestion on how to set it up. We have added both of the requested results.

Cross-input scores for full SFT, DC-RL without spatial consistency, and full DC-RL

The revised Table 3 reports multi-image (MI) and multimodal (MM) scores for the three variants. Note that in the revised manuscript, Overjet-Bench has been extended to include more multi-image and multimodal questions.

VariantMIMM
+ SFT (full)62.361.0
+ DC-RL, no s63.061.8
+ DC-RL73.671.4

On multi-image and multimodal questions, DC-RL with no spatial consistency improves over full SFT by 0.7 and 0.8 points, and full DC-RL adds a further 10.6 and 9.6 points. This pattern suggests that spatial consistency is particularly useful for questions requiring tooth identities and locations to be linked across images.

Independent per-view targets compared with joint cross-view targets

To isolate the training contribution of joint supervision, we trained models under matched conditions that differ only in the targets: per-view targets or joint cross-view targets. Both conditions use the same images, tooth labels, coordinates, initialization and training budget, and shared tooth labels remain available in both, so the comparison tests the added benefit of joint supervision over the association cues that shared labels provide. All variants use the same 10% subset of training cases. We ran this comparison at each training stage (SFT, DC-RL with no spatial consistency, and full DC-RL), giving six models in total. The results are reported in Section H.4 and Table 21 of the revised manuscript:

StageTargetsMMOral ClosedS Loc. F1S Ident. F1MI/MM Loc. F1MI/MM Ident. F1
SFTPer‑view54.460.465.344.549.7
SFTJoint55.863.667.953.359.0
DC‑RL, no spatialPer‑view57.268.071.749.355.8
DC‑RL, no spatialJoint58.570.674.059.065.1
DC‑RL, fullPer‑view60.170.274.854.360.4
DC‑RL, fullJoint61.572.177.065.171.0

Under matched conditions, joint supervision improves every reported metric at all three training stages. On single-input questions, it improves localization and identification F1 by 1.9–3.2 and 2.2–2.6 points. On cross-input questions, it improves them by 8.8–10.8 and 9.3–10.6 points. The larger gain on cross-input questions shows that most of the benefit comes from cross-view association, which is what training the model on multiple views of the same tooth provides. The gain on single-input questions also shows that joint supervision improves the model's understanding of individual images.

This shows the benefit of training the model to observe multiple instances of the same tooth simultaneously, and that the gain comes from cross-view association. We believe it has strengthened the key contribution of the paper, which is now more clearly presented.

AuthorsOpenReview ↗

Response to Reviewer o6hQ — Part 2/3

Q2. Evidence reward and attribution

Could the authors ablate evidence reward e with other rewards and training conditions fixed, and report wrong-tooth and wrong-view attribution rates on cross-input cases? D.3 retains the source view for reference geometry but defines clinical claims as tooth–finding–polarity tuples. How does the evidence parser or reward catch a correct finding assigned to the wrong image?

We have added the matched ablation of e and the attribution rates, and we clarify below how a finding assigned to the wrong image is handled.

Roles of the evidence reward components e

Appendix D.3 (paragraph “Evidence and format scores e and f”) separates the roles of the reward components. In particular, evidence reward e matches complete graph-trace patterns to the Dental Graph, including source-image IDs for image-bound claims. Table 3 reports results with and without e; the table below also includes Ground and pooled MI+MM scores:

VariantMMOral ClosedOverjet-Bench OverallGroundMI+MM
Without e68.067.877.465.2
Full DC-RL71.573.579.372.4

Wrong-tooth and wrong-view attribution rates

In order to quantify the effectiveness of e, we further report wrong-tooth and wrong-view attribution rates on cross-input cases during training. Table 18 reports the rates at matched training checkpoints on the same held-out MI/MM cases. Attribution is determined by parsing predicted tooth and source-view IDs.

Wrong-tooth and wrong-view rates (%).

CheckpointNo e: toothNo e: viewWith e: toothWith e: view
0%22.518.722.518.7
25%20.019.416.414.7
50%18.317.116.810.3
75%19.116.211.111.0
100%17.316.89.06.9

Mechanics of evidence parser

DC-RL assigns evidence rewards by extracting task-specific graph-trace patterns using key tokens, such as “tooth #N in image X.” Image-bound patterns retain tooth and source-image IDs together with any required finding and polarity. Only a fully extracted pattern that matches the Dental Graph receives evidence credit. Therefore, a claim assigned to the wrong image receives no evidence credit. Table 18 measures wrong-tooth and wrong-view attribution from the explicit predicted IDs.

These points clarify the contribution and mechanics of evidence reward e, and we showed that this has a measurable impact on the final score.

Q3. Reliability of supervision

What fraction of Gemini-generated targets was rejected or corrected, and what errors remain in a representative clinical audit of generated traces and benchmark references? Please distinguish rewriting errors from inherited annotation errors and clarify how the checks detect changes in clinical meaning.

We want to clarify that the original construction already fixes the clinical content before language rewriting: the Dental Graph query selects the teeth, findings, relations, and geometry; short answers are rendered deterministically; and Gemini rewrites the supplied facts and reasoning traces (Sec. 3.1; Appendix C.2). For most tasks, we prompt the teacher model with only the structured facts, without the source image, so that it could focus on diversifying wording while keeping clinical information and coordinates consistent with source annotations. Appendix C.2 and Table 7 describe checks on the schema, tooth identities, coordinates, clinical claims, and graph alignment. For Overjet-Bench, the test set was annotated by 31 trained dentists (17 specialists and 14 general dentists).

To further quantify this process, the annotation team additionally audits the fractions of generated targets during the rewrite. On 500 randomly sampled Gemini-diversified text pairs, only 0.4% (2) of the pairs were flagged as not completely based on the source facts. In one, Gemini described the teeth as “permanent” without supporting information from the source facts. In the other, it added a description of the crown. In both cases the question and answers are still correct.

AuthorsOpenReview ↗

Response to Reviewer o6hQ — Part 3/3

Q4. Localization metrics

F.1 specifies within-view recall at IoU 0.5. Does tooth localization additionally require correct tooth identity? Please report precision and F1 alongside the 90.8 recall score using the same matching rules, and align G.3's description with the metrics actually reported.

We apologize for not making the localization criteria clear in the manuscript. Correct tooth identity is a prerequisite for matching in tooth-localization tasks: a predicted box counts as a correct match only when it identifies the correct tooth, belongs to the same source view, and has IoU ≥ 0.5 with the reference box under one-to-one matching. This identity requirement was already applied in the Overjet-Bench results reported in Tables 2 and 3.

Joint semantic accuracy (JSA) evaluates whether all requested fields of a structured answer are jointly correct; it is separate from box recall. Extra predictions affect precision and may invalidate JSA, but do not directly reduce recall.

We have updated the scoring table and descriptions in Appendix F.1 and aligned Appendix G.3 so that these criteria are explicit. The revised descriptions give the task-specific scoring criteria for the benchmark, which our annotation team has also checked. We now also report question-averaged precision, recall, and F1 using the same matching rules in Appendix F.1 and the table below. The results shown here cover the 810 grounding questions. Precision counts every unmatched or duplicate prediction as a false positive, so it penalizes unsupported extra regions, as the reviewer suggested.

Model*RecallPrecisionF1
Overjet-184.793.887.0
GPT-5.6 Sol21.334.224.5

*only top 2 performing models shown.

Overjet-1 reaches 93.8 precision and 87.0 F1, compared with 34.2 and 24.5 for the second-best model.

Concluding Remark

Thank you again for your careful review and constructive suggestions. Your questions led us to make the contribution of cross-view supervision, the role of the evidence reward and the reliability of our supervision much more explicit in the paper. We would be grateful if you could take another look at the revised version. We would be glad to clarify or expand on any part of our response.

Reviewer_f8JhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Overjet-1, a 26B-parameter dental vision-language model for reasoning over multiple images and modalities from one examination. The authors introduce a Dental Graph that links tooth identities, findings, measurements, and spatial geometry, and use it to derive grounded supervision and annotation-based reasoning traces. Dental Consistency Reinforcement Learning (DC-RL) adds deterministic rewards for clinical correctness, evidence support, localization, and cross-view consistency. Evaluations on MMOral-OPG-Bench and the proposed Overjet-Bench report strong gains over open and proprietary baselines. The benchmark spans localization, identification, reports, and cross-input clinical reasoning. Multi-modal shuffling studies, external anomaly transfer, and specialist-model comparisons further examine its generality.

Critical strengths
    1. This paper uses a Dental Graph to organize records. The Dental Graph integrates multifaceted information from multimodal data, connects different views through spatial coordinates, and constructs CoT reasoning path. Based on this, a generalist dental VLM, Overjet-1, is trained.
  1. This paper proposes Dental Consistency Reinforcement Learning (DC-RL), which extends GRPO and simultaneously optimizes diagnostic accuracy and spatial localization.
  2. This paper proposes Overjet-Bench, a more comprehensive benchmark covering nine task categories under single-image, multi-image, and multimodal settings.
Critical weaknesses
  1. The paper targets multi-image, multimodal reasoning, but Overjet-Bench is dominated by single-image records: 825 out of 1,000, versus 85 multi-image and 90 multimodal. Therefore, the existing evaluation results do not adequately support the paper’s central claim.
  2. Baseline comparisons are not fully controlled. Overjet-1 is trained with DC-RL on the dataset proposed in this paper, whereas the comparison methods are typically zero-shot or use other methods on different datasets. This makes it difficult to isolate the advantages of the proposed dataset and the DC-RL method over the baselines.
  3. There is a lack of further ablation studies on DC-RL, including ablations of g and e in the reward proposed in the paper.
  4. DC-RL involves many parameters, which makes parameter selection difficult. The paper lacks a sensitivity analysis of these parameters.
  5. The paper used the Gemini model in dataset construction, but the Open split of MMOral-OPG-Bench also uses a Gemini model as the judge. This may lead to bias toward models from the same family.
Questions for authors

1.What are the experimental results of the model and the comparison algorithms specifically on multi-image, multimodal data? 2. How can the advantages of the newly constructed dataset and DC-RL over other datasets and methods be demonstrated separately? 3. What are the ablation results for each individual component of DC-RL? What about the parameter sensitivity analysis? 4.What are the results on the Open split of MMOral-OPG-Bench when models from other families are used as judges?

4 comments
AuthorsOpenReview ↗

Response to Reviewer f8Jh — Part 1/3

Thank you very much for the review, feedback and suggestions. We greatly appreciate that you recognized the value of the Dental Graph, DC-RL and Overjet-Bench, and we have taken this opportunity to improve the paper on each of the points you raised. In summary:

  • Q1 (multi-image and multimodal results): Overjet-Bench is extended to 3,000 expert-validated QA pairs, and results for all models are now reported separately for single-image, multi-image and multimodal questions (Section 5, Table 2).

  • Q2 (contribution of dataset and DC-RL): Ablations on two backbones separate the gain from the dataset and the gain from DC-RL (Section 5).

  • Q3 (DC-RL subcomponents and parameters): An ablation of each reward term and a sensitivity analysis are added (Section H.4).

  • Q4 (choice of judge): The Open split is re-evaluated with judges from two other model families (Appendix H.1, Table 17).

Q1. Multi-image and multimodal results

What are the experimental results of the model and the comparison algorithms specifically on multi-image, multimodal data?

As the reviewer pointed out, the original Overjet-Bench included a large proportion of single-image QA pairs, which made it difficult to interpret the efficacy of the proposed framework on multi-image and multimodal evaluation. In order to clarify this, we have revised and substantially extended Overjet-Bench as follows:

For task families, we have added single-image (S), multi-image (MI) and multimodal (MM) questions. MI and MM further contain inherently multi-image and multimodal tasks that have no single-image counterpart (e.g., orthodontic qualification).

The benchmark now consists of 3,000 expert-validated QA pairs: 825 S, 1,040 MI and 1,135 MM. The three cohorts are now comparable in size, and 72.5% of the questions are MI or MM, so the final results are predominantly driven by performance on multi-image and multimodal tasks.

Results on MI/MM tasks

The results are shown in Table 2. The proposed model achieved 73.6 on MI tasks, 71.4 on MM tasks, and 72.4 for MI/MM combined, leading the second-best model (Muse Glimmer-30B) by +35.0, +34.3, +34.6 points, respectively.

Comparing per-view and joint supervision also enabled us to evaluate the benefit of joint cross-view targets. In particular, the per-view model predominantly failed when answering required using a second image or comparing and reasoning across images (failed 131 out of 200 cases). Overall S/MI/MM results are reported in Section 5, Table 2; illustrative cross-view outputs appear in Figure 4 and Appendix J (Figure J4). It shows a gap that is not yet addressed by the existing literature, and the importance of using multi-image, multimodal context for dental tasks.

We are grateful to you for specifically highlighting this point. We believe it has strengthened the key contribution of the paper, which is now more clearly presented.

AuthorsOpenReview ↗

Response to Reviewer f8Jh — Part 2/3

Q2. Contributions of the dataset and DC-RL

How can the advantages of the newly constructed dataset and DC-RL over other datasets and methods be demonstrated separately

In order to separately demonstrate the advantages of the dataset construction and of DC-RL, we have revised the manuscript as follows. The table below summarizes the experiments.

Gemma 4 26B-A4B

VariantMMOral closedOJ Overall
A1. Base (no training)46.418.6
A2. SFT (no grounding, no graph-traced reasoning)53.232.8
A3. SFT (grounding only)60.751.1
A4. SFT (full: grounding and reasoning traces)65.464.0
A5. DC-RL (no spatial consistency)67.065.9
A6. DC-RL (full)71.573.5

Qwen2.5-VL-7B

VariantMMOral closedOJ Overall
B1. OralGPT (Hao et al., 2025a)46.211.8
B2. SFT on Dental Graph records59.952.1
B3. DC-RL (full)62.154.9

Advantage of Dental Graph Dataset Construction

The revised manuscript dedicates Section 5. - “Ablations” and Appendix H.4 to the contribution of the dataset to the overall performance. By utilizing the dataset alone (trained directly on image and answer, A1→A2), the model gained +14.2 points, and by adding the proposed grounding and reasoning traces (A2→A4), it gained a further +31.2 points on the Gemma 4 backbone. On the OralGPT backbone, training on the Dental Graph records gained +40.3 points over OralGPT (B1→B2), even though our model was not trained on any MMOral images or question templates. This ablation is valuable because it shows how much of the improvement comes from the dataset alone and how much comes from the improved supervision signal (grounding and reasoning).

In the extended test set (see our response to Q1), 22.3% of the questions use a question style that does not appear in training. On these questions, the SFT model improves over the base model in all task families: by 36.4 points on average, with the largest gain on Ground (65.0 points) and the smallest on Modality (4.3 points).

Advantage of DC-RL

The revised manuscript dedicates Section 5 - “Effect of DC-RL” and Appendix H.2 and H.4 to the contribution of DC-RL. By utilizing DC-RL, the Gemma 4 and OralGPT backbones gained a further +9.5 and +2.8 points, respectively (A4→A6, B2→B3). Compared with standard GRPO using a correctness-only reward (see Q3), DC-RL gains +7.3 points, which shows that the reward design adds value on top of RL itself. We saw the most benefit of DC-RL in cross-input reasoning, tooth identification, and numbering, where the SFT model frequently made errors on adjacent bounding boxes and DC-RL reduced such errors by 46.1%.

We note that DC-RL is built on the proposed dataset by design and is not completely isolatable. It uses the dense entity bounding boxes and the graph relations to reward correctness and consistency, which suits dental tasks because a single task often requires an understanding of the surrounding anatomical structures. To show the value of the consistency reward beyond our data, Appendix I (Generality of the Consistency Reward; Table 22) applies it with reward-ranked fine-tuning to a separate dataset and problem.

In conclusion, the above ablation studies show that both the dataset construction and DC-RL have a measurable impact on the final result, and that the improvement can be attributed quantitatively to each.

AuthorsOpenReview ↗

Response to Reviewer f8Jh — Part 3/3

Q3. Reward components and parameter sensitivity

What are the ablation results for each individual component of DC-RL? What about the parameter sensitivity analysis?

We acknowledge that the ablation study in the original manuscript did not give a full understanding of DC-RL and its parameters. In order to address this, we have added an evaluation of each DC-RL component and a sensitivity analysis of the reward weights.

Contribution of each component

DC-RL has five components: (1) the correctness reward c, (2) the grounding reward g, (3) the graph-trace evidence reward e, (4) the spatial-consistency term s (within g) and (5) the format reward f. To examine their contributions more directly, we extended the ablations to include correctness-only RL and variants that each remove one component. Each RL variant starts from the same full-SFT checkpoint and uses the same training data and full training schedule. The results are shown in Table 3 and the table below.

VariantMMOral closedOJ OverallGroundMI+MM
Full SFT (starting point)65.464.070.761.6
Correctness only (c)67.666.270.063.7
Remove g (grounding)68.469.043.969.1
Remove e (graph-trace evidence)68.067.877.465.2
Remove s (spatial consistency)67.065.979.362.4
Remove f (format)71.171.375.269.8
Full DC-RL71.573.579.372.4

Compared with full DC-RL, removing the grounding reward g lowered Ground by 35.4 points, removing the graph-trace evidence reward e lowered MI+MM by 7.2 points, and removing the spatial-consistency term s lowered MI+MM by 10.0 points. The grounding reward g directly evaluates geometric accuracy, whereas c checks tooth identities or the presence of the requested coordinates on localization tasks. These ablations show the complementary contributions of the reward components.

Sensitivity Analysis

The revised manuscript now includes, in Section H.4, a sensitivity analysis of the reward weights (λ_g, λ_e, μ, α) and of the coefficient controlling the spatial-consistency contribution, with the results shown in Table 20 and Figure H1. Across the tested range, the OJ Overall score varied by at most 1.6 points, and every setting remained above full SFT. This indicates that training is relatively insensitive to the choice of weights and that the reward formulation is robust. We believe these additional experiments clarify the contribution and sensitivity of each part of DC-RL.

Q4. Alternative judges

What are the results on the Open split of MMOral-OPG-Bench when models from other families are used as judges?

We thank the reviewer for raising the potential bias from using the same model family for dataset construction and for evaluation. To this end, we make two remarks. First, we clarify the role of the LLM judge in the Open split. Second, we report evaluations with judges from other model families.

Open split and role of the LLM judge

The revised manuscript explains the role of the LLM judge in Appendix F.1. The judge receives the question, reference answer, and model prediction, and assigns a score from 0 to 1 using the benchmark’s original nine-example rubric, which illustrates fully correct (1.0), incorrect (0.0), and partially correct answers with intermediate scores. We assess sensitivity to the choice of judge by rescoring the same saved predictions with two additional judge models, as reported below.

Results from alternative judges

In order to evaluate the impact of the choice of judge, and to validate the point above, we re-evaluated the Open split using two additional judges from different model families: GPT-5.6 Sol and Claude Opus 5. We scored the same saved predictions from all evaluated models, keeping the reference answers, scoring prompt and aggregation procedure unchanged. The results are presented in Appendix H.1, Table 17

ModelGemini 3.7 FlashGPT-5.6 SolClaude Opus 5
Overjet-175.172.1478.45
Gemini 3.1 Pro63.657.5361.88
GPT-5.6 Sol60.956.8167.69
Opus-559.3060.0762.67

Overjet-1 ranks first under all three judges, leading the second-best model by 11.50, 12.07, and 10.76 points under Gemini 3.7 Flash, GPT-5.6 Sol, and Claude Opus 5, respectively. Rankings are positively correlated across judges (Kendall's τ = 0.697-0.879). The closed split, which does not use a judge, also ranks Overjet-1 first. Together, these two points show that the conclusions are robust to the choice of judge.

Concluding remark

Thank you again for your time and for the constructive review. Your questions have led to a clearer and more complete paper, and we would be grateful if you could take another look at the revised version. We would be glad to clarify or expand on any part of our response.

The author has addressed most of my questions, and I intend to improve my score.

Comments on the paper

1 comment

Dear Reviewers,

The rebuttal period is coming to an end. Please read the authors' responses, follow-up responses and update your reviews if appropriate. It would be helpful to share your thoughts as soon as possible to allow enough time for a productive discussion. Thank you!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.