AgenTeeth: A Model-Agnostic Framework for Suppressing Hallucination in Frozen Vision-Language Models on Dental X-Rays via Tool Evidence Injection
Ahmed Rafid, Fariya Ahmed, Rumman Adib, Mehedi Ahamed, Ajwad Abrar, Tareque Mohmud Chowdhury
cs.CV
Quick verdict
Injecting evidence from seven dental detectors into frozen VLMs gains 12.9-23.0 pts on MMOral-OPG-Bench; strong controlled design, but single-run, no variance.
Summary
The paper presents AgenTeeth, a training-free, model-agnostic agentic framework that suppresses spatial hallucination in frozen vision-language models (VLMs) reading panoramic dental radiographs. Seven RF-DETR-based experts (FDI tooth numbering, alveolar bone loss, jaw structures, dental history, periapical lesions, caries, impacted teeth) are invoked by a question-aware dispatcher; a deterministic grounding layer binds detections to FDI tooth numbers; a fresh synthesis call then answers from a structured evidence packet plus annotated overlays, instructed to prioritize tool evidence. On MMOral-OPG-Bench (491 closed, 578 open questions over 100 radiographs), four backbones gain 12.9-23.0 mean percentage points over a matched detector-free pipeline; the best configuration reaches 74.1% closed and 65.7% open, above OralGPT-Plus (45.4 open), with a frozen Qwen2.5-VL-7B at 48.1. Ablations show downstream accuracy tracks detector quality, with FDI grounding disproportionately influential, and overlays adding ~2.6 open points. Key limitations: lesion detectors are weak (54.6-66.5 mAP@50) and their errors can propagate under evidence primacy; evaluation is single-run with stochastic decoding and no variance estimates; cross-paper comparisons mix image-resolution protocols and judge rubrics; code and model release is promised upon acceptance rather than available.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Injecting structured detections from seven dental experts into frozen VLMs improves MMOral-OPG-Bench accuracy by 12.9-23.0 mean percentage points across four backbones, without any fine-tuning of the reasoning layer.
- At matched 7B scale, frozen Qwen2.5-VL-7B-Instruct with AgenTeeth (48.1 open-ended) exceeds OralGPT-Plus-7B (45.4) despite the latter's supervised fine-tuning and reinforcement learning for tool use.
- Framework performance tracks detector quality, and the FDI grounding tool has disproportionate downstream impact: a 4.3-point mAP drop causes more VQA loss than a 28.6-point drop in the dental-history tool.
- Annotated image overlays add a modest +2.6 open-ended points over structured text evidence alone, concentrated on the open split.
Strengths
Rigorous matched-baseline causal design
The baseline keeps the identical two-pass pipeline, parser, report prompt, decoding and call count, removing only the seven experts (Sec. 5, B.2, A.6); 43/44 category comparisons (Tab. 2) and 28/28 under the GPT-4o judge (Tab. 7) favor the tool-aided arm, making the attribution of gains to tool evidence unusually clean for an agentic-systems paper.
Informative ablations with an internal control
Degrading FDI from 92.4 to 88.1 mAP@50 costs 3.3 mean points while Jaw is exactly unchanged (79.1/89.1), precisely as the dependency structure predicts since jaw findings bypass FDI binding (Sec. 7.2, Tab. 4); notably a 4.3-point FDI drop hurts more than a 28.6-point dental-history drop, identifying grounding as the bottleneck.
Judge robustness and honest evaluation bounds
Transparent reporting of weak components
New annotated resource
Weaknesses
No variance quantification despite stochastic decoding
Headline cross-paper comparisons are protocol-confounded
Evidence primacy enforced by hand-tuned prompt guardrails
Model-agnostic claim validated on a single benchmark
Release inconsistency and minor internal discrepancies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is AgenTeeth and does it require fine-tuning?
AgenTeeth is a training-free framework that improves vision-language models on panoramic dental X-rays by feeding them evidence from seven specialized detector models (tooth numbering, caries, periapical lesions, alveolar bone loss, jaw structures, dental history, impacted teeth). The underlying VLM stays frozen and interchangeable; only the detectors are trained, so the same pipeline works with different VLM backbones.
How much does AgenTeeth improve performance on dental X-ray questions?
On the MMOral-OPG-Bench benchmark (100 panoramic radiographs, 491 multiple-choice and 578 open-ended questions), four different VLM backbones gained 12.9-23.0 percentage points on average over the same pipeline without tools. The best configuration reached 74.1% on multiple-choice and 65.7% on open-ended questions, above the fine-tuned OralGPT-Plus system (45.4% open-ended); a frozen 7B model with AgenTeeth also beat it at matched scale.
Is the AgenTeeth code, models, or dataset publicly available?
The paper states the framework, the seven expert models, and a dentist-annotated alveolar bone-loss dataset (943 TUFTS panoramic radiographs) will be released upon acceptance, and the abstract phrased this as already released; no public link appears in the reviewed text. The supplementary material does publish all prompts verbatim and full dataset documentation, but exact reproduction additionally depends on hosted Roboflow and Ollama Cloud services.
What are the main limitations of AgenTeeth?
The small-lesion detectors are weak (caries 54.6, bone loss 57.5, periapical 66.5 mAP@50), and because the VLM is instructed to trust tool evidence, detector false positives can propagate into answers - visible as negligible or negative gains on closed-ended jaw questions. Evaluation was single-run with stochastic decoding (no variance estimates), comparisons to published systems mix evaluation protocols and judge rubrics, and the framework performs no contralateral (left-right symmetry) comparison.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 are available only as captions in the extracted text; the annotated-overlay visuals central to the method could not be inspected.
- The judge-prompt few-shot examples (Supp. A.7) contain garbled JSON in the extraction, so the exact rubric formatting was only partially verifiable.
- Backbones (gemma4:31b, qwen3.5, mistral-large-3, minimax-3 via Ollama Cloud) could not be independently verified beyond the paper's naming; some table values were reconstructed from inline text (e.g., Tab. 2 column alignment).
- The extraction appears complete through Supplementary Sec. G with no obvious truncation, supporting high review confidence.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.