Med-Banana: Learning Agentic Quality-Controlled Medical Image Editing from Success-and-Failure Trajectories 🔗

Zhihui Chen, Qingyuan Lei, Mengling Feng

Conference on Empirical Methods in Natural Language Processing (EMNLP) 2026 Accepted · Main Conference

Text-guided medical image editing must deliver the requested pathology while preserving anatomy, modality-specific appearance, and clinical plausibility. Existing datasets supervise editors only with accepted final edits and discard the failed attempts produced during generation. We argue that failures are indispensable supervision for agentic post-training: they specify what should be rejected, why an edit is medically or visually invalid, and how the instruction should be revised. We present Med-Banana, a trajectory-supervised framework that jointly trains an editor, a verifier, and a refiner for an edit-verify-refine inference loop, together with Med-Banana-80K, a large-scale resource of success-and-failure editing trajectories comprising candidate images, verification outcomes, rejection reasons, and prompt refinements (50,635 accepted + 37,822 rejected trajectories). Experiments with MLLM judges, blind expert assessment, and separability probes show consistent gains over open medical image editors. Code and data are publicly available.

[Paper] [Code] [Dataset] [Project Page]

Looking Ahead to Stay Grounded: Future-Discounted Visual Credit Assignment for VLM Reinforcement Learning

Zhihui Chen, Yike Yun, et al.

ICLR 2026 submission Under Review First Author

Farsight targets a core failure mode of RLVR post-training for vision-language models: token-level credit assignment that only rewards instant visual dependence, which lets models drift into ungrounded reasoning over long horizons. We extend credit assignment to future-discounted visual grounding — a token earns visual credit not only for the evidence it uses now, but for how that evidence supports discounted future reasoning steps — encouraging policies that stay anchored to the image across multi-step chains of thought. Developed during my Agentic Post-Training research internship at ByteDance (Singapore).

MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning 🔗

Zhihui Chen, Kai He, Qingyuan Lei, Bin Pu, Jian Zhang, Yuling Xu, Mengling Feng^# (^# corresponding author 通讯作者)

Annual Meeting of the Association for Computational Linguistics (ACL) 2026 Main Conference

As generative models improve, medical deepfakes that implant or remove lesions while staying visually plausible pose growing risks to clinical safety and the integrity of medical evidence. Most prior work reduces detection to binary real-vs-fake scoring with little insight into where manipulation occurs or why. We present MedForge, an interpretable framework that introduces MedForge-90K—the first large-scale explainable medical deepfake dataset spanning CT, MRI, and X-ray, covering 19 lesion types with forgeries from 10 state-of-the-art deepfake models, each paired with expert-guided localization and clinical-grade explanations—and MedForge-Reasoner, a detector trained with a Localize-then-Analyze chain-of-thought paradigm and Forgery-aware GSPO reinforcement learning. MedForge-Reasoner achieves state-of-the-art detection while producing localized, verifiable medical rationales.

[Paper] [Project Page] [Model] [Dataset]

DivScore: Zero-Shot Detection of LLM-Generated Text in Specialized Domains 🔗

Zhihui Chen, Kai He, Yucheng Huang, Yunxiao Zhu, Mengling Feng

Conference on Empirical Methods in Natural Language Processing (EMNLP) 2025 Main Conference

Detecting LLM-generated text in specialized and high-stakes domains like medicine and law is crucial for combating misinformation and ensuring authenticity. We propose DivScore, a zero-shot detection framework using normalized entropy-based scoring and domain knowledge distillation to robustly identify LLM-generated text in specialized domains. Experiments show that DivScore consistently outperforms state-of-the-art detectors, with 14.4% higher AUROC and 64.0% higher recall at 0.1% false positive rate threshold.

[Paper] [Code] [Project Page]