Medical vision-language models (VLMs) are often evaluated by report retrieval, grounding, and explanation quality. We show that these metrics can create a metric illusion: a model may appear clinically grounded while exploiting acquisition template, crop, orientation, or protocol rather than patient anatomy. We formalize this failure as template-induced ambiguity and introduce TIA-VL, a counterfactual benchmark with source, template-family, severity, and compound protocol shifts. We also propose Rationale-Grounded Template Ambiguity Calibration (R-GTAC), a plug-in calibration layer that discounts template-dominated evidence while preserving anatomy and textual rationale support. Across strong medical VLM-style baselines, including BioViL-T-style frozen features, adapter features, and anatomy-teacher features, R-GTAC reduces wrong same-template errors to 0.0000 on the main protocol splits. We then test whether systems know when to abstain through readiness and counterfactual rationale deferral. Counterfactual rationale readiness filtering further identifies a reliable low-risk 25% subset reaching 0.8500 Top-1 on the hardest compound shift, while revealing that coupled protocol shifts pose critical frontiers for global risk calibration. Thus, TIA-VL supports a conservative message for medical VLM evaluation: models should be judged not only by whether they answer, but by whether their apparent rationales survive anti-shortcut counterfactuals. The code and benchmark will be available at https://github.com/YifanLuo0103/TIA-VL.

Vision-Language Models (VLMs) show immense promise in medicine, yet surgical scene understanding remains formidable due to complex visual obstructions and rigorous multi-step clinical logic. Standard Supervised Fine-Tuning (SFT) imparts domain knowledge, but we identify a critical flaw—the "Reasoning Collapse" phenomenon: during answer-only SFT, models progressively lose explicit step-by-step reasoning capabilities, regressing into "black boxes" that generate uninterpretable shortcut answers—a severe risk for clinical traceability. Constructing high-quality Chain-of-Thought (CoT) datasets to mitigate this is prohibitively expensive in the surgical domain. We propose SurgMind-R1, a surgical VLM with a three-stage training pipeline. We construct Cholec-Think, a reasoning-enhanced dataset combining expert and VLM-assisted CoT annotations; apply domain-adaptive SFT and reasoning-enhanced SFT on Cholec-Think to internalize surgical logic; and introduce Reinforcement Learning via Group Relative Policy Optimization (GRPO) with rule-based rewards to enforce structured thinking paths and eliminate reasoning collapse. On SurgLite-Bench, our comprehensive benchmark, SurgMind-R1 significantly outperforms state-of-the-art VLMs on complex inference tasks. It internalizes structured thinking, maintaining high accuracy in a direct no-think mode that resolves the high-latency bottleneck of reasoning models for real-time clinical applications, and exhibits strong zero-shot out-of-distribution (OOD) generalization across unseen surgical scenarios and novel tasks, paving the way for transparent, trustworthy computer-assisted interventions. The code, benchmark, and models will be available at https://github.com/YifanLuo0103/SurgMind-R1.

Multimodal Large Language Models (MLLMs) exhibit strong generalization and reasoning abilities due to large-scale multimodal pre-training. However, fine-tuning these models on downstream tasks often leads to catastrophic forgetting, where newly learned task-specific knowledge degrades previously acquired capabilities. This issue arises because gradient updates for new tasks overwrite parameters critical to prior knowledge, limiting the practical deployment of MLLMs. To address this challenge, we propose **Activation-Weighted Adaptive REtention (AWARe)**, a fine-tuning method that mitigates catastrophic forgetting by dynamically controlling parameter updates based on activation patterns. AWARe assigns activation-based importance scores to parameters, selectively freezing those essential for preserving prior capabilities while allowing less important parameters to adapt to new tasks. Importantly, AWARe operates without modifying model architectures, ensuring compatibility with existing inference engines. Extensive experiments demonstrate that AWARe effectively preserves upstream capabilities while achieving superior downstream performance compared to existing methods. Code is available at https://github.com/kaln27/AWARe.

Leveraging Large Language Models (LLMs) to build domain-specific conversational agents, especially for e-commerce customer service chatbots, is a growing focus. While existing methods enhance dialogue performance by extracting core patterns from dialogue data and integrating them into models, two key challenges persist: (1) heavy reliance on human experts for dialogue strategy induction, and (2) LLM-based automatic extraction often focuses on summarizing specific behaviors, neglecting the underlying thought processes behind strategy selection. In this paper, we present ChatMap, which focuses on enhancing customer service chatbots by mining thought processes using a Multi-Agent aPproach. Specifically, the process begins by extracting customer requests and solutions from a raw dialogue dataset, followed by clustering similar requests, analyzing the thought processes behind solutions, and refining service thoughts. Through a quality inspection and reflection mechanism, the final service thought dataset is generated, helping chatbots provide more appropriate responses. Offline experimental results show that ChatMap performs comparably to manually annotated thought processes and significantly outperforms other baselines, demonstrating its ability to automate human annotation and enhance dialogue capabilities through strategic understanding. Online A/B tests on Taobao, a popular e-commerce platform in China reveal that ChatMap can better improve customer satisfaction and address customer requests from a business perspective.

Inspired by human cognitive behavior, we introduce visual modality to enhance the performance of pure text-based question-answering tasks with the development of multimodal models. However, obtaining corresponding images through manual annotation often entails high costs. Faced with this challenge, an intuitive strategy is to use search engines or use web scraping techniques to automatically obtain relevant image information. However, the images obtained by this strategy may be of low quality and may not match the context of the original task, which could fail to improve or even decrease performance on downstream tasks. In this paper, we propose a novel framework named “ITERATE”, aimed at retrieving and optimizing the quality of images to improve the alignment between text and images. Inspired by evolutionary algorithms in reinforcement learning and driven by the synergy of large language models (LLMs) and multimodal models, ITERATE employs a series of strategic actions such as filtering, optimizing, and retrieving to acquire higher quality images, and repeats this process over multiple generations to enhance the quality of the entire image cluster. Our experimental results on the ScienceQA, ARC-Easy, and OpenDataEval datasets also verify the effectiveness of our method, showing improvements of 3.5%, 5%, and 7%, respectively.