Large language models (LLMs) have acquired substantial medical knowledge and reasoning skills during pretraining, yet eliciting these capabilities typically requires scarce rationales or costly post-training. To address this gap, we introduce **RIPT**, a lightweight framework for data-efficient endogenous capability activation. It explicitly decomposes this objective into two complementary components: reasoning enhancement and medical knowledge elicitation. For the former, we extract steering vectors from hidden activations on a small set of high-quality paired reasoning and vanilla responses to shape LLMs’ reasoning behavior. For the latter, we obtain prefix vectors via prefix tuning on simple medical QA pairs to elicit domain-specific knowledge. At inference, RIPT freezes the backbone LLM and combines steering-vector injection with prefix-vector conditioning. Rather than injecting external knowledge, RIPT aims to unlock medical reasoning capabilities already acquired by LLMs during pretraining. Experiments under limited-resource settings show that RIPT consistently improves diverse backbone LLMs and benchmarks, suggesting a generalizable pathway for lightweight medical reasoning enhancement.

Retrieval-Augmented Generation (RAG) improves generation quality by incorporating evidence retrieved from large external corpora. However, most existing methods rely on statically selecting top-k passages based on individual relevance, which fails to exploit combinatorial gains among passages and often introduces substantial redundancy. To address this limitation, we propose OptiSet, a set-centric framework that unifies set selection and set-level ranking for RAG. OptiSet adopts an “Expand-then-Refine” paradigm: it first expands a query into multiple perspectives to enable a diverse candidate pool and then refines the candidate pool via re-selection to form a compact evidence set. We then devise a self-synthesis strategy without strong LLM supervision to derive preference labels from the set conditional utility changes of the generator, thereby identifying complementary and redundant evidence. Finally, we introduce a set-list wise training strategy that jointly optimizes set selection and set-level ranking, enabling the model to favor compact, high-gain evidence sets. Extensive experiments demonstrate that OptiSet improves performance on complex combinatorial problems and makes generation more efficient.

Text-guided medical image editing must satisfy the requested pathology while preserving anatomy, modality-specific appearance, and clinical plausibility. However, existing datasets largely supervise editors with final accepted edits and discard the failed attempts produced during generation. We argue that these failures provide essential supervision for quality control: 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 for quality-controlled medical image editing. We introduce Med-Banana-80K, a large-scale resource of success-and-failure editing trajectories with candidate images, verification outcomes, rejection reasons, and prompt refinements. Building on it, Med-Banana jointly trains an editor, verifier, and refiner, enabling edit–verify–refine inference from accepted and rejected attempts. Experiments across MLLM judges, blind expert assessment, source-preservation and real–synthetic separability probes demonstrate consistent improvements over open medical image editors. Code and data are publicly available.

Instruction Fine-Tuning (IFT) has emerged as a critical technique for customizing Large Language Models (LLMs) to meet diverse downstream applications. However, recent studies have revealed that IFT can compromise the built-in security mechanisms of LLMs, thereby posing significant security risks. Although defense methods targeting various training stages have been proposed, they either face challenges in practical deployment or exhibit instability and limited performance gains. In our study, we propose a novel SWAT method that introduces a key idea: shifting more of the learning burden onto security-robust parameters. To this end, our study investigates how module-level parameters affect LLMs’ internal security feature space, aiming to uncover robustness patterns in parameters. Guided by this analysis, we identify a robust module set (Mods_Rob) that exhibits minimal effects on LLMs’ security feature space. Leveraging this insight, SWAT proceeds in two phases: (1) a warm-up phase that preferentially trains Mods_Rob to learn low-level features with minimal security risk, followed by (2) standard tuning to achieve optimal task performance. Across diverse knowledge-intensive datasets, scenarios, and LLMs, SWAT substantially reduces security risks without sacrificing task performance gains.

Explaining black-box model behavior with natural language has achieved impressive results in various NLP tasks. Recent research has explored the utilization of subsequences from the input text as a rationale, providing users with evidence to support the model decision. Although existing frameworks excel in generating high-quality rationales while achieving high task performance, they neglect to account for the unreliable link between the generated rationale and model decision. In simpler terms, a model may make correct decisions while attributing wrong rationales, or make poor decisions while attributing correct rationales. To mitigate this issue, we propose a unified two-stage framework known as Self-Attribution and Decision-Making (SADM). Through extensive experiments on five reasoning datasets from the ERASER benchmark, we demonstrate that our framework not only establishes a more reliable link between the generated rationale and model decision but also achieves competitive results in task performance and the quality of rationale. Furthermore, we explore the potential of our framework in semi-supervised scenarios.