DEAL: Disentangling Transformer Head Activations for LLM Steering

Inference-time steering aims to alter the response characteristics of large language models (LLMs) without modifying their underlying parameters. A critical step in this process is the identification of internal modules within LLMs that are associated with the target behavior. However, current approaches to module selection often depend on superficial cues or ad-hoc heuristics, which can result in suboptimal or unintended outcomes. In this work, we propose a principled causal-attribution framework for identifying behavior-relevant attention heads in transformers. For each head, we train a vector-quantized autoencoder (VQ-AE) on its attention activations, partitioning the latent space into behavior-relevant and behavior-irrelevant subspaces, each quantized with a shared learnable codebook. We assess the behavioral relevance of each head by quantifying the separability of VQ-AE encodings for behavior-aligned versus behavior-violating responses using a binary classification metric. This yields a behavioral relevance score that reflects each head discriminative capacity with respect to the target behavior, guiding both selection and importance weighting. Experiments on seven LLMs from two model families and five behavioral steering datasets demonstrate that our method enables more accurate inference-time interventions, achieving superior performance on the truthfulness-steering task. Furthermore, the heads selected by our approach exhibit strong zero-shot generalization in cross-domain truthfulness-steering scenarios.

GEMeX: A Large-Scale, Groundable, and Explainable Medical VQA Benchmark for Chest X-ray Diagnosis

Medical Visual Question Answering (VQA) is an essential technology that integrates computer vision and natural language processing to automatically respond to clinical inquiries about medical images. However, current medical VQA datasets exhibit two significant limitations: (1) they often lack visual and textual explanations for answers, which impedes their ability to satisfy the comprehension needs of patients and junior doctors; (2) they typically offer a narrow range of question formats, inadequately reflecting the diverse requirements encountered in clinical scenarios. These limitations pose significant challenges to the development of a reliable and user-friendly Med-VQA system. To address these challenges, we introduce a large-scale, Groundable, and Explainable Medical VQA benchmark for chest X-ray diagnosis (GEMeX), featuring several innovative components: (1) A multi-modal explainability mechanism that offers detailed visual and textual explanations for each question-answer pair, thereby enhancing answer comprehensibility; (2) Four distinct question types, open-ended, closed-ended, single-choice, and multiple-choice, that better reflect diverse clinical needs. We evaluated 10 representative large vision language models on GEMeX and found that they underperformed, highlighting the dataset's complexity. However, after fine-tuning a baseline model using the training set, we observed a significant performance improvement, demonstrating the dataset's effectiveness. The project is available at www.med-vqa.com/GEMeX.

Diversity-grounded Channel Prototypical Learning for Out-of-Distribution Intent Detection

In the realm of task-oriented dialogue systems, a robust intent detection mechanism must effectively handle malformed utterances encountered in real-world scenarios. This study presents a novel fine-tuning framework for large language models (LLMs) aimed at enhancing in-distribution (ID) intent classification and out-of-distribution (OOD) intent detection, which utilizes semantic matching with prototypes derived from ID class names. By harnessing the highly distinguishable representations of LLMs, we construct semantic prototypes for each ID class using a diversity-grounded prompt tuning approach. We rigorously test our framework in a challenging OOD context, where ID and OOD classes are semantically close yet distinct, referred to as \emph{near} OOD detection. For a thorough assessment, we benchmark our method against the prevalent fine-tuning approaches. The experimental findings reveal that our method demonstrates superior performance in both few-shot ID intent classification and near-OOD intent detection tasks.