Neural Information Processing Systems http://nips.cc/

Recent works have extended notions of feature importance to semantic concepts that are inherently interpretable to the users interacting with a black-box predictive model. Yet, precise statistical guarantees such as false positive rate and false discovery rate control are needed to communicate findings transparently, and to avoid unintended consequences in real-world scenarios. In this paper, we formalize the global (i.e., over a population) and local (i.e., for a sample) statistical importance of semantic concepts for the predictions of opaque models by means of conditional independence, which allows for rigorous testing. We use recent ideas of sequential kernelized independence testing to induce a rank of importance across concepts, and we showcase the effectiveness and flexibility of our framework on synthetic datasets as well as on image classification using several vision-language models.

As generative AI systems become competent and democratized in science, business, and government, deeper insight into their failure modes now poses an acute need. The occasional volatility in their behavior, such as the propensity of transformer models to hallucinate, impedes trust and adoption of emerging AI solutions in high-stakes areas. In the present work, we establish how and when hallucinations arise in pre-trained transformer models through concept representations captured by sparse autoencoders, under scenarios with experimentally controlled uncertainty in the input space. Our systematic experiments reveal that the number of semantic concepts used by the transformer model grows as the input information becomes increasingly unstructured. In the face of growing uncertainty in the input space, the transformer model becomes prone to activate coherent yet input-insensitive semantic features, leading to hallucinated output. At its extreme, for pure-noise inputs, we identify a wide variety of robustly triggered and meaningful concepts in the intermediate activations of pre-trained transformer models, whose functional integrity we confirm through targeted steering. We also show that hallucinations in the output of a transformer model can be reliably predicted from the concept patterns embedded in transformer layer activations. This collection of insights on transformer internal processing mechanics has immediate consequences for aligning AI models with human values, AI safety, opening the attack surface for potential adversarial attacks, and providing a basis for automatic quantification of a model's hallucination risk.

In this paper we address the text to scene image generation problem.

With the rapid advancement of image generation techniques, robust forgery detection has become increasingly imperative to ensure the trustworthiness of digital media. Recent research indicates that the learned semantic concepts of pre-trained models are critical for identifying fake images. However, the misalignment between the forgery and semantic concept spaces hinders the model's forgery detection performance. To address this problem, we propose a novel Semantic Discrepancy-aware Detector (SDD) that leverages reconstruction learning to align the two spaces at a fine-grained visual level. By exploiting the conceptual knowledge embedded in the pre-trained vision language model, we specifically design a semantic token sampling module to mitigate the space shifts caused by features irrelevant to both forgery traces and semantic concepts. A concept-level forgery discrepancy learning module, built upon a visual reconstruction paradigm, is proposed to strengthen the interaction between visual semantic concepts and forgery traces, effectively capturing discrepancies under the concepts' guidance. Finally, the low-level forgery feature enhancemer integrates the learned concept level forgery discrepancies to minimize redundant forgery information. Experiments conducted on two standard image forgery datasets demonstrate the efficacy of the proposed SDD, which achieves superior results compared to existing methods. The code is available at https://github.com/wzy1111111/SSD.

Multimodal representation learning produces high-dimensional embeddings that align diverse modalities in a shared latent space. While this enables strong generalization, it also introduces scalability challenges, both in terms of storage and downstream processing. A key open problem is how to achieve semantic compression, reducing the memory footprint of multimodal embeddings while preserving their ability to represent shared semantic content across modalities. In this paper, we prove a strong connection between reducing the modality gap, which is the residual separation of embeddings from different modalities, and the feasibility of post-training semantic compression. When the gap is sufficiently reduced, embeddings from different modalities but expressing the same semantics share a common portion of the space. Therefore, their centroid is a faithful representation of such a semantic concept. This enables replacing multiple embeddings with a single centroid, yielding significant memory savings. We propose a novel approach for semantic compression grounded on the latter intuition, operating directly on pretrained encoders. We demonstrate its effectiveness across diverse large-scale multimodal downstream tasks. Our results highlight that modality alignment is a key enabler for semantic compression, showing that the proposed approach achieves significant compression without sacrificing performance.

Concept Bottleneck Models (CBMs) provide interpretable decision-making through explicit, human-understandable concepts. However, existing generative CBMs often rely on auxiliary visual cues at the bottleneck, which undermines interpretability and intervention capabilities. We propose EnCoBo, a post-hoc concept bottleneck for generative models that eliminates auxiliary cues by constraining all representations to flow solely through explicit concepts. Unlike autoencoder-based approaches that inherently rely on black-box decoders, EnCoBo leverages a decoder-free, energy-based framework that directly guides generation in the latent space. Guided by diffusion-scheduled energy functions, EnCoBo supports robust post-hoc interventions-such as concept composition and negation-across arbitrary concepts. Experiments on CelebA-HQ and CUB datasets showed that EnCoBo improved concept-level human intervention and interpretability while maintaining competitive visual quality.

Vision-Language-Action (VLA) models are a promising path to realizing generalist embodied agents that can quickly adapt to new tasks, modalities, and environments. However, methods for interpreting and steering VLAs fall far short of classical robotics pipelines, which are grounded in explicit models of kinematics, dynamics, and control. This lack of mechanistic insight is a central challenge for deploying learned policies in real-world robotics, where robustness and explainability are critical. Motivated by advances in mechanistic interpretability for large language models, we introduce the first framework for interpreting and steering VLAs via their internal representations, enabling direct intervention in model behavior at inference time. We project feedforward activations within transformer layers onto the token embedding basis, identifying sparse semantic directions - such as speed and direction - that are causally linked to action selection. Leveraging these findings, we introduce a general-purpose activation steering method that modulates behavior in real time, without fine-tuning, reward signals, or environment interaction. We evaluate this method on two recent open-source VLAs, Pi0 and OpenVLA, and demonstrate zero-shot behavioral control in simulation (LIBERO) and on a physical robot (UR5). This work demonstrates that interpretable components of embodied VLAs can be systematically harnessed for control - establishing a new paradigm for transparent and steerable foundation models in robotics.