Text guided diffusion models are used by millions of users, but can be easily exploited to produce harmful content. Concept unlearning methods aim at reducing the models' likelihood of generating harmful content. Traditionally, this has been tackled at an individual concept level, with only a handful of recent works considering more realistic concept combinations. However, state of the art methods depend on full finetuning, which is computationally expensive. Concept localisation methods can facilitate selective finetuning, but existing techniques are static, resulting in suboptimal utility. In order to tackle these challenges, we propose TRUST (Targeted Robust Selective fine Tuning), a novel approach for dynamically estimating target concept neurons and unlearning them through selective finetuning, empowered by a Hessian based regularization. We show experimentally, against a number of SOTA baselines, that TRUST is robust against adversarial prompts, preserves generation quality to a significant degree, and is also significantly faster than the SOTA. Our method achieves unlearning of not only individual concepts but also combinations of concepts and conditional concepts, without a
Standard Sparse Autoencoders (SAEs) excel at discovering a dictionary of a model's learned features, providing a powerful lens for passive feature discovery. However, this passive nature makes it difficult to systematically evaluate or analyze concepts that users explicitly care about. We introduce Concept-SAE, a framework that augments SAEs with a structured and controllable interface for probing user-defined concepts. Concept-SAE decomposes an activation subspace into two orthogonal components: Concept Tokens, which are aligned to externally specified semantics through dual supervision on both concept existence and spatial localization, and Free Tokens, which operate like standard SAEs to capture all remaining information. This hybrid disentanglement strategy ensures that Concept Tokens are faithful, spatially grounded, and cleanly separated from the residual subspace while preserving the ability of SAEs for open-ended concept discovery. We conduct extensive experiments demonstrating that Concept-SAE yields high-fidelity, well-localized, and strongly disentangled concept representations, outperforming alternatives in interface quality. Finally, we validate the utility of this con
The customization of text-to-image models has seen significant advancements, yet generating multiple personalized concepts remains a challenging task. Current methods struggle with attribute leakage and layout confusion when handling multiple concepts, leading to reduced concept fidelity and semantic consistency. In this work, we introduce a novel training-free framework, Concept Conductor, designed to ensure visual fidelity and correct layout in multi-concept customization. Concept Conductor isolates the sampling processes of multiple custom models to prevent attribute leakage between different concepts and corrects erroneous layouts through self-attention-based spatial guidance. Additionally, we present a concept injection technique that employs shape-aware masks to specify the generation area for each concept. This technique injects the structure and appearance of personalized concepts through feature fusion in the attention layers, ensuring harmony in the final image. Extensive qualitative and quantitative experiments demonstrate that Concept Conductor can consistently generate composite images with accurate layouts while preserving the visual details of each concept. Compared
Concept bottleneck models predict outcomes from high-level concepts detected in inputs. Although concepts provide a simple way to reap benefits from interpretability, very few datasets include concept labels. This limits researchers' ability to determine which problems are suitable for these models, isolate the factors that drive their performance or lead to failures, or uncover which algorithms perform well. In this paper, we develop synthetic benchmarks for concept-bottleneck models, focusing on their two main use cases: decision support, in which models assist humans in making better decisions, and automation, in which models handle routine tasks without supervision. Our benchmarks can generate labeled datasets while controlling for properties that affect performance, including data modality, concept choice, annotation quality, and completeness. We demonstrate how the benchmarks can be used to evaluate representative classes of concept bottleneck models. Our demonstrations show how the benchmarks can diagnose failure modes and guide follow-up testing.
Vision Transformers (ViTs) often degrade under distribution shifts because they rely on spurious correlations, such as background cues, rather than semantically meaningful features. Existing regularization methods, typically relying on simple foreground-background masks, which fail to capture the fine-grained semantic concepts that define an object (e.g., ``long beak'' and ``wings'' for a ``bird''). As a result, these methods provide limited robustness to distribution shifts. To address this limitation, we introduce a novel finetuning framework that steers model reasoning toward concept-level semantics. Our approach optimizes the model's internal relevance maps to align with spatially grounded concept masks. These masks are generated automatically, without manual annotation: class-relevant concepts are first proposed using an LLM-based, label-free method, and then segmented using a VLM. The finetuning objective aligns relevance with these concept regions while simultaneously suppressing focus on spurious background areas. Notably, this process requires only a minimal set of images and uses half of the dataset classes. Extensive experiments on five out-of-distribution benchmarks dem
Large Language Models (LLMs) are being used for a wide variety of tasks. While they are capable of generating human-like responses, they can also produce undesirable output including potentially harmful information, racist or sexist language, and hallucinations. Alignment methods are designed to reduce such undesirable outputs via techniques such as fine-tuning, prompt engineering, and representation engineering. However, existing methods face several challenges: some require costly fine-tuning for every alignment task; some do not adequately remove undesirable concepts, failing alignment; some remove benign concepts, lowering the linguistic capabilities of LLMs. To address these issues, we propose Parsimonious Concept Engineering (PaCE), a novel activation engineering framework for alignment. First, to sufficiently model the concepts, we construct a large-scale concept dictionary in the activation space, in which each atom corresponds to a semantic concept. Given any alignment task, we instruct a concept partitioner to efficiently annotate the concepts as benign or undesirable. Then, at inference time, we decompose the LLM activations along the concept dictionary via sparse coding
In recent years, the black-box nature of deep learning models has limited their application in high-stakes domains such as medical diagnosis and finance, where interpretability is essential. To address this, we propose a novel approach using influence functions to enhance interpretability in NLP models at both the sample and concept levels. Experiments on CEBaB and Yelp datasets show that influence functions effectively identify the most impactful training samples, both helpful and harmful, on model predictions. By adjusting the labels and weights of these samples, we demonstrate that model performance can be restored to baseline levels without retraining, confirming the value of influence functions for efficient data debugging. Furthermore, our concept-level analysis identifies key concepts within Concept Bottleneck Models (CBM) that significantly affect predictions. Modifying these concepts alters model behavior observably, providing clear insights into the decision process.
With the growing popularity of general-purpose Large Language Models (LLMs), comes a need for more global explanations of model behaviors. Concept-based explanations arise as a promising avenue for explaining high-level patterns learned by LLMs. Yet their evaluation poses unique challenges, especially due to their non-local nature and high dimensional representation in a model's hidden space. Current methods approach concepts from different perspectives, lacking a unified formalization. This makes evaluating the core measures of concepts, namely faithfulness or readability, challenging. To bridge the gap, we introduce a formal definition of concepts generalizing to diverse concept-based explanations' settings. Based on this, we quantify the faithfulness of a concept explanation via perturbation. We ensure adequate perturbation in the high-dimensional space for different concepts via an optimization problem. Readability is approximated via an automatic and deterministic measure, quantifying the coherence of patterns that maximally activate a concept while aligning with human understanding. Finally, based on measurement theory, we apply a meta-evaluation method for evaluating these m
The rapid proliferation of large-scale text-to-image diffusion (T2ID) models has raised serious concerns about their potential misuse in generating harmful content. Although numerous methods have been proposed for erasing undesired concepts from T2ID models, they often provide a false sense of security; concept-erased models (CEMs) can still be manipulated via adversarial attacks to regenerate the erased concept. While a few robust concept erasure methods based on adversarial training have emerged recently, they compromise on utility (generation quality for benign concepts) to achieve robustness and/or remain vulnerable to advanced embedding space attacks. These limitations stem from the failure of robust CEMs to thoroughly search for "blind spots" in the embedding space. To bridge this gap, we propose STEREO, a novel two-stage framework that employs adversarial training as a first step rather than the only step for robust concept erasure. In the first stage, STEREO employs adversarial training as a vulnerability identification mechanism to search thoroughly enough. In the second robustly erase once stage, STEREO introduces an anchor-concept-based compositional objective to robustl
Do large language models (LLMs) represent concepts abstractly, i.e., independent of input format? We revisit Function Vectors (FVs), compact representations of in-context learning (ICL) tasks that causally drive task performance. Across multiple LLMs, we show that FVs are not fully invariant: FVs are nearly orthogonal when extracted from different input formats (e.g., open-ended vs. multiple-choice), even if both target the same concept. We identify Concept Vectors (CVs), which carry more stable concept representations. Like FVs, CVs are composed of attention head outputs; however, unlike FVs, the constituent heads are selected using Representational Similarity Analysis (RSA) based on whether they encode concepts consistently across input formats. While these heads emerge in similar layers to FV-related heads, the two sets are largely distinct, suggesting different underlying mechanisms. Steering experiments reveal that FVs excel in-distribution, when extraction and application formats match (e.g., both open-ended in English), while CVs generalize better out-of-distribution across both question types (open-ended vs. multiple-choice) and languages. Our results show that LLMs do cont
Our understanding of the visual world is centered around various concept axes, characterizing different aspects of visual entities. While different concept axes can be easily specified by language, e.g. color, the exact visual nuances along each axis often exceed the limitations of linguistic articulations, e.g. a particular style of painting. In this work, our goal is to learn a language-informed visual concept representation, by simply distilling large pre-trained vision-language models. Specifically, we train a set of concept encoders to encode the information pertinent to a set of language-informed concept axes, with an objective of reproducing the input image through a pre-trained Text-to-Image (T2I) model. To encourage better disentanglement of different concept encoders, we anchor the concept embeddings to a set of text embeddings obtained from a pre-trained Visual Question Answering (VQA) model. At inference time, the model extracts concept embeddings along various axes from new test images, which can be remixed to generate images with novel compositions of visual concepts. With a lightweight test-time finetuning procedure, it can also generalize to novel concepts unseen at
Proofs of Concept (PoCs) are widely adopted practices in software engineering. Despite their relevance, PoCs remain conceptually underdefined and methodologically ad hoc in both research and industry, with definitions and implementation approaches that often lack clarity and consistency. This paper investigates the concept of PoCs with two complementary goals: (1) to provide a refined definition and astructured framework for PoC development grounded in a systematic review of academic and grey literature; and (2) to position PoCs as first-class architectural decision instruments rather than informal experiments or disposable artifacts. Through a systematic review of academic and grey literature we identify the key characteristics, processes, associated with PoCs and expose a significant gap the academic literature describes PoC outcomes but rarely its process. By synthesizing insights from diverse sources we propose a refined definition and a lightweight, three-phase framework (planning, execution, decision-making) that encompasses technical validation and explicit decision traceability. We also introduce the Undocumented Architectural Experiment anti-pattern, arising when PoCs infl
Concept erasure in text-to-image diffusion models aims to disable pre-trained diffusion models from generating images related to a target concept. To perform reliable concept erasure, the properties of robustness and locality are desirable. The former refrains the model from producing images associated with the target concept for any paraphrased or learned prompts, while the latter preserves its ability in generating images with non-target concepts. In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler). It learns a lightweight Eraser to perform concept erasing while satisfying the above desirable properties through the proposed concept-localized regularization and adversarial prompt learning scheme. Experiments with various concepts verify the superiority of Receler over previous methods.
Embedding models group text by semantic content, what text is about. We show that temporal co-occurrence within texts discovers a different kind of structure: recurrent transition-structure concepts or what text does. We train a 29.4M-parameter contrastive model on 373 million co-occurrence pairs from 9,766 Project Gutenberg texts (24.96 million passages), mapping pre-trained embeddings into an association space where passages with similar transition structure cluster together. Under capacity constraint (42.75% accuracy), the model must compress across recurring patterns rather than memorise individual co-occurrences. Clustering at six granularities (k=50 to k=2,000) produces a multi-resolution concept map; from broad modes like "direct confrontation" and "lyrical meditation" to precise registers and scene templates like "sailor dialect" and "courtroom cross-examination." At k=100, clusters average 4,508 books each (of 9,766), confirming corpus-wide patterns. Direct comparison with embedding-similarity clustering shows that raw embeddings group by topic while association-space clusters group by function, register, and literary tradition. Unseen novels are assigned to existing clust
Universal medical image segmentation seeks to use a single foundational model to handle diverse tasks across multiple imaging modalities. However, existing approaches often rely heavily on manual visual prompts or retrieved reference images, which limits their automation and robustness. In addition, naive joint training across modalities often fails to address large domain shifts. To address these limitations, we propose Concept-to-Pixel (C2P), a novel prompt-free universal segmentation framework. C2P explicitly separates anatomical knowledge into two components: Geometric and Semantic representations. It leverages Multimodal Large Language Models (MLLMs) to distill abstract, high-level medical concepts into learnable Semantic Tokens and introduces explicitly supervised Geometric Tokens to enforce universal physical and structural constraints. These disentangled tokens interact deeply with image features to generate input-specific dynamic kernels for precise mask prediction. Furthermore, we introduce a Geometry-Aware Inference Consensus mechanism, which utilizes the model's predicted geometric constraints to assess prediction reliability and suppress outliers. Extensive experiments
This work frames the first three publications around the development of the Flight Physics Concept Inventory (FliP-CoIn), and elaborates on many aspects in more detail. FliP-CoIn is a multiple-choice conceptual assessment instrument for improving fluid dynamics learning and teaching. I give insights into why and how FliP-CoIn was developed and how it is best used for improving conceptual learning. Further, this work presents evidence for several dimensions of FliP-CoIn's validity and reliability. Finally, I discuss key insights from the development process, the data analysis, and give recommendations for future research. This is a pre print version of the following book: Florian Genz, The Flight Physics Concept Inventory, 2025, Springer Spektrum, published with permission of Springer Fachmedien Wiesbaden GmbH. The final authenticated version is available online at: http://doi.org/10.1007/978-3-658-47515-4 and https://link.springer.com/book/9783658475147
Visual concept learning, also known as Text-to-image personalization, is the process of teaching new concepts to a pretrained model. This has numerous applications from product placement to entertainment and personalized design. Here we show that many existing methods can be substantially augmented by adding a personalization step that is (1) specific to the prompt and noise seed, and (2) using two loss terms based on the self- and cross- attention, capturing the identity of the personalized concept. Specifically, we leverage PDM features -- previously designed to capture identity -- and show how they can be used to improve personalized semantic similarity. We evaluate the benefit that our method gains on top of six different personalization methods, and several base text-to-image models (both UNet- and DiT-based). We find significant improvements even over previous per-query personalization methods.
Consider this prompt "Draw a unicorn with two horns". Should large language models (LLMs) recognize that a unicorn has only one horn by definition and ask users for clarifications, or proceed to generate something anyway? We introduce concept incongruence to capture such phenomena where concept boundaries clash with each other, either in user prompts or in model representations, often leading to under-specified or mis-specified behaviors. In this work, we take the first step towards defining and analyzing model behavior under concept incongruence. Focusing on temporal boundaries in the Role-Play setting, we propose three behavioral metrics--abstention rate, conditional accuracy, and answer rate--to quantify model behavior under incongruence due to the role's death. We show that models fail to abstain after death and suffer from an accuracy drop compared to the Non-Role-Play setting. Through probing experiments, we identify two main causes: (i) unreliable encoding of the "death" state across different years, leading to unsatisfactory abstention behavior, and (ii) role playing causes shifts in the model's temporal representations, resulting in accuracy drops. We leverage these insigh
In this article we present a new modelling framework for structured concepts using a category-theoretic generalisation of conceptual spaces, and show how the conceptual representations can be learned automatically from data, using two very different instantiations: one classical and one quantum. A contribution of the work is a thorough category-theoretic formalisation of our framework. We claim that the use of category theory, and in particular the use of string diagrams to describe quantum processes, helps elucidate some of the most important features of our approach. We build upon Gardenfors' classical framework of conceptual spaces, in which cognition is modelled geometrically through the use of convex spaces, which in turn factorise in terms of simpler spaces called domains. We show how concepts from the domains of shape, colour, size and position can be learned from images of simple shapes, where concepts are represented as Gaussians in the classical implementation, and quantum effects in the quantum one. In the classical case we develop a new model which is inspired by the Beta-VAE model of concepts, but is designed to be more closely connected with language, so that the name
World models are often evaluated by native frame cadence, but higher nominal frame rate can trade away long-horizon scene stability. This article reports an independent proof of concept implemented using Overworld's Waypoint-1.5 family and WorldEngine runtime on a Windows fallback stack with ONNX Runtime + DirectML and an FSR4 DX12 bridge. The tested coherence-first branch generates higher-context anchor frames at a 15 FPS presentation-timeline cadence and reconstructs presentation to 30 FPS using latent-delta motion guidance and synthesized depth. It is compared against a lower-context cadence-first baseline that generates about 30 FPS natively under the same seed, route, control script, target presentation duration, and local time-scaling regime. Across forest, sword, desert, and snow scenes, the coherence-first branch preserves path geometry, object identity, large silhouettes, and depth layering longer, while the baseline degrades earlier into brightness drift and geometric distortion. Lightweight temporal metrics and paired videos support the visual comparison, with LPIPS favoring the coherence-first branch across all tested scenes. Here compute-normalized means approximately