Consistency Training while Mitigating Obfuscation via Rate Matching

Large language models are often influenced by extraneous input features, such as cues revealing a user's preferred answer. Consistency training reduces this influence by training models to behave similarly across inputs with and without the extraneous feature. However, existing methods train for consistency over entire responses or internal activations, which also constrains whether the model verbalises said extraneous features. We show this leads to obfuscation, where the model learns not to mention a cue while remaining influenced by it, which may undermine monitorability. To address this, we introduce Rate Matching Consistency Training (RMCT), which trains for consistency over selected behavioural properties without constraining how this behaviour is expressed. RMCT matches the rate at which the model exhibits a target behaviour (e.g., following a bias cue) across input perturbations, rather than requiring paired inputs with and without the extraneous feature, extending consistency training to settings where the extraneous features cannot be removed. We evaluate RMCT on sycophancy reduction in two open-weight language models, achieving reductions in bias-following comparable to a standard consistency-training baseline on held-out bias types, while largely preserving the model's tendency to verbalise the bias cue. Further, we find that RMCT is more data-efficient at the expense of being less compute-efficient in our experiments. Overall, RMCT shows that consistency training can improve behavioural robustness without directly trading off against monitorability.

Evaluating and Understanding Scheming Propensity in LLM Agents

As frontier language models are increasingly deployed as autonomous agents pursuing complex, long-term objectives, there is increased risk of scheming: agents covertly pursuing misaligned goals. Prior work has focused on showing agents are capable of scheming, but their propensity to scheme in realistic scenarios remains underexplored. To understand when agents scheme, we decompose scheming incentives into agent factors and environmental factors. We develop realistic settings allowing us to systematically vary these factors, each with scheming opportunities for agents that pursue instrumentally convergent goals such as self-preservation, resource acquisition, and goal-guarding. We find only minimal instances of scheming despite high environmental incentives, and show this is unlikely due to evaluation awareness. While inserting adversarially-designed prompt snippets that encourage agency and goal-directedness into an agent's system prompt can induce high scheming rates, snippets used in real agent scaffolds rarely do. Surprisingly, in model organisms (Hubinger et al., 2023) built with these snippets, scheming behavior is remarkably brittle: removing a single tool can drop the scheming rate from 59% to 3%, and increasing oversight can raise rather than deter scheming by up to 25%. Our incentive decomposition enables systematic measurement of scheming propensity in settings relevant for deployment, which is necessary as agents are entrusted with increasingly consequential tasks.

The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence

The safety alignment of large language models (LLMs) can be circumvented through adversarially crafted inputs, yet the mechanisms by which these attacks bypass safety barriers remain poorly understood. Prior work suggests that a single refusal direction in the model's activation space determines whether an LLM refuses a request. In this study, we propose a novel gradient-based approach to representation engineering and use it to identify refusal directions. Contrary to prior work, we uncover multiple independent directions and even multi-dimensional concept cones that mediate refusal. Moreover, we show that orthogonality alone does not imply independence under intervention, motivating the notion of representational independence that accounts for both linear and non-linear effects. Using this framework, we identify mechanistically independent refusal directions. We show that refusal mechanisms in LLMs are governed by complex spatial structures and identify functionally independent directions, confirming that multiple distinct mechanisms drive refusal behavior. Our gradient-based approach uncovers these mechanisms and can further serve as a foundation for future work on understanding LLMs.

Simplified Learning of CAD Features Leveraging a Deep Residual Autoencoder

In the domain of computer vision, deep residual neural networks like EfficientNet have set new standards in terms of robustness and accuracy. One key problem underlying the training of deep neural networks is the immanent lack of a sufficient amount of training data. The problem worsens especially if labels cannot be generated automatically, but have to be annotated manually. This challenge occurs for instance if expert knowledge related to 3D parts should be externalized based on example models. One way to reduce the necessary amount of labeled data may be the use of autoencoders, which can be learned in an unsupervised fashion without labeled data. In this work, we present a deep residual 3D autoencoder based on the EfficientNet architecture, intended for transfer learning tasks related to 3D CAD model assessment. For this purpose, we adopted EfficientNet to 3D problems like voxel models derived from a STEP file. Striving to reduce the amount of labeled 3D data required, the networks encoder can be utilized for transfer training.

Feature Visualization within an Automated Design Assessment leveraging Explainable Artificial Intelligence Methods

Not only automation of manufacturing processes but also automation of automation procedures itself become increasingly relevant to automation research. In this context, automated capability assessment, mainly leveraged by deep learning systems driven from 3D CAD data, have been presented. Current assessment systems may be able to assess CAD data with regards to abstract features, e.g. the ability to automatically separate components from bulk goods, or the presence of gripping surfaces. Nevertheless, they suffer from the factor of black box systems, where an assessment can be learned and generated easily, but without any geometrical indicator about the reasons of the system's decision. By utilizing explainable AI (xAI) methods, we attempt to open up the black box. Explainable AI methods have been used in order to assess whether a neural network has successfully learned a given task or to analyze which features of an input might lead to an adversarial attack. These methods aim to derive additional insights into a neural network, by analyzing patterns from a given input and its impact to the network output. Within the NeuroCAD Project, xAI methods are used to identify geometrical features which are associated with a certain abstract feature. Within this work, a sensitivity analysis (SA), the layer-wise relevance propagation (LRP), the Gradient-weighted Class Activation Mapping (Grad-CAM) method as well as the Local Interpretable Model-Agnostic Explanations (LIME) have been implemented in the NeuroCAD environment, allowing not only to assess CAD models but also to identify features which have been relevant for the network decision. In the medium run, this might enable to identify regions of interest supporting product designers to optimize their models with regards to assembly processes.