arXiv subjects

Roland S. Zimmermann

Publications and source records attributed to Roland S. Zimmermann.

At least 19 recordsLinked to original sources

arXiv2026

CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs

AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model. This partially tampers with the untrusted model's trajectory. If the acting model detects such an intervention, it may infer properties of the monitor and adapt to evade the control protocol. We introduce CIAware-Bench, a benchmark for measuring control intervention (CI) awareness across frontier models. CIAware-Bench tests whether models can distinguish their own trajectories from those modified by a control intervention. The benchmark comprises four task domains (essay writing, BigCodeBench, Bash Arena, and SHADE-Arena), with options to vary trajectory watermarking, side-task presence, and the control protocol. Evaluating thirteen acting models with eight intervening models shows substantial variation between settings and model pairs. CI awareness rises sharply for GPT-6 Astra and the Claude 5 models (Fable 5 and Opus 5). When explicitly probed, Astra reaches mean AUROC of 0.90 on Essay, 0.91 on BigCodeBench, 0.86 on Bash Arena and 0.77 on SHADE-Arena. Fable 5 and Opus 5 both reach 0.77 on Essay, with less consistent gains in the other environments. On SHADE-Arena, we demonstrate that telling an acting model that an action was replaced and asking it to adapt leads to moderate improvements in monitor evasion rates. In summary, our results highlight that control evaluations should already assume perfect CI awareness for conservative safety estimates, and that protocol design should explore countermeasures that make interventions harder to detect.

cs.AI↗

arXiv2026

LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models

Out-of-distribution (OOD) robustness is a desired property of computer vision models. Improving model robustness requires high-quality signals from robustness benchmarks to quantify progress. While various benchmark datasets such as ImageNet-C were proposed in the ImageNet era, most ImageNet-C corruption types are no longer OOD relative to today's large, web-scraped datasets, which already contain common corruptions such as blur or JPEG compression artifacts. Consequently, these benchmarks are no longer well-suited for evaluating OOD robustness in the era of web-scale datasets. Indeed, recent models show saturating scores on ImageNet-era OOD benchmarks, indicating that it is unclear whether models trained on web-scale datasets truly become better at OOD generalization or whether they have simply been exposed to the test distortions during training. To address this, we introduce LAION-C as a benchmark alternative for ImageNet-C. LAION-C consists of six novel distortion types specifically designed to be OOD, even for web-scale datasets such as LAION. In a comprehensive evaluation of state-of-the-art models, we find that the LAION-C dataset poses significant challenges to contemporary models, including MLLMs such as Gemini and GPT-4o. We additionally conducted a psychophysical experiment to evaluate the difficulty of our corruptions for human observers, enabling a comparison of models to lab-quality human robustness data. We observe a paradigm shift in OOD generalization: from humans outperforming models, to the best models now matching or outperforming the best human observers.

cs.CV↗

arXiv2026

Exploration Hacking: Can LLMs Learn to Resist RL Training?

Reinforcement learning (RL) has become essential to the post-training of large language models (LLMs) for reasoning, agentic capabilities and alignment. Successful RL relies on sufficient exploration of diverse actions by the model during training, which creates a potential failure mode: a model could strategically alter its exploration during training to influence the subsequent training outcome. In this paper we study this behavior, called exploration hacking. First, we create model organisms of selective RL resistance by fine-tuning LLMs to follow specific underperformance strategies; these models can successfully resist our RL-based capability elicitation in agentic biosecurity and AI R&D environments while maintaining performance on related tasks. We then use our model organisms to evaluate detection and mitigation strategies, including monitoring, weight noising, and SFT-based elicitation. Finally, we show that current frontier models can exhibit explicit reasoning about suppressing their exploration when provided with sufficient information about their training context, with higher rates when this information is acquired indirectly through the environment. Together, our results suggest exploration hacking is a possible failure mode of RL on sufficiently capable LLMs.

cs.LG↗

arXiv2026

Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?

Chain-of-Thought (CoT) monitoring, in which automated systems monitor the CoT of an LLM, is a promising approach for effectively overseeing AI systems. However, the extent to which a model's CoT helps us oversee the model - the monitorability of the CoT - can be affected by training, for instance by the model learning to hide important features of its reasoning. We propose and empirically validate a conceptual framework for predicting when and why this occurs. We model LLM post-training as an RL environment where the reward decomposes into two terms: one term depending on final outputs and another term depending on the CoT. Our framework allows us to classify these two terms as "aligned", "orthogonal", or "in-conflict" before training. We predict that training with in-conflict terms will reduce monitorability, orthogonal terms will not affect it, and aligned terms will improve it. To validate our framework, we use it to classify a set of RL environments, train LLMs within those environments, and evaluate how training affects CoT monitorability. We find that (1) training with "in-conflict" reward terms reduces CoT monitorability and (2) optimizing in-conflict reward terms is difficult.

cs.LG↗

arXiv2025

Practical challenges of control monitoring in frontier AI deployments

Automated control monitors could play an important role in overseeing highly capable AI agents that we do not fully trust. Prior work has explored control monitoring in simplified settings, but scaling monitoring to real-world deployments introduces additional dynamics: parallel agent instances, non-negligible oversight latency, incremental attacks between agent instances, and the difficulty of identifying scheming agents based on individual harmful actions. In this paper, we analyse design choices to address these challenges, focusing on three forms of monitoring with different latency-safety trade-offs: synchronous, semi-synchronous, and asynchronous monitoring. We introduce a high-level safety case sketch as a tool for understanding and comparing these monitoring protocols. Our analysis identifies three challenges -- oversight, latency, and recovery -- and explores them in four case studies of possible future AI deployments.

cs.CR↗

arXiv2025

A Pragmatic Way to Measure Chain-of-Thought Monitorability

While Chain-of-Thought (CoT) monitoring offers a unique opportunity for AI safety, this opportunity could be lost through shifts in training practices or model architecture. To help preserve monitorability, we propose a pragmatic way to measure two components of it: legibility (whether the reasoning can be followed by a human) and coverage (whether the CoT contains all the reasoning needed for a human to also produce the final output). We implement these metrics with an autorater prompt that enables any capable LLM to compute the legibility and coverage of existing CoTs. After sanity-checking our prompted autorater with synthetic CoT degradations, we apply it to several frontier models on challenging benchmarks, finding that they exhibit high monitorability. We present these metrics, including our complete autorater prompt, as a tool for developers to track how design decisions impact monitorability. While the exact prompt we share is still a preliminary version under ongoing development, we are sharing it now in the hopes that others in the community will find it useful. Our method helps measure the default monitorability of CoT - it should be seen as a complement, not a replacement, for the adversarial stress-testing needed to test robustness against deliberately evasive models.

cs.LG↗

arXiv2025

Early Signs of Steganographic Capabilities in Frontier LLMs

Monitoring Large Language Model (LLM) outputs is crucial for mitigating risks from misuse and misalignment. However, LLMs could evade monitoring through steganography: Encoding hidden information within seemingly benign generations. In this paper, we evaluate the steganography capabilities in frontier LLMs to better understand the risk they pose. We focus on two types of steganography: passing encoded messages and performing encoded reasoning. We find that current models are unable to encode short messages in their outputs without a monitor noticing under standard affordances. They can succeed, however, if given additional affordances like using an unmonitored scratchpad and coordinating on what encoding scheme to use. We additionally find early signs that models can perform basic encoded reasoning in a simple state-tracking problem. This includes some ability to reason with their own and pre-defined schemes, including encoding schemes such as Hexadecimal. Despite this, they can rarely hide reasoning subtly within a cover task to fool a monitor. Overall, our results indicate that current LLMs exhibit nascent steganographic capabilities. While these capabilities are likely insufficient to bypass well-designed monitors at present, this could change in the future.

cs.CR↗

arXiv2025

Evaluating Frontier Models for Stealth and Situational Awareness

Recent work has demonstrated the plausibility of frontier AI models scheming -- knowingly and covertly pursuing an objective misaligned with its developer's intentions. Such behavior could be very hard to detect, and if present in future advanced systems, could pose severe loss of control risk. It is therefore important for AI developers to rule out harm from scheming prior to model deployment. In this paper, we present a suite of scheming reasoning evaluations measuring two types of reasoning capabilities that we believe are prerequisites for successful scheming: First, we propose five evaluations of ability to reason about and circumvent oversight (stealth). Second, we present eleven evaluations for measuring a model's ability to instrumentally reason about itself, its environment and its deployment (situational awareness). We demonstrate how these evaluations can be used as part of a scheming inability safety case: a model that does not succeed on these evaluations is almost certainly incapable of causing severe harm via scheming in real deployment. We run our evaluations on current frontier models and find that none of them show concerning levels of either situational awareness or stealth.

cs.LG↗

arXiv2025

In Search of Forgotten Domain Generalization

Out-of-Domain (OOD) generalization is the ability of a model trained on one or more domains to generalize to unseen domains. In the ImageNet era of computer vision, evaluation sets for measuring a model's OOD performance were designed to be strictly OOD with respect to style. However, the emergence of foundation models and expansive web-scale datasets has obfuscated this evaluation process, as datasets cover a broad range of domains and risk test domain contamination. In search of the forgotten domain generalization, we create large-scale datasets subsampled from LAION -- LAION-Natural and LAION-Rendition -- that are strictly OOD to corresponding ImageNet and DomainNet test sets in terms of style. Training CLIP models on these datasets reveals that a significant portion of their performance is explained by in-domain examples. This indicates that the OOD generalization challenges from the ImageNet era still prevail and that training on web-scale data merely creates the illusion of OOD generalization. Furthermore, through a systematic exploration of combining natural and rendition datasets in varying proportions, we identify optimal mixing ratios for model generalization across these domains. Our datasets and results re-enable meaningful assessment of OOD robustness at scale -- a crucial prerequisite for improving model robustness.

cs.CV↗

arXiv2025

InfoNCE: Identifying the Gap Between Theory and Practice

Prior theory work on Contrastive Learning via the InfoNCE loss showed that, under certain assumptions, the learned representations recover the ground-truth latent factors. We argue that these theories overlook crucial aspects of how CL is deployed in practice. Specifically, they either assume equal variance across all latents or that certain latents are kept invariant. However, in practice, positive pairs are often generated using augmentations such as strong cropping to just a few pixels. Hence, a more realistic assumption is that all latent factors change with a continuum of variability across all factors. We introduce AnInfoNCE, a generalization of InfoNCE that can provably uncover the latent factors in this anisotropic setting, broadly generalizing previous identifiability results in CL. We validate our identifiability results in controlled experiments and show that AnInfoNCE increases the recovery of previously collapsed information in CIFAR10 and ImageNet, albeit at the cost of downstream accuracy. Finally, we discuss the remaining mismatches between theoretical assumptions and practical implementations.

cs.LG↗

arXiv2024

Don't trust your eyes: on the (un)reliability of feature visualizations

How do neural networks extract patterns from pixels? Feature visualizations attempt to answer this important question by visualizing highly activating patterns through optimization. Today, visualization methods form the foundation of our knowledge about the internal workings of neural networks, as a type of mechanistic interpretability. Here we ask: How reliable are feature visualizations? We start our investigation by developing network circuits that trick feature visualizations into showing arbitrary patterns that are completely disconnected from normal network behavior on natural input. We then provide evidence for a similar phenomenon occurring in standard, unmanipulated networks: feature visualizations are processed very differently from standard input, casting doubt on their ability to "explain" how neural networks process natural images. This can be used as a sanity check for feature visualizations. We underpin our empirical findings by theory proving that the set of functions that can be reliably understood by feature visualization is extremely small and does not include general black-box neural networks. Therefore, a promising way forward could be the development of networks that enforce certain structures in order to ensure more reliable feature visualizations.

cs.CV↗

arXiv2024

Scale Alone Does not Improve Mechanistic Interpretability in Vision Models

In light of the recent widespread adoption of AI systems, understanding the internal information processing of neural networks has become increasingly critical. Most recently, machine vision has seen remarkable progress by scaling neural networks to unprecedented levels in dataset and model size. We here ask whether this extraordinary increase in scale also positively impacts the field of mechanistic interpretability. In other words, has our understanding of the inner workings of scaled neural networks improved as well? We use a psychophysical paradigm to quantify one form of mechanistic interpretability for a diverse suite of nine models and find no scaling effect for interpretability - neither for model nor dataset size. Specifically, none of the investigated state-of-the-art models are easier to interpret than the GoogLeNet model from almost a decade ago. Latest-generation vision models appear even less interpretable than older architectures, hinting at a regression rather than improvement, with modern models sacrificing interpretability for accuracy. These results highlight the need for models explicitly designed to be mechanistically interpretable and the need for more helpful interpretability methods to increase our understanding of networks at an atomic level. We release a dataset containing more than 130'000 human responses from our psychophysical evaluation of 767 units across nine models. This dataset facilitates research on automated instead of human-based interpretability evaluations, which can ultimately be leveraged to directly optimize the mechanistic interpretability of models.

cs.CV↗

arXiv2023

Sensitivity of Slot-Based Object-Centric Models to their Number of Slots

Self-supervised methods for learning object-centric representations have recently been applied successfully to various datasets. This progress is largely fueled by slot-based methods, whose ability to cluster visual scenes into meaningful objects holds great promise for compositional generalization and downstream learning. In these methods, the number of slots (clusters) $K$ is typically chosen to match the number of ground-truth objects in the data, even though this quantity is unknown in real-world settings. Indeed, the sensitivity of slot-based methods to $K$, and how this affects their learned correspondence to objects in the data has largely been ignored in the literature. In this work, we address this issue through a systematic study of slot-based methods. We propose using analogs to precision and recall based on the Adjusted Rand Index to accurately quantify model behavior over a large range of $K$. We find that, especially during training, incorrect choices of $K$ do not yield the desired object decomposition and, in fact, cause substantial oversegmentation or merging of separate objects (undersegmentation). We demonstrate that the choice of the objective function and incorporating instance-level annotations can moderately mitigate this behavior while still falling short of fully resolving this issue. Indeed, we show how this issue persists across multiple methods and datasets and stress its importance for future slot-based models.

cs.CV↗

arXiv2023

Provably Learning Object-Centric Representations

Learning structured representations of the visual world in terms of objects promises to significantly improve the generalization abilities of current machine learning models. While recent efforts to this end have shown promising empirical progress, a theoretical account of when unsupervised object-centric representation learning is possible is still lacking. Consequently, understanding the reasons for the success of existing object-centric methods as well as designing new theoretically grounded methods remains challenging. In the present work, we analyze when object-centric representations can provably be learned without supervision. To this end, we first introduce two assumptions on the generative process for scenes comprised of several objects, which we call compositionality and irreducibility. Under this generative process, we prove that the ground-truth object representations can be identified by an invertible and compositional inference model, even in the presence of dependencies between objects. We empirically validate our results through experiments on synthetic data. Finally, we provide evidence that our theory holds predictive power for existing object-centric models by showing a close correspondence between models' compositionality and invertibility and their empirical identifiability.

cs.LG↗

arXiv2022

Increasing Confidence in Adversarial Robustness Evaluations

Hundreds of defenses have been proposed to make deep neural networks robust against minimal (adversarial) input perturbations. However, only a handful of these defenses held up their claims because correctly evaluating robustness is extremely challenging: Weak attacks often fail to find adversarial examples even if they unknowingly exist, thereby making a vulnerable network look robust. In this paper, we propose a test to identify weak attacks, and thus weak defense evaluations. Our test slightly modifies a neural network to guarantee the existence of an adversarial example for every sample. Consequentially, any correct attack must succeed in breaking this modified network. For eleven out of thirteen previously-published defenses, the original evaluation of the defense fails our test, while stronger attacks that break these defenses pass it. We hope that attack unit tests - such as ours - will be a major component in future robustness evaluations and increase confidence in an empirical field that is currently riddled with skepticism.

cs.LG↗

arXiv2022

Contrastive Learning Inverts the Data Generating Process

Contrastive learning has recently seen tremendous success in self-supervised learning. So far, however, it is largely unclear why the learned representations generalize so effectively to a large variety of downstream tasks. We here prove that feedforward models trained with objectives belonging to the commonly used InfoNCE family learn to implicitly invert the underlying generative model of the observed data. While the proofs make certain statistical assumptions about the generative model, we observe empirically that our findings hold even if these assumptions are severely violated. Our theory highlights a fundamental connection between contrastive learning, generative modeling, and nonlinear independent component analysis, thereby furthering our understanding of the learned representations as well as providing a theoretical foundation to derive more effective contrastive losses.

cs.LG↗

arXiv2021

Score-Based Generative Classifiers

The tremendous success of generative models in recent years raises the question whether they can also be used to perform classification. Generative models have been used as adversarially robust classifiers on simple datasets such as MNIST, but this robustness has not been observed on more complex datasets like CIFAR-10. Additionally, on natural image datasets, previous results have suggested a trade-off between the likelihood of the data and classification accuracy. In this work, we investigate score-based generative models as classifiers for natural images. We show that these models not only obtain competitive likelihood values but simultaneously achieve state-of-the-art classification accuracy for generative classifiers on CIFAR-10. Nevertheless, we find that these models are only slightly, if at all, more robust than discriminative baseline models on out-of-distribution tasks based on common image corruptions. Similarly and contrary to prior results, we find that score-based are prone to worst-case distribution shifts in the form of adversarial perturbations. Our work highlights that score-based generative models are closing the gap in classification accuracy compared to standard discriminative models. While they do not yet deliver on the promise of adversarial and out-of-domain robustness, they provide a different approach to classification that warrants further research.

stat.ML↗

arXiv2021

How Well do Feature Visualizations Support Causal Understanding of CNN Activations?

A precise understanding of why units in an artificial network respond to certain stimuli would constitute a big step towards explainable artificial intelligence. One widely used approach towards this goal is to visualize unit responses via activation maximization. These synthetic feature visualizations are purported to provide humans with precise information about the image features that cause a unit to be activated - an advantage over other alternatives like strongly activating natural dataset samples. If humans indeed gain causal insight from visualizations, this should enable them to predict the effect of an intervention, such as how occluding a certain patch of the image (say, a dog's head) changes a unit's activation. Here, we test this hypothesis by asking humans to decide which of two square occlusions causes a larger change to a unit's activation. Both a large-scale crowdsourced experiment and measurements with experts show that on average the extremely activating feature visualizations by Olah et al. (2017) indeed help humans on this task ($68 \pm 4$% accuracy; baseline performance without any visualizations is $60 \pm 3$%). However, they do not provide any substantial advantage over other visualizations (such as e.g. dataset samples), which yield similar performance ($66\pm3$% to $67 \pm3$% accuracy). Taken together, we propose an objective psychophysical task to quantify the benefit of unit-level interpretability methods for humans, and find no evidence that a widely-used feature visualization method provides humans with better "causal understanding" of unit activations than simple alternative visualizations.

cs.CV↗