Search arXiv⌕ Search

arXiv subjects

Simon Schrodi

Publications and source records attributed to Simon Schrodi.

15 recordsLinked to original sources

Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect

We study homeopathic learning, a phenomenon where poisoned training data with little effect on its own can create a hidden backdoor once clean, contrasting data is added. In our experiments, a model fine-tuned only on a "poison" dataset associated with a trait T (such as speaking German or being misaligned) does not learn T. Surprisingly, adding similarly formatted "clean" data not associated with T causes the model to learn T. This effect requires that the poison data have an identifying "trigger" feature that is absent from the clean data. We demonstrate this effect in a variety of settings: subliminal learning of backdoor behaviors, inoculation prompting of reward hacking transcripts, generalization from archaic bird names, and an MNIST classifier with mislabeled data. We propose a simple regression model as an explanation. In the poison data, the task and trigger always co-occur, and regularization favors attributing T to the task. Clean data without the trigger causes T to be attributed to the trigger instead. Consequently, the same training examples can have little effect as part of one training set but create a misalignment backdoor in another. This finding exposes a potential limitation of safety methods that audit or filter training data in isolation.

cs.AI↗

What Drives Compositional Generalization in Visual Generative Models? The Importance of Continuous Training Objectives

Compositional generalization, the ability to generate novel combinations of known concepts, is a key ingredient for visual generative models. Yet, not all mechanisms that enable or inhibit it are fully understood. In this work, we conduct a systematic study of which design choices critically determine compositional generalization in image and video generation. By isolating independent design axes, we identify two key factors strongly associated with compositional success: (i) whether the training objective operates on a discrete or continuous distribution, and (ii) the completeness of conditioning information about constituent factors during training. We also show that relaxing the discrete loss with an auxiliary continuous latent objective can partially recover compositional performance in discrete models like MaskGIT. Our findings, corroborated by diverse compositional tasks and preliminary evidence in world models and LLMs, motivate a shift toward continuous objectives for compositional generalization.

cs.CV↗

AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors

We introduce AuditBench, an alignment auditing benchmark. AuditBench consists of 56 language models with implanted hidden behaviors. Each model has one of 14 concerning behaviors--such as sycophantic deference, opposition to AI regulation, or secret geopolitical loyalties--which it does not confess to when directly asked. AuditBench models are highly diverse--some are subtle, while others are overt, and we use varying training techniques both for implanting behaviors and training models not to confess. To demonstrate AuditBench's utility, we develop an investigator agent that autonomously employs a configurable set of auditing tools. By measuring investigator agent success using different tools, we can evaluate their efficacy. Notably, we observe a tool-to-agent gap, where tools that perform well in standalone non-agentic evaluations fail to translate into improved performance when used with our investigator agent. We find that our most effective tools involve scaffolded calls to auxiliary models that generate diverse prompts for the target. White-box interpretability tools can be helpful, but the agent performs best with black-box tools. We also find that audit success varies greatly across training techniques: models trained on synthetic documents are easier to audit than models trained on demonstrations, with better adversarial training further increasing auditing difficulty. We release our models, agent, and evaluation framework to support future quantitative, iterative science on alignment auditing.

cs.CL↗

Towards Understanding Subliminal Learning: When and How Hidden Biases Transfer

Language models can transfer hidden biases during distillation. For example, a teacher that "likes owls" can make its student "like owls" too, even when the training data consists only of lists of numbers. This surprising phenomenon is called subliminal learning. Subliminal learning can be expected under soft distillation, where the student is trained on the teacher's full next-token distribution. But the fact that this also occurs under hard distillation-where the student only sees sampled tokens-raises a deeper question: when and how does subliminal learning actually occur? We answer this question through controlled experiments and mechanistic analysis. Our results show that subliminal learning does not need (global) token entanglement or logit leakage. Instead, it comes down to a small set of divergence tokens-rare cases where teachers with different biases would predict different tokens. Masking out these tokens mostly removes the hidden bias transfer. Mechanistically, divergence tokens reveal that early layers are critical. Surprisingly, finetuning even a single such early layer is sufficient for subliminal learning. Finally, we find that subliminal learning is fragile. Even small changes, like prompt paraphrasings, are usually sufficient to suppress it.

cs.LG↗

Simple LLM Baselines are Competitive for Model Diffing

Standard LLM evaluations only test capabilities or dispositions that evaluators designed them for, missing unexpected differences such as behavioral shifts between model revisions or emergent misaligned tendencies. Model diffing addresses this limitation by automatically surfacing systematic behavioral differences. Recent approaches include LLM-based methods that generate natural language descriptions and sparse autoencoder (SAE)-based methods that identify interpretable features. However, no systematic comparison of these approaches exists nor are there established evaluation criteria. We address this gap by proposing evaluation metrics for key desiderata (generalization, interestingness, and abstraction level) and use these to compare existing methods. Our results show that an improved LLM-based baseline performs comparably to the SAE-based method while typically surfacing more abstract behavioral differences.

cs.LG↗

When and How Does CLIP Enable Domain and Compositional Generalization?

The remarkable generalization performance of contrastive vision-language models like CLIP is often attributed to the diversity of their training distributions. However, key questions remain unanswered: Can CLIP generalize to an entirely unseen domain when trained on a diverse mixture of domains (domain generalization)? Can it generalize to unseen classes within partially seen domains (compositional generalization)? What factors affect such generalization? To answer these questions, we trained CLIP models on systematically constructed training distributions with controlled domain diversity and object class exposure. Our experiments show that domain diversity is essential for both domain and compositional generalization, yet compositional generalization can be surprisingly weaker than domain generalization when the training distribution contains a suboptimal subset of the test domain. Through data-centric and mechanistic analyses, we find that successful generalization requires the learning of sufficiently shared representations in intermediate layers and circuits.

cs.LG↗

Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models

Contrastive vision-language models (VLMs), like CLIP, have gained popularity for their versatile applicability to various downstream tasks. Despite their successes in some tasks, like zero-shot object recognition, they perform surprisingly poor on other tasks, like attribute recognition. Previous work has attributed these challenges to the modality gap, a separation of image and text in the shared representation space, and to a bias towards objects over other factors, such as attributes. In this analysis paper, we investigate both phenomena thoroughly. We evaluated off-the-shelf VLMs and while the gap's influence on performance is typically overshadowed by other factors, we find indications that closing the gap indeed leads to improvements. Moreover, we find that, contrary to intuition, only few embedding dimensions drive the gap and that the embedding spaces are differently organized. To allow for a clean study of object bias, we introduce a definition and a corresponding measure of it. Equipped with this tool, we find that object bias does not lead to worse performance on other concepts, such as attributes per se. However, why do both phenomena, modality gap and object bias, emerge in the first place? To answer this fundamental question and uncover some of the inner workings of contrastive VLMs, we conducted experiments that allowed us to control the amount of shared information between the modalities. These experiments revealed that the driving factor behind both the modality gap and the object bias, is an information imbalance between images and captions, and unveiled an intriguing connection between the modality gap and entropy of the logits.

cs.CV↗

Concept Bottleneck Models Without Predefined Concepts

There has been considerable recent interest in interpretable concept-based models such as Concept Bottleneck Models (CBMs), which first predict human-interpretable concepts and then map them to output classes. To reduce reliance on human-annotated concepts, recent works have converted pretrained black-box models into interpretable CBMs post-hoc. However, these approaches predefine a set of concepts, assuming which concepts a black-box model encodes in its representations. In this work, we eliminate this assumption by leveraging unsupervised concept discovery to automatically extract concepts without human annotations or a predefined set of concepts. We further introduce an input-dependent concept selection mechanism that ensures only a small subset of concepts is used across all classes. We show that our approach improves downstream performance and narrows the performance gap to black-box models, while using significantly fewer concepts in the classification. Finally, we demonstrate how large vision-language models can intervene on the final model weights to correct model errors.

cs.LG↗

Eureka-Moments in Transformers: Multi-Step Tasks Reveal Softmax Induced Optimization Problems

In this work, we study rapid improvements of the training loss in transformers when being confronted with multi-step decision tasks. We found that transformers struggle to learn the intermediate task and both training and validation loss saturate for hundreds of epochs. When transformers finally learn the intermediate task, they do this rapidly and unexpectedly. We call these abrupt improvements Eureka-moments, since the transformer appears to suddenly learn a previously incomprehensible concept. We designed synthetic tasks to study the problem in detail, but the leaps in performance can be observed also for language modeling and in-context learning (ICL). We suspect that these abrupt transitions are caused by the multi-step nature of these tasks. Indeed, we find connections and show that ways to improve on the synthetic multi-step tasks can be used to improve the training of language modeling and ICL. Using the synthetic data we trace the problem back to the Softmax function in the self-attention block of transformers and show ways to alleviate the problem. These fixes reduce the required number of training steps, lead to higher likelihood to learn the intermediate task, to higher final accuracy and training becomes more robust to hyper-parameters.

cs.LG↗

Is Mamba Capable of In-Context Learning?

State of the art foundation models such as GPT-4 perform surprisingly well at in-context learning (ICL), a variant of meta-learning concerning the learned ability to solve tasks during a neural network forward pass, exploiting contextual information provided as input to the model. This useful ability emerges as a side product of the foundation model's massive pretraining. While transformer models are currently the state of the art in ICL, this work provides empirical evidence that Mamba, a newly proposed state space model which scales better than transformers w.r.t. the input sequence length, has similar ICL capabilities. We evaluated Mamba on tasks involving simple function approximation as well as more complex natural language processing problems. Our results demonstrate that, across both categories of tasks, Mamba closely matches the performance of transformer models for ICL. Further analysis reveals that, like transformers, Mamba appears to solve ICL problems by incrementally optimizing its internal representations. Overall, our work suggests that Mamba can be an efficient alternative to transformers for ICL tasks involving long input sequences. This is an exciting finding in meta-learning and may enable generalizations of in-context learned AutoML algorithms (like TabPFN or Optformer) to long input sequences.

cs.LG↗

Construction of Hierarchical Neural Architecture Search Spaces based on Context-free Grammars

The discovery of neural architectures from simple building blocks is a long-standing goal of Neural Architecture Search (NAS). Hierarchical search spaces are a promising step towards this goal but lack a unifying search space design framework and typically only search over some limited aspect of architectures. In this work, we introduce a unifying search space design framework based on context-free grammars that can naturally and compactly generate expressive hierarchical search spaces that are 100s of orders of magnitude larger than common spaces from the literature. By enhancing and using their properties, we effectively enable search over the complete architecture and can foster regularity. Further, we propose an efficient hierarchical kernel design for a Bayesian Optimization search strategy to efficiently search over such huge spaces. We demonstrate the versatility of our search space design framework and show that our search strategy can be superior to existing NAS approaches. Code is available at https://github.com/automl/hierarchical_nas_construction.

cs.LG↗

Latent Diffusion Counterfactual Explanations

Counterfactual explanations have emerged as a promising method for elucidating the behavior of opaque black-box models. Recently, several works leveraged pixel-space diffusion models for counterfactual generation. To handle noisy, adversarial gradients during counterfactual generation -- causing unrealistic artifacts or mere adversarial perturbations -- they required either auxiliary adversarially robust models or computationally intensive guidance schemes. However, such requirements limit their applicability, e.g., in scenarios with restricted access to the model's training data. To address these limitations, we introduce Latent Diffusion Counterfactual Explanations (LDCE). LDCE harnesses the capabilities of recent class- or text-conditional foundation latent diffusion models to expedite counterfactual generation and focus on the important, semantic parts of the data. Furthermore, we propose a novel consensus guidance mechanism to filter out noisy, adversarial gradients that are misaligned with the diffusion model's implicit classifier. We demonstrate the versatility of LDCE across a wide spectrum of models trained on diverse datasets with different learning paradigms. Finally, we showcase how LDCE can provide insights into model errors, enhancing our understanding of black-box model behavior.

cs.LG↗

Climate-sensitive Urban Planning through Optimization of Tree Placements

Climate change is increasing the intensity and frequency of many extreme weather events, including heatwaves, which results in increased thermal discomfort and mortality rates. While global mitigation action is undoubtedly necessary, so is climate adaptation, e.g., through climate-sensitive urban planning. Among the most promising strategies is harnessing the benefits of urban trees in shading and cooling pedestrian-level environments. Our work investigates the challenge of optimal placement of such trees. Physical simulations can estimate the radiative and thermal impact of trees on human thermal comfort but induce high computational costs. This rules out optimization of tree placements over large areas and considering effects over longer time scales. Hence, we employ neural networks to simulate the point-wise mean radiant temperatures--a driving factor of outdoor human thermal comfort--across various time scales, spanning from daily variations to extended time scales of heatwave events and even decades. To optimize tree placements, we harness the innate local effect of trees within the iterated local search framework with tailored adaptations. We show the efficacy of our approach across a wide spectrum of study areas and time scales. We believe that our approach is a step towards empowering decision-makers, urban designers and planners to proactively and effectively assess the potential of urban trees to mitigate heat stress.

cs.CV↗

Towards Understanding Adversarial Robustness of Optical Flow Networks

Recent work demonstrated the lack of robustness of optical flow networks to physical patch-based adversarial attacks. The possibility to physically attack a basic component of automotive systems is a reason for serious concerns. In this paper, we analyze the cause of the problem and show that the lack of robustness is rooted in the classical aperture problem of optical flow estimation in combination with bad choices in the details of the network architecture. We show how these mistakes can be rectified in order to make optical flow networks robust to physical patch-based attacks. Additionally, we take a look at global white-box attacks in the scope of optical flow. We find that targeted white-box attacks can be crafted to bias flow estimation models towards any desired output, but this requires access to the input images and model weights. However, in the case of universal attacks, we find that optical flow networks are robust. Code is available at https://github.com/lmb-freiburg/understanding_flow_robustness.

cs.CV↗

Bag of Baselines for Multi-objective Joint Neural Architecture Search and Hyperparameter Optimization

Neural architecture search (NAS) and hyperparameter optimization (HPO) make deep learning accessible to non-experts by automatically finding the architecture of the deep neural network to use and tuning the hyperparameters of the used training pipeline. While both NAS and HPO have been studied extensively in recent years, NAS methods typically assume fixed hyperparameters and vice versa - there exists little work on joint NAS + HPO. Furthermore, NAS has recently often been framed as a multi-objective optimization problem, in order to take, e.g., resource requirements into account. In this paper, we propose a set of methods that extend current approaches to jointly optimize neural architectures and hyperparameters with respect to multiple objectives. We hope that these methods will serve as simple baselines for future research on multi-objective joint NAS + HPO. To facilitate this, all our code is available at https://github.com/automl/multi-obj-baselines.

cs.LG↗