Matthias Bethge

Karsten Roth, Vishaal Udandarao, Sebastian Dziadzio et al. · cambridge

Multimodal foundation models serve numerous applications at the intersection of vision and language. Still, despite being pretrained on extensive data, they become outdated over time. To keep models updated, research into continual pretraining mainly explores scenarios with either (1) infrequent, indiscriminate updates on large-scale new data, or (2) frequent, sample-level updates. However, practical model deployment often operates in the gap between these two limit cases, as real-world applications often demand adaptation to specific subdomains, tasks or concepts -- spread over the entire, varying life cycle of a model. In this work, we complement current perspectives on continual pretraining through a research test bed as well as provide comprehensive guidance for effective continual model updates in such scenarios. We first introduce FoMo-in-Flux, a continual multimodal pretraining benchmark with realistic compute constraints and practical deployment requirements, constructed over 63 datasets with diverse visual and semantic coverage. Using FoMo-in-Flux, we explore the complex landscape of practical continual pretraining through multiple perspectives: (1) A data-centric investigation of data mixtures and stream orderings that emulate real-world deployment situations, (2) a method-centric investigation ranging from simple fine-tuning and traditional continual learning strategies to parameter-efficient updates and model merging, (3) meta learning rate schedules and mechanistic design choices, and (4) the influence of model and compute scaling. Together, our insights provide a practitioner's guide to continual multimodal pretraining for real-world deployment. Our benchmark and code is here: https://github.com/ExplainableML/fomo_in_flux.

Eli Verwimp, Rahaf Aljundi, Shai Ben-David et al. · deepmind

Continual learning is a subfield of machine learning, which aims to allow machine learning models to continuously learn on new data, by accumulating knowledge without forgetting what was learned in the past. In this work, we take a step back, and ask: "Why should one care about continual learning in the first place?". We set the stage by examining recent continual learning papers published at four major machine learning conferences, and show that memory-constrained settings dominate the field. Then, we discuss five open problems in machine learning, and even though they might seem unrelated to continual learning at first sight, we show that continual learning will inevitably be part of their solution. These problems are model editing, personalization and specialization, on-device learning, faster (re-)training and reinforcement learning. Finally, by comparing the desiderata from these unsolved problems and the current assumptions in continual learning, we highlight and discuss four future directions for continual learning research. We hope that this work offers an interesting perspective on the future of continual learning, while displaying its potential value and the paths we have to pursue in order to make it successful. This work is the result of the many discussions the authors had at the Dagstuhl seminar on Deep Continual Learning, in March 2023.

Vishaal Udandarao, Max F. Burg, Samuel Albanie et al. · cambridge

Recent advances in the development of vision-language models (VLMs) are yielding remarkable success in recognizing visual semantic content, including impressive instances of compositional image understanding. Here, we introduce the novel task of Visual Data-Type Identification, a basic perceptual skill with implications for data curation (e.g., noisy data-removal from large datasets, domain-specific retrieval) and autonomous vision (e.g., distinguishing changing weather conditions from camera lens staining). We develop two datasets consisting of animal images altered across a diverse set of 27 visual data-types, spanning four broad categories. An extensive zero-shot evaluation of 39 VLMs, ranging from 100M to 80B parameters, shows a nuanced performance landscape. While VLMs are reasonably good at identifying certain stylistic \textit{data-types}, such as cartoons and sketches, they struggle with simpler data-types arising from basic manipulations like image rotations or additive noise. Our findings reveal that (i) model scaling alone yields marginal gains for contrastively-trained models like CLIP, and (ii) there is a pronounced drop in performance for the largest auto-regressively trained VLMs like OpenFlamingo. This finding points to a blind spot in current frontier VLMs: they excel in recognizing semantic content but fail to acquire an understanding of visual data-types through scaling. By analyzing the pre-training distributions of these models and incorporating data-type information into the captions during fine-tuning, we achieve a significant enhancement in performance. By exploring this previously uncharted task, we aim to set the stage for further advancing VLMs to equip them with visual data-type understanding. Code and datasets are released at https://github.com/bethgelab/DataTypeIdentification.

Thaddäus Wiedemer, Prasanna Mayilvahanan, Matthias Bethge et al.

Leveraging the compositional nature of our world to expedite learning and facilitate generalization is a hallmark of human perception. In machine learning, on the other hand, achieving compositional generalization has proven to be an elusive goal, even for models with explicit compositional priors. To get a better handle on compositional generalization, we here approach it from the bottom up: Inspired by identifiable representation learning, we investigate compositionality as a property of the data-generating process rather than the data itself. This reformulation enables us to derive mild conditions on only the support of the training distribution and the model architecture, which are sufficient for compositional generalization. We further demonstrate how our theoretical framework applies to real-world scenarios and validate our findings empirically. Our results set the stage for a principled theoretical study of compositional generalization.

Ori Press, Steffen Schneider, Matthias Kümmerer et al.

Test-Time Adaptation (TTA) allows to update pre-trained models to changing data distributions at deployment time. While early work tested these algorithms for individual fixed distribution shifts, recent work proposed and applied methods for continual adaptation over long timescales. To examine the reported progress in the field, we propose the Continually Changing Corruptions (CCC) benchmark to measure asymptotic performance of TTA techniques. We find that eventually all but one state-of-the-art methods collapse and perform worse than a non-adapting model, including models specifically proposed to be robust to performance collapse. In addition, we introduce a simple baseline, "RDumb", that periodically resets the model to its pretrained state. RDumb performs better or on par with the previously proposed state-of-the-art in all considered benchmarks. Our results show that previous TTA approaches are neither effective at regularizing adaptation to avoid collapse nor able to outperform a simplistic resetting strategy.

Ori Press, Andreas Hochlehnert, Ameya Prabhu et al. · uw

Thousands of new scientific papers are published each month. Such information overload complicates researcher efforts to stay current with the state-of-the-art as well as to verify and correctly attribute claims. We pose the following research question: Given a text excerpt referencing a paper, could an LM act as a research assistant to correctly identify the referenced paper? We advance efforts to answer this question by building a benchmark that evaluates the abilities of LMs in citation attribution. Our benchmark, CiteME, consists of text excerpts from recent machine learning papers, each referencing a single other paper. CiteME use reveals a large gap between frontier LMs and human performance, with LMs achieving only 4.2-18.5% accuracy and humans 69.7%. We close this gap by introducing CiteAgent, an autonomous system built on the GPT-4o LM that can also search and read papers, which achieves an accuracy of 35.3\% on CiteME. Overall, CiteME serves as a challenging testbed for open-ended claim attribution, driving the research community towards a future where any claim made by an LM can be automatically verified and discarded if found to be incorrect.

Ilze Amanda Auzina, Çağatay Yıldız, Sara Magliacane et al.

Neural ordinary differential equations (NODEs) have been proven useful for learning non-linear dynamics of arbitrary trajectories. However, current NODE methods capture variations across trajectories only via the initial state value or by auto-regressive encoder updates. In this work, we introduce Modulated Neural ODEs (MoNODEs), a novel framework that sets apart dynamics states from underlying static factors of variation and improves the existing NODE methods. In particular, we introduce $\textit{time-invariant modulator variables}$ that are learned from the data. We incorporate our proposed framework into four existing NODE variants. We test MoNODE on oscillating systems, videos and human walking trajectories, where each trajectory has trajectory-specific modulation. Our framework consistently improves the existing model ability to generalize to new dynamic parameterizations and to perform far-horizon forecasting. In addition, we verify that the proposed modulator variables are informative of the true unknown factors of variation as measured by $R^2$ scores.

Prasanna Mayilvahanan, Thaddäus Wiedemer, Evgenia Rusak et al.

Foundation models like CLIP are trained on hundreds of millions of samples and effortlessly generalize to new tasks and inputs. Out of the box, CLIP shows stellar zero-shot and few-shot capabilities on a wide range of out-of-distribution (OOD) benchmarks, which prior works attribute mainly to today's large and comprehensive training dataset (like LAION). However, it is questionable how meaningful terms like out-of-distribution generalization are for CLIP as it seems likely that web-scale datasets like LAION simply contain many samples that are similar to common OOD benchmarks originally designed for ImageNet. To test this hypothesis, we retrain CLIP on pruned LAION splits that replicate ImageNet's train-test similarity with respect to common OOD benchmarks. While we observe a performance drop on some benchmarks, surprisingly, CLIP's overall performance remains high. This shows that high train-test similarity is insufficient to explain CLIP's OOD performance, and other properties of the training data must drive CLIP to learn more generalizable representations. Additionally, by pruning data points that are dissimilar to the OOD benchmarks, we uncover a 100M split of LAION ($\frac{1}{4}$th of its original size) on which CLIP can be trained to match its original OOD performance.

Thaddäus Wiedemer, Jack Brady, Alexander Panfilov et al.

Learning representations that generalize to novel compositions of known concepts is crucial for bridging the gap between human and machine perception. One prominent effort is learning object-centric representations, which are widely conjectured to enable compositional generalization. Yet, it remains unclear when this conjecture will be true, as a principled theoretical or empirical understanding of compositional generalization is lacking. In this work, we investigate when compositional generalization is guaranteed for object-centric representations through the lens of identifiability theory. We show that autoencoders that satisfy structural assumptions on the decoder and enforce encoder-decoder consistency will learn object-centric representations that provably generalize compositionally. We validate our theoretical result and highlight the practical relevance of our assumptions through experiments on synthetic image data.

Max F. Burg, Thomas Zenkel, Michaela Vystrčilová et al.

Identifying cell types and understanding their functional properties is crucial for unraveling the mechanisms underlying perception and cognition. In the retina, functional types can be identified by carefully selected stimuli, but this requires expert domain knowledge and biases the procedure towards previously known cell types. In the visual cortex, it is still unknown what functional types exist and how to identify them. Thus, for unbiased identification of the functional cell types in retina and visual cortex, new approaches are needed. Here we propose an optimization-based clustering approach using deep predictive models to obtain functional clusters of neurons using Most Discriminative Stimuli (MDS). Our approach alternates between stimulus optimization with cluster reassignment akin to an expectation-maximization algorithm. The algorithm recovers functional clusters in mouse retina, marmoset retina and macaque visual area V4. This demonstrates that our approach can successfully find discriminative stimuli across species, stages of the visual system and recording techniques. The resulting most discriminative stimuli can be used to assign functional cell types fast and on the fly, without the need to train complex predictive models or show a large natural scene dataset, paving the way for experiments that were previously limited by experimental time. Crucially, MDS are interpretable: they visualize the distinctive stimulus patterns that most unambiguously identify a specific type of neuron.

Philipp Davydov, Ameya Prabhu, Matthias Bethge et al.

Understanding how language-model outputs relate to the pretraining corpus is central to studying model behavior. Most training data attribution (TDA) methods ask which training examples causally influence a given output, often using leave-one-out tests. We invert the question: which outputs cannot be attributed to any pretraining example? We introduce un-attributability as an operational measure of semantic novelty: an output is novel if the pretraining corpus contains no semantically similar context. We approximate this with a simple two-stage retrieval pipeline: index the corpus with lightweight GIST embeddings, retrieve the top-n candidates, then rerank with ColBERTv2. If the nearest corpus item is less attributable than a human-generated text reference, we consider the output of the model as novel. We evaluate on SmolLM and SmolLM2 and report three findings: (1) models draw on pretraining data across much longer spans than previously reported; (2) some domains systematically promote or suppress novelty; and (3) instruction tuning not only alters style but also increases novelty. Reframing novelty assessment around un-attributability enables efficient analysis at pretraining scale. We release ~20 TB of corpus chunks and index artifacts to support replication and large-scale extension of our analysis at https://huggingface.co/datasets/stai-tuebingen/faiss-smollm

Luca M. Schulze Buschoff, Elif Akata, Matthias Bethge et al.

A chief goal of artificial intelligence is to build machines that think like people. Yet it has been argued that deep neural network architectures fail to accomplish this. Researchers have asserted these models' limitations in the domains of causal reasoning, intuitive physics, and intuitive psychology. Yet recent advancements, namely the rise of large language models, particularly those designed for visual processing, have rekindled interest in the potential to emulate human-like cognitive abilities. This paper evaluates the current state of vision-based large language models in the domains of intuitive physics, causal reasoning, and intuitive psychology. Through a series of controlled experiments, we investigate the extent to which these modern models grasp complex physical interactions, causal relationships, and intuitive understanding of others' preferences. Our findings reveal that, while some of these models demonstrate a notable proficiency in processing and interpreting visual data, they still fall short of human capabilities in these areas. Our results emphasize the need for integrating more robust mechanisms for understanding causality, physical dynamics, and social cognition into modern-day, vision-based language models, and point out the importance of cognitively-inspired benchmarks.

Anne-Sofie Maerten, Juliane Verwiebe, Shyamgopal Karthik et al.

Modern text-to-image (T2I) models generate high-fidelity visuals but remain indifferent to individual user preferences. While existing reward models optimize for "average" human appeal, they fail to capture the inherent subjectivity of aesthetic judgment. In this work, we introduce a novel dataset and predictive framework, called PAMELA, designed to model personalized image evaluations. Our dataset comprises 70,000 ratings across 5,000 diverse images generated by state-of-the-art models (Flux 2 and Nano Banana). Each image is evaluated by 15 unique users, providing a rich distribution of subjective preferences across domains such as art, design, fashion, and cinematic photography. Leveraging this data, we propose a personalized reward model trained jointly on our high-quality annotations and existing aesthetic assessment subsets. We demonstrate that our model predicts individual liking with higher accuracy than the majority of current state-of-the-art methods predict population-level preferences. Using our personalized predictor, we demonstrate how simple prompt optimization methods can be used to steer generations towards individual user preferences. Our results highlight the importance of data quality and personalization to handle the subjectivity of user preferences. We release our dataset and model to facilitate standardized research in personalized T2I alignment and subjective visual quality assessment.

Ronald Skorobogat, Ameya Prabhu, Matthias Bethge

Multilingual benchmarks guide the development of frontier models. Yet multilingual evaluations reported by frontier models are structured similar to popular reasoning and knowledge benchmarks, but across many languages. We show such benchmarks, and consequently multilingual evaluations, measure mathematical reasoning and factual recall, not multilingual proficiency. For example, thinking variants dramatically outperform instruct variants on these benchmarks, yet often perform worse on real-world multilingual tasks, such as LMArena. We propose a simple alternative: evaluate multilingual capability via round-trip translation. Given text in a source language, translate it to a target language and back; semantic gaps between the original and result expose failures in multilingual generation capabilities. Round-trip translation correlates almost perfectly (\r{ho} = 0.94) with user ratings on LMArena with our benchmark, requires no human reference translations, and does not require a more capable multilingual judge than tested models. Lastly, we introduce Lost in Translation (LiT), a challenging round-trip translation benchmark spanning widely spoken languages worldwide, for realistic evaluation of multilingual frontier models.

Ori Press, Ravid Shwartz-Ziv, Yann LeCun et al.

Entropy minimization (EM) is frequently used to increase the accuracy of classification models when they're faced with new data at test time. EM is a self-supervised learning method that optimizes classifiers to assign even higher probabilities to their top predicted classes. In this paper, we analyze why EM works when adapting a model for a few steps and why it eventually fails after adapting for many steps. We show that, at first, EM causes the model to embed test images close to training images, thereby increasing model accuracy. After many steps of optimization, EM makes the model embed test images far away from the embeddings of training images, which results in a degradation of accuracy. Building upon our insights, we present a method for solving a practical problem: estimating a model's accuracy on a given arbitrary dataset without having access to its labels. Our method estimates accuracy by looking at how the embeddings of input images change as the model is optimized to minimize entropy. Experiments on 23 challenging datasets show that our method sets the SoTA with a mean absolute error of $5.75\%$, an improvement of $29.62\%$ over the previous SoTA on this task. Our code is available at https://github.com/oripress/EntropyEnigma

Ilze Amanda Auzina, Joschka Strüber, Sergio Hernández-Gutiérrez et al.

How can we train agents to navigate uncertainty over long horizons? In this work, we propose ΔBelief-RL, which leverages a language model's own intrinsic beliefs to reward intermediate progress. Our method utilizes the change in the probability an agent assigns to the target solution for credit assignment. By training on synthetic interaction data, ΔBelief-RL teaches information-seeking capabilities that consistently outperform purely outcome-based rewards for Reinforcement Learning, with improvements generalizing to out-of-distribution applications ranging from customer service to personalization. Notably, the performance continues to improve as we scale test-time interactions beyond the training horizon, with interaction-efficiency increasing even on Pass@k metrics. Overall, our work introduces a scalable training strategy for navigating uncertainty over a long-horizon, by enabling credit assignment to intermediate actions via intrinsic ΔBelief rewards.

Alexander Rubinstein, Ameya Prabhu, Matthias Bethge et al.

Object-centric learning (OCL) seeks to learn representations that only encode an object, isolated from other objects or background cues in a scene. This approach underpins various aims, including out-of-distribution (OOD) generalization, sample-efficient composition, and modeling of structured environments. Most research has focused on developing unsupervised mechanisms that separate objects into discrete slots in the representation space, evaluated using unsupervised object discovery. However, with recent sample-efficient segmentation models, we can separate objects in the pixel space and encode them independently. This achieves remarkable zero-shot performance on OOD object discovery benchmarks, is scalable to foundation models, and can handle a variable number of slots out-of-the-box. Hence, the goal of OCL methods to obtain object-centric representations has been largely achieved. Despite this progress, a key question remains: How does the ability to separate objects within a scene contribute to broader OCL objectives, such as OOD generalization? We address this by investigating the OOD generalization challenge caused by spurious background cues through the lens of OCL. We propose a novel, training-free probe called Object-Centric Classification with Applied Masks (OCCAM), demonstrating that segmentation-based encoding of individual objects significantly outperforms slot-based OCL methods. However, challenges in real-world applications remain. We provide the toolbox for the OCL community to use scalable object-centric representations, and focus on practical applications and fundamental questions, such as understanding object perception in human cognition. Our code is available here: https://github.com/AlexanderRubinstein/OCCAM.

Ori Press, Brandon Amos, Haoyu Zhao et al. · princeton, uw

Despite progress in language model (LM) capabilities, evaluations have thus far focused on models' performance on tasks that humans have previously solved, including in programming (Jimenez et al., 2024) and mathematics (Glazer et al., 2024). We therefore propose testing models' ability to design and implement algorithms in an open-ended benchmark: We task LMs with writing code that efficiently solves computationally challenging problems in computer science, physics, and mathematics. Our AlgoTune benchmark consists of 154 coding tasks collected from domain experts and a framework for validating and timing LM-synthesized solution code, which is compared to reference implementations from popular open-source packages. In addition, we develop a baseline LM agent, AlgoTuner, and evaluate its performance across a suite of frontier models. AlgoTuner uses a simple, budgeted loop that edits code, compiles and runs it, profiles performance, verifies correctness on tests, and selects the fastest valid version. AlgoTuner achieves an average 1.72x speedup against our reference solvers, which use libraries such as SciPy, sk-learn and CVXPY. However, we find that current models fail to discover algorithmic innovations, instead preferring surface-level optimizations. We hope that AlgoTune catalyzes the development of LM agents exhibiting creative problem solving beyond state-of-the-art human performance.

Matthias Tangemann, Matthias Kümmerer, Matthias Bethge

Humans excel at detecting and segmenting moving objects according to the Gestalt principle of "common fate". Remarkably, previous works have shown that human perception generalizes this principle in a zero-shot fashion to unseen textures or random dots. In this work, we seek to better understand the computational basis for this capability by evaluating a broad range of optical flow models and a neuroscience inspired motion energy model for zero-shot figure-ground segmentation of random dot stimuli. Specifically, we use the extensively validated motion energy model proposed by Simoncelli and Heeger in 1998 which is fitted to neural recordings in cortex area MT. We find that a cross section of 40 deep optical flow models trained on different datasets struggle to estimate motion patterns in random dot videos, resulting in poor figure-ground segmentation performance. Conversely, the neuroscience-inspired model significantly outperforms all optical flow models on this task. For a direct comparison to human perception, we conduct a psychophysical study using a shape identification task as a proxy to measure human segmentation performance. All state-of-the-art optical flow models fall short of human performance, but only the motion energy model matches human capability. This neuroscience-inspired model successfully addresses the lack of human-like zero-shot generalization to random dot stimuli in current computer vision models, and thus establishes a compelling link between the Gestalt psychology of human object perception and cortical motion processing in the brain. Code, models and datasets are available at https://github.com/mtangemann/motion_energy_segmentation

Jana Zeller, Thaddäus Wiedemer, Fanfei Li et al.

Frontier models are transitioning from multimodal large language models (MLLMs) that merely ingest visual information to unified multimodal models (UMMs) capable of native interleaved generation. This shift has sparked interest in using intermediate visualizations as a reasoning aid, akin to human mental imagery. Central to this idea is the ability to form, maintain, and manipulate visual representations in a goal-oriented manner. To evaluate and probe this capability, we develop MentisOculi, a procedural, stratified suite of multi-step reasoning problems amenable to visual solution, tuned to challenge frontier models. Evaluating visual strategies ranging from latent tokens to explicit generated imagery, we find they generally fail to improve performance. Analysis of UMMs specifically exposes a critical limitation: While they possess the textual reasoning capacity to solve a task and can sometimes generate correct visuals, they suffer from compounding generation errors and fail to leverage even ground-truth visualizations. Our findings suggest that despite their inherent appeal, visual thoughts do not yet benefit model reasoning. MentisOculi establishes the necessary foundation to analyze and close this gap across diverse model families.

Vishaal Udandarao, Shyamgopal Karthik, Surabhi S. Nath et al.

Cambrian-S aims to take the first steps towards improving video world models with spatial supersensing by introducing (i) two benchmarks, VSI-Super-Recall (VSR) and VSI-Super-Counting (VSC), and (ii) bespoke predictive sensing inference strategies tailored to each benchmark. In this work, we conduct a critical analysis of Cambrian-S across both these fronts. First, we introduce a simple baseline, NoSense, which discards almost all temporal structure and uses only a bag-of-words SigLIP model, yet near-perfectly solves VSR, achieving 95% accuracy even on 4-hour videos. This shows benchmarks like VSR can be nearly solved without spatial cognition, world modeling or spatial supersensing. Second, we hypothesize that the tailored inference methods proposed by Cambrian-S likely exploit shortcut heuristics in the benchmark. We illustrate this with a simple sanity check on the VSC benchmark, called VSC-Repeat: We concatenate each video with itself 1-5 times, which does not change the number of unique objects. However, this simple perturbation entirely collapses the mean relative accuracy of Cambrian-S from 42% to 0%. A system that performs spatial supersensing and integrates information across experiences should recognize views of the same scene and keep object-count predictions unchanged; instead, Cambrian-S inference algorithm relies largely on a shortcut in the VSC benchmark that rooms are never revisited. Taken together, our findings suggest that (i) current VSI-Super benchmarks do not yet reliably measure spatial supersensing, and (ii) predictive-sensing inference recipes used by Cambrian-S improve performance by inadvertently exploiting shortcuts rather than from robust spatial supersensing. We include the response from the Cambrian-S authors (in Appendix A) to provide a balanced perspective alongside our claims. We release our code at: https://github.com/bethgelab/supersanity

Adhiraj Ghosh, Vishaal Udandarao, Thao Nguyen et al.

What data should a vision-language model be trained on? To answer this question, many data curation efforts center on the quality of a dataset. However, most of these existing methods are (i) offline, i.e. they produce a static dataset from a set of predetermined filtering criteria, and (ii) concept-agnostic, i.e. they use model-based filters which induce additional data biases. In this work, we go beyond such offline, concept-agnostic methods and advocate for more flexible, task-adaptive online concept-based curation. Our first contribution is DataConcept, a collection of 128M web-crawled image-text pairs annotated with fine-grained details about their concept composition. Building on DataConcept, we introduce Concept-Aware Batch Sampling (CABS), a simple yet effective batch sampling framework that flexibly constructs batches on-the-fly based on specific target distributions. We propose two variants: (i) Diversity Maximization (CABS-DM) to curate batches with a broad coverage of available concepts, and (ii) Frequency Maximization (CABS-FM) to curate batches with high object multiplicity. Through extensive evaluations across 28 benchmarks, we demonstrate that our CABS method significantly benefits CLIP/SigLIP model classes and yields highly performant models. Overall, CABS represents a strong open-source alternative to proprietary online data curation algorithms, enabling practitioners to define custom concept distributions that optimize for specific downstream tasks.

Christoph Schuhmann, Gollam Rabby, Ameya Prabhu et al.

Paywalls, licenses and copyright rules often restrict the broad dissemination and reuse of scientific knowledge. We take the position that it is both legally and technically feasible to extract the scientific knowledge in scholarly texts. Current methods, like text embeddings, fail to reliably preserve factual content, and simple paraphrasing may not be legally sound. We propose a new idea for the community to adopt: convert scholarly documents into knowledge preserving, but style agnostic representations we term Knowledge Units using LLMs. These units use structured data capturing entities, attributes and relationships without stylistic content. We provide evidence that Knowledge Units (1) form a legally defensible framework for sharing knowledge from copyrighted research texts, based on legal analyses of German copyright law and U.S. Fair Use doctrine, and (2) preserve most (~95\%) factual knowledge from original text, measured by MCQ performance on facts from the original copyrighted text across four research domains. Freeing scientific knowledge from copyright promises transformative benefits for scientific research and education by allowing language models to reuse important facts from copyrighted text. To support this, we share open-source tools for converting research documents into Knowledge Units. Overall, our work posits the feasibility of democratizing access to scientific knowledge while respecting copyright.

Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus et al.

A few years ago, the first CNN surpassed human performance on ImageNet. However, it soon became clear that machines lack robustness on more challenging test cases, a major obstacle towards deploying machines "in the wild" and towards obtaining better computational models of human visual perception. Here we ask: Are we making progress in closing the gap between human and machine vision? To answer this question, we tested human observers on a broad range of out-of-distribution (OOD) datasets, recording 85,120 psychophysical trials across 90 participants. We then investigated a range of promising machine learning developments that crucially deviate from standard supervised CNNs along three axes: objective function (self-supervised, adversarially trained, CLIP language-image training), architecture (e.g. vision transformers), and dataset size (ranging from 1M to 1B). Our findings are threefold. (1.) The longstanding distortion robustness gap between humans and CNNs is closing, with the best models now exceeding human feedforward performance on most of the investigated OOD datasets. (2.) There is still a substantial image-level consistency gap, meaning that humans make different errors than models. In contrast, most models systematically agree in their categorisation errors, even substantially different ones like contrastive self-supervised vs. standard supervised models. (3.) In many cases, human-to-model consistency improves when training dataset size is increased by one to three orders of magnitude. Our results give reason for cautious optimism: While there is still much room for improvement, the behavioural difference between human and machine vision is narrowing. In order to measure future progress, 17 OOD datasets with image-level human behavioural data and evaluation code are provided as a toolbox and benchmark at: https://github.com/bethgelab/model-vs-human/

Jonas Rauber, Matthias Bethge, Wieland Brendel

EagerPy is a Python framework that lets you write code that automatically works natively with PyTorch, TensorFlow, JAX, and NumPy. Library developers no longer need to choose between supporting just one of these frameworks or reimplementing the library for each framework and dealing with code duplication. Users of such libraries can more easily switch frameworks without being locked in by a specific 3rd party library. Beyond multi-framework support, EagerPy also brings comprehensive type annotations and consistent support for method chaining to any framework. The latest documentation is available online at https://eagerpy.jonasrauber.de and the code can be found on GitHub at https://github.com/jonasrauber/eagerpy.

Claudio Michaelis, Ivan Ustyuzhaninov, Matthias Bethge et al.

We tackle the problem of one-shot instance segmentation: Given an example image of a novel, previously unknown object category, find and segment all objects of this category within a complex scene. To address this challenging new task, we propose Siamese Mask R-CNN. It extends Mask R-CNN by a Siamese backbone encoding both reference image and scene, allowing it to target detection and segmentation towards the reference category. We demonstrate empirical results on MS Coco highlighting challenges of the one-shot setting: while transferring knowledge about instance segmentation to novel object categories works very well, targeting the detection network towards the reference category appears to be more difficult. Our work provides a first strong baseline for one-shot instance segmentation and will hopefully inspire further research into more powerful and flexible scene analysis algorithms. Code is available at: https://github.com/bethgelab/siamese-mask-rcnn

Wieland Brendel, Jonas Rauber, Matthias Bethge

Many machine learning algorithms are vulnerable to almost imperceptible perturbations of their inputs. So far it was unclear how much risk adversarial perturbations carry for the safety of real-world machine learning applications because most methods used to generate such perturbations rely either on detailed model information (gradient-based attacks) or on confidence scores such as class probabilities (score-based attacks), neither of which are available in most real-world scenarios. In many such cases one currently needs to retreat to transfer-based attacks which rely on cumbersome substitute models, need access to the training data and can be defended against. Here we emphasise the importance of attacks which solely rely on the final model decision. Such decision-based attacks are (1) applicable to real-world black-box models such as autonomous cars, (2) need less knowledge and are easier to apply than transfer-based attacks and (3) are more robust to simple defences than gradient- or score-based attacks. Previous attacks in this category were limited to simple models or simple datasets. Here we introduce the Boundary Attack, a decision-based attack that starts from a large adversarial perturbation and then seeks to reduce the perturbation while staying adversarial. The attack is conceptually simple, requires close to no hyperparameter tuning, does not rely on substitute models and is competitive with the best gradient-based attacks in standard computer vision tasks like ImageNet. We apply the attack on two black-box algorithms from Clarifai.com. The Boundary Attack in particular and the class of decision-based attacks in general open new avenues to study the robustness of machine learning models and raise new questions regarding the safety of deployed machine learning systems. An implementation of the attack is available as part of Foolbox at https://github.com/bethgelab/foolbox .

Jonas Rauber, Wieland Brendel, Matthias Bethge

Even todays most advanced machine learning models are easily fooled by almost imperceptible perturbations of their inputs. Foolbox is a new Python package to generate such adversarial perturbations and to quantify and compare the robustness of machine learning models. It is build around the idea that the most comparable robustness measure is the minimum perturbation needed to craft an adversarial example. To this end, Foolbox provides reference implementations of most published adversarial attack methods alongside some new ones, all of which perform internal hyperparameter tuning to find the minimum adversarial perturbation. Additionally, Foolbox interfaces with most popular deep learning frameworks such as PyTorch, Keras, TensorFlow, Theano and MXNet and allows different adversarial criteria such as targeted misclassification and top-k misclassification as well as different distance measures. The code is licensed under the MIT license and is openly available at https://github.com/bethgelab/foolbox . The most up-to-date documentation can be found at http://foolbox.readthedocs.io .

Vishaal Udandarao, Ameya Prabhu, Adhiraj Ghosh et al. · cambridge

Web-crawled pretraining datasets underlie the impressive "zero-shot" evaluation performance of multimodal models, such as CLIP for classification/retrieval and Stable-Diffusion for image generation. However, it is unclear how meaningful the notion of "zero-shot" generalization is for such multimodal models, as it is not known to what extent their pretraining datasets encompass the downstream concepts targeted for during "zero-shot" evaluation. In this work, we ask: How is the performance of multimodal models on downstream concepts influenced by the frequency of these concepts in their pretraining datasets? We comprehensively investigate this question across 34 models and five standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting "zero-shot" generalization, multimodal models require exponentially more data to achieve linear improvements in downstream "zero-shot" performance, following a sample inefficient log-linear scaling trend. This trend persists even when controlling for sample-level similarity between pretraining and downstream datasets, and testing on purely synthetic data distributions. Furthermore, upon benchmarking models on long-tailed data sampled based on our analysis, we demonstrate that multimodal models across the board perform poorly. We contribute this long-tail test set as the "Let it Wag!" benchmark to further research in this direction. Taken together, our study reveals an exponential need for training data which implies that the key to "zero-shot" generalization capabilities under large-scale training paradigms remains to be found.

Andreas Hochlehnert, Hardik Bhatnagar, Vishaal Udandarao et al. · cambridge

Reasoning has emerged as the next major frontier for language models (LMs), with rapid advances from both academic and industrial labs. However, this progress often outpaces methodological rigor, with many evaluations relying on benchmarking practices that lack transparency, robustness, or statistical grounding. In this work, we conduct a comprehensive empirical study and find that current mathematical reasoning benchmarks are highly sensitive to subtle implementation choices--including decoding parameters, random seeds, prompt formatting, and even hardware and software configurations. Performance gains reported in recent studies frequently hinge on unclear comparisons or unreported sources of variance. To address these issues, we propose a standardized evaluation framework with clearly defined best practices and reporting standards. Using this framework, we reassess recent methods and find that most reinforcement learning (RL) approaches yield only modest improvements--far below prior claims--and are prone to overfitting, especially on small-scale benchmarks like AIME'24. In contrast, supervised finetuning (SFT) methods show consistently stronger generalization in the settings we study. To foster reproducibility, we release all code, prompts, and model outputs, for reasoning benchmarks, establishing more rigorous foundations for future work.

Marcel Binz, Elif Akata, Matthias Bethge et al. · princeton

Establishing a unified theory of cognition has been a major goal of psychology. While there have been previous attempts to instantiate such theories by building computational models, we currently do not have one model that captures the human mind in its entirety. A first step in this direction is to create a model that can predict human behavior in a wide range of settings. Here we introduce Centaur, a computational model that can predict and simulate human behavior in any experiment expressible in natural language. We derived Centaur by finetuning a state-of-the-art language model on a novel, large-scale data set called Psych-101. Psych-101 reaches an unprecedented scale, covering trial-by-trial data from over 60,000 participants performing over 10,000,000 choices in 160 experiments. Centaur not only captures the behavior of held-out participants better than existing cognitive models, but also generalizes to new cover stories, structural task modifications, and entirely new domains. Furthermore, we find that the model's internal representations become more aligned with human neural activity after finetuning. Taken together, our results demonstrate that it is possible to discover computational models that capture human behavior across a wide range of domains. We believe that such models provide tremendous potential for guiding the development of cognitive theories and present a case study to demonstrate this.

Çağatay Yıldız, Nishaanth Kanna Ravichandran, Nitin Sharma et al.

Continual learning (CL) in large language models (LLMs) is an evolving domain that focuses on developing efficient and sustainable training strategies to adapt models to emerging knowledge and achieve robustness in dynamic environments. Our primary emphasis is on continual domain-adaptive pretraining, a process designed to equip LLMs with the ability to integrate new information from various domains while retaining previously learned knowledge. Since existing works concentrate mostly on continual fine-tuning for a limited selection of downstream tasks or training domains, we introduce a new benchmark designed to measure the adaptability of LLMs to changing pretraining data landscapes. We further examine the impact of model size on learning efficacy and forgetting, as well as how the progression and similarity of emerging domains affect the knowledge transfer within these models. Our findings uncover several key insights: (i) continual pretraining consistently improves <1.5B models studied in this work and is also superior to domain adaptation, (ii) larger models always achieve better perplexity than smaller ones when continually pretrained on the same corpus, (iii) smaller models are particularly sensitive to continual pretraining, showing the most significant rates of both learning and forgetting, (iv) continual pretraining boosts downstream task performance of GPT-2 family, (v) continual pretraining enables LLMs to specialize better when the sequence of domains shows semantic similarity while randomizing training domains leads to better transfer and final performance otherwise. We posit that our research establishes a new benchmark for CL in LLMs, providing a more realistic evaluation of knowledge retention and transfer across diverse domains.

Shashwat Goel, Joschka Struber, Ilze Amanda Auzina et al.

As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as ''AI Oversight''. We study how model similarity affects both aspects of AI oversight by proposing Chance Adjusted Probabilistic Agreement (CAPA): a metric for LM similarity based on overlap in model mistakes. Using CAPA, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from ''weak-to-strong generalization''. As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend -- model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

Sebastian Dziadzio, Vishaal Udandarao, Karsten Roth et al. · cambridge

Model merging combines multiple expert models - finetuned from a base foundation model on diverse tasks and domains - into a single, more capable model. However, most existing model merging approaches assume that all experts are available simultaneously. In reality, new tasks and domains emerge progressively over time, requiring strategies to integrate the knowledge of expert models as they become available: a process we call temporal model merging. The temporal dimension introduces unique challenges not addressed in prior work, raising new questions such as: when training for a new task, should the expert model start from the merged past experts or from the original base model? Should we merge all models at each time step? Which merging techniques are best suited for temporal merging? Should different strategies be used to initialize the training and deploy the model? To answer these questions, we propose a unified framework called TIME - Temporal Integration of Model Expertise - which defines temporal model merging across three axes: (1) Initialization Phase, (2) Deployment Phase, and (3) Merging Technique. Using TIME, we study temporal model merging across model sizes, compute budgets, and learning horizons on the FoMo-in-Flux benchmark. Our comprehensive suite of experiments across TIME allows us to uncover key insights for temporal model merging, offering a better understanding of current challenges and best practices for effective temporal model merging.

Prasanna Mayilvahanan, Thaddäus Wiedemer, Sayak Mallick et al.

Scaling laws guide the development of large language models (LLMs) by offering estimates for the optimal balance of model size, tokens, and compute. More recently, loss-to-loss scaling laws that relate losses across pretraining datasets and downstream tasks have emerged as a powerful tool for understanding and improving LLM performance. In this work, we investigate which factors most strongly influence loss-to-loss scaling. Our experiments reveal that the pretraining data and tokenizer determine the scaling trend. In contrast, model size, optimization hyperparameters, and even significant architectural differences, such as between transformer-based models like Llama and state-space models like Mamba, have limited impact. Consequently, practitioners should carefully curate suitable pretraining datasets for optimal downstream performance, while architectures and other settings can be freely optimized for training efficiency.

Lukas Thede, Karsten Roth, Matthias Bethge et al.

Keeping large language models factually up-to-date is crucial for deployment, yet costly retraining remains a challenge. Knowledge editing offers a promising alternative, but methods are only tested on small-scale or synthetic edit benchmarks. In this work, we aim to bridge research into lifelong knowledge editing to real-world edits at a practically relevant scale. We first introduce WikiBigEdit; a large-scale benchmark of real-world Wikidata edits, built to automatically extend lifelong for future-proof benchmarking. In its first instance, it includes over 500K question-answer pairs for knowledge editing alongside a comprehensive evaluation pipeline. Finally, we use WikiBigEdit to study existing knowledge editing techniques' ability to incorporate large volumes of real-world facts and contrast their capabilities to generic modification techniques such as retrieval augmentation and continual finetuning to acquire a complete picture of the practical extent of current lifelong knowledge editing.

Adhiraj Ghosh, Sebastian Dziadzio, Ameya Prabhu et al. · cambridge

Traditional fixed test sets fall short in evaluating open-ended capabilities of foundation models. To address this, we propose ONEBench(OpeN-Ended Benchmarking), a new testing paradigm that consolidates individual evaluation datasets into a unified, ever-expanding sample pool. ONEBench allows users to generate custom, open-ended evaluation benchmarks from this pool, corresponding to specific capabilities of interest. By aggregating samples across test sets, ONEBench enables the assessment of diverse capabilities beyond those covered by the original test sets, while mitigating overfitting and dataset bias. Most importantly, it frames model evaluation as a collective process of selecting and aggregating sample-level tests. The shift from task-specific benchmarks to ONEBench introduces two challenges: (1)heterogeneity and (2)incompleteness. Heterogeneity refers to the aggregation over diverse metrics, while incompleteness describes comparing models evaluated on different data subsets. To address these challenges, we explore algorithms to aggregate sparse measurements into reliable model scores. Our aggregation algorithm ensures identifiability(asymptotically recovering ground-truth scores) and rapid convergence, enabling accurate model ranking with less data. On homogenous datasets, we show our aggregation algorithm provides rankings that highly correlate with those produced by average scores. We also demonstrate robustness to ~95% of measurements missing, reducing evaluation cost by up to 20x with little-to-no change in model rankings. We introduce ONEBench-LLM for language models and ONEBench-LMM for vision-language models, unifying evaluations across these domains. Overall, we present a technique for open-ended evaluation, which can aggregate over incomplete, heterogeneous sample-level measurements to continually grow a benchmark alongside the rapidly developing foundation models.

Ameya Prabhu, Vishaal Udandarao, Philip Torr et al. · cambridge

Standardized benchmarks drive progress in machine learning. However, with repeated testing, the risk of overfitting grows as algorithms over-exploit benchmark idiosyncrasies. In our work, we seek to mitigate this challenge by compiling ever-expanding large-scale benchmarks called Lifelong Benchmarks. These benchmarks introduce a major challenge: the high cost of evaluating a growing number of models across very large sample sets. To address this challenge, we introduce an efficient framework for model evaluation, Sort & Search (S&S)}, which reuses previously evaluated models by leveraging dynamic programming algorithms to selectively rank and sub-select test samples. To test our approach at scale, we create Lifelong-CIFAR10 and Lifelong-ImageNet, containing 1.69M and 1.98M test samples for classification. Extensive empirical evaluations across over 31,000 models demonstrate that S&S achieves highly-efficient approximate accuracy measurement, reducing compute cost from 180 GPU days to 5 GPU hours (about 1000x reduction) on a single A100 GPU, with low approximation error and memory cost of <100MB. Our work also highlights issues with current accuracy prediction metrics, suggesting a need to move towards sample-level evaluation metrics. We hope to guide future research by showing our method's bottleneck lies primarily in generalizing Sort beyond a single rank order and not in improving Search.

Thaddäus Wiedemer, Yash Sharma, Ameya Prabhu et al.

We investigate the success conditions for compositional generalization of CLIP models on real-world data through performance prediction. Prior work shows that CLIP requires exponentially more pretraining data for linear performance gains on individual concepts. This sample-inefficient scaling could be mitigated if CLIP systematically understood new inputs as compositions of learned components, allowing rare observation to be mapped to common concepts. To explore CLIP's compositional generalization ability, we filter retrieval corpora for samples with object combinations not present in the pretraining corpus. We show that CLIP's performance on these samples can be accurately predicted from the pretraining frequencies of individual objects. Our findings demonstrate that CLIP learns to disentangle objects observed in its pretraining data and can recompose them straightforwardly. Additionally, we are the first to show how this ability scales with pretraining data. For data curation in practice, our results suggest that balancing object occurrences improves generalization, which should benefit CLIP's efficiency and accuracy without scaling data volume.

Shiven Sinha, Ameya Prabhu, Ponnurangam Kumaraguru et al.

Proving geometric theorems constitutes a hallmark of visual reasoning combining both intuitive and logical skills. Therefore, automated theorem proving of Olympiad-level geometry problems is considered a notable milestone in human-level automated reasoning. The introduction of AlphaGeometry, a neuro-symbolic model trained with 100 million synthetic samples, marked a major breakthrough. It solved 25 of 30 International Mathematical Olympiad (IMO) problems whereas the reported baseline based on Wu's method solved only ten. In this note, we revisit the IMO-AG-30 Challenge introduced with AlphaGeometry, and find that Wu's method is surprisingly strong. Wu's method alone can solve 15 problems, and some of them are not solved by any of the other methods. This leads to two key findings: (i) Combining Wu's method with the classic synthetic methods of deductive databases and angle, ratio, and distance chasing solves 21 out of 30 methods by just using a CPU-only laptop with a time limit of 5 minutes per problem. Essentially, this classic method solves just 4 problems less than AlphaGeometry and establishes the first fully symbolic baseline strong enough to rival the performance of an IMO silver medalist. (ii) Wu's method even solves 2 of the 5 problems that AlphaGeometry failed to solve. Thus, by combining AlphaGeometry with Wu's method we set a new state-of-the-art for automated theorem proving on IMO-AG-30, solving 27 out of 30 problems, the first AI method which outperforms an IMO gold medalist.

Daniil Zverev, Thaddäus Wiedemer, Ameya Prabhu et al.

The emergence of audio-visual foundation models underscores the importance of reliably assessing their multi-modal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric.

Sebastian Dziadzio, Çağatay Yıldız, Gido M. van de Ven et al.

The ability of machine learning systems to learn continually is hindered by catastrophic forgetting, the tendency of neural networks to overwrite previously acquired knowledge when learning a new task. Existing methods mitigate this problem through regularization, parameter isolation, or rehearsal, but they are typically evaluated on benchmarks comprising only a handful of tasks. In contrast, humans are able to learn over long time horizons in dynamic, open-world environments, effortlessly memorizing unfamiliar objects and reliably recognizing them under various transformations. To make progress towards closing this gap, we introduce Infinite dSprites, a parsimonious tool for creating continual classification and disentanglement benchmarks of arbitrary length and with full control over generative factors. We show that over a sufficiently long time horizon, the performance of all major types of continual learning methods deteriorates on this simple benchmark. This result highlights an important and previously overlooked aspect of continual learning: given a finite modelling capacity and an arbitrarily long learning horizon, efficient learning requires memorizing class-specific information and accumulating knowledge about general mechanisms. In a simple setting with direct supervision on the generative factors, we show how learning class-agnostic transformations offers a way to circumvent catastrophic forgetting and improve classification accuracy over time. Our approach sets the stage for continual learning over hundreds of tasks with explicit control over memorization and forgetting, emphasizing open-set classification and one-shot generalization.

Vishaal Udandarao, Mehdi Cherti, Shyamgopal Karthik et al. · cambridge

We investigate 17 benchmarks (e.g. SugarCREPE, VALSE) commonly used for measuring compositional understanding capabilities of vision-language models (VLMs). We scrutinize design choices in their construction, including data source (e.g. MS-COCO) and curation procedures (e.g. constructing negative images/captions), uncovering several inherent biases across most benchmarks. We find that blind heuristics (e.g. token-length, log-likelihood under a language model) perform on par with CLIP models, indicating that these benchmarks do not effectively measure compositional understanding. We demonstrate that the underlying factor is a distribution asymmetry between positive and negative images/captions, induced by the benchmark construction procedures. To mitigate these issues, we provide a few key recommendations for constructing more robust vision-language compositional understanding benchmarks, that would be less prone to such simple attacks.

Shiven Sinha, Shashwat Goel, Ponnurangam Kumaraguru et al.

There is growing excitement about the potential of Language Models (LMs) to accelerate scientific discovery. Falsifying hypotheses is key to scientific progress, as it allows claims to be iteratively refined over time. This process requires significant researcher effort, reasoning, and ingenuity. Yet current benchmarks for LMs predominantly assess their ability to generate solutions rather than challenge them. We advocate for developing benchmarks that evaluate this inverse capability - creating counterexamples for subtly incorrect solutions. To demonstrate this approach, we start with the domain of algorithmic problem solving, where counterexamples can be evaluated automatically using code execution. Specifically, we introduce REFUTE, a dynamically updating benchmark that includes recent problems and incorrect submissions from programming competitions, where human experts successfully identified counterexamples. Our analysis finds that the best reasoning agents, even OpenAI o3-mini (high) with code execution feedback, can create counterexamples for only <9% of incorrect solutions in REFUTE, even though ratings indicate its ability to solve up to 48% of these problems from scratch. We hope our work spurs progress in evaluating and enhancing LMs' ability to falsify incorrect solutions - a capability that is crucial for both accelerating research and making models self-improve through reliable reflective reasoning.

Tobias Schmidt, Steffen Schneider, Matthias Bethge

We propose Equivariance by Contrast (EbC) to learn equivariant embeddings from observation pairs $(\mathbf{y}, g \cdot \mathbf{y})$, where $g$ is drawn from a finite group acting on the data. Our method jointly learns a latent space and a group representation in which group actions correspond to invertible linear maps -- without relying on group-specific inductive biases. We validate our approach on the infinite dSprites dataset with structured transformations defined by the finite group $G:= (R_m \times \mathbb{Z}_n \times \mathbb{Z}_n)$, combining discrete rotations and periodic translations. The resulting embeddings exhibit high-fidelity equivariance, with group operations faithfully reproduced in latent space. On synthetic data, we further validate the approach on the non-abelian orthogonal group $O(n)$ and the general linear group $GL(n)$. We also provide a theoretical proof for identifiability. While broad evaluation across diverse group types on real-world data remains future work, our results constitute the first successful demonstration of general-purpose encoder-only equivariant learning from group action observations alone, including non-trivial non-abelian groups and a product group motivated by modeling affine equivariances in computer vision.

Jackson Harmon, Andreas Hochlehnert, Matthias Bethge et al.

Scaled post-training now drives many of the largest capability gains in language models (LMs), yet its effect on pretrained knowledge remains poorly understood. Not all forgetting is equal: Forgetting one fact (e.g., a U.S. president or an API call) does not "average out" by recalling another. Hence, we propose a sample-wise paradigm to measure what is forgotten and when backward transfer occurs. Our metric counts 1->0 transitions (correct before post-training, incorrect after) to quantify forgetting and 0->1 transitions to quantify backward transfer. Traditional task averages conflate these effects and obscure large changes. For multiple-choice benchmarks, we add chance-adjusted variants that subtract the expected contribution of random guessing from pre- and post-training accuracies. We apply this framework across post-training stages, model sizes, and data scales. Our large-scale analysis shows that: (1) Domain-continual pretraining induces moderate forgetting with low-to-moderate backward transfer; (2) RL/SFT post-training applied to base models and Instruction tuning yields moderate-to-large backward transfer on math and logic with overall low-to-moderate forgetting; (3) Applying RL/SFT to instruction-tuned models is sensitive on data scale: at small scales, both forgetting and backward transfer are small; at larger scales, effects are mixed and warrant further study with better controls; (4) Model merging does not reliably mitigate forgetting. Overall, our framework offers a practical yardstick for mapping how post-training alters pretrained knowledge at scale -- enabling progress towards generally capable AI systems.

Alexander Panfilov, Evgenii Kortukov, Kristina Nikolić et al.

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are crafted to be subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using them as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

Matthias Kümmerer, Harneet Singh Khanuja, Matthias Bethge

Recent advances in image-based saliency prediction are approaching gold standard performance levels on existing benchmarks. Despite this success, we show that predicting fixations across multiple saliency datasets remains challenging due to dataset bias. We find a significant performance drop (around 40%) when models trained on one dataset are applied to another. Surprisingly, increasing dataset diversity does not resolve this inter-dataset gap, with close to 60% attributed to dataset-specific biases. To address this remaining generalization gap, we propose a novel architecture extending a mostly dataset-agnostic encoder-decoder structure with fewer than 20 dataset-specific parameters that govern interpretable mechanisms such as multi-scale structure, center bias, and fixation spread. Adapting only these parameters to new data accounts for more than 75% of the generalization gap, with a large fraction of the improvement achieved with as few as 50 samples. Our model sets a new state-of-the-art on all three datasets of the MIT/Tuebingen Saliency Benchmark (MIT300, CAT2000, and COCO-Freeview), even when purely generalizing from unrelated datasets, but with a substantial boost when adapting to the respective training datasets. The model also provides valuable insights into spatial saliency properties, revealing complex multi-scale effects that combine both absolute and relative sizes.

Luca M. Schulze Buschoff, Konstantinos Voudouris, Elif Akata et al.

Pre-trained vision language models still fall short of human visual cognition. In an effort to improve visual cognition and align models with human behavior, we introduce visual stimuli and human judgments on visual cognition tasks, allowing us to systematically evaluate performance across cognitive domains under a consistent environment. We fine-tune models on ground truth data for intuitive physics and causal reasoning and find that this improves model performance in the respective fine-tuning domain. Furthermore, it can improve model alignment with human behavior. However, we find that task-specific fine-tuning does not contribute to robust human-like generalization to data with other visual characteristics or to tasks in other cognitive domains.

Lukas Thede, Karsten Roth, Olivier J. Hénaff et al.

With the advent and recent ubiquity of foundation models, continual learning (CL) has recently shifted from continual training from scratch to the continual adaptation of pretrained models, seeing particular success on rehearsal-free CL benchmarks (RFCL). To achieve this, most proposed methods adapt and restructure parameter-efficient finetuning techniques (PEFT) to suit the continual nature of the problem. Based most often on input-conditional query-mechanisms or regularizations on top of prompt- or adapter-based PEFT, these PEFT-style RFCL (P-RFCL) approaches report peak performances; often convincingly outperforming existing CL techniques. However, on the other end, critical studies have recently highlighted competitive results by training on just the first task or via simple non-parametric baselines. Consequently, questions arise about the relationship between methodological choices in P-RFCL and their reported high benchmark scores. In this work, we tackle these questions to better understand the true drivers behind strong P-RFCL performances, their placement w.r.t. recent first-task adaptation studies, and their relation to preceding CL standards such as EWC or SI. In particular, we show: (1) P-RFCL techniques relying on input-conditional query mechanisms work not because, but rather despite them by collapsing towards standard PEFT shortcut solutions. (2) Indeed, we show how most often, P-RFCL techniques can be matched by a simple and lightweight PEFT baseline. (3) Using this baseline, we identify the implicit bound on tunable parameters when deriving RFCL approaches from PEFT methods as a potential denominator behind P-RFCL efficacy. Finally, we (4) better disentangle continual versus first-task adaptation, and (5) motivate standard RFCL techniques s.a. EWC or SI in light of recent P-RFCL methods.