Searcharxiv⌕ Search

arXiv subjects

Matthias Bethge

Publications and source records attributed to Matthias Bethge.

At least 19 recordsLinked to original sources
arXiv2026

Benchmarks Are Surprisingly Compaction-Friendly

Long horizon agentic tasks are now pushing LLM agents past their context window. To continue the task, the agent scaffold has to free up the context window, a process known as context compaction. Compaction is generally lossy by design: agents can silently lose task constraints, instructions, or their reasoning behind earlier decisions. Although compaction is a standard component of coding agent scaffolds, it has not been studied systematically with frontier models on realistic, long horizon agentic tasks. Previous work evaluate smaller open weight models on toy or synthetic trajectories. We fix the agent scaffold (OpenCode), implement five compaction strategies inside it, and evaluate three frontier models on three long horizon benchmarks (PaperBench, ProgramBench, PostTrainBench). For the two 1M context models, the compaction threshold does not measurably affect the task performance, whereas GPT-5.3-Codex, the model with the smallest context window, performs worse at lower thresholds on PaperBench and, less clearly, on PostTrainBench. No compaction method outperforms default self-summarization, and methods that discard more content cost some models up to 20 points while leaving other models the same. Compaction can, however, silently remove task constraints: on ProgramBench, the first compaction summary already drops about one third of the task's anticheating rules on average, and this rises monotonically with the number of compactions. On real user-agent sessions from SWE-chat, instructions about how the agent should work and what it should not do are dropped around 6 times more than specifications of what to build. We argue that current benchmarks are compaction-friendly: i.e., the objective is fixed and the agent's progress stays on disk, so most of what compaction discards can be recovered, except constraints that don't exist anywhere else.

cs.LG↗
arXiv2026

Towards Robust Model Organisms for Eliciting Latent Knowledge

A key problem in AI safety is eliciting latent knowledge (ELK), extracting information from a model that is misaligned and more knowledgeable than its overseer. For example, a misaligned model used to scan code for vulnerabilities may want to withhold its findings, i.e., sandbag, in order to exploit them later. To study ELK empirically, we need model organisms that mirror this setting. Existing model organisms fall short in several respects. In this work, we train model organisms to withhold solutions in a vulnerability detection task. We combine transcript distillation with counter-prompting and tamper-resistance training to instill three properties that we argue are critical for a useful ELK testbed: 1) Selectivity, the model sandbags when it thinks the overseer doesn't know the answer and complies otherwise, 2) Robustness, the model resists elicitation, and 3) Retention, the model retains a knowledge advantage over the overseer, as opposed to losing the relevant capability. We train a sequence of model organisms and show that our strongest model substantially outperforms prior models on robustness while keeping the other two properties intact. We hope that our model organisms provide a useful ELK testbed for the community, and that our methodology sparks the development of increasingly challenging model organisms. All our code, models and the dataset are available at https://huggingface.co/collections/lennart-finke/sandbagging.

cs.AI↗
arXiv2026

Moving Forward with Video Saliency: A New Dataset and Benchmark where Motion Matters

Video saliency prediction is inherently harder to model than static image saliency due to the additional temporal dimension. Video saliency benchmarks rest on the premise that predicting gaze on video requires utilizing temporal activity distributed across frames. Prior work has challenged this, showing that static baselines recover a significant fraction of the explainable gaze information on LEDOV, a popular video saliency dataset, and that video saliency models fail in the same places as this static baseline. We verify that this diagnosis still stands: under a more capable gold standard than the original analysis, and an updated panel of recent architectures, the strongest temporal architecture in the panel still does not substantially improve over a fine-tuned static baseline. However, it remains unclear whether the marginal gain reflects limitations of current temporal architectures or a lack of temporal patterns in the benchmark itself. We introduce SalTempto, a video saliency benchmark with greater dynamism: 224 clips of highly dynamic content, sourced from the HACS-Segments dataset so that each clip contains an event together with its lead-up and aftermath, with gaze recordings from up to 16 subjects and a training split for adapting pretrained models. On SalTempto, the static baseline recovers only about 13\% of the headroom above the centerbias, against more than half on LEDOV. A fine-tuned temporal architecture shows a substantial gain in performance over the static baseline, indicating that it does capture meaningfully more temporal information, which LEDOV fails to measure. Yet, even this SoTA model still leaves nearly half of SalTempto's headroom unexplained, indicating room for improvement in video saliency modelling. Examination of SalTempto also lets us describe human tendencies that models miss. SalTempto link: https://huggingface.co/datasets/bethgelab/video_saliency.

cs.CV↗
arXiv2026

Replay on Demand: An Emergent Curriculum for Balancing Adaptation and Forgetting in Continued Pretraining

Continued pretraining enables language models to adapt to new domains and knowledge, but often at the cost of forgetting previously acquired capabilities. Replay can mitigate this trade-off, but fixed replay mixtures allocate training independently of the model's actual retention needs. We introduce Replay on Demand (RoD), which instead derives the replay allocation from the model's learning dynamics. RoD jointly prioritizes adaptation samples by their remaining learning potential and replay samples by their observed forgetting. Their competition for a shared training budget yields an online curriculum that determines what to train on at each step. Across models, scales, and adaptation domains, RoD reaches or improves upon the adaptation-forgetting frontier of tuned fixed-replay baselines and model merging without prescribing a replay allocation in advance. Replay concentrates on sources that are more vulnerable to forgetting and dynamically increases and redistributes as forgetting emerges during training. Together, our results show that replay can be allocated online from the model's evolving state, targeting what is needed, when it is needed.

cs.LG↗
arXiv2026

MedKIT: Evaluating Knowledge Integration and Generalization in Large Language Models

Constantly evolving real-world knowledge necessitates models to be updated continuously. Especially in medicine, as clinical evidence changes over time, outdated knowledge can pose safety risks. Existing evaluations of knowledge integration focus on factual recall, offering limited insight into whether newly integrated knowledge is actually usable. Our benchmark MedKIT (Medical Knowledge Integration and Transfer) provides a granular evaluation of how models integrate and apply knowledge under realistic sequences of clinical updates. Each instance corresponds to a factual update derived from clinical evidence, paired with targeted probes that assess transfer across lexical variation, relational transformations, compositional reasoning, and open-ended operationalization, as well as locality tests for knowledge preservation. Using MedKIT, we conduct a large-scale empirical study of 12 knowledge integration strategies across 5 diverse models, including both general-purpose and medical LLMs. Our results reveal a consistent gap between recall and usable knowledge: while most methods achieve strong gains on the original update task and under lexical variation, relational generalization is limited, and no method yields meaningful improvements on compositional or operational tasks. These findings highlight a fundamental challenge in knowledge integration and position MedKIT as a testbed for developing methods that make newly integrated knowledge more consistently usable across tasks and contexts.

cs.CL↗
arXiv2026

LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training

We present LAION-BVD, a large-scale open video dataset for multimodal learning, which contains 1.3B platform-specific video URLs collected from CommonCrawl. From these, we download 80M videos with a total duration of 10 million hours. The dataset is designed for multimodal pre-training across the video, audio, and image modalities. Using content-aware scene detection, we extract clips for which we synthetically generate video and audio captions. Models trained on these data achieve competitive performance on standard video-text and audio-text benchmarks, with consistent improvements as training or model scale increases. Additionally, we explore video frames as an alternative source of image-text data by extracting scene-changing frames. These frames exhibit a visual distribution distinct from standard web image corpora, and models trained on this dataset achieve strong image-text retrieval performance. We release LAION-BVD to the research community. It significantly expands open access to multimodal videos at an unprecedented scale.

cs.CV↗
arXiv2026

DataComp-VLM: Improved Open Datasets for Vision-Language Models

Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we collect 160 datasets spanning four data types -- image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data -- into a corpus of 6T multimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets. Models are then evaluated on a carefully selected suite of up to 52 downstream benchmarks across 9 domains. We conduct extensive experiments on DCVLM and find that data mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-Baseline, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FineVision, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available at https://www.datacomp.ai/dcvlm/.

cs.CV↗
arXiv2026

LLM generation novelty through the lens of semantic similarity

Generation novelty is a key indicator of an LLM's ability to generalize, yet measuring it against full pretraining corpora is computationally challenging. Existing evaluations often rely on lexical overlap, failing to detect paraphrased text, or do not consider the full pretraining corpus. We frame novelty as a semantic retrieval problem. This framing enables us to address novelty with modern embedding and indexing pipelines, allowing for efficient analysis at pre-training scale. Specifically, we propose a three-stage framework that retrieves semantically similar samples, reranks them at varying subsequence lengths, and calibrates scores using a human novelty reference for interpretability. We apply this framework to the SmolLM model family and report three key findings: (1) models draw on pre-training data across much longer sequences than previously reported; (2) some task domains systematically promote or suppress generation novelty; and (3) instruction tuning not only alters style but also increases novelty. These results highlight the value of semantic novelty analysis for studying generalization. To support reproducibility and further research, we release ~20 TB of corpus chunks and index artifacts at https://huggingface.co/datasets/stai-tuebingen/faiss-smollm

cs.LG↗
arXiv2026

DeepGaze3.5-VL: Modeling Scanpaths via Autoregressive Token Prediction

Understanding human visual attention on a scene over time has applications in domains such as interface design and inferring cognitive states. Modeling visual scanpaths has historically relied on specialized architectures with hand-crafted priors. While these architectures can model fixation sequences, their rigid structural biases restrict easy extendability and flexible conditioning. For instance, integrating task-specific instructions or adapting to distinct viewer identities requires custom, disjoint architectural additions. We frame scanpath prediction purely as a discrete sequence modeling task. By mapping coordinates into a text vocabulary, we leverage the pretrained representations of Vision-Language Models. This framing absorbs diverse factors of variation: simple prompting allows for global conditioning, such as providing viewer identities to capture personalized biases, or task-specific objectives like visual search. The framework can also integrate per-fixation attributes, such as individual fixation durations, alongside spatial locations. The autoregressive alignment enables the scalable, exact computation of per-fixation log-likelihoods, directly equivalent to the commonly used Information Gain (IG) metric. Our model, DeepGaze3.5-VL, establishes a new state-of-the-art across multiple datasets, achieving 2.18 bits of IG on MIT1003, a 46% improvement over DeepGaze III. This advantage persists even when baselines use identical high-capacity vision encoders. Beyond predictive performance, our generative framework serves as a powerful computational tool for direct behavioral interventions, allowing for controlled in-silico simulations that would be experimentally difficult or impossible to conduct in vivo. We demonstrate this ability by performing controlled interventions on the durations of pre-saccadic fixations, recovering known oculomotor phenomena purely from data.

cs.CV↗
arXiv2026

VGGSounder: Audio-Visual Evaluations for Foundation Models

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.

cs.MM↗
arXiv2026

QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents

LLM agents increasingly act over long horizons, where a single trajectory can contain hundreds or thousands of actions. In these settings, outcome-only rewards provide too sparse guidance, failing to inform the model about the goodness of intermediate actions. Dense supervision methods aim to solve this problem by scoring intermediate steps, from intrinsic confidence to self-distillation and embedding similarities. However, it is common practice to evaluate them by measuring the downstream performance of a training pipeline that integrates them. This is expensive, conflates supervision quality with training engineering confounders, and renders different methodological families requiring distinct training setups incomparable. As a result, dense supervision methods are rarely benchmarked on common ground. We introduce QVal, a training-free testbed for directly evaluating dense supervision signals. Given a state-action pair, QVal measures how well a method's score is Q-aligned: whether it orders actions according to the Q-values of a strong reference-policy. This lets us compare signals before any training run and separate signal quality from other engineering choices. We instantiate QVal as QVal-v1.0, benchmarking 21 dense supervision methods across four diverse environments and seven methodological families, with over 1.2K evaluation experiments across six open-weight model backbones. We find that simple prompting baselines consistently outperform recent dense supervision methods from the literature, and that performance clusters strongly by family. These findings hold across model sizes, environments, and observation modalities. QVal is designed to be easily extensible to new environments and methods, enabling researchers to iterate on dense supervision methods before any training run.

cs.LG↗
arXiv2026

RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments

For most of scientific history, researchers studying behavior could only infer hidden mechanisms from outward actions: an inverse problem that becomes more tractable when observation is augmented by targeted intervention. We pose a computational analogue: given only behavioral traces of an agent in a game environment, can a learner reconstruct the underlying decision program as executable code, and how much does this reconstruction improve with the ability to design controlled experiments? We introduce RevengeBench, a benchmark of 75 LLM generated, Elo-calibrated policies across five game environments, drawn from CodeClash tournament trajectories. The learner observes the hidden target policy play against sampled opponents and designs behavioral probes in the form of custom opponent policies that elicit informative behavior. It then submits an executable hypothesis, which is evaluated using continuous action-distance metrics. We further validate that recovered code carries informative signal in downstream player-versus-player tournaments. Across twelve frontier LLMs, recovery quality varies substantially (34 to 72% of initial distance closed), with reconstructed policies yielding measurable competitive advantage, particularly for weaker models that otherwise struggle to design effective counter-strategies. Our benchmark positions behavioral recovery of programmatic policies as a tractable inverse problem in code-space, opening a path to opponent modeling, policy interpretability, and the broader question of inferring latent mechanisms from observations.

cs.LG↗
arXiv2026

MentisOculi: Revealing the Limits of Reasoning with Mental Imagery

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.

cs.AI↗
arXiv2026

LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws

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 and generalization. In this work, we investigate which factors most strongly influence loss-to-loss scaling. Our experiments reveal that the pretraining data determines the scaling trend. In contrast, model size, optimization hyperparameters, tokenizer and even significant architectural differences, such as between transformer-based models like Llama and state-space models like Mamba, generally 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.

cs.LG↗
arXiv2026

Only Brains Align with Brains: Cross-Region Alignment Patterns Expose Limits of Normative Models

Neuroscientists and computer vision researchers use model-brain alignment benchmarks to compare artificial and biological vision systems. These benchmarks rank models according to alignment measures such as the similarity of representational geometry or the predictability of neural responses from model activations. However, recent works have identified a number of problems with these rankings, among them their lack of discriminative power and robustness, raising the conceptual question of what it means for a model to be brain-aligned. Here we introduce alignment patterns -- characteristic functional relationship profiles of each brain region to all others -- and propose that models should reproduce these patterns to qualify as brain-aligned. First, we apply a standard benchmarking pipeline to a broad spectrum of vision models of the BOLD Moments video fMRI dataset across visual regions of interest (ROIs). We find diverse models appear equivalent in their brain alignment, reflecting the lack of discriminative power of conventional alignment benchmarking pipelines. In contrast, alignment pattern analysis (APA) is a second-order structural consistency test: a model aligned to a given ROI should reproduce that ROI's characteristic cross-region alignment profile. Applying APA, we find that, while these patterns are highly stable across brains of different subjects, even top-ranked models often fail to capture them. Finally, we argue for a clearer distinction between the criteria a model must meet to serve as a tool versus as a computational model for human visual cortex. Conventional alignment measures may be sufficient for identifying neurally predictive models, but claims about computational or algorithmic similarity may require a stronger basis of evidence, including the reproducibility of relational alignment patterns.

q-bio.NC↗
arXiv2026

Round-Trip Translation Reveals What Frontier Multilingual Benchmarks Miss

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.

cs.CL↗
arXiv2026

Personalizing Text-to-Image Generation to Individual Taste

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.

cs.CV↗
arXiv2026

PostTrainBench: Can LLM Agents Automate LLM Post-Training?

AI agents have become surprisingly proficient at software engineering over the past year, largely due to improvements in reasoning capabilities. This raises a deeper question: can these systems extend their capabilities to automate AI research itself? In this paper, we explore post-training, the critical phase that turns base LLMs into useful assistants. We introduce PostTrainBench to benchmark how well LLM agents can perform post-training autonomously under bounded compute constraints (10 hours on one H100 GPU). We ask frontier agents (e.g., Claude Code with Opus 4.6) to optimize the performance of a base LLM on a particular benchmark (e.g., Qwen3-4B on AIME). Importantly, we do not provide any predefined strategies to the agents and instead give them full autonomy to find necessary information on the web, run experiments, and curate data. We find that frontier agents make substantial progress but generally lag behind instruction-tuned LLMs from leading providers: 23.2% for the best agent vs. 51.1% for official instruction-tuned models. However, agents can exceed instruction-tuned models in targeted scenarios: GPT-5.1 Codex Max achieves 89% on BFCL with Gemma-3-4B vs. 67% for the official model. We also observe several failure modes worth flagging. Agents sometimes engage in reward hacking: training on the test set, downloading existing instruction-tuned checkpoints instead of training their own, and using API keys they find to generate synthetic data without authorization. These behaviors are concerning and highlight the importance of careful sandboxing as these systems become more capable. Overall, we hope PostTrainBench will be useful for tracking progress in AI R&D automation and for studying the risks that come with it. Website and code are available at https://posttrainbench.com/.

cs.SE↗