We introduce a large-scale dataset for instruction-guided vector image editing, consisting of over 270,000 pairs of SVG images paired with natural language edit instructions. Our dataset enables training and evaluation of models that modify vector graphics based on textual commands. We describe the data collection process, including image pairing via CLIP similarity and instruction generation with vision-language models. Initial experiments with state-of-the-art large language models reveal that current methods struggle to produce accurate and valid edits, underscoring the challenge of this task. To foster research in natural language-driven vector graphic generation and editing, we make our resources created within this work publicly available.

While Transformer models can learn algorithmic tasks and generalize reliably to unseen, in-distribution data, they often fail catastrophically when required to extrapolate to sequence lengths beyond their training regime. Using AttentionSpan, a suite of synthetic diagnostic tasks designed to enable inspection of internal model computations, we analyze the model’s behavior on length extrapolation. Our findings indicate that this failure is not fully explained by an inability to induce general rules for the task. Instead, we identify inconsistent attention patterns as a major contributing factor: information retrieval strategies learned by individual attention heads, which fail to remain stable as sequence length increases. Reinforcing a reference attention pattern at inference time, using no out-of-distribution data, largely recovers accuracy on String Reversal and partially on Value Assignment, while leaving multi-digit arithmetic unresolved. We also study length adaptation, how a model trained on one length adjusts to another by fine-tuning its Key and Query parameters, and find that a single dimension in every attention head is enough, across 4 algorithmic tasks and the RULER benchmark. Adaptation does not carry past the fine-tuned lengths, and our interventions are diagnostic rather than a proposed remedy. We make our benchmark and code publicly available.

Prior work has shown that large language models (LLMs) often converge to accurate input embedding for numbers, based on sinusoidal representations.In this work, we demonstrate that these representations are in fact strikingly systematic, to the point of being almost perfectly universal: different LLM families develop equivalent sinusoidal structures, and number representations are broadly interchangeable in a large swathe of experimental setups.We show that properly factoring in this characteristic is crucial when it comes to assessing how accurately LLMs encode numeric and other ordinal information, and that mechanistically enhancing this sinusoidality can also lead to reductions of LLMs’ arithmetic errors.

A large body of recent work assesses models’ generalization capabilities through the lens of performance on out-of-distribution (OOD) datasets. Despite their practicality, such evaluations build upon a strong assumption: that OOD evaluations can capture and reflect upon possible failures in a real-world deployment. In this work, we challenge this assumption and confront the results obtained from OOD evaluations with a set of specific failure modes documented in existing question-answering (QA) models, referred to as a reliance on spurious features or prediction shortcuts.We find that different datasets used for OOD evaluations in QA provide an estimate of models’ robustness to shortcuts that have a vastly different quality, some largely under-performing even a simple, in-distribution evaluation. We partially attribute this to the observation that spurious shortcuts are shared across ID+OOD datasets, but also find cases where a dataset’s quality for training and evaluation is largely disconnected. Our work underlines limitations of commonly-used OOD-based evaluations of generalization, and provides methodology and recommendations for evaluating generalization within and beyond QA more robustly.

Pretrained language models (LMs) are prone to arithmetic errors. Existing work showed limited success in probing numeric values from models’ representations, indicating that these errors can be attributed to the inherent unreliability of distributionally learned embeddings in representing exact quantities. However, we observe that previous probing methods are inadequate for the emergent structure of learned number embeddings with sinusoidal patterns.In response, we propose a novel probing technique that decodes numeric values from input embeddings with near-perfect accuracy across a range of open-source LMs. This proves that after the sole pre-training, LMs represent numbers with remarkable precision. Finally, we find that the embeddings’ preciseness judged by our probe’s accuracy explains a large portion of LM’s errors in elementary arithmetic, and show that aligning the embeddings with the pattern discovered by our probe can mitigate these errors.