Code Agent development is an extremely active research area, where a reliable performance metric is critical for tracking progress and guiding new developments. This demand is underscored by the meteoric rise in popularity of SWE-Bench. This benchmark challenges code agents to generate patches addressing GitHub issues given the full repository as context. The correctness of generated patches is then evaluated by executing a human-written test suite extracted from the repository after the issue's resolution. However, constructing benchmarks like SWE-Bench requires substantial manual effort to set up historically accurate execution environments for testing. Crucially, this severely limits the number of considered repositories, e.g., just 12 for SWE-Bench. Considering so few repositories, selected for their popularity runs the risk of leading to a distributional mismatch, i.e., the measured performance may not be representative of real-world scenarios potentially misguiding development efforts. In this work, we address this challenge and introduce SetUpAgent, a fully automated system capable of historically accurate dependency setup, test execution, and result parsing. Using SetUpAgent, we generate two new datasets: (i) SWEE-Bench an extended version of SWE-Bench encompassing hundreds of repositories, and (ii) SWA-Bench a benchmark focusing on applications rather than libraries. Comparing these datasets to SWE-Bench with respect to their characteristics and code agent performance, we find significant distributional differences, including lower issue description quality and detail level, higher fix complexity, and most importantly up to 40% lower agent success rates.

Randomized smoothing is a popular approach for providing certified robustness guarantees against adversarial attacks, and has become a very active area of research. Over the past years, the average certified radius (ACR) has emerged as the single most important metric for comparing methods and tracking progress in the field. However, in this work, we show that ACR is an exceptionally poor metric for evaluating robustness guarantees provided by randomized smoothing. We theoretically show not only that a trivial classifier can have arbitrarily large ACR, but also that ACR is much more sensitive to improvements on easy samples than on hard ones. Empirically, we confirm that existing training strategies that improve ACR reduce the model's robustness on hard samples. Further, we show that by focusing on easy samples, we can effectively replicate the increase in ACR. We develop strategies, including explicitly discarding hard samples, reweighing the dataset with certified radius, and extreme optimization for easy samples, to achieve state-of-the-art ACR, although these strategies ignore robustness for the general data distribution. Overall, our results suggest that ACR has introduced a strong undesired bias to the field, and better metrics are required to holistically evaluate randomized smoothing.

Quickly resolving issues reported in industrial applications is crucial to minimize economic impact. However, the required data analysis makes diagnosing the underlying root causes a challenging and time-consuming task, even for experts. In contrast, large language models (LLMs) excel at analyzing large amounts of data. Indeed, prior work in AI-Ops demonstrates their effectiveness in analyzing IT systems. Here, we extend this work to the challenging and largely unexplored domain of robotics systems. To this end, we create SYSDIAGBENCH, a proprietary system diagnostics benchmark for robotics, containing over 2500 reported issues. We leverage SYSDIAGBENCH to investigate the performance of LLMs for root cause analysis, considering a range of model sizes and adaptation techniques. Our results show that QLoRA finetuning can be sufficient to let a 7B-parameter model outperform GPT-4 in terms of diagnostic accuracy while being significantly more cost-effective. We validate our LLM-as-a-judge results with a human expert study and find that our best model achieves similar approval ratings as our reference labels.

As open-weight large language models (LLMs) achieve ever more impressive performances across a wide range of tasks in English, practitioners aim to adapt these models to different languages. However, such language adaptation is often accompanied by catastrophic forgetting of the base model's capabilities, severely limiting the usefulness of the resulting model. We address this issue by proposing Branch-and-Merge (BaM), a new adaptation method based on iteratively merging multiple models, fine-tuned on a subset of the available training data. BaM is based on the insight that this yields lower magnitude but higher quality weight changes, reducing forgetting of the source domain while maintaining learning on the target domain. We demonstrate in an extensive empirical study on Bulgarian and German that BaM can significantly reduce forgetting while matching or even improving target domain performance compared to both standard continued pretraining and instruction finetuning across different model architectures.

Federated learning is a framework for collaborative machine learning where clients only share gradient updates and not their private data with a server. However, it was recently shown that gradient inversion attacks can reconstruct this data from the shared gradients. In the important honest-but-curious setting, existing attacks enable exact reconstruction only for a batch size of $b=1$, with larger batches permitting only approximate reconstruction. In this work, we propose SPEAR, the first algorithm reconstructing whole batches with $b >1$ exactly. SPEAR combines insights into the explicit low-rank structure of gradients with a sampling-based algorithm. Crucially, we leverage ReLU-induced gradient sparsity to precisely filter out large numbers of incorrect samples, making a final reconstruction step tractable. We provide an efficient GPU implementation for fully connected networks and show that it recovers high-dimensional ImageNet inputs in batches of up to $b \lesssim 25$ exactly while scaling to large networks. Finally, we show theoretically that much larger batches can be reconstructed with high probability given exponential time.

Public benchmarks play an essential role in the evaluation of large language models. However, data contamination can lead to inflated performance, rendering them unreliable for model comparison. It is therefore crucial to detect contamination and estimate its impact on measured performance. Unfortunately, existing detection methods can be easily evaded and fail to quantify contamination. To overcome these limitations, we propose a novel definition of contamination as artificially inflated and non-generalizing benchmark performance instead of the inclusion of benchmark samples in the training data. This perspective enables us to detect any model with inflated performance, i.e., performance that does not generalize to rephrased samples, synthetic samples from the same distribution, or different benchmarks for the same task. Based on this insight, we develop ConStat, a statistical method that reliably detects and quantifies contamination by comparing performance between a primary and reference benchmark relative to a set of reference models. We demonstrate the effectiveness of ConStat in an extensive evaluation of diverse model architectures, benchmarks, and contamination scenarios and find high levels of contamination in multiple popular models including Mistral, Llama, Yi, and the top-3 Open LLM Leaderboard models.

Federated learning works by aggregating locally computed gradients from multiple clients, thus enabling collaborative training without sharing private client data. However, prior work has shown that the data can actually be recovered by the server using so-called gradient inversion attacks. While these attacks perform well when applied on images, they are limited in the text domain and only permit approximate reconstruction of small batches and short input sequences. In this work, we propose DAGER, the first algorithm to recover whole batches of input text exactly. DAGER leverages the low-rank structure of self-attention layer gradients and the discrete nature of token embeddings to efficiently check if a given token sequence is part of the client data. We use this check to exactly recover full batches in the honest-but-curious setting without any prior on the data for both encoder- and decoder-based architectures using exhaustive heuristic search and a greedy approach, respectively. We provide an efficient GPU implementation of DAGER and show experimentally that it recovers full batches of size up to 128 on large language models (LLMs), beating prior attacks in speed (20x at same batch size), scalability (10x larger batches), and reconstruction quality (ROUGE-1/2 > 0.99).

Large language models are widespread, with their performance on benchmarks frequently guiding user preferences for one model over another. However, the vast amount of data these models are trained on can inadvertently lead to contamination with public benchmarks, thus compromising performance measurements. While recently developed contamination detection methods try to address this issue, they overlook the possibility of deliberate contamination by malicious model providers aiming to evade detection. We argue that this setting is of crucial importance as it casts doubt on the reliability of public benchmarks. To more rigorously study this issue, we propose a categorization of both model providers and contamination detection methods. This reveals vulnerabilities in existing methods that we exploit with EAL, a simple yet effective contamination technique that significantly inflates benchmark performance while completely evading current detection methods.

While the ImageNet dataset has been driving computer vision research over the past decade, significant label noise and ambiguity have made top-1 accuracy an insufficient measure of further progress. To address this, new label-sets and evaluation protocols have been proposed for ImageNet showing that state-of-the-art models already achieve over 95% accuracy and shifting the focus on investigating why the remaining errors persist. Recent work in this direction employed a panel of experts to manually categorize all remaining classification errors for two selected models. However, this process is time-consuming, prone to inconsistencies, and requires trained experts, making it unsuitable for regular model evaluation thus limiting its utility. To overcome these limitations, we propose the first automated error classification framework, a valuable tool to study how modeling choices affect error distributions. We use our framework to comprehensively evaluate the error distribution of over 900 models. Perhaps surprisingly, we find that across model architectures, scales, and pre-training corpora, top-1 accuracy is a strong predictor for the portion of all error types. In particular, we observe that the portion of severe errors drops significantly with top-1 accuracy indicating that, while it underreports a model's true performance, it remains a valuable performance metric.

Many recent prompting strategies for large language models (LLMs) query the model multiple times sequentially -- first to produce intermediate results and then the final answer. However, using these methods, both decoder and model are unaware of potential follow-up prompts, leading to disconnected and undesirably wordy intermediate responses. In this work, we address this issue by proposing prompt sketching, a new prompting paradigm in which an LLM does not only respond by completing a prompt, but by predicting values for multiple variables in a template. This way, sketching grants users more control over the generation process, e.g., by providing a reasoning framework via intermediate instructions, leading to better overall results. The key idea enabling sketching with existing, autoregressive models is to adapt the decoding procedure to also score follow-up instructions during text generation, thus optimizing overall template likelihood in inference. Our experiments show that in a zero-shot setting, prompt sketching outperforms existing, sequential prompting schemes such as direct asking or chain-of-thought on 7 out of 8 LLM benchmarking tasks, including state tracking, arithmetic reasoning, and general question answering. To facilitate future use, we release a number of generic, yet effective sketches applicable to many tasks, and an open source library called dclib, powering our sketch-aware decoders.