ICLR 2025 Papers

With the goal of benchmarking generative systems beyond expert software development ability, we introduce Commit0, a benchmark that challenges AI agents to write libraries from scratch. Agents are provided with a specification document outlining the library’s API as well as a suite of interactive unit tests, with the goal of producing an implementation of this API accordingly. The implementation is validated through running these unit tests. As a benchmark, Commit0 is designed to move beyond static one-shot code generation towards agents that must process long-form natural language specifications, adapt to multi-stage feedback, and generate code with complex dependencies. Commit0 also offers an interactive environment where models receive static analysis and execution feedback on the code they generate. Our experiments demonstrate that while current agents can pass some unit tests, none can yet fully reproduce full libraries. Results also show that interactive feedback is quite useful for models to generate code that passes more unit tests, validating the benchmarks that facilitate its use. We publicly release the benchmark, the interactive environment, and the leaderboard.

We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs. For automated evaluation with WildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo. WildBench evaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments. WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie. Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation. Additionally, we propose a simple method to mitigate length bias, by converting outcomes of “slightly better/worse” to “tie” if the winner response exceeds the loser one by more than K characters. WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric. WildBench results demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard’s 0.91 and AlpacaEval2.0’s 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates.

Language models (LMs) can “memorize” information, i.e., encode training data in their weights in such a way that inference-time queries can lead to verbatim regurgitation of that data. This ability to extract training data can be problematic, for example, when data are private or sensitive. In this work, we investigate methods to mitigate memorization: three regularizer-based, three fine-tuning-based, and eleven machine unlearning-based methods, with five of the latter being new methods that we introduce. We also introduce TinyMem, a suite of small, computationally-efficient LMs for the rapid development and evaluation of memorization-mitigation methods. We demonstrate that the mitigation methods that we develop using TinyMem can successfully be applied to production-grade LMs, and we determine via experiment that: regularizer-based mitigation methods are slow and ineffective at curbing memorization; fine-tuning-based methods are effective at curbing memorization, but overly expensive, especially for retaining higher accuracies; and unlearning-based methods are faster and more effective, allowing for the precise localization and removal of memorized information from LM weights prior to inference. We show, in particular, that our proposed unlearning method BalancedSubnet outperforms other mitigation methods at removing memorized information while preserving performance on target tasks.

Offline goal-conditioned reinforcement learning (GCRL) is a major problem in reinforcement learning (RL) because it provides a simple, unsupervised, and domain-agnostic way to acquire diverse behaviors and representations from unlabeled data without rewards. Despite the importance of this setting, we lack a standard benchmark that can systematically evaluate the capabilities of offline GCRL algorithms. In this work, we propose OGBench, a new, high-quality benchmark for algorithms research in offline goal-conditioned RL. OGBench consists of 8 types of environments, 85 datasets, and reference implementations of 6 representative offline GCRL algorithms. We have designed these challenging and realistic environments and datasets to directly probe different capabilities of algorithms, such as stitching, long-horizon reasoning, and the ability to handle high-dimensional inputs and stochasticity. While representative algorithms may rank similarly on prior benchmarks, our experiments reveal stark strengths and weaknesses in these different capabilities, providing a strong foundation for building new algorithms. Project page: https://seohong.me/projects/ogbench

Large language models (LLMs) are driving substantial advancements in software engineering, with successful applications like Copilot and Cursor transforming real-world development practices. However, current research predominantly focuses on the early stages of development, such as code generation, while overlooking the post-development phases that are crucial to user experience. To explore the potential of LLMs in this direction, we propose OpenRCA, a benchmark dataset and evaluation framework for assessing LLMs’ ability to identify the root cause of software failures. OpenRCA includes 335 failures from three enterprise software systems, along with over 68 GB of telemetry data (logs, metrics, and traces). Given a failure case and its associated telemetry, the LLM is tasked to identify the root cause that triggered the failure, requiring comprehension of software dependencies and reasoning over heterogeneous, long-context telemetry data. Our results show substantial room for improvement, as current models can only handle the simplest cases. Even with the specially designed RCA-agent, the best-performing model, Claude 3.5, solved only 11.34% failure cases. Our work paves the way for future research in this direction.

Entropy-based objectives are widely used to perform state space exploration in reinforcement learning (RL) and dataset generation for offline RL. Behavioral entropy (BE), a rigorous generalization of classical entropies that incorporates cognitive and perceptual biases of agents, was recently proposed for discrete settings and shown to be a promising metric for robotic exploration problems. In this work, we propose using BE as a principled exploration objective for systematically generating datasets that provide diverse state space coverage in complex, continuous, potentially high-dimensional domains. To achieve this, we extend the notion of BE to continuous settings, derive tractable $k$-nearest neighbor estimators, provide theoretical guarantees for these estimators, and develop practical reward functions that can be used with standard RL methods to learn BE-maximizing policies. Using standard MuJoCo environments, we experimentally compare the performance of offline RL algorithms for a variety of downstream tasks on datasets generated using BE, R\'{e}nyi, and Shannon entropy-maximizing policies, as well as the SMM and RND algorithms. We find that offline RL algorithms trained on datasets collected using BE outperform those trained on datasets collected using Shannon entropy, SMM, and RND on all tasks considered, and on 80\% of the tasks compared to datasets collected using Renyi entropy.

Imitation learning from large multi-task demonstration datasets has emerged as a promising path for building generally-capable robots. As a result, 1000s of hours have been spent on building such large-scale datasets around the globe. Despite the continuous growth of such efforts, we still lack a systematic understanding of what data should be collected to improve the utility of a robotics dataset and facilitate downstream policy learning. In this work, we conduct a large-scale dataset composition study to answer this question. We develop a data generation framework to procedurally emulate common sources of diversity in existing datasets (such as sensor placements and object types and arrangements), and use it to generate large-scale robot datasets with controlled compositions, enabling a suite of dataset composition studies that would be prohibitively expensive in the real world. We focus on two practical settings: (1) what types of diversity should be emphasized when future researchers collect large-scale datasets for robotics, and (2) how should current practitioners retrieve relevant demonstrations from existing datasets to maximize downstream policy performance on tasks of interest. Our study yields several critical insights -- for example, we find that camera poses and spatial arrangements are crucial dimensions for both diversity in collection and alignment in retrieval. In real-world robot learning settings, we find that not only do our insights from simulation carry over, but our retrieval strategies on existing datasets such as DROID allow us to consistently outperform existing training strategies by up to 70\%.

Advances in machine learning research drive progress in real-world applications. To ensure this progress, it is important to understand the potential pitfalls on the way from a novel method's success on academic benchmarks to its practical deployment. In this work, we analyze existing tabular deep learning benchmarks and find two common characteristics of tabular data in typical industrial applications that are underrepresented in the datasets usually used for evaluation in the literature. First, in real-world deployment scenarios, distribution of data often changes over time. To account for this distribution drift, time-based train/test splits should be used in evaluation. However, existing academic tabular datasets often lack timestamp metadata to enable such evaluation. Second, a considerable portion of datasets in production settings stem from extensive data acquisition and feature engineering pipelines. This can have an impact on the absolute and relative number of predictive, uninformative, and correlated features compared to academic datasets. In this work, we aim to understand how recent research advances in tabular deep learning transfer to these underrepresented conditions. To this end, we introduce TabReD -- a collection of eight industry-grade tabular datasets. We reassess a large number of tabular ML models and techniques on TabReD. We demonstrate that evaluation on both time-based data splits and richer feature sets leads to different methods ranking, compared to evaluation on random splits and smaller number of features, which are common in academic benchmarks. Furthermore, simple MLP-like architectures and GBDT show the best results on the TabReD datasets, while other methods are less effective in the new setting.

Large Language Models (LLMs) have become instrumental across various applications, with the customization of these models to specific scenarios becoming increasingly critical. System message, a fundamental component of LLMs, is consist of carefully crafted instructions that guide the behavior of model to meet intended goals. Despite the recognized potential of system messages to optimize AI-driven solutions, there is a notable absence of a comprehensive benchmark for evaluating how well LLMs follow system messages. To fill this gap, we introduce SysBench, a benchmark that systematically analyzes system message following ability in terms of three limitations of existing LLMs: constraint violation, instruction misjudgement and multi-turn instability. Specifically, we manually construct evaluation dataset based on six prevalent types of constraints, including 500 tailor-designed system messages and multi-turn user conversations covering various interaction relationships. Additionally, we develop a comprehensive evaluation protocol to measure model performance. Finally, we conduct extensive evaluation across various existing LLMs, measuring their ability to follow specified constraints given in system messages. The results highlight both the strengths and weaknesses of existing models, offering key insights and directions for future research.

Real-world formal theorem proving often depends on a wealth of context, including definitions, lemmas, comments, file structure, and other information. We introduce $\texttt{miniCTX}$, which tests a model's ability to prove formal mathematical theorems that depend on new context that is not seen during training. $\texttt{miniCTX}$ contains theorems sourced from real Lean projects and textbooks, each associated with a context that can span tens of thousands of tokens. Models are tasked with proving a theorem given access to code from the theorem's repository, which contains context that is needed for the proof. As a baseline for $\texttt{miniCTX}$, we tested fine-tuning and prompting methods that condition theorem proving on preceding context. Both approaches substantially outperform traditional methods that rely solely on state information. We found that this ability to use context is not captured by previous benchmarks such as $\texttt{miniF2F}$. Alongside $\texttt{miniCTX}$, we offer $\texttt{ntp-toolkit}$ for automatically extracting and annotating theorem proving data, making it easy to add new projects into $\texttt{miniCTX}$ to ensure that contexts are not seen during training. $\texttt{miniCTX}$ offers a challenging and realistic evaluation of neural theorem provers.

We argue that the negative transfer problem occurring when the new task to learn arrives is an important problem that needs not be overlooked when developing effective Continual Reinforcement Learning (CRL) algorithms. Through comprehensive experimental validation, we demonstrate that such issue frequently exists in CRL and cannot be effectively addressed by several recent work on either mitigating plasticity loss of RL agents or enhancing the positive transfer in CRL scenario. To that end, we develop Reset & Distill (R&D), a simple yet highly effective baseline method, to overcome the negative transfer problem in CRL. R&D combines a strategy of resetting the agent's online actor and critic networks to learn a new task and an offline learning step for distilling the knowledge from the online actor and previous expert's action probabilities. We carried out extensive experiments on long sequence of Meta World tasks and show that our simple baseline method consistently outperforms recent approaches, achieving significantly higher success rates across a range of tasks. Our findings highlight the importance of considering negative transfer in CRL and emphasize the need for robust strategies like R&D to mitigate its detrimental effects.

Multimodal visual language models are gaining prominence in open-world applications, driven by advancements in model architectures, training techniques, and high-quality data. However, their performance is often limited by insufficient task-specific data, leading to poor generalization and biased outputs. Existing efforts to increase task diversity in fine-tuning datasets are hindered by the labor-intensive process of manual task labeling, which typically produces only a few hundred task types. To address this, we propose TaskGalaxy, a large-scale multimodal instruction fine-tuning dataset comprising 19,227 hierarchical task types and 413,648 samples. TaskGalaxy utilizes GPT-4o to enrich task diversity by expanding from a small set of manually defined tasks, with CLIP and GPT-4o filtering those that best match open-source images, and generating relevant question-answer pairs. Multiple models are employed to ensure sample quality. This automated process enhances both task diversity and data quality, reducing manual intervention. Incorporating TaskGalaxy into LLaVA-v1.5 and InternVL-Chat-v1.0 models shows substantial performance improvements across 16 benchmarks, demonstrating the critical importance of task diversity. TaskGalaxy is publicly released at https://github.com/Kwai-YuanQi/TaskGalaxy.

In the field of industrial inspection, Multimodal Large Language Models (MLLMs) have a high potential to renew the paradigms in practical applications due to their robust language capabilities and generalization abilities. However, despite their impressive problem-solving skills in many domains, MLLMs' ability in industrial anomaly detection has not been systematically studied. To bridge this gap, we present MMAD, a full-spectrum MLLM benchmark in industrial Anomaly Detection. We defined seven key subtasks of MLLMs in industrial inspection and designed a novel pipeline to generate the MMAD dataset with 39,672 questions for 8,366 industrial images. With MMAD, we have conducted a comprehensive, quantitative evaluation of various state-of-the-art MLLMs. The commercial models performed the best, with the average accuracy of GPT-4o models reaching 74.9\%. However, this result falls far short of industrial requirements. Our analysis reveals that current MLLMs still have significant room for improvement in answering questions related to industrial anomalies and defects. We further explore two training-free performance enhancement strategies to help models improve in industrial scenarios, highlighting their promising potential for future research. The code and data are available at https://github.com/jam-cc/MMAD.

Tool-Augmented Language Models (TALMs) leverage external APIs to answer user queries across various domains. However, existing benchmark datasets for TALM research often feature simplistic dialogues that do not reflect real-world scenarios, such as the need for models to ask clarifying questions or proactively call additional APIs when essential information is missing. To address these limitations, we construct and release ToolDial, a dataset comprising 11,111 multi-turn dialogues, with an average of 8.95 turns per dialogue, based on APIs from RapidAPI. ToolDial has two key characteristics. First, the dialogues incorporate 16 user and system actions (e.g., request, clarify, fail inform) to capture the rich dynamics of real-world interactions. Second, we simulate dialogues where the system requests necessary information from the user based on API documentation and seeks additional APIs if the user fails to provide the required information. To facilitate this process, we introduce a method for generating an API graph that represents input and output compatibility between APIs. Using ToolDial, we evaluate a suite of language models on their ability to predict correct actions and extract input parameter values for API calls from the dialogue history. Modern language models achieve accuracy scores below 70\%, indicating substantial room for improvement. We provide a detailed analysis of the areas where these models fall short.

A unique aspect of human visual understanding is the ability to flexibly interpret abstract concepts: acquiring lifted rules explaining what they symbolize, grounding them across familiar and unfamiliar contexts, and making predictions or reasoning about them. While off-the-shelf vision-language models excel at making literal interpretations of images (e.g., recognizing object categories such as tree branches), they still struggle to make sense of such visual abstractions (e.g., how an arrangement of tree branches may form the walls of a maze). To address this challenge, we introduce Deep Schema Grounding (DSG), a framework that leverages explicit structured representations of visual abstractions for grounding and reasoning. At the core of DSG are schemas—dependency graph descriptions of abstract concepts that decompose them into more primitive-level symbols. DSG uses large language models to extract schemas, then hierarchically grounds concrete to abstract components of the schema onto images with vision-language models. The grounded schema is used to augment visual abstraction understanding. We systematically evaluate DSG and different methods in reasoning on our new Visual Abstractions Benchmark, which consists of diverse, real-world images of abstract concepts and corresponding question-answer pairs labeled by humans. We show that DSG significantly improves the abstract visual reasoning performance of vision-language models, and is a step toward human-aligned understanding of visual abstractions.

This paper introduces MedTrinity-25M, a comprehensive, large-scale multimodal dataset for medicine, covering over 25 million images across 10 modalities with multigranular annotations for more than 65 diseases. These multigranular annotations encompass both global information, such as modality and organ detection, and local information like ROI analysis, lesion texture, and region-wise correlations. Unlike the existing multimodal datasets, which are limited by the availability of image-text pairs, we have developed the first automated pipeline that scales up multimodal data by generating multigranular visual and textual annotations in the form of image-ROI-description triplets without the need for any paired text descriptions. Specifically, data from over 30 different sources have been collected, preprocessed, and grounded using domain-specific expert models to identify ROIs related to abnormal regions. We then build a comprehensive knowledge base and prompt multimodal large language models to perform retrieval-augmented generation with the identified ROIs as guidance, resulting in multigranular textual descriptions. Compared to existing datasets, MedTrinity-25M provides the most enriched annotations, supporting a comprehensive range of multimodal tasks such as captioning and report generation, as well as vision-centric tasks like classification and segmentation. We propose LLaVA-Tri by pretraining LLaVA on MedTrinity-25M, achieving state-of-the-art performance on VQA-RAD, SLAKE, and PathVQA, surpassing representative SOTA multimodal large language models. Furthermore, MedTrinity-25M can also be utilized to support large-scale pre-training of multimodal medical AI models, contributing to the development of future foundation models in the medical domain. We will make our dataset available. The dataset is publicly available at https://yunfeixie233.github.io/MedTrinity-25M/.

While text-to-image (T2I) generative models have become ubiquitous, they do not necessarily generate images that align with a given prompt. While many metrics and benchmarks have been proposed to evaluate T2I models and alignment metrics, the impact of the evaluation components (prompt sets, human annotations, evaluation task) has not been systematically measured. We find that looking at only *one slice of data*, i.e. one set of capabilities or human annotations, is not enough to obtain stable conclusions that generalise to new conditions or slices when evaluating T2I models or alignment metrics. We address this by introducing an evaluation suite of $>$100K annotations across four human annotation templates that comprehensively evaluates models' capabilities across a range of methods for gathering human annotations and comparing models. In particular, we propose (1) a carefully curated set of prompts -- *Gecko2K*; (2) a statistically grounded method of comparing T2I models; and (3) how to systematically evaluate metrics under three *evaluation tasks* -- *model ordering, pair-wise instance scoring, point-wise instance scoring*. Using this evaluation suite, we evaluate a wide range of metrics and find that a metric may do better in one setting but worse in another. As a result, we introduce a new, interpretable auto-eval metric that is consistently better correlated with human ratings than such existing metrics on our evaluation suite--across different human templates and evaluation settings--and on TIFA160.

Fine-tuning is widely used to adapt language models for specific goals, often leveraging real-world data such as patient records, customer-service interactions, or web content in languages not covered in pre-training. These datasets are typically massive, noisy, and often confidential, making their direct inspection challenging. However, understanding them is essential for guiding model deployment and informing decisions about data cleaning or suppressing any harmful behaviors learned during fine-tuning. In this study, we introduce the task of novelty discovery through generation, which aims to identify novel domains of a fine-tuning dataset by generating examples that illustrate these properties. Our approach － Contrastive Generative Exploration (CGE) － assumes no direct access to the data but instead relies on a pre-trained model and the same model after fine-tuning. By contrasting the predictions of these two models, CGE can generate examples that highlight novel domains of the fine-tuning data. However, this simple approach may produce examples that are too similar to one another, failing to capture the full range of novel domains present in the dataset. We address this by introducing an iterative version of CGE, where the previously generated examples are used to update the pre-trained model, and this updated model is then contrasted with the fully fine-tuned model to generate the next example, promoting diversity in the generated outputs. Our experiments demonstrate the effectiveness of CGE in detecting novel domains, such as toxic language, as well as new natural and programming languages. Furthermore, we show that CGE remains effective even when models are fine-tuned using differential privacy techniques.

In this paper, we introduce group convolutional neural networks (GCNNs) equivariant to color variation. GCNNs have been designed for a variety of geometric transformations from 2D and 3D rotation groups, to semi-groups such as scale. Despite the improved interpretability, accuracy and generalizability of these architectures, GCNNs have seen limited application in the context of perceptual quantities. Notably, the recent CEConv network uses a GCNN to achieve equivariance to hue transformations by convolving input images with a hue rotated RGB filter. However, this approach leads to invalid RGB values which break equivariance and degrade performance. We resolve these issues with a lifting layer that transforms the input image directly, thereby circumventing the issue of invalid RGB values and improving equivariance error by over three orders of magnitude. Moreover, we extend the notion of color equivariance to include equivariance to saturation and luminance shift. Our hue-, saturation-, luminance- and color-equivariant networks achieve strong generalization to out-of-distribution perceptual variations and improved sample efficiency over conventional architectures. We demonstrate the utility of our approach on synthetic and real world datasets where we consistently outperform competitive baselines.