ICLR 2025 Papers

Large language models (LLMs) have proven invaluable for code generation, particularly in interactive settings. However, existing code generation benchmarks fail to capture the diverse feedback encountered in multi-turn interactions, limiting our ability to evaluate LLMs in these contexts. To address this gap, we present a set of novel benchmarks that explicitly model the quality of feedback provided to code generation LLMs. Our contributions are threefold: **First**, we introduce CONVCODEWORLD, a novel and reproducible environment for benchmarking interactive code generation. CONVCODEWORLD simulates 9 distinct interactive code generation scenarios while systematically combining three types of feedback: (a) compilation feedback; (b) execution feedback with varying test coverage; (c) verbal feedback generated by GPT-4o with different levels of expertise. **Second**, we introduce CONVCODEBENCH, a fast, static version of benchmark that uses pre-generated feedback logs, eliminating the need for costly dynamic verbal feedback generation while maintaining strong Spearman’s rank correlations (0.82 to 0.99) with CONVCODEWORLD. **Third**, extensive evaluations of both closed-source and open-source LLMs including R1-Distill on CONVCODEWORLD reveal key insights: (a) LLM performance varies significantly based on the feedback provided; (b) Weaker LLMs, with sufficient feedback, can outperform single-turn results of state-of-the-art LLMs without feedback; (c) Training on a specific feedback combination can limit an LLM’s ability to utilize unseen combinations; (d) LLMs solve problems in fewer turns (high MRR) may not solve as many problems overall (high Recall), and vice versa. All implementations and benchmarks will be made publicly available at https://huggingface.co/spaces/ConvCodeWorld/ConvCodeWorld

Existing benchmarks for language agents do not set them up to interact with human users or follow domain-specific rules, both of which are vital to safe and realistic deployment. We propose $\tau$-bench, a benchmark with two domains (retail and airline) emulating dynamic conversations between a user (simulated by language models) and a customer service agent provided with domain-specific API tools and policy guidelines. We employ a efficient and faithful evaluation process that compares the database state at the end of a conversation with the annotated goal state, and propose a new metric (pass^k) to evaluate the reliability of agent behavior over multiple trials. Our experiments show that even state-of-the-art function calling agents (gpt-4o) succeed on $<50\%$ of the tasks, and are terribly inconsistent (pass^8 < 25\% in retail). Our findings point to the need for methods that can improve the ability of agents to act consistently and reliably follow rules.

Autonomous systems for software engineering are now capable of fixing bugs and developing features. These systems are commonly evaluated on SWE-bench (Jimenez et al., 2024a), which assesses their ability to solve software issues from GitHub repositories. However, SWE-bench uses only Python repositories, with problem statements presented predominantly as text and lacking visual elements such as images. This limited coverage motivates our inquiry into how existing systems might perform on unrepresented software engineering domains (e.g., front-end, game development, DevOps), which use different programming languages and paradigms. Therefore, we propose SWE-bench Multimodal (SWE-bench M), to evaluate systems on their ability to fix bugs in visual, user-facing JavaScript software. SWE-bench M features 617 task instances collected from 17 JavaScript libraries used for web interface design, diagramming, data visualization, syntax highlighting, and interactive mapping. Each SWE-bench M task instance contains at least one image in its problem statement or unit tests. Our analysis finds that top-performing SWE-bench systems struggle with SWE-bench M, revealing limitations in visual problem-solving and cross-language generalization. Lastly, we show that SWE-agent’s flexible language-agnostic features enable it to substantially outperform alternatives on SWE-bench M, resolving 12% of task instances compared to 6% for the next best system.

Artificial neural networks often struggle with _catastrophic forgetting_ when learning multiple tasks sequentially, as training on new tasks degrades the performance on previously learned tasks. Recent theoretical work has addressed this issue by analysing learning curves in synthetic frameworks under predefined training protocols. However, these protocols relied on heuristics and lacked a solid theoretical foundation assessing their optimality. In this paper, we fill this gap by combining exact equations for training dynamics, derived using statistical physics techniques, with optimal control methods. We apply this approach to teacher-student models for continual learning and multi-task problems, obtaining a theory for task-selection protocols maximising performance while minimising forgetting. Our theoretical analysis offers non-trivial yet interpretable strategies for mitigating catastrophic forgetting, shedding light on how optimal learning protocols modulate established effects, such as the influence of task similarity on forgetting. Finally, we validate our theoretical findings with experiments on real-world data.

This paper introduces PhyloLM, a method adapting phylogenetic algorithms to Large Language Models (LLMs) to explore whether and how they relate to each other and to predict their performance characteristics. Our method calculates a phylogenetic distance metric based on the similarity of LLMs' output. The resulting metric is then used to construct dendrograms, which satisfactorily capture known relationships across a set of 111 open-source and 45 closed models. Furthermore, our phylogenetic distance predicts performance in standard benchmarks, thus demonstrating its functional validity and paving the way for a time and cost-effective estimation of LLM capabilities. To sum up, by translating population genetic concepts to machine learning, we propose and validate a tool to evaluate LLM development, relationships and capabilities, even in the absence of transparent training information.

Temporal Graph Neural Networks (TGNNs) lay emphasis on capturing node interactions over time but often overlook evolution in node classes and dynamic data distributions triggered by the continuous emergence of new class labels, known as the open-set problem. This problem poses challenges for existing TGNNs in preserving learned classes while rapidly adapting to new, unseen classes. To address this, this paper identifies two primary factors affecting model performance on the open temporal graph, backed by a theoretical guarantee: (1) the forgetting of prior knowledge and (2) distribution discrepancies between successive tasks. Building on these insights, we propose N-ForGOT, which incorporates two plug-in modules into TGNNs to preserve prior knowledge and enhance model generalizability for new classes simultaneously. The first module preserves previously established inter-class connectivity and decision boundaries during the training of new classes to mitigate the forgetting caused by temporal evolutions of class characteristics. The second module introduces an efficient method for measuring distribution discrepancies with designed temporal Weisfeiler-Lehman subtree patterns, effectively addressing both structural and temporal shifts while reducing time complexity. Experimental results on four public datasets demonstrate that our method significantly outperforms state-of-the-art approaches in prediction accuracy, prevention of forgetting, and generalizability.

Current backdoor defense methods are evaluated against a single attack at a time. This is unrealistic, as powerful machine learning systems are trained on large datasets scraped from the internet, which may be attacked multiple times by one or more attackers. We demonstrate that multiple backdoors can be simultaneously installed in a single model through parallel data poisoning attacks without substantially degrading clean accuracy. Furthermore, we show that existing backdoor defense methods do not effectively defend against multiple simultaneous attacks. Finally, we leverage insights into the nature of backdoor attacks to develop a new defense, BaDLoss (**Ba**ckdoor **D**etection via **Loss** Dynamics), that is effective in the multi-attack setting. With minimal clean accuracy degradation, BaDLoss attains an average attack success rate in the multi-attack setting of 7.98% in CIFAR-10, 10.29% in GTSRB, and 19.17% in Imagenette, compared to the average of other defenses at 63.44%, 74.83%, and 41.74% respectively. BaDLoss scales to ImageNet-1k, reducing the average attack success rate from 88.57% to 15.61%.

Graph Neural Networks (GNNs) have become important tools for machine learning on graph-structured data. In this paper, we explore the synergistic combination of graph encoding, graph rewiring, and graph attention, by introducing Graph Attention with Stochastic Structures (GRASS), a novel GNN architecture. GRASS utilizes relative random walk probabilities (RRWP) encoding and a novel decomposed variant (D-RRWP) to efficiently capture structural information. It rewires the input graph by superimposing a random regular graph to enhance long-range information propagation. It also employs a novel additive attention mechanism tailored for graph-structured data. Our empirical evaluations demonstrate that GRASS achieves state-of-the-art performance on multiple benchmark datasets, including a 20.3% reduction in mean absolute error on the ZINC dataset.

Many real-world user queries (e.g. *"How do to make egg fried rice?"*) could benefit from systems capable of generating responses with both textual steps with accompanying images, similar to a cookbook. Models designed to generate interleaved text and images face challenges in ensuring consistency within and across these modalities. To address these challenges, we present ISG, a comprehensive evaluation framework for interleaved text-and-image generation. ISG leverages a scene graph structure to capture relationships between text and image blocks, evaluating responses on four levels of granularity: holistic, structural, block-level, and image-specific. This multi-tiered evaluation allows for a nuanced assessment of consistency, coherence, and accuracy, and provides interpretable question-answer feedback. In conjunction with ISG, we introduce a benchmark, ISG-Bench, encompassing 1,150 samples across 8 categories and 21 subcategories. This benchmark dataset includes complex language-vision dependencies and golden answers to evaluate models effectively on vision-centric tasks such as style transfer, a challenging area for current models. Using ISG-Bench, we demonstrate that recent unified vision-language models perform poorly on generating interleaved content. While compositional approaches that combine separate language and image models show a 111% improvement over unified models at the holistic level, their performance remains suboptimal at both block and image levels. To facilitate future work, we develop ISG-Agent, a baseline agent employing a *"plan-execute-refine"* pipeline to invoke tools, achieving a 122% performance improvement.

Post-training quantization (PTQ) has emerged as a widely adopted technique for compressing and accelerating Large Language Models (LLMs). The major challenge in LLM quantization is that uneven and heavy-tailed data distributions can expand the quantization range, thereby reducing bit precision for most values. Recent methods attempt to eliminate outliers and balance inter-channel differences by employing linear transformations; however, they remain heuristic and are often overlook optimizing the data distribution across the entire quantization space. In this paper, we introduce Quantization Space Utilization Rate (QSUR), a novel metric that effectively assesses the quantizability of transformed data by measuring the space utilization of the data in the quantization space. We complement QSUR with mathematical derivations that examine the effects and limitations of various transformations, guiding our development of Orthogonal and Scaling Transformation-based Quantization (OSTQuant). OSTQuant employs a learnable equivalent transformation, consisting of an orthogonal transformation and a scaling transformation, to optimize the distributions of weights and activations across the entire quantization space. Futhermore, we propose the KL-Top loss function, designed to mitigate noise during optimization while retaining richer semantic information within the limited calibration data imposed by PTQ. OSTQuant outperforms existing work on various LLMs and benchmarks. In the W4-only setting, it retains 99.5\% of the floating-point accuracy. In the more challenging W4A4KV4 configuration, OSTQuant reduces the performance gap by 32\% on the LLaMA-3-8B model compared to state-of-the-art methods. Code will be available.

Forecasting is a task that is difficult to evaluate: the ground truth can only be known in the future. Recent work showing LLM forecasters rapidly approaching human-level performance begs the question: how can we benchmark and evaluate these forecasters *instantaneously*? Following the consistency check framework, we measure the performance of forecasters in terms of the consistency of their predictions on different logically-related questions. We propose a new, general consistency metric based on *arbitrage*: for example, if a forecasting AI illogically predicts that both the Democratic and Republican parties have 60\% probability of winning the 2024 US presidential election, an arbitrageur could trade against the forecaster's predictions and make a profit. We build an automated evaluation system that generates a set of base questions, instantiates consistency checks from these questions, elicits the predictions of the forecaster, and measures the consistency of the predictions. We then build a standard, proper-scoring-rule forecasting benchmark, and show that our (instantaneous) consistency metrics correlate strongly with LLM forecasters' ground truth Brier scores (which are only known in the future). We also release a consistency benchmark that resolves in 2028, providing a long-term evaluation tool for forecasting.

The proliferation of deepfakes and AI-generated content has led to a surge in media forgeries and misinformation, necessitating robust detection systems. However, current datasets lack diversity across modalities, languages, and real-world scenarios. To address this gap, we present ILLUSION (Integration of Life-Like Unique Synthetic Identities and Objects from Neural Networks), a large-scale, multi-modal deepfake dataset comprising 1.3 million samples spanning audio-visual forgeries, 26 languages, challenging noisy environments, and various manipulation protocols. Generated using 28 state-of-the-art generative techniques, ILLUSION includes faceswaps, audio spoofing, synchronized audio-video manipulations, and synthetic media while ensuring a balanced representation of gender and skin tone for unbiased evaluation. Using Jaccard Index and UpSet plot analysis, we demonstrate ILLUSION’s distinctiveness and minimal overlap with existing datasets, emphasizing its novel generative coverage. We benchmarked image, audio, video, and multi-modal detection models, revealing key challenges such as performance degradation in multilingual and multi-modal contexts, vulnerability to real-world distortions, and limited generalization to zero-day attacks. By bridging synthetic and real-world complexities, ILLUSION provides a challenging yet essential platform for advancing deepfake detection research. The dataset is publicly available at https://www.iab-rubric.org/illusion-database.

Offline reinforcement learning (RL) aims to learn from static datasets and thus faces the challenge of value estimation errors for out-of-distribution actions. The in-sample learning scheme addresses this issue by performing implicit TD backups that does not query the values of unseen actions. However, pre-existing in-sample value learning and policy extraction methods suffer from over-regularization, limiting their performance on suboptimal or compositional datasets. In this paper, we analyze key factors in in-sample learning that might potentially hinder the use of a milder constraint. We propose Actor-Critic with Temperature adjustment and In-sample Value Ensemble (ACTIVE), a novel in-sample offline RL algorithm that leverages an ensemble of $V$-functions for critic training and adaptively adjusts the constraint level using dual gradient descent. We theoretically show that the $V$-ensemble suppresses the accumulation of initial value errors, thereby mitigating overestimation. Our experiments on the D4RL benchmarks demonstrate that ACTIVE alleviates overfitting of value functions and outperforms existing in-sample methods in terms of learning stability and policy optimality.

Multimodal foundation models (MMFMs) play a crucial role in various applications, including autonomous driving, healthcare, and virtual assistants. However, several studies have revealed vulnerabilities in these models, such as generating unsafe content by text-to-image models. Existing benchmarks on multimodal models either predominantly assess the helpfulness of these models, or only focus on limited perspectives such as fairness and privacy. In this paper, we present the first unified platform, MMDT (Multimodal DecodingTrust), designed to provide a comprehensive safety and trustworthiness evaluation for MMFMs. Our platform assesses models from multiple perspectives, including safety, hallucination, fairness/bias, privacy, adversarial robustness, and out-of-distribution (OOD) generalization. We have designed various evaluation scenarios and red teaming algorithms under different tasks for each perspective to generate challenging data, forming a high-quality benchmark. We evaluate a range of multimodal models using MMDT, and our findings reveal a series of vulnerabilities and areas for improvement across these perspectives. This work introduces the first comprehensive and unique safety and trustworthiness evaluation platform for MMFMs, paving the way for developing safer and more reliable MMFMs and systems. Our platform and benchmark are available at https://mmdecodingtrust.github.io/.

Quantifying model generalization to out-of-distribution data has been a longstanding challenge in machine learning. Addressing this issue is crucial for leveraging machine learning in scientific discovery, where models must generalize to new molecules or materials. Current methods typically split data into train and test sets using various criteria — temporal, sequence identity, scaffold, or random cross-validation — before evaluating model performance. However, with so many splitting criteria available, existing approaches offer limited guidance on selecting the most appropriate one, and they do not provide mechanisms for incorporating prior knowledge about the target deployment distribution(s). To tackle this problem, we have developed a novel metric, AU-GOOD, which quantifies expected model performance under conditions of increasing dissimilarity between train and test sets, while also accounting for prior knowledge about the target deployment distribution(s), when available. This metric is broadly applicable to biochemical entities, including proteins, small molecules, nucleic acids, or cells; as long as a relevant similarity function is defined for them. Recognizing the wide range of similarity functions used in biochemistry, we propose criteria to guide the selection of the most appropriate metric for partitioning. We also introduce a new partitioning algorithm that generates more challenging test sets, and we propose statistical methods for comparing models based on AU-GOOD. Finally, we demonstrate the insights that can be gained from this framework by applying it to two different use cases: developing predictors for pharmaceutical properties of small molecules, and using protein language models as embeddings to build biophysical property predictors.

Multimodal Large Language Models (MLLMs) have become a powerful tool for integrating visual and textual information. Despite their exceptional performance on visual understanding benchmarks, measuring their ability to reason abstractly across multiple images remains a significant challenge. To address this, we introduce VOILA, a large-scale, open-ended, dynamic benchmark designed to evaluate MLLMs' perceptual understanding and abstract relational reasoning. VOILA employs an analogical mapping approach in the visual domain, requiring models to generate an image that completes an analogy between two given image pairs, reference and application, without relying on predefined choices. Our experiments demonstrate that the analogical reasoning tasks in VOILA present a challenge to MLLMs. Through multi-step analysis, we reveal that current MLLMs struggle to comprehend inter-image relationships and exhibit limited capabilities in high-level relational reasoning. Notably, we observe that performance improves when following a multi-step strategy of least-to-most prompting. Comprehensive evaluations on open-source models and GPT-4o show that on text-based answers, the best accuracy for challenging scenarios is 13% (LLaMa 3.2) and even for simpler tasks is only 29% (GPT-4o), while human performance is significantly higher at 70% across both difficulty levels.

While Large Language Models (LLMs) excel in question-answering (QA) tasks, their real reasoning abilities on multiple evidence retrieval and integration on Multi-hop QA tasks remain less explored. Firstly, LLMs sometimes generate answers that rely on internal memory rather than retrieving evidence and reasoning in the given context, which brings concerns about the evaluation quality of real reasoning abilities. Although previous counterfactual QA benchmarks can separate the internal memory of LLMs, they focus solely on final QA performance, which is insufficient for reporting LLMs' real reasoning abilities. Because LLMs are expected to engage in intricate reasoning processes that involve evidence retrieval and answering a series of sub-questions from given passages. Moreover, current factual Multi-hop QA (MHQA) benchmarks are annotated on open-source corpora such as Wikipedia, although useful for multi-step reasoning evaluation, they show limitations due to the potential data contamination in LLMs' pre-training stage. To address these issues, we introduce the Step-wise and Counterfactual benchmark (CofCA), a novel evaluation benchmark consisting of factual data and counterfactual data that reveals LLMs' real reasoning abilities on multi-step reasoning and reasoning chain evaluation. Our experimental results reveal a significant performance gap of several LLMs between Wikipedia-based factual data and counterfactual data, deeming data contamination issues in existing benchmarks. Moreover, we observe that LLMs usually bypass the correct reasoning chain, showing an inflated multi-step reasoning performance. We believe that our CofCA benchmark will enhance and facilitate the evaluations of trustworthy LLMs.

Graphs are ubiquitous in social networks, chemical molecules, and financial data, where Graph Neural Networks (GNNs) achieve superior predictive accuracy. Graphs can be evolving, while understanding how GNN predictions respond to the evolution provides significant insight and trust. We explore the problem of explaining evolving GNN predictions due to continuously changing edge weights. We introduce a layer edge-based explanation to balance explanation fidelity and interpretability. We propose a novel framework to address the challenges of axiomatic attribution and the entanglement of multiple computational graph paths due to continuous change of edge weights. We first design an axiomatic attribution of the evolution of the model prediction to message flows, then develop Shapley value to fairly map message flow contributions to layer edges. We formulate a novel optimization problem to find the critical layer edges based on KL-divergence minimization. Extensive experiments on eight datasets for node classification, link prediction, and graph classification tasks with evolving graphs demonstrate the better fidelity and interpretability of the proposed method over the baseline methods. The code is available at https://github.com/yazhengliu/Axiomatic-Layer-Edges/tree/main.

Learning holistic computational representations in physical, chemical or biological systems requires the ability to process information from different distributions and modalities within the same model. Thus, the demand for multimodal machine learning models has sharply risen for modalities that go beyond vision and language, such as sequences, graphs, time series, or tabular data. While there are many available multimodal fusion and alignment approaches, most of them require end-to-end training, scale quadratically with the number of modalities, cannot handle cases of high modality imbalance in the training set, or are highly topology-specific, making them too restrictive for many biomedical learning tasks. This paper presents Multimodal Lego (MM-Lego), a general-purpose fusion framework to turn any set of encoders into a competitive multimodal model with no or minimal fine-tuning. We achieve this by introducing a wrapper for any unimodal encoder that enforces shape consistency between modality representations. It harmonises these representations by learning features in the frequency domain to enable model merging with little signal interference. We show that MM-Lego 1) can be used as a model merging method which achieves competitive performance with end-to-end fusion models without any fine-tuning, 2) can operate on any unimodal encoder, and 3) is a model fusion method that, with minimal fine-tuning, surpasses all benchmarks in five out of seven datasets.