ICLR 2025 Papers

We present MEGA-Bench, an evaluation suite that scales multimodal evaluation to over 500 real-world tasks, to address the highly heterogeneous daily use cases of end users. Our objective is to optimize for a set of high-quality data samples that cover a highly diverse and rich set of multimodal tasks, while enabling cost-effective and accurate model evaluation. In particular, we collected 505 realistic tasks encompassing over 8,000 samples from 16 expert annotators to extensively cover the multimodal task space. Instead of unifying these problems into standard multi-choice questions (like MMMU, MM-Bench, and MMT-Bench), we embrace a wide range of output formats like numbers, phrases, code, \LaTeX, coordinates, JSON, free-form, etc. To accommodate these formats, we developed over 40 metrics to evaluate these tasks. Unlike existing benchmarks, MEGA-Bench offers a fine-grained capability report across multiple dimensions (e.g., application, input type, output format, skill), allowing users to interact with and visualize model capabilities in depth. We evaluate a wide variety of frontier vision-language models on MEGA-Bench to understand their capabilities across these dimensions.

This paper introduces LAFT, a novel feature transformation method designed to incorporate user knowledge and preferences into anomaly detection using natural language. Accurately modeling the boundary of normality is crucial for distinguishing abnormal data, but this is often challenging due to limited data or the presence of nuisance attributes. While unsupervised methods that rely solely on data without user guidance are common, they may fail to detect anomalies of specific interest. To address this limitation, we propose Language-Assisted Feature Transformation (LAFT), which leverages the shared image-text embedding space of vision-language models to transform visual features according to user-defined requirements. Combined with anomaly detection methods, LAFT effectively aligns visual features with user preferences, allowing anomalies of interest to be detected. Extensive experiments on both toy and real-world datasets validate the effectiveness of our method.

Vision-and-Language Navigation (VLN) tasks mainly evaluate agents based on one-time execution of individual instructions across multiple environments, aiming to develop agents capable of functioning in any environment in a zero-shot manner. However, real-world navigation robots often operate in persistent environments with relatively consistent physical layouts, visual observations, and language styles from instructors. Such a gap in the task setting presents an opportunity to improve VLN agents by incorporating continuous adaptation to specific environments. To better reflect these real-world conditions, we introduce GSA-VLN (General Scene Adaptation for VLN), a novel task requiring agents to execute navigation instructions within a specific scene and simultaneously adapt to it for improved performance over time. To evaluate the proposed task, one has to address two challenges in existing VLN datasets: the lack of out-of-distribution (OOD) data, and the limited number and style diversity of instructions for each scene. Therefore, we propose a new dataset, GSA-R2R, which significantly expands the diversity and quantity of environments and instructions for the Room-to-Room (R2R) dataset to evaluate agent adaptability in both ID and OOD contexts. Furthermore, we design a three-stage instruction orchestration pipeline that leverages large language models (LLMs) to refine speaker-generated instructions and apply role-playing techniques to rephrase instructions into different speaking styles. This is motivated by the observation that each individual user often has consistent signatures or preferences in their instructions, taking the use case of home robotic assistants as an example. We conducted extensive experiments on GSA-R2R to thoroughly evaluate our dataset and benchmark various methods, revealing key factors enabling agents to adapt to specific environments. Based on our findings, we propose a novel method, Graph-Retained DUET (GR-DUET), which incorporates memory-based navigation graphs with an environment-specific training strategy, achieving state-of-the-art results on all GSA-R2R splits.

We introduce EURUS, a suite of large language models (LLMs) optimized for reasoning. Finetuned from Mistral-7B, Llama-3-8B, and Mixtral-8x22B, EURUS models achieve state-of-the-art results among open-source models on a diverse set of benchmarks covering mathematics, code generation, and logical reasoning problems. Notably, EURUX-8X22B outperforms GPT-3.5 Turbo in reasoning through a comprehensive benchmarking across 12 test sets covering five tasks. The strong performance of EURUS can be primarily attributed to ULTRAINTERACT, our newly-curated large-scale, high-quality training data dataset specifically designed for complex reasoning tasks. ULTRAINTERACT can be used in both supervised fine-tuning, preference learning, and reward modeling. It pairs each instruction with a preference tree consisting of (1) reasoning chains with diverse planning strategies in a unified format, (2) multi-turn interaction trajectories with the environment and the critique, and (3) pairwise positive and negative responses to facilitate preference learning. ULTRAINTERACT allows us to conduct an in-depth exploration of preference learning for reasoning tasks. Our investigation reveals that some well-established preference learning algorithms may be less suitable for reasoning tasks compared to their effectiveness in general conversations. The hypothesis is that in reasoning tasks, the space of correct answers is much smaller than that of incorrect ones, so it is necessary to explicitly increase the reward of chosen data. Therefore, in addition to increasing the reward margin as many preference learning algorithms do, the absolute values of positive responses’ rewards should be positive and may serve as a proxy for performance. Inspired by this, we derive a novel reward modeling objective and empirically that it leads to a stable reward modeling curve and better performance. Together with ULTRAINTERACT, we obtain a strong reward model.

Recent studies have introduced various approaches for prompt-tuning black-box vision-language models, referred to as black-box prompt-tuning (BBPT). While BBPT has demonstrated considerable potential, it is often found that many existing methods require an excessive number of queries (i.e., function evaluations), which poses a significant challenge in real-world scenarios where the number of allowed queries is limited. To tackle this issue, we propose Zeroth-order Intrinsic-dimensional Prompt-tuning (ZIP), a novel approach that enables efficient and robust prompt optimization in a purely black-box setting. The key idea of ZIP is to reduce the problem dimensionality and the variance of zeroth-order gradient estimates, such that the training is done fast with far less queries. We achieve this by re-parameterizing prompts in low-rank representations and designing intrinsic-dimensional clipping of estimated gradients. We evaluate ZIP on 13+ vision-language tasks in standard benchmarks and show that it achieves an average improvement of approximately 6% in few-shot accuracy and 48% in query efficiency compared to the best-performing alternative BBPT methods, establishing a new state of the art. Our ablation analysis further shows that the proposed clipping mechanism is robust and nearly optimal, without the need to manually select the clipping threshold, matching the result of expensive hyperparameter search.

We propose Few-Class Arena (FCA), as a unified benchmark with focus on testing efficient image classification models for few classes. A wide variety of benchmark datasets with many classes (80-1000) have been created to assist Computer Vision architectural evolution. An increasing number of vision models are evaluated with these many-class datasets. However, real-world applications often involve substantially fewer classes of interest (2-10). This gap between many and few classes makes it difficult to predict performance of the few-class applications using models trained on the available many-class datasets. To date, little has been offered to evaluate models in this Few-Class Regime. We conduct a systematic evaluation of the ResNet family trained on ImageNet subsets from 2 to 1000 classes, and test a wide spectrum of Convolutional Neural Networks and Transformer architectures over ten datasets by using our newly proposed FCA tool. Furthermore, to aid an up-front assessment of dataset difficulty and a more efficient selection of models, we incorporate a difficulty measure as a function of class similarity. FCA offers a new tool for efficient machine learning in the Few-Class Regime, with goals ranging from a new efficient class similarity proposal, to lightweight model architecture design, to a new scaling law. FCA is user-friendly and can be easily extended to new models and datasets, facilitating future research work. Our benchmark is available at https://github.com/bryanbocao/fca.

Existing benchmarks like NLGraph and GraphQA evaluate LLMs on graphs by focusing mainly on pairwise relationships, overlooking the high-order correlations found in real-world data. Hypergraphs, which can model complex beyond-pairwise relationships, offer a more robust framework but are still underexplored in the context of LLMs. To address this gap, we introduce LLM4Hypergraph, the first comprehensive benchmark comprising 21,500 problems across eight low-order, five high-order, and two isomorphism tasks, utilizing both synthetic and real-world hypergraphs from citation networks and protein structures. We evaluate six prominent LLMs, including GPT-4o, demonstrating our benchmark’s effectiveness in identifying model strengths and weaknesses. Our specialized prompt- ing framework incorporates seven hypergraph languages and introduces two novel techniques, Hyper-BAG and Hyper-COT, which enhance high-order reasoning and achieve an average 4% (up to 9%) performance improvement on structure classification tasks. This work establishes a foundational testbed for integrating hypergraph computational capabilities into LLMs, advancing their comprehension.

Do large language models (LLMs) have theory of mind? A plethora of papers and benchmarks have been introduced to evaluate if current models have been able to develop this key ability of social intelligence. However, all rely on limited datasets with simple patterns that can potentially lead to problematic blind spots in evaluation and an overestimation of model capabilities. We introduce ExploreToM, the first framework to allow large-scale generation of diverse and challenging theory of mind data for robust training and evaluation. Our approach leverages an A* search over a custom domain-specific language to produce complex story structures and novel, diverse, yet plausible scenarios to stress test the limits of LLMs. Our evaluation reveals that state-of-the-art LLMs, such as Llama-3.1-70B and GPT-4o, show accuracies as low as 0% and 9% on ExploreToM-generated data, highlighting the need for more robust theory of mind evaluation. As our generations are a conceptual superset of prior work, fine-tuning on our data yields a 27-point accuracy improvement on the classic ToMi benchmark (Le et al., 2019). ExploreToM also enables uncovering underlying skills and factors missing for models to show theory of mind, such as unreliable state tracking or data imbalances, which may contribute to models' poor performance on benchmarks.

The capabilities and limitations of Large Language Models (LLMs) have been sketched out in great detail in recent years, providing an intriguing yet conflicting picture. On the one hand, LLMs demonstrate a general ability to solve problems. On the other hand, they show surprising reasoning gaps when compared to humans, casting doubt on the robustness of their generalisation strategies. The sheer volume of data used in the design of LLMs has precluded us from applying the method traditionally used to measure generalisation: train-test set separation. To overcome this, we study what kind of generalisation strategies LLMs employ when performing reasoning tasks by investigating the pretraining data they rely on. For two models of different sizes (7B and 35B) and 2.5B of their pretraining tokens, we identify what documents influence the model outputs for three simple mathematical reasoning tasks and contrast this to the data that are influential for answering factual questions. We find that, while the models rely on mostly distinct sets of data for each factual question, a document often has a similar influence across different reasoning questions within the same task, indicating the presence of procedural knowledge. We further find that the answers to factual questions often show up in the most influential data. However, for reasoning questions the answers usually do not show up as highly influential, nor do the answers to the intermediate reasoning steps. When we characterise the top ranked documents for the reasoning questions qualitatively, we confirm that the influential documents often contain procedural knowledge, like demonstrating how to obtain a solution using formulae or code. Our findings indicate that the approach to reasoning the models use is unlike retrieval, and more like a generalisable strategy that synthesises procedural knowledge from documents doing a similar form of reasoning.

Long context compression is a critical research problem due to its significance in reducing the high computational and memory costs associated with LLMs. In this paper, we propose Activation Beacon, a plug-in module for transformer-based LLMs that targets effective, efficient, and flexible compression of long contexts. To achieve this, our method introduces the following technical designs. 1) We directly compress the activations (i.e. keys and values at every layer), rather than leveraging soft prompts to relay information (which constitute a major bottleneck to encapsulate the complex information within long contexts). 2) We tailor the compression workflow, where each fine-grained input unit is progressively compressed, enabling high-quality compression and efficient computation during both training and inference. 3) We train the model through compression-based auto-regression, making full use of plain texts and instructional data to optimize the model's compression performance. 4) During training, we randomly sample a compression ratio at each step, teaching the model to support a wide range of compression configurations. Extensive evaluations are conducted on various long-context tasks whose lengths (e.g., 128K) may far exceed the maximum training length (20K), such as document understanding, few-shot learning, and Needle-in-a-Haystack. Whilst existing methods struggle to handle these challenging tasks, Activation Beacon maintains a comparable performance to the uncompressed baseline across various scenarios, achieving a 2x acceleration in inference time and an 8x reduction of memory costs for KV cache.

Vision-language models such as CLIP have demonstrated strong zero-shot performance, but their considerable size and inefficient inference limit customizable deployment for users. While knowledge distillation is a solution, it still requires the original data, which is not always available due to copyrights and privacy concerns. For many users seeking open-vocabulary customization, Data-Free Knowledge Distillation (DFKD) emerges as a promising direction. Upon rethinking DFKD, we find that existing methods fail on CLIP due to their heavy reliance on BatchNorm layers, which are unexpectedly unusable in CLIP. Based on our findings, we adopt image-text matching to achieve DFKD for CLIP, enabling customization based on arbitrary class texts. This involves (i) inversing a surrogate dataset from CLIP based on text prompts; and (ii) distilling a student model from CLIP using the surrogate dataset. Specifically, we introduce style dictionary diversification to enhance the diversity of synthetic images. To prevent uncontrollable semantics introduced by diversification, we propose a class consistency maintaining strategy to ensure the consistency of synthetic images. Based on synthetic images with various styles, we further propose meta knowledge distillation to train the student model with good generalization ability. Moreover, we introduce a simple yet effective method to enable customization based on few example images. Comprehensive experiments showcase the superiority of our approach across twelve customized tasks, achieving a 9.33\% improvement compared to existing DFKD methods.

Audio is essential for multimodal video understanding. On the one hand, video inherently contains audio, which supplies complementary information to vision. Besides, video large language models (Video-LLMs) can encounter many audio-centric settings. However, existing Video-LLMs and Audio-Visual Large Language Models (AV-LLMs) exhibit deficiencies in exploiting audio information, leading to weak understanding and hallucinations. To solve the issues, we delve into the model architecture and dataset. (1) From the architectural perspective, we propose a fine-grained AV-LLM, namely Dolphin. The concurrent alignment of audio and visual modalities in both temporal and spatial dimensions ensures a comprehensive and accurate understanding of videos. Specifically, we devise an audio-visual multi-scale adapter for multi-scale information aggregation, which achieves spatial alignment. For temporal alignment, we propose audio-visual interleaved merging. (2) From the dataset perspective, we curate an audio-visual caption \& instruction-tuning dataset, called AVU. It comprises 5.2 million diverse, open-ended data tuples (video, audio, question, answer) and introduces a novel data partitioning strategy. Extensive experiments show our model not only achieves remarkable performance in audio-visual understanding, but also mitigates potential hallucinations.

Surrogate models are used to predict the behavior of complex energy systems that are too expensive to simulate with traditional numerical methods. Our work introduces the use of language descriptions, which we call "system captions" or SysCaps, to interface with such surrogates. We argue that interacting with surrogates through text, particularly natural language, makes these models more accessible for both experts and non-experts. We introduce a lightweight multimodal text and timeseries regression model and a training pipeline that uses large language models (LLMs) to synthesize high-quality captions from simulation metadata. Our experiments on two real-world simulators of buildings and wind farms show that our SysCaps-augmented surrogates have better accuracy on held-out systems than traditional methods while enjoying new generalization abilities, such as handling semantically related descriptions of the same test system. Additional experiments also highlight the potential of SysCaps to unlock language-driven design space exploration and to regularize training through prompt augmentation.

How do multi-modal generative models describe images of recent historical events and figures, whose legacies may be nuanced, multifaceted, or contested? This task necessitates not only accurate visual recognition, but also socio-cultural knowledge and cross-modal reasoning. To address this evaluation challenge, we introduce Century -- a novel dataset of sensitive historical images. This dataset consists of 1,500 images from recent history, created through an automated method combining knowledge graphs and language models with quality and diversity criteria created from the practices of museums and digital archives. We demonstrate through automated and human evaluation that this method produces a set of images that depict events and figures that are diverse across topics and represents all regions of the world. We additionally propose an evaluation framework for evaluating the historical contextualisation capabilities along dimensions of accuracy, thoroughness, and objectivity. We demonstrate this approach by using Century to evaluate four foundation models, scoring performance using both automated and human evaluation. We find that historical contextualisation of sensitive images poses a significant challenge for modern multi-modal foundation models, and offer practical recommendations for how developers can use Century to evaluate improvements to models and applications.

Character image animation, which generates high-quality videos from a reference image and target pose sequence, has seen significant progress in recent years. However, most existing methods only apply to human figures, which usually do not generalize well on anthropomorphic characters commonly used in industries like gaming and entertainment. Our in-depth analysis suggests to attribute this limitation to their insufficient modeling of motion, which is unable to comprehend the movement pattern of the driving video, thus imposing a pose sequence rigidly onto the target character. To this end, this paper proposes $\texttt{Animate-X}$, a universal animation framework based on LDM for various character types (collectively named $\texttt{X}$), including anthropomorphic characters. To enhance motion representation, we introduce the Pose Indicator, which captures comprehensive motion pattern from the driving video through both implicit and explicit manner. The former leverages CLIP visual features of a driving video to extract its gist of motion, like the overall movement pattern and temporal relations among motions, while the latter strengthens the generalization of LDM by simulating possible inputs in advance that may arise during inference. Moreover, we introduce a new Animated Anthropomorphic Benchmark ($\texttt{$A^2$Bench}$) to evaluate the performance of $\texttt{Animate-X}$ on universal and widely applicable animation images. Extensive experiments demonstrate the superiority and effectiveness of $\texttt{Animate-X}$ compared to state-of-the-art methods.

Code has been shown to be effective in enhancing the mathematical reasoning abilities of large language models due to its precision and accuracy. Previous works involving continued mathematical pretraining often include code that utilizes math-related packages, which are primarily designed for fields such as engineering, machine learning, signal processing, or module testing, rather than being directly focused on mathematical reasoning. In this paper, we introduce a novel method for generating mathematical code accompanied with corresponding reasoning steps for continued pretraining. Our approach begins with the construction of a high-quality mathematical continued pretraining dataset by incorporating math-related web data, code using mathematical packages, math textbooks, and synthetic data. Next, we construct reasoning steps by extracting LaTeX expressions, the conditions needed for the expressions, and the results of the expressions from the previously collected dataset. Based on this extracted information, we generate corresponding code to accurately capture the mathematical reasoning process. Appending the generated code to each reasoning step results in data consisting of paired natural language reasoning steps and their corresponding code. Combining this data with the original dataset results in a 19.2B-token high-performing mathematical pretraining corpus, which we name MathCode-Pile. Training several popular base models with this corpus significantly improves their mathematical abilities, leading to the creation of the MathCoder2 family of models. All of our data processing and training code is open-sourced, ensuring full transparency and easy reproducibility of the entire data collection and training pipeline.

While Large Language Models (LLMs) excel at in-context learning (ICL) using just a few demonstrations, their performances are sensitive to demonstration orders. The reasons behind this sensitivity remain poorly understood. In this paper, we investigate the prompt embedding space to bridge the gap between the order sensitivity of ICL with inner workings of decoder-only LLMs, uncovering the clustering property: prompts sharing the first and last demonstrations have closer embeddings, with first-demonstration clustering usually being stronger in practice. We explain this property through extensive theoretical analyses and empirical evidences. Our finding suggests that the positional encoding and the causal attention mask are key contributors to the clustering phenomenon. Leveraging this clustering insight, we introduce Cluster-based Search, a novel method that accelerates the selection and ordering of demonstrations in self-adaptive ICL settings. Our approach substantially decreases the time complexity from factorial to quadratic, saving 92% to nearly 100% execution time while maintaining comparable performance to exhaustive search.

In this work, we investigate whether small language models can determine high-quality subsets of large-scale text datasets that improve the performance of larger language models. While existing work has shown that pruning based on the perplexity of a larger model can yield high-quality data, we investigate whether smaller models can be used for perplexity-based pruning and how pruning is affected by the domain composition of the data being pruned. We demonstrate that for multiple dataset compositions, perplexity-based pruning of pretraining data can significantly improve downstream task performance: pruning based on perplexities computed with a 125 million parameter model improves the average performance on downstream tasks of a 3 billion parameter model by up to 2.04 and achieves up to a 1.45× reduction in pretraining steps to reach commensurate baseline performance. Furthermore, we demonstrate that such perplexity-based data pruning also yields downstream performance gains in the over-trained and data-constrained regimes.

Large Language Models (LLMs) have demonstrated strong reasoning and memorization capabilities via pretraining on massive textual corpora. However, this poses risk of privacy and copyright violations, highlighting the need for efficient machine unlearning methods that remove sensitive data without retraining from scratch. While Gradient Ascent (GA) is commonly used to unlearn by reducing the likelihood of generating unwanted content, it leads to unstable optimization and catastrophic forgetting of retrained knowledge. We find that combining GA with low-rank adaptation results in poor trade-offs between computational cost and generative performance. To address these challenges, we propose Low-rank Knowledge Unlearning (LoKU), a novel framework that enables robust and efficient unlearning for LLMs. First, we introduce Inverted Hinge Loss, which suppresses unwanted tokens while maintaining fluency by boosting the probability of the next most likely token. Second, we develop a data-adaptive initialization for LoRA adapters via low-rank approximation weighted with relative Fisher information, thereby focusing updates on parameters critical for removing targeted knowledge. Experiments on the Training Data Extraction Challenge dataset using GPT-Neo models as well as on the TOFU benchmark with Phi-1.5B and Llama2-7B models demonstrate that our approach effectively removes sensitive information while maintaining reasoning and generative capabilities with minimal impact. Our implementation can be found in https://github.com/csm9493/efficient-llm-unlearning.