ICLR 2025 Papers

We investigate the mechanisms of self-distillation in multi-class classification, particularly in the context of linear probing with fixed feature extractors where traditional feature learning explanations do not apply. Our theoretical analysis reveals that multi-round self-distillation effectively performs label averaging among instances with high feature correlations, governed by the eigenvectors of the Gram matrix derived from input features. This process leads to clustered predictions and improved generalization, mitigating the impact of label noise by reducing the model's reliance on potentially corrupted labels. We establish conditions under which multi-round self-distillation achieves 100\% population accuracy despite label noise. Furthermore, we introduce a novel, efficient single-round self-distillation method using refined partial labels from the teacher's top two softmax outputs, referred to as the PLL student model. This approach replicates the benefits of multi-round distillation in a single round, achieving comparable or superior performance--especially in high-noise scenarios--while significantly reducing computational cost.

Pretraining data has a direct impact on the behaviors and quality of language models (LMs), but we only understand the most basic principles of this relationship. While most work focuses on pretraining data's effect on downstream task behavior, we investigate its relationship to LM representations. Previous work has discovered that, in language models, some concepts are encoded "linearly" in the representations, but what factors cause these representations to form (or not)? We study the connection between pretraining data frequency and models' linear representations of factual relations (e.g., mapping France to Paris in a capital prediction task). We find evidence that the formation of linear representations is strongly connected to pretraining term frequencies; specifically for subject-relation-object fact triplets, both subject-object co-occurrence frequency and in-context learning accuracy for the relation are highly correlated with linear representations. This is the case across all phases of pretraining, i.e., it is not affected by the model's underlying capability. In OLMo-7B and GPT-J (6B), we discover that a linear representation consistently (but not exclusively) forms when the subjects and objects within a relation co-occur at least 1k and 2k times, respectively, regardless of when these occurrences happen during pretraining (and around 4k times for OLMo-1B). Finally, we train a regression model on measurements of linear representation quality in fully-trained LMs that can predict how often a term was seen in pretraining. Our model achieves low error even on inputs from a different model with a different pretraining dataset, providing a new method for estimating properties of the otherwise-unknown training data of closed-data models. We conclude that the strength of linear representations in LMs contains signal about the models' pretraining corpora that may provide new avenues for controlling and improving model behavior: particularly, manipulating the models' training data to meet specific frequency thresholds. We release our code to support future work.

Boosting has garnered significant interest across both machine learning and statistical communities. Traditional boosting algorithms, designed for fully observed random samples, often struggle with real-world problems, particularly with interval-censored data. This type of data is common in survival analysis and time-to-event studies where exact event times are unobserved but fall within known intervals. Effective handling of such data is crucial in fields like medical research, reliability engineering, and social sciences. In this work, we introduce novel nonparametric boosting methods for regression and classification tasks with interval-censored data. Our approaches leverage censoring unbiased transformations to adjust loss functions and impute transformed responses while maintaining model accuracy. Implemented via functional gradient descent, these methods ensure scalability and adaptability. We rigorously establish their theoretical properties, including optimality and mean squared error trade-offs. Our proposed methods not only offer a robust framework for enhancing predictive accuracy in domains where interval-censored data are common but also complement existing work, expanding the applicability of existing boosting techniques. Empirical studies demonstrate robust performance across various finite-sample scenarios, highlighting the practical utility of our approaches.

In this study, we investigate the problem of dynamic multi-product selection and pricing by introducing a novel framework based on a *censored multinomial logit* (C-MNL) choice model. In this model, sellers present a set of products with prices, and buyers filter out products priced above their valuation, purchasing at most one product from the remaining options based on their preferences. The goal is to maximize seller revenue by dynamically adjusting product offerings and prices, while learning both product valuations and buyer preferences through purchase feedback. To achieve this, we propose a Lower Confidence Bound (LCB) pricing strategy. By combining this pricing strategy with either an Upper Confidence Bound (UCB) or Thompson Sampling (TS) product selection approach, our algorithms achieve regret bounds of $\tilde{O}(d^{\frac{3}{2}}\sqrt{T/\kappa})$ and $\tilde{O}(d^{2}\sqrt{T/\kappa})$, respectively. Finally, we demonstrate the performance of our methods through simulations.

Strength estimation and adjustment are crucial in designing human-AI interactions, particularly in games where AI surpasses human players. This paper introduces a novel strength system, including a *strength estimator* (SE) and an SE-based Monte Carlo tree search, denoted as *SE-MCTS*, which predicts strengths from games and offers different playing strengths with human styles. The strength estimator calculates strength scores and predicts ranks from games without direct human interaction. SE-MCTS utilizes the strength scores in a Monte Carlo tree search to adjust playing strength and style. We first conduct experiments in Go, a challenging board game with a wide range of ranks. Our strength estimator significantly achieves over 80% accuracy in predicting ranks by observing 15 games only, whereas the previous method reached 49% accuracy for 100 games. For strength adjustment, SE-MCTS successfully adjusts to designated ranks while achieving a 51.33% accuracy in aligning to human actions, outperforming a previous state-of-the-art, with only 42.56% accuracy. To demonstrate the generality of our strength system, we further apply SE and SE-MCTS to chess and obtain consistent results. These results show a promising approach to strength estimation and adjustment, enhancing human-AI interactions in games. Our code is available at https://rlg.iis.sinica.edu.tw/papers/strength-estimator.

The goal of the AlgoPerf: Training Algorithms competition is to evaluate practical speed-ups in neural network training achieved solely by improving the underlying training algorithms. In the external tuning ruleset, submissions must provide workload-agnostic hyperparameter search spaces, while in the self-tuning ruleset they must be completely hyperparameter-free. In both rulesets, submissions are compared on time-to-result across multiple deep learning workloads, training on fixed hardware. This paper presents the inaugural AlgoPerf competition's results, which drew 18 diverse submissions from 10 teams. Our investigation reveals several key findings: (1) The winning submission in the external tuning ruleset, using Distributed Shampoo, demonstrates the effectiveness of non-diagonal preconditioning over popular methods like Adam, even when compared on wall-clock runtime. (2) The winning submission in the self-tuning ruleset, based on the Schedule Free AdamW algorithm, demonstrates a new level of effectiveness for completely hyperparameter-free training algorithms. (3) The top-scoring submissions were surprisingly robust to workload changes. We also discuss the engineering challenges encountered in ensuring a fair comparison between different training algorithms. These results highlight both the significant progress so far, and the considerable room for further improvements.

In modern deep neural networks, the learning dynamics of individual neurons are often obscure, as the networks are trained via global optimization. Conversely, biological systems build on self-organized, local learning, achieving robustness and efficiency with limited global information. Here, we show how self-organization between individual artificial neurons can be achieved by designing abstract bio-inspired local learning goals. These goals are parameterized using a recent extension of information theory, Partial Information Decomposition (PID), which decomposes the information that a set of information sources holds about an outcome into unique, redundant and synergistic contributions. Our framework enables neurons to locally shape the integration of information from various input classes, i.e., feedforward, feedback, and lateral, by selecting which of the three inputs should contribute uniquely, redundantly or synergistically to the output. This selection is expressed as a weighted sum of PID terms, which, for a given problem, can be directly derived from intuitive reasoning or via numerical optimization, offering a window into understanding task-relevant local information processing. Achieving neuron-level interpretability while enabling strong performance using local learning, our work advances a principled information-theoretic foundation for local learning strategies.

Event-guided imaging has received significant attention due to its potential to revolutionize instant imaging systems. However, the prior methods primarily focus on enhancing RGB images in a post-processing manner, neglecting the challenges of image signal processor (ISP) dealing with event sensor and the benefits events provide for reforming the ISP process. To achieve this, we conduct the first research on event-guided ISP. First, we present a new event-RAW paired dataset, collected with a novel but still confidential sensor that records pixel-level aligned events and RAW images. This dataset includes 3373 RAW images with $2248\times 3264$ resolution and their corresponding events, spanning 24 scenes with 3 exposure modes and 3 lenses. Second, we propose a convential ISP pipeline to generate good RGB frames as reference. This convential ISP pipleline performs basic ISP operations, e.g., demosaicing, white balancing, denoising and color space transforming, with a ColorChecker as reference. Third, we classify the existing learnable ISP methods into 3 classes, and select multiple methods to train and evaluate on our new dataset. Lastly, since there is no prior work for reference, we propose a simple event-guided ISP method and test it on our dataset. We further put forward key technical challenges and future directions in RGB-Event ISP. In summary, to the best of our knowledge, this is the very first research focusing on event-guided ISP, and we hope it will inspire the community.

Large Vision-Language Models (LVLMs) exhibit impressive capabilities in complex visual tasks but are prone to hallucination, especially in open-ended generation tasks. This paper explores why LVLMs tend to hallucinate and how to mitigate it. First, we conduct causal mediation analysis through counterfactual edits on specific modules in LVLMs. Our results disclose that Multi-Head Attention (MHA) modules contribute more to the probability of generating hallucination words than multi-layer perceptron modules. We then identify specific heads that are responsible for hallucination, referred to as hallucination heads. Second, we examine the behavior of hallucination heads. We find that they are concentrated in the middle and deeper layers, displaying a strong attention bias toward text tokens. Further, we show that the attention patterns of certain hallucination heads exhibit greater similarity to the base language model and change slowly during the instruction tuning process. Finally, we propose two simple yet effective methods to mitigate hallucination: one is training-free and can be applied directly during decoding, while the other involves fine-tuning. Both methods are targeted for hallucination heads to reduce their reliance on text tokens. Notably, our methods achieve up to 1.7x reduction in hallucination rate for the LLaVA-v1.5-7B model in COCO captioning task, outperforming existing baselines. Overall, our findings suggest that hallucinations in LVLMs are likely to stem from certain modules, and targeted interventions can effectively mitigate these issues.

Expressivity and generalization are two critical aspects of graph neural networks (GNNs). While significant progress has been made in studying the expressivity of GNNs, much less is known about their generalization capabilities, particularly when dealing with the inherent complexity of graph-structured data. In this work, we address the intricate relationship between expressivity and generalization in GNNs. Theoretical studies conjecture a trade-off between the two: highly expressive models risk overfitting, while those focused on generalization may sacrifice expressivity. However, empirical evidence often contradicts this assumption, with expressive GNNs frequently demonstrating strong generalization. We explore this contradiction by introducing a novel framework that connects GNN generalization to the variance in graph structures they can capture. This leads us to propose a $k$-variance margin-based generalization bound that characterizes the structural properties of graph embeddings in terms of their upper-bounded expressive power. Our analysis does not rely on specific GNN architectures, making it broadly applicable across GNN models. We further uncover a trade-off between intra-class concentration and inter-class separation, both of which are crucial for effective generalization. Through case studies and experiments on real-world datasets, we demonstrate that our theoretical findings align with empirical results, offering a deeper understanding of how expressivity can enhance GNN generalization.

Prototypical parts networks combine the power of deep learning with the explainability of case-based reasoning to make accurate, interpretable decisions. They follow the this looks like that reasoning, representing each prototypical part with patches from training images. However, a single image patch comprises multiple visual features, such as color, shape, and texture, making it difficult for users to identify which feature is important to the model. To reduce this ambiguity, we introduce the Lucid Prototypical Parts Network (LucidPPN), a novel prototypical parts network that separates color prototypes from other visual features. Our method employs two reasoning branches: one for non-color visual features, processing grayscale images, and another focusing solely on color information. This separation allows us to clarify whether the model's decisions are based on color, shape, or texture. Additionally, LucidPPN identifies prototypical parts corresponding to semantic parts of classified objects, making comparisons between data classes more intuitive, e.g., when two bird species might differ primarily in belly color. Our experiments demonstrate that the two branches are complementary and together achieve results comparable to baseline methods. More importantly, LucidPPN generates less ambiguous prototypical parts, enhancing user understanding.

Since the success of GPT, large language models (LLMs) have revolutionized machine learning and have initiated the so-called *LLM prompting* paradigm. In the era of LLMs, people train a single general-purpose LLM and provide the LLM with different *prompts* to perform different tasks. However, such empirical success largely lacks theoretical understanding. Here, we present the first theoretical study on the LLM prompting paradigm to the best of our knowledge. In this work, we show that prompting is in fact Turing-complete: there exists a finite-size Transformer such that for any computable function, there exists a corresponding prompt following which the Transformer computes the function. Furthermore, we show that even though we use only a single finite-size Transformer, it can still achieve nearly the same complexity bounds as that of the class of all unbounded-size Transformers. Overall, our result reveals that prompting can enable a single finite-size Transformer to be efficiently universal, which establishes a theoretical underpinning for prompt engineering in practice.

Algorithmic predictions are increasingly informing societal resource allocations by identifying individuals for targeting. Policymakers often build these systems with the assumption that by gathering more observations on individuals, they can improve predictive accuracy and, consequently, allocation efficiency. An overlooked yet consequential aspect of prediction-driven allocations is that of timing. The planner has to trade off relying on earlier and potentially noisier predictions to intervene before individuals experience undesirable outcomes, or they may wait to gather more observations to make more precise allocations. We examine this tension using a simple mathematical model, where the planner collects observations on individuals to improve predictions over time. We analyze both the ranking induced by these predictions and optimal resource allocation. We show that though individual prediction accuracy improves over time, counter-intuitively, the average ranking loss can worsen. As a result, the planner's ability to improve social welfare can decline. We identify inequality as a driving factor behind this phenomenon. Our findings provide a nuanced perspective and challenge the conventional wisdom that it is preferable to wait for more accurate predictions to ensure the most efficient allocations.

Understanding human locomotion is crucial for AI agents such as robots, particularly in complex indoor home environments. Modeling human trajectories in these spaces requires insight into how individuals maneuver around physical obstacles and manage social navigation dynamics. These dynamics include subtle behaviors influenced by proxemics - the social use of space, such as stepping aside to allow others to pass or choosing longer routes to avoid collisions. Previous research has developed datasets of human motion in indoor scenes, but these are often limited in scale and lack the nuanced social navigation dynamics common in home environments. To address this, we present LocoVR, a dataset of 7000+ two-person trajectories captured in virtual reality from over 130 different indoor home environments. LocoVR provides accurate trajectory and precise spatial information, along with rich examples of socially-motivated movement behaviors. For example, the dataset captures instances of individuals navigating around each other in narrow spaces, adjusting paths to respect personal boundaries in living areas, and coordinating movements in high-traffic zones like entryways and kitchens. Our evaluation shows that LocoVR significantly enhances model performance in three practical indoor tasks utilizing human trajectories, and demonstrates predicting socially-aware navigation patterns in home environments.

The influence of contextual input on the behavior of large language models (LLMs) has prompted the development of context attribution methods that aim to quantify each context span's effect on an LLM's generations. The leave-one-out (LOO) error, which measures the change in the likelihood of the LLM's response when a given span of the context is removed, provides a principled way to perform context attribution, but can be prohibitively expensive to compute for large models. In this work, we introduce AttriBoT, a series of novel techniques for efficiently computing an approximation of the LOO error for context attribution. Specifically, AttriBoT uses cached activations to avoid redundant operations, performs hierarchical attribution to reduce computation, and emulates the behavior of large target models with smaller proxy models. Taken together, AttriBoT can provide a 300x speedup while remaining more faithful to a target model's LOO error than prior context attribution methods. This stark increase in performance makes computing context attributions for a given response $30\times$ faster than generating the response itself, empowering real-world applications that require computing attributions at scale. We release a user-friendly and efficient implementation of AttriBoT to enable efficient LLM interpretability as well as encourage future development of efficient context attribution methods.

A common goal of mechanistic interpretability is to decompose the activations of neural networks into features: interpretable properties of the input computed by the model. Sparse autoencoders (SAEs) are a popular method for finding these features in LLMs, and it has been postulated that they can be used to find a canonical set of units: a unique and complete list of atomic features. We cast doubt on this belief using two novel techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic. SAE stitching involves inserting or swapping latents from a larger SAE into a smaller one. Latents from the larger SAE can be divided into two categories: novel latents, which improve performance when added to the smaller SAE, indicating they capture novel information, and reconstruction latents, which can replace corresponding latents in the smaller SAE that have similar behavior. The existence of novel features indicates incompleteness of smaller SAEs. Using meta-SAEs - SAEs trained on the decoder matrix of another SAE - we find that latents in SAEs often decompose into combinations of latents from a smaller SAE, showing that larger SAE latents are not atomic. The resulting decompositions are often interpretable; e.g. a latent representing "Einstein" decomposes into "scientist", "Germany", and "famous person". To train meta-SAEs we introduce BatchTopK SAEs, an improved variant of the popular TopK SAE method, that only enforces a fixed average sparsity. Even if SAEs do not find canonical units of analysis, they may still be useful tools. We suggest that future research should either pursue different approaches for identifying such units, or pragmatically choose the SAE size suited to their task. We provide an interactive dashboard to explore meta-SAEs: https://metasaes.streamlit.app/

Aligning large language models (LLMs) with human preferences has gained significant attention, with Proximal Policy Optimization (PPO) as a standard yet computationally expensive method and Direct Preference Optimization (DPO) as a more efficient alternative. While DPO offers simplicity, it remains underutilized in state-of-the-art LLMs, suggesting potential limitations. In this work, we revisit DPO, analyzing its theoretical foundations and empirical performance to bridge this gap. We identify three key properties—termed \textbf{3D}-properties—that emerge from DPO’s learning process: \textbf{D}rastic drop in rejected response likelihood, \textbf{D}egradation into response suppression, and \textbf{D}ispersion effect on unseen responses. We show that these issues arise from DPO’s optimization dynamics, where the interaction between chosen and rejected response gradients leads to instability. Our findings are supported by experiments on both a controlled toy model and real-world LLM tasks, including mathematical problem-solving and instruction following. To address these challenges, we propose simple regularization techniques that improve training stability and performance. Additionally, we examine how preference data distribution impacts DPO’s effectiveness, offering insights into how alignment models handle out-of-domain (OOD) data. Our work connects these observations to broader research and provides a theoretical explanation for DPO’s limitations. We hope these insights will guide future advancements in reward-model-free preference learning, bringing it closer to reward-model-based approaches.

As diffusion models are deployed in real-world settings and their performance driven by training data, appraising the contribution of data contributors is crucial to creating incentives for sharing quality data and to implementing policies for data compensation. Depending on the use case, model performance corresponds to various global properties of the distribution learned by a diffusion model (e.g., overall aesthetic quality). Hence, here we address the problem of attributing global properties of diffusion models to data contributors. The Shapley value provides a principled approach to valuation by uniquely satisfying game-theoretic axioms of fairness. However, estimating Shapley values for diffusion models is computationally impractical because it requires retraining and rerunning inference on many subsets of data contributors. We introduce a method to efficiently retrain and rerun inference for Shapley value estimation, by leveraging model pruning and fine-tuning. We evaluate the utility of our method with three use cases: (i) image quality for a DDPM trained on a CIFAR dataset, (ii) demographic diversity for an LDM trained on CelebA-HQ, and (iii) aesthetic quality for a Stable Diffusion model LoRA-finetuned on Post-Impressionist artworks. Our results empirically demonstrate that our framework can identify important data contributors across global properties, outperforming existing attribution methods for diffusion models.

We investigate the internal representations of vision-language models (VLMs) to address hallucinations, a persistent challenge despite advances in model size and training. We project VLMs’ internal image representations to their language vocabulary and observe more confident output probabilities on real objects than hallucinated objects. We additionally use these output probabilities to spatially localize real objects. Building on this approach, we introduce a knowledge erasure algorithm that removes hallucinations by linearly orthogonalizing image features with respect to hallucinated object features. We show that targeted edits to a model’s latent representations can reduce hallucinations by up to 25.7% on the COCO2014 dataset while preserving performance. Our findings demonstrate how a deeper understanding of VLMs’ latent representations can enhance reliability and enable novel capabilities, such as zero-shot segmentation.