ICLR 2025 Papers

In this work, we investigate how a model's tendency to broadly integrate its parametric knowledge evolves throughout pretraining, and how this behavior affects overall performance, particularly in terms of knowledge acquisition and forgetting. We introduce the concept of knowledge entropy, which quantifies the range of memory sources the model engages with; high knowledge entropy indicates that the model utilizes a wide range of memory sources, while low knowledge entropy suggests reliance on specific sources with greater certainty. Our analysis reveals a consistent decline in knowledge entropy as pretraining advances. We also find that the decline is closely associated with a reduction in the model's ability to acquire and retain knowledge, leading us to conclude that diminishing knowledge entropy (smaller number of active memory sources) impairs the model's knowledge acquisition and retention capabilities. We find further support for this by demonstrating that increasing the activity of inactive memory sources enhances the model's capacity for knowledge acquisition and retention.

While large language models (LLMs) excel on generation tasks, their decoder-only architecture often limits their potential as embedding models if no further representation finetuning is applied. Does this contradict their claim of generalists? To answer the question, we take a closer look at Mixture-of-Experts (MoE) LLMs. Our study shows that the expert routers in MoE LLMs can serve as an off-the-shelf embedding model with promising performance on a diverse class of embedding-focused tasks, without requiring any finetuning. Moreover, our extensive analysis shows that the MoE routing weights (RW) is complementary to the hidden state (HS) of LLMs, a widely-used embedding. Compared to HS, we find that RW is more robust to the choice of prompts and focuses on high-level semantics. Motivated by the analysis, we propose MoEE combining RW and HS, which achieves better performance than using either separately. Our exploration of their combination and prompting strategy shed several novel insights, e.g., a weighted sum of RW and HS similarities outperforms the similarity on their concatenation. Our experiments are conducted on 6 embedding tasks with 20 datasets from the Massive Text Embedding Benchmark (MTEB). The results demonstrate the significant improvement brought by MoEE to LLM-based embedding without further finetuning.

The rise of large language models (LLMs) has created a significant disparity: industrial research labs with their computational resources, expert teams, and advanced infrastructures, can effectively fine-tune LLMs, while individual developers and small organizations face barriers due to limited resources to effectively explore the experiment space. In this paper, we aim to bridge this gap by presenting a comprehensive study on supervised fine-tuning of LLMs using instruction-tuning datasets spanning diverse knowledge domains and skills. We focus on small-sized LLMs (3B to 7B parameters) for their cost-efficiency and accessibility. We explore various training configurations and strategies across four open-source pre-trained models. We provide detailed documentation of these configurations, revealing findings that challenge several common training practices, including hyperparameter recommendations from TULU and phased training recommended by Orca. The code used for the experiments can be found here: https://github.com/instructlab/training. Key insights from our work include: (i) larger batch sizes paired with lower learning rates lead to improved model performance on benchmarks such as MMLU, MTBench, and Open LLM Leaderboard; (ii) early-stage training dynamics, such as lower gradient norms and higher loss values, are strong indicators of better final model performance, allowing for early termination of sub-optimal runs and significant computational savings; (iii) through a thorough exploration of hyperparameters like warmup steps and learning rate schedules, we provide guidance for practitioners and find that certain simplifications do not compromise performance; and (iv) we observe no significant difference in performance between phased (sequentially training on data divided into phases) and stacked (training on the entire dataset at once) strategies, but stacked training is simpler and more sample efficient. With these findings holding robustly across datasets as well as model families and sizes, we hope this study serves as a guide for practitioners fine-tuning small LLMs and promotes a more inclusive research environment for LLM development.

Aligning with personalized preferences, which vary significantly across cultural, educational, and political differences, poses a significant challenge due to the computational costs and data demands of traditional alignment methods. In response, this paper presents Personalized Alignment at Decoding-time (PAD), a novel framework designed to align LLM outputs with diverse personalized preferences during the inference phase, eliminating the need for additional training. By introducing a unique personalized reward modeling strategy, this framework decouples the text generation process from personalized preferences, facilitating the generation of generalizable token-level personalized rewards. The PAD algorithm leverages these rewards to guide the decoding process, dynamically tailoring the base model’s predictions to personalized preferences. Extensive experimental results demonstrate that PAD not only outperforms existing training-based alignment methods in terms of aligning with diverse preferences but also shows significant generalizability to preferences unseen during training and scalability across different base models. This work advances the capability of LLMs to meet user needs in real-time applications, presenting a substantial step forward in personalized LLM alignment.

Deep learning for time series forecasting has seen significant advancements over the past decades. However, despite the success of large-scale pre-training in language and vision domains, pre-trained time series models remain limited in scale and operate at a high cost, hindering the development of larger capable forecasting models in real-world applications. In response, we introduce Time-MoE, a scalable and unified architecture designed to pre-train larger, more capable forecasting foundation models while reducing inference costs. By leveraging a sparse mixture-of-experts (MoE) design, Time-MoE enhances computational efficiency by activating only a subset of networks for each prediction, reducing computational load while maintaining high model capacity. This allows Time-MoE to scale effectively without a corresponding increase in inference costs. Time-MoE comprises a family of decoder-only transformer models that operate in an auto-regressive manner and support flexible forecasting horizons with varying input context lengths. We pre-trained these models on our newly introduced large-scale data Time-300B, which spans over 9 domains and encompassing over 300 billion time points. For the first time, we scaled a time series foundation model up to 2.4 billion parameters, achieving significantly improved forecasting precision. Our results validate the applicability of scaling laws for training tokens and model size in the context of time series forecasting. Compared to dense models with the same number of activated parameters or equivalent computation budgets, our models consistently outperform them by large margin. These advancements position Time-MoE as a state-of-the-art solution for tackling real-world time series forecasting challenges with superior capability, efficiency, and flexibility. Code is available at https://github.com/Time-MoE/Time-MoE

Motivated by deep learning regimes with multiple interacting yet distinct model components, we introduce learning diagrams, graphical depictions of training setups that capture parameterized learning as data rather than code. A learning diagram compiles to a unique loss function on which component models are trained. The result of training on this loss is a collection of models whose predictions ``agree" with one another. We show that a number of popular learning setups such as few-shot multi-task learning, knowledge distillation, and multi-modal learning can be depicted as learning diagrams. We further implement learning diagrams in a library that allows users to build diagrams of PyTorch and Flux.jl models. By implementing some classic machine learning use cases, we demonstrate how learning diagrams allow practitioners to build complicated models as compositions of smaller components, identify relationships between workflows, and manipulate models during or after training. Leveraging a category theoretic framework, we introduce a rigorous semantics for learning diagrams that puts such operations on a firm mathematical foundation.

This work investigates the selection of high-quality pre-training data from massive corpora to enhance LMs' capabilities for downstream usage. We formulate data selection as a generalized Optimal Control problem, which can be solved theoretically by Pontryagin's Maximum Principle (PMP), yielding a set of necessary conditions that characterize the relationship between optimal data selection and LM training dynamics. Based on these theoretical results, we introduce **P**MP-based **D**ata **S**election (**PDS**), a framework that approximates optimal data selection by solving the PMP conditions. In our experiments, we adopt PDS to select data from CommmonCrawl and show that the PDS-selected corpus accelerates the learning of LMs and constantly boosts their performance on a wide range of downstream tasks across various model sizes. Moreover, the benefits of PDS extend to ~400B models trained on ~10T tokens, as evidenced by the extrapolation of the test loss curves according to the Scaling Laws. PDS also improves data utilization when the pre-training data is limited, by reducing the data demand by 1.8 times, which helps mitigate the quick exhaustion of available web-crawled corpora. Our code, model, and data can be found at https://github.com/microsoft/LMOps/tree/main/data_selection.

Self-adaptive large language models (LLMs) aim to solve the challenges posed by traditional fine-tuning methods, which are often computationally intensive and static in their ability to handle diverse tasks. We introduce Transformer-Squared, a novel self-adaptation framework that adapts LLMs for unseen tasks in real-time by selectively adjusting only the singular components of their weight matrices. During inference, Transformer-Squared employs a two-pass mechanism: first, a dispatch system identifies the task properties, and then task-specific 'expert' vectors, trained using reinforcement learning, are dynamically mixed to obtain targeted behavior for the incoming prompt. Our method consistently outperforms ubiquitous approaches such as LoRA, with fewer parameters and greater efficiency. Furthermore, Transformer-Squared demonstrates versatility across different LLM architectures and modalities, including vision-language tasks. Transformer-Squared represents a significant leap forward, offering a scalable, efficient solution for enhancing the adaptability and task-specific performance of LLMs, paving the way for truly dynamic, self-organizing AI systems.

It is generally perceived that Dynamic Sparse Training opens the door to a new era of scalability and efficiency for artificial neural networks at, perhaps, some costs in accuracy performance for the classification task. At the same time, Dense Training is widely accepted as being the "de facto" approach to train artificial neural networks if one would like to maximize their robustness against image corruption. In this paper, we question this general practice. Consequently, \textit{we claim that}, contrary to what is commonly thought, the Dynamic Sparse Training methods can consistently outperform Dense Training in terms of robustness accuracy, particularly if the efficiency aspect is not considered as a main objective (i.e., sparsity levels between 10\% and up to 50\%), without adding (or even reducing) resource cost. We validate our claim on two types of data, images and videos, using several traditional and modern deep learning architectures for computer vision and three widely studied Dynamic Sparse Training algorithms. Our findings reveal a new yet-unknown benefit of Dynamic Sparse Training and open new possibilities in improving deep learning robustness beyond the current state of the art.

Large language models show impressive abilities in memorizing world knowledge, which leads to concerns regarding memorization of private information, toxic or sensitive knowledge, and copyrighted content. We introduce the problem of Large Scale Knowledge Washing, focusing on unlearning an extensive amount of factual knowledge. Previous unlearning methods usually define the reverse loss and update the model via backpropagation, which may affect the model's fluency and reasoning ability or even destroy the model due to extensive training with the reverse loss. Existing works introduce additional data from downstream tasks to prevent the model from losing capabilities, which requires downstream task awareness. Controlling the tradeoff of unlearning existing knowledge while maintaining existing capabilities is also challenging. To this end, we propose LaW (Large Scale Washing), where we update the MLP layers in decoder-only large language models to perform knowledge washing, as inspired by model editing methods. We derive a new objective with the knowledge to be unlearned to update the weights of certain MLP layers. Experimental results demonstrate the effectiveness of LaW in forgetting target knowledge while maximally maintaining reasoning ability. The code will be open-sourced.

Activation sparsity can enable practical inference speedups in large language models (LLMs) by reducing the compute and memory-movement required for matrix multiplications during the forward pass. However, existing methods face limitations that inhibit widespread adoption. Some approaches are tailored towards older models with ReLU-based sparsity, while others require extensive continued pre-training on up to hundreds of billions of tokens. This paper describes TEAL (**T**raining-Fre**e** **A**ctivation Sparsity in **L**LMs), a simple training-free method that applies magnitude-based activation sparsity to hidden states throughout the entire model. TEAL achieves 40-50\% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families, with sizes varying from 7B to 70B. We improve existing sparse kernels and demonstrate wall-clock decoding speed-ups of up to 1.53× and 1.8× at 40\% and 50\% model-wide sparsity. TEAL is compatible with weight quantization, enabling further efficiency gains.

Self-supervised learning has driven significant progress in learning from single-subject, _iconic_ images. However, there are still unanswered questions about the use of minimally-curated, naturalistic video data, which contain _dense_ scenes with many independent objects, imbalanced class distributions, and varying object sizes. In this paper, we propose PooDLe, a self-supervised learning method that combines an invariance-based objective on pooled representations with a dense SSL objective that enforces equivariance to optical flow warping. Our results show that a unified objective applied at multiple feature scales is essential for learning effective image representations from naturalistic videos. We validate our method with experiments on the BDD100K driving video dataset and the Walking Tours first-person video dataset, demonstrating its ability to capture spatial understanding from a dense objective and semantic understanding via a pooled representation objective.

The prohibitive training costs of Large Language Models (LLMs) have emerged as a significant bottleneck in the development of next-generation LLMs. In this paper, we show that it is possible to significantly reduce the training costs of LLMs without sacrificing their performance. Specifically, we introduce patch-level training for LLMs, in which multiple tokens are aggregated into a unit of higher information density, referred to as a `patch', to serve as the fundamental text unit for training LLMs. During patch-level training, we feed the language model shorter sequences of patches and train it to predict the next patch, thereby processing the majority of the training data at a significantly reduced cost. Following this, the model continues token-level training on the remaining training data to align with the inference mode. Experiments on a diverse range of models (370M-2.7B parameters) demonstrate that patch-level training can reduce the overall training costs to 0.5$\times$, without compromising the model performance compared to token-level training. Source code: \url{https://github.com/shaochenze/PatchTrain}.

Despite the remarkable success of large language models (LLMs) on traditional natural language processing tasks, their planning ability remains a critical bottleneck in tackling complex multi-step reasoning tasks. Existing approaches mainly rely on prompting or task-specific fine-tuning, often suffering from weak robustness and cross-task generalization. To address the limitation, we introduce CodePlan, a scalable paradigm that empowers LLMs to generate and follow code-form plans---pseudocode that outlines high-level, structured reasoning processes. By leveraging the structured and versatile nature of code, CodePlan effectively captures the rich semantics and control flows inherent to sophisticated reasoning. Importantly, CodePlan allows the automatic extraction of code-form plans from massive, wide-ranging text corpora without the need for curated, task-specific datasets. This enables it to scale up efficiently and improve reasoning capabilities across diverse scenarios. To train CodePlan, we construct a large-scale dataset of 2M examples that integrate code-form plans with standard prompt-response pairs from existing corpora. With minimal computation overhead during both training and inference, CodePlan achieves a 25.1\% relative improvement compared with directly generating responses, averaged across 13 challenging multi-step reasoning benchmarks, spanning mathematical reasoning, symbolic reasoning, instruction-following, multi-hop QA, and decision-making tasks. Further analysis reveals CodePlan's increasing performance gains on more complex reasoning tasks, as well as significant data efficiency thanks to its generalization ability.

Tracking dense 3D motion from monocular videos remains challenging, particularly when aiming for pixel-level precision over long sequences. We introduce DELTA, a novel method that efficiently tracks every pixel in 3D space, enabling accurate motion estimation across entire videos. Our approach leverages a joint global-local attention mechanism for reduced-resolution tracking, followed by a transformer-based upsampler to achieve high-resolution predictions. Unlike existing methods, which are limited by computational inefficiency or sparse tracking, DELTA delivers dense 3D tracking at scale, running over 8x faster than previous methods while achieving state-of-the-art accuracy. Furthermore, we explore the impact of depth representation on tracking performance and identify log-depth as the optimal choice. Extensive experiments demonstrate the superiority of DELTA on multiple benchmarks, achieving new state-of-the-art results in both 2D and 3D dense tracking tasks. Our method provides a robust solution for applications requiring fine-grained, long-term motion tracking in 3D space.

Masked image modeling (MIM) has emerged as a promising approach for pre-training Vision Transformers (ViTs). MIMs predict masked tokens token-wise to recover target signals that are tokenized from images or generated by pre-trained models like vision-language models. While using tokenizers or pre-trained models is viable, they often offer spatially inconsistent supervision even for neighboring tokens, hindering models from learning discriminative representations. Our pilot study identifies spatial inconsistency in supervisory signals and suggests that addressing it can improve representation learning. Building upon this insight, we introduce Dynamic Token Morphing (DTM), a novel method that dynamically aggregates tokens while preserving context to generate contextualized targets, thereby likely reducing spatial inconsistency. DTM is compatible with various SSL frameworks; we showcase significantly improved MIM results, barely introducing extra training costs. Our method facilitates MIM training by using more spatially consistent targets, resulting in improved training trends as evidenced by lower losses. Experiments on ImageNet-1K and ADE20K demonstrate DTM's superiority, which surpasses complex state-of-the-art MIM methods. Furthermore, the evaluation of transfer learning on downstream tasks like iNaturalist, along with extensive empirical studies, supports DTM's effectiveness.

Measuring biodiversity is crucial for understanding ecosystem health. While prior works have developed machine learning models for taxonomic classification of photographic images and DNA separately, in this work, we introduce a multi-modal approach combining both, using CLIP-style contrastive learning to align images, barcode DNA, and text-based representations of taxonomic labels in a unified embedding space. This allows for accurate classification of both known and unknown insect species without task-specific fine-tuning, leveraging contrastive learning for the first time to fuse DNA and image data. Our method surpasses previous single-modality approaches in accuracy by over 8% on zero-shot learning tasks, showcasing its effectiveness in biodiversity studies.

Image retrieval is crucial in robotics and computer vision, with downstream applications in robot place recognition and vision-based product recommendations. Modern retrieval systems face two key challenges: scalability and efficiency. State-of-the-art image retrieval systems train specific neural networks for each dataset, an approach that lacks scalability. Furthermore, since retrieval speed is directly proportional to embedding size, existing systems that use large embeddings lack efficiency. To tackle scalability, recent works propose using off-the-shelf foundation models. However, these models, though applicable across datasets, fall short in achieving performance comparable to that of dataset-specific models. Our key observation is that, while foundation models capture necessary subtleties for effective retrieval, the underlying distribution of their embedding space can negatively impact cosine similarity searches. We introduce Autoencoders with Strong Variance Constraints (AE-SVC), which, when used for projection, significantly improves the performance of foundation models. We provide an in-depth theoretical analysis of AE-SVC. Addressing efficiency, we introduce Single-Shot Similarity Space Distillation ((SS)2D), a novel approach to learn embeddings with adaptive sizes that offers a better trade-off between size and performance. We conducted extensive experiments on four retrieval datasets, including Stan- ford Online Products (SoP) and Pittsburgh30k, using four different off-the-shelf foundation models, including DinoV2 and CLIP. AE-SVC demonstrates up to a 16% improvement in retrieval performance, while (SS)2D shows a further 10% improvement for smaller embedding sizes.

Large language models (LLMs) must often respond to highly ambiguous user requests. In such cases, the LLM's best response may be to ask a clarifying question to elicit more information. Existing LLMs often respond by presupposing a single interpretation of such ambiguous requests, frustrating users who intended a different interpretation. We speculate this is caused by current preference data labeling practice, where LLM responses are evaluated only on their prior contexts. To address this, we assign preference labels by simulating their expected outcomes in future turns. This allows LLMs to learn to ask clarifying questions when it can generate responses that are tailored to each user interpretation in future turns. On open-domain QA datasets with multiple annotations, we evaluate systems based on their ability to ask clarifying questions to recover each user's interpretation and expected answer. We compare systems trained using our proposed preference labeling methods against standard methods, which assign preferences based on only prior context. Our method achieves a 5% improvement in F1 measured against the answer set from different interpretations of each query, showing the value of modeling future conversation turns. We further demonstrate that our method can be used to train models to judiciously determine when to ask clarifying questions, directly answering the question when clarification is unnecessary. In our experiments, we find that our method achives a 3% improvement in accuracy of such judgments over existing methods.