ICLR 2025 Papers

When solving challenging problems, language models (LMs) are able to identify relevant information from long and complicated contexts. To study how LMs solve retrieval tasks in diverse situations, we introduce ORION, a collection of structured retrieval tasks spanning six domains, from text understanding to coding. Each task in ORION can be represented abstractly by a request (e.g. a question) that retrieves an attribute (e.g. the character name) from a context (e.g. a story). We apply causal analysis on 18 open-source language models with sizes ranging from 125 million to 70 billion parameters. We find that LMs internally decompose retrieval tasks in a modular way: middle layers at the last token position process the request, while late layers retrieve the correct entity from the context. After causally enforcing this decomposition, models are still able to solve the original task, preserving 70% of the original correct token probability in 98 of the 106 studied model-task pairs. We connect our macroscopic decomposition with a microscopic description by performing a fine-grained case study of a question-answering task on Pythia-2.8b. Building on our high-level understanding, we demonstrate a proof of concept application for scalable internal oversight of LMs to mitigate prompt-injection while requiring human supervision on only a single input. Our solution improves accuracy drastically (from 15.5% to 97.5% on Pythia-12b). This work presents evidence of a universal emergent modular processing of tasks across varied domains and models and is a pioneering effort in applying interpretability for scalable internal oversight of LMs.

Improving social welfare is a complex challenge requiring policymakers to optimize objectives across multiple time horizons. Evaluating the impact of such policies presents a fundamental challenge, as those that appear suboptimal in the short run may yield significant long-term benefits. We tackle this challenge by analyzing the long-term dynamics of two prominent policy frameworks: Rawlsian policies, which prioritize those with the greatest need, and utilitarian policies, which maximize immediate welfare gains. Conventional wisdom suggests these policies are at odds, as Rawlsian policies are assumed to come at the cost of reducing the average social welfare, which their utilitarian counterparts directly optimize. We challenge this assumption by analyzing these policies in a sequential decision-making framework where individuals' welfare levels stochastically decay over time, and policymakers can intervene to prevent this decay. Under reasonable assumptions, we prove that interventions following Rawlsian policies can outperform utilitarian policies in the long run, even when the latter dominate in the short run. We characterize the exact conditions under which Rawlsian policies can outperform utilitarian policies. We further illustrate our theoretical findings using simulations, which highlight the risks of evaluating policies based solely on their short-term effects. Our results underscore the necessity of considering long-term horizons in designing and evaluating welfare policies; the true efficacy of even well-established policies may only emerge over time.

Embodiment co-design aims to optimize a robot's morphology and control policy simultaneously. While prior work has demonstrated its potential for generating environment-adaptive robots, this field still faces persistent challenges in optimization efficiency due to the (i) combinatorial nature of morphological search spaces and (ii) intricate dependencies between morphology and control. We prove that the ineffective morphology representation and unbalanced reward signals between the design and control stages are key obstacles to efficiency. To advance towards efficient embodiment co-design, we propose **BodyGen**, which utilizes (1) topology-aware self-attention for both design and control, enabling efficient morphology representation with lightweight model sizes; (2) a temporal credit assignment mechanism that ensures balanced reward signals for optimization. With our findings, BodyGen achieves an average **60.03%** performance improvement against state-of-the-art baselines. We provide codes and more results on the website: https://genesisorigin.github.io.

Recent works have demonstrated promising performances of neural networks on hyperbolic spaces and symmetric positive definite (SPD) manifolds. These spaces belong to a family of Riemannian manifolds referred to as symmetric spaces of noncompact type. In this paper, we propose a novel approach for developing neural networks on such spaces. Our approach relies on a unified formulation of the distance from a point to a hyperplane on the considered spaces. We show that some existing formulations of the point-to-hyperplane distance can be recovered by our approach under specific settings. Furthermore, we derive a closed-form expression for the point-to-hyperplane distance in higher-rank symmetric spaces of noncompact type equipped with G-invariant Riemannian metrics. The derived distance then serves as a tool to design fully-connected (FC) layers and an attention mechanism for neural networks on the considered spaces. Our approach is validated on challenging benchmarks for image classification, electroencephalogram (EEG) signal classification, image generation, and natural language inference.

Circuit representation learning has become pivotal in electronic design automation, enabling critical tasks such as testability analysis, logic reasoning, power estimation, and SAT solving. However, existing models face significant challenges in scaling to large circuits due to limitations like over-squashing in graph neural networks and the quadratic complexity of transformer-based models. To address these issues, we introduce \textbf{DeepGate4}, a scalable and efficient graph transformer specifically designed for large-scale circuits. DeepGate4 incorporates several key innovations: (1) an update strategy tailored for circuit graphs, which reduce memory complexity to sub-linear and is adaptable to any graph transformer; (2) a GAT-based sparse transformer with global and local structural encodings for AIGs; and (3) an inference acceleration CUDA kernel that fully exploit the unique sparsity patterns of AIGs. Our extensive experiments on the ITC99 and EPFL benchmarks show that DeepGate4 significantly surpasses state-of-the-art methods, achieving 15.5\% and 31.1\% performance improvements over the next-best models. Furthermore, the Fused-DeepGate4 variant reduces runtime by 35.1\% and memory usage by 46.8\%, making it highly efficient for large-scale circuit analysis. These results demonstrate the potential of DeepGate4 to handle complex EDA tasks while offering superior scalability and efficiency.

Sharpness-Aware Minimization (SAM) has substantially improved the generalization of neural networks under various settings. Despite the success, its effectiveness remains poorly understood. In this work, we discover an intriguing phenomenon in the training dynamics of SAM, shedding light on understanding its implicit bias towards flatter minima over Stochastic Gradient Descent (SGD). Specifically, we find that *SAM efficiently selects flatter minima late in training*. Remarkably, even a few epochs of SAM applied at the end of training yield nearly the same generalization and solution sharpness as full SAM training. Subsequently, we delve deeper into the underlying mechanism behind this phenomenon. Theoretically, we identify two phases in the learning dynamics after applying SAM late in training: i) SAM first escapes the minimum found by SGD exponentially fast; and ii) then rapidly converges to a flatter minimum within the same valley. Furthermore, we empirically investigate the role of SAM during the early training phase. We conjecture that the optimization method chosen in the late phase is more crucial in shaping the final solution's properties. Based on this viewpoint, we extend our findings from SAM to Adversarial Training.

We introduce Concept Bottleneck Protein Language Models (CB-pLM), a generative masked language model with a layer where each neuron corresponds to an interpretable concept. Our architecture offers three key benefits: i) Control: We can intervene on concept values to precisely control the properties of generated proteins, achieving a 3$\times$ larger change in desired concept values compared to baselines. ii) Interpretability: A linear mapping between concept values and predicted tokens allows transparent analysis of the model's decision-making process. iii) Debugging: This transparency facilitates easy debugging of trained models. Our models achieve pre-training perplexity and downstream task performance comparable to traditional masked protein language models, demonstrating that interpretability does not compromise performance. While adaptable to any language model, we focus on masked protein language models due to their importance in drug discovery and the ability to validate our model's capabilities through real-world experiments and expert knowledge. We scale our CB-pLM from 24 million to 3 billion parameters, making them the largest Concept Bottleneck Models trained and the first capable of generative language modeling.

The ability to accurately interpret complex visual information is a crucial topic of multimodal large language models (MLLMs). Recent work indicates that enhanced visual perception significantly reduces hallucinations and improves performance on resolution-sensitive tasks, such as optical character recognition and document analysis. A number of recent MLLMs achieve this goal using a mixture of vision encoders. Despite their success, there is a lack of systematic comparisons and detailed ablation studies addressing critical aspects, such as expert selection and the integration of multiple vision experts. This study provides an extensive exploration of the design space for MLLMs using a mixture of vision encoders and resolutions. Our findings reveal several underlying principles common to various existing strategies, leading to a streamlined yet effective design approach. We discover that simply concatenating visual tokens from a set of complementary vision encoders is as effective as more complex mixing architectures or strategies. We additionally introduce Pre-Alignment to bridge the gap between vision-focused encoders and language tokens, enhancing model coherence. The resulting family of MLLMs, Eagle, surpasses other leading open-source models on major MLLM benchmarks.

Deep ensembles improve the performance of the models by taking the average predictions of a group of ensemble members. However, the origin of these capabilities remains a mystery and deep ensembles are used as a reliable “black box” to improve the performance. Existing studies typically attribute such improvement to Jensen gaps of the deep ensemble method, where the loss of the mean does not exceed the mean of the loss for any convex loss metric. In this work, we demonstrate that Jensen’s inequality is not responsible for the effectiveness of deep ensembles, and convexity is not a necessary condition. Instead, Jensen Gap focuses on the “average loss” of individual models, which provides no practical meaning. Thus it fails to explain the core phenomena of deep ensembles such as the superiority to any single ensemble member, the decreasing loss with the number of ensemble members, etc. Regarding this mystery, we provide theoretical analysis and comprehensive empirical results from a statistical perspective that reveal the true mechanism of deep ensembles. Our results highlight that deep ensembles originate from the homogeneous output distribution across all ensemble members. Specifically, the predictions of homogeneous models (Abe et al., 2022b) have the distributional equivalence property – Although the predictions of independent ensemble members are point-wise different, they form an identical distribution. Such agreement and disagreement contribute to deep ensembles’ “magical power”. Based on this discovery, we provide rigorous proof of the effectiveness of deep ensembles and analytically quantify the extent to which ensembles improve performance. The derivations not only theoretically quantify the effectiveness of deep ensembles for the first time, but also enable estimation schemes that foresee the performance of ensembles with different capacities. Furthermore, different from existing studies, our results also point out that deep ensembles work in a different mechanism from model scaling a single model, even though significant correlations between them have been observed.

Understanding the dynamics of neural networks in different width regimes is crucial for improving their training and performance. We present an exact solution for the learning dynamics of a one-hidden-layer linear network, with one-dimensional data, across any finite width, uniquely exhibiting both kernel and feature learning phases. This study marks a technical advancement by enabling the analysis of the training trajectory from any initialization and a detailed phase diagram under varying common hyperparameters such as width, layer-wise learning rates, and scales of output and initialization. We identify three novel prototype mechanisms specific to the feature learning regime: (1) learning by alignment, (2) learning by disalignment, and (3) learning by rescaling, which contrast starkly with the dynamics observed in the kernel regime. Our theoretical findings are substantiated with empirical evidence showing that these mechanisms also manifest in deep nonlinear networks handling real-world tasks, enhancing our understanding of neural network training dynamics and guiding the design of more effective learning strategies.

Recent work in language modeling has raised the possibility of “self-improvement,” where an LLM evaluates and refines its own generations to achieve higher performance without external feedback. It is impossible for this self-improvement to create information that is not already in the model, so why should we expect that this will lead to improved capabilities? We offer a new theoretical perspective on the capabilities of self-improvement through a lens we refer to as “sharpening.” Motivated by the observation that language models are often better at verifying response quality than they are at generating correct responses, we formalize self-improvement as using the model itself as a verifier during post-training in order to ‘sharpen’ the model to one placing large mass on high-quality sequences, thereby amortizing the expensive inference-time computation of generating good sequences. We begin by introducing a new statistical framework for sharpening in which the learner has sample access to a pre-trained base policy. Then, we analyze two natural families of self improvement algorithms based on SFT and RLHF. We find that (i) the SFT-based approach is minimax optimal whenever the initial model has sufficient coverage, but (ii) the RLHF-based approach can improve over SFT-based self- improvement by leveraging online exploration, bypassing the need for coverage. We view these findings as a starting point toward a foundational understanding that can guide the design and evaluation of self-improvement algorithms.

In gradient descent dynamics of neural networks, the top eigenvalue of the Hessian of the loss (sharpness) displays a variety of robust phenomena throughout training. This includes early time regimes where the sharpness may decrease during early periods of training (sharpness reduction), and later time behavior such as progressive sharpening and edge of stability. We demonstrate that a simple -layer linear network (UV model) trained on a single training example exhibits all of the essential sharpness phenomenology observed in real-world scenarios. By analyzing the structure of dynamical fixed points in function space and the vector field of function updates, we uncover the underlying mechanisms behind these sharpness trends. Our analysis reveals (i) the mechanism behind early sharpness reduction and progressive sharpening, (ii) the required conditions for edge of stability, and (iii) a period-doubling route to chaos on the edge of stability manifold as learning rate is increased. Finally, we demonstrate that various predictions from this simplified model generalize to real-world scenarios and discuss its limitations.

Model-based reinforcement learning (MBRL) seeks to enhance data efficiency by learning a model of the environment and generating synthetic rollouts from it. However, accumulated model errors during these rollouts can distort the data distribution, negatively impacting policy learning and hindering long-term planning. Thus, the accumulation of model errors is a key bottleneck in current MBRL methods. We propose Infoprop, a model-based rollout mechanism that separates aleatoric from epistemic model uncertainty and reduces the influence of the latter on the data distribution. Further, Infoprop keeps track of accumulated model errors along a model rollout and provides termination criteria to limit data corruption. We demonstrate the capabilities of Infoprop in the Infoprop-Dyna algorithm, reporting state-of-the-art performance in Dyna-style MBRL on common MuJoCo benchmark tasks while substantially increasing rollout length and data quality.

Designing mechanically efficient geometry for architectural structures like shells, towers, and bridges, is an expensive iterative process. Existing techniques for solving such inverse problems rely on traditional optimization methods, which are slow and computationally expensive, limiting iteration speed and design exploration. Neural networks would seem to offer a solution via data-driven amortized optimization, but they often require extensive fine-tuning and cannot ensure that important design criteria, such as mechanical integrity, are met. In this work, we combine neural networks with a differentiable mechanics simulator to develop a model that accelerates the solution of shape approximation problems for architectural structures represented as bar systems. This model explicitly guarantees compliance with mechanical constraints while generating designs that closely match target geometries. We validate our approach in two tasks, the design of masonry shells and cable-net towers. Our model achieves better accuracy and generalization than fully neural alternatives, and comparable accuracy to direct optimization but in real time, enabling fast and reliable design exploration. We further demonstrate its advantages by integrating it into 3D modeling software and fabricating a physical prototype. Our work opens up new opportunities for accelerated mechanical design enhanced by neural networks for the built environment.

Recent advances in structure-based drug design (SBDD) have produced surprising results, with models often generating molecules that achieve better Vina docking scores than actual ligands. However, these results are frequently overly optimistic due to the limitations of docking score accuracy and the challenges of wet-lab validation. While generated molecules may demonstrate high QED (drug-likeness) and SA (synthetic accessibility) scores, they often lack true drug-like properties or synthesizability. To address these limitations, we propose a model-level evaluation framework that emphasizes practical metrics aligned with real-world applications. Inspired by recent findings on the utility of generated molecules in ligand-based virtual screening, our framework evaluates SBDD models by their ability to produce molecules that effectively retrieve active compounds from chemical libraries via similarity-based searches. This approach provides a direct indication of therapeutic potential, bridging the gap between theoretical performance and real-world utility. Our experiments reveal that while SBDD models may excel in theoretical metrics like Vina scores, they often fall short in these practical metrics. By introducing this new evaluation strategy, we aim to enhance the relevance and impact of SBDD models for pharmaceutical research and development.

In modern large language models (LLMs), increasing the context length is crucial for improving comprehension and coherence in long-context, multi-modal, and retrieval-augmented language generation. While many recent transformer models attempt to extend their context length over a million tokens, they remain impractical due to the quadratic time and space complexities. Although recent works on linear and sparse attention mechanisms can achieve this goal, their real-world applicability is often limited by the need to re-train from scratch and significantly worse performance. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which reduces the time complexity of the attention mechanism to $O(T \log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length. We notice a pattern in the attention scores of pretrained LLMs where tokens close together tend to have similar scores, which we call "attention locality". Based on this observation, we utilize a novel tree-search-like algorithm that estimates the top-$k$ key tokens for a given query on the fly, which is mathematically guaranteed to have better performance than random attention pruning. In addition to improving the time complexity of the attention mechanism, we further optimize GPU memory usage by implementing KV cache offloading, which stores only $O(\log T)$ tokens on the GPU while maintaining similar decoding throughput. Experiments on benchmarks show that HiP, with its training-free nature, significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation. HiP enables pretrained LLMs to scale up to millions of tokens on commodity GPUs, potentially unlocking long-context LLM applications previously deemed infeasible.

Protein family design emerges as a promising alternative by combining the advantages of de novo protein design and mutation-based directed evolution.In this paper, we propose ProfileBFN, the Profile Bayesian Flow Networks, for specifically generative modeling of protein families. ProfileBFN extends the discrete Bayesian Flow Network from an MSA profile perspective, which can be trained on single protein sequences by regarding it as a degenerate profile, thereby achieving efficient protein family design by avoiding large-scale MSA data construction and training. Empirical results show that ProfileBFN has a profound understanding of proteins. When generating diverse and novel family proteins, it can accurately capture the structural characteristics of the family. The enzyme produced by this method is more likely than the previous approach to have the corresponding function, offering better odds of generating diverse proteins with the desired functionality.

Articulated object manipulation is a critical capability for robots to perform various tasks in real-world scenarios. Composed of multiple parts connected by joints, articulated objects are endowed with diverse functional mechanisms through complex relative motions. For example, a safe consists of a door, a handle, and a lock, where the door can only be opened when the latch is unlocked. The internal structure, such as the state of a lock or joint angle constraints, cannot be directly observed from visual observation. Consequently, successful manipulation of these objects requires adaptive adjustment based on trial and error rather than a one-time visual inference. However, previous datasets and simulation environments for articulated objects have primarily focused on simple manipulation mechanisms where the complete manipulation process can be inferred from the object's appearance. To enhance the diversity and complexity of adaptive manipulation mechanisms, we build a novel articulated object manipulation environment and equip it with 9 categories of objects. Based on the environment and objects, we further propose an adaptive demonstration collection and 3D visual diffusion-based imitation learning pipeline that learns the adaptive manipulation policy. The effectiveness of our designs and proposed method is validated through both simulation and real-world experiments.

Generative machine learning models are increasingly being used to design novel proteins. However, their major limitation is the inability to account for protein flexibility, a property crucial for protein function. Learning to engineer flexibility is difficult because the relevant data is scarce, heterogeneous, and costly to obtain using computational and experimental methods. Our contributions are three-fold. First, we perform a comprehensive comparison of methods for evaluating protein flexibility and identify relevant data for learning. Second, we overcome the data scarcity issue by leveraging a pre-trained protein language model. We design and train flexibility predictors utilizing either only sequential or both sequential and structural information on the input. Third, we introduce a method for fine-tuning a protein inverse folding model to make it steerable toward desired flexibility at specified regions. We demonstrate that our method Flexpert enables guidance of inverse folding models toward increased flexibility. This opens up a transformative possibility of engineering protein flexibility.