ICLR 2025 论文列表

Unsupervised reinforcement learning (RL) aims at pre-training models that can solve a wide range of downstream tasks in complex environments. Despite recent advancements, existing approaches suffer from several limitations: they may require running an RL process on each task to achieve a satisfactory performance, they may need access to datasets with good coverage or well-curated task-specific samples, or they may pre-train policies with unsupervised losses that are poorly correlated with the downstream tasks of interest. In this paper, we introduce FB-CPR, which regularizes unsupervised zero-shot RL based on the forward-backward (FB) method towards imitating trajectories from unlabeled behaviors. The resulting models learn useful policies imitating the behaviors in the dataset, while retaining zero-shot generalization capabilities. We demonstrate the effectiveness of FB-CPR in a challenging humanoid control problem. Training FB-CPR online with observation-only motion capture datasets, we obtain the first humanoid behavioral foundation model that can be prompted to solve a variety of whole-body tasks, including motion tracking, goal reaching, and reward optimization. The resulting model is capable of expressing human-like behaviors and it achieves competitive performance with task-specific methods while outperforming state-of-the-art unsupervised RL and model-based baselines.

How can we automatically select an out-of-distribution (OOD) detection model for various underlying tasks? This is crucial for maintaining the reliability of open-world applications by identifying data distribution shifts, particularly in critical domains such as online transactions, autonomous driving, and real-time patient diagnosis. Despite the availability of numerous OOD detection methods, the challenge of selecting an optimal model for diverse tasks remains largely underexplored, especially in scenarios lacking ground truth labels. In this work, we introduce MetaOOD, the first zero-shot, unsupervised framework that utilizes meta-learning to select an OOD detection model automatically. As a meta-learning approach, MetaOOD leverages historical performance data of existing methods across various benchmark OOD detection datasets, enabling the effective selection of a suitable model for new datasets without the need for labeled data at the test time. To quantify task similarities more accurately, we introduce language model-based embeddings that capture the distinctive OOD characteristics of both datasets and detection models. Through extensive experimentation with 24 unique test dataset pairs to choose from among 11 OOD detection models, we demonstrate that MetaOOD significantly outperforms existing methods and only brings marginal time overhead. Our results, validated by Wilcoxon statistical tests, show that MetaOOD surpasses a diverse group of 11 baselines, including established OOD detectors and advanced unsupervised selection methods.

The dynamic nature of proteins, influenced by ligand interactions, is essential for comprehending protein function and progressing drug discovery. Traditional structure-based drug design (SBDD) approaches typically target binding sites with rigid structures, limiting their practical application in drug development. While molecular dynamics simulation can theoretically capture all the biologically relevant conformations, the transition rate is dictated by the intrinsic energy barrier between them, making the sampling process computationally expensive. To overcome the aforementioned challenges, we propose to use generative modeling for SBDD considering conformational changes of protein pockets. We curate a dataset of apo and multiple holo states of protein-ligand complexes, simulated by molecular dynamics, and propose a full-atom flow model (and a stochastic version), named DynamicFlow, that learns to transform apo pockets and noisy ligands into holo pockets and corresponding 3D ligand molecules. Our method uncovers promising ligand molecules and corresponding holo conformations of pockets. Additionally, the resultant holo-like states provide superior inputs for traditional SBDD approaches, playing a significant role in practical drug discovery.

Linear temporal logic (LTL) has recently been adopted as a powerful formalism for specifying complex, temporally extended tasks in multi-task reinforcement learning (RL). However, learning policies that efficiently satisfy arbitrary specifications not observed during training remains a challenging problem. Existing approaches suffer from several shortcomings: they are often only applicable to finite-horizon fragments of LTL, are restricted to suboptimal solutions, and do not adequately handle safety constraints. In this work, we propose a novel learning approach to address these concerns. Our method leverages the structure of Büchi automata, which explicitly represent the semantics of LTL specifications, to learn policies conditioned on sequences of truth assignments that lead to satisfying the desired formulae. Experiments in a variety of discrete and continuous domains demonstrate that our approach is able to zero-shot satisfy a wide range of finite- and infinite-horizon specifications, and outperforms existing methods in terms of both satisfaction probability and efficiency. Code available at: https://deep-ltl.github.io/

In this study, we examine the representation learning abilities of Denoising Diffusion Models (DDM) that were originally purposed for image generation. Our philosophy is to deconstruct a DDM, gradually transforming it into a classical Denoising Autoencoder (DAE). This deconstructive process allows us to explore how various components of modern DDMs influence self-supervised representation learning. We observe that only a very few modern components are critical for learning good representations, while many others are nonessential. Our study ultimately arrives at an approach that is highly simplified and to a large extent resembles a classical DAE. We hope our study will rekindle interest in a family of classical methods within the realm of modern self-supervised learning.

Self-driving laboratories have begun to replace human experimenters in performing single experimental skills or predetermined experimental protocols. However, as the pace of idea iteration in scientific research has been intensified by Artificial Intelligence, the demand for rapid design of new protocols for new discoveries become evident. Efforts to automate protocol design have been initiated, but the capabilities of knowledge-based machine designers, such as Large Language Models, have not been fully elicited, probably for the absence of a systematic representation of experimental knowledge, as opposed to isolated, flatten pieces of information. To tackle this issue, we propose a multi-faceted, multi-scale representation, where instance actions, generalized operations, and product flow models are hierarchically encapsulated using Domain-Specific Languages. We further develop a data-driven algorithm based on non-parametric modeling that autonomously customizes these representations for specific domains. The proposed representation is equipped with various machine designers to manage protocol design tasks, including planning, modification, and adjustment. The results demonstrate that the proposed method could effectively complement Large Language Models in the protocol design process, serving as an auxiliary module in the realm of machine-assisted scientific exploration.

Pareto front profiling in multi-objective optimization (MOO), i.e., finding a diverse set of Pareto optimal solutions, is challenging, especially with expensive objectives that require training a neural network. Typically, in MOO for neural architecture search (NAS), we aim to balance performance and hardware metrics across devices. Prior NAS approaches simplify this task by incorporating hardware constraints into the objective function, but profiling the Pareto front necessitates a computationally expensive search for each constraint. In this work, we propose a novel NAS algorithm that encodes user preferences to trade-off performance and hardware metrics, yielding representative and diverse architectures across multiple devices in just a single search run. To this end, we parameterize the joint architectural distribution across devices and multiple objectives via a hypernetwork that can be conditioned on hardware features and preference vectors, enabling zero-shot transferability to new devices. Extensive experiments involving up to 19 hardware devices and 3 different objectives demonstrate the effectiveness and scalability of our method. Finally, we show that, without any additional costs, our method outperforms existing MOO NAS methods across a broad range of qualitatively different search spaces and datasets, including MobileNetV3 on ImageNet-1k, an encoder-decoder transformer space for machine translation and a decoder-only space for language modelling.

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.

Neural activity in the brain is known to encode low-dimensional, time-evolving, behaviour-related variables. A long-standing goal of neural data analysis has been to identify these variables and their mapping to neural activity. A productive and canonical approach has been to simply visualise neural "tuning curves" as a function of behaviour. However, significant discrepancies between behaviour and the true latent variables -- such as an agent thinking of position Y whilst located at position X -- distort and blur the tuning curves, decreasing their interpretability. To address this, latent variable models propose to learn the latent variable from data; these are typically expensive, hard to tune, or scale poorly, complicating their adoption. Here we propose SIMPL (Scalable Iterative Maximization of Population-coded Latents), an EM-style algorithm which iteratively optimises latent variables and tuning curves. SIMPL is fast, scalable and exploits behaviour as an initial condition to further improve convergence and identifiability. It can accurately recover latent variables in spatial and non-spatial tasks. When applied to a large hippocampal dataset SIMPL converges on smaller, more numerous, and more uniformly sized place fields than those based on behaviour, suggesting the brain may encode space with greater resolution than previously thought.

Large Language Models (LLMs) have recently been used as experts to infer causal graphs, often by repeatedly applying a pairwise prompt that asks about the causal relationship of each variable pair. However, such experts, including human domain experts, cannot distinguish between direct and indirect effects given a pairwise prompt. Therefore, instead of the graph, we propose that causal order be used as a more stable output interface for utilizing expert knowledge. When querying a perfect expert with a pairwise prompt, we show that the inferred graph can have significant errors whereas the causal order is always correct. In practice, however, LLMs are imperfect experts and we find that pairwise prompts lead to multiple cycles and do not yield a valid order. Hence, we propose a prompting strategy that introduces an auxiliary variable for every variable pair and instructs the LLM to avoid cycles within this triplet. We show, both theoretically and empirically, that such a triplet prompt leads to fewer cycles than the pairwise prompt. Across multiple real-world graphs, the triplet prompt yields a more accurate order using both LLMs and human annotators as experts. By querying the expert with different auxiliary variables for the same variable pair, it also increases robustness---triplet method with much smaller models such as Phi-3 and Llama-3 8B outperforms a pairwise prompt with GPT-4. For practical usage, we show how the estimated causal order from the triplet method can be used to reduce error in downstream discovery and effect inference tasks.

Accurate prediction of RNA properties, such as stability and interactions, is crucial for advancing our understanding of biological processes and developing RNA-based therapeutics. RNA structures can be represented as 1D sequences, 2D topological graphs, or 3D all-atom models, each offering different insights into its function. Existing works predominantly focus on 1D sequence-based models, which overlook the geometric context provided by 2D and 3D geometries. This study presents the first systematic evaluation of incorporating explicit 2D and 3D geometric information into RNA property prediction, considering not only performance but also real-world challenges such as limited data availability, partial labeling, sequencing noise, and computational efficiency. To this end, we introduce a newly curated set of RNA datasets with enhanced 2D and 3D structural annotations, providing a resource for model evaluation on RNA data. Our findings reveal that models with explicit geometry encoding generally outperform sequence-based models, with an average prediction RMSE reduction of around 12% across all various RNA tasks and excelling in low-data and partial labeling regimes, underscoring the value of explicitly incorporating geometric context. On the other hand, geometry-unaware sequence-based models are more robust under sequencing noise but often require around 2-5x training data to match the performance of geometry-aware models. Our study offers further insights into the trade-offs between different RNA representations in practical applications and addresses a significant gap in evaluating deep learning models for RNA tasks.

In cross-device federated learning (FL) with millions of mobile clients, only a small subset of clients participate in training in every communication round, and Federated Averaging (FedAvg) is the most popular algorithm in practice. Existing analyses of FedAvg usually assume the participating clients are independently sampled in each round from a uniform distribution, which does not reflect real-world scenarios. This paper introduces a theoretical framework that models client participation in FL as a Markov chain to study optimization convergence when clients have non-uniform and correlated participation across rounds. We apply this framework to analyze a more practical pattern: every client must wait a minimum number of $R$ rounds (minimum separation) before re-participating. We theoretically prove and empirically observe that increasing minimum separation reduces the bias induced by intrinsic non-uniformity of client availability in cross-device FL systems. Furthermore, we develop an effective debiasing algorithm for FedAvg that provably converges to the unbiased optimal solution under arbitrary minimum separation and unknown client availability distribution.

Training embodied agents to perform complex robotic tasks presents significant challenges due to the entangled factors of task compositionality, environmental diversity, and dynamic changes. In this work, we introduce a novel imitation learning framework to train closed-loop concept-guided policies that enhance long-horizon task performance by leveraging discovered manipulation concepts. Unlike methods that rely on predefined skills and human-annotated labels, our approach allows agents to autonomously abstract manipulation concepts from their proprioceptive states, thereby alleviating misalignment due to ambiguities in human semantics and environmental complexity. Our framework comprises two primary components: an *Automatic Concept Discovery* module that identifies meaningful and consistent manipulation concepts, and a *Concept-Guided Policy Learning* module that effectively utilizes these manipulation concepts for adaptive task execution, including a *Concept Selection Transformer* for concept-based guidance and a *Concept-Guided Policy* for action prediction with the selected concepts. Experiments demonstrate that our approach significantly outperforms baseline methods across a range of tasks and environments, while showcasing emergent consistency in motion patterns associated with the discovered manipulation concepts. Codes are available at: https://github.com/PeiZhou26/AutoCGP.

Instruction-following is a fundamental capability of language models, requiring the model to recognize even the most subtle requirements in the instructions and accurately reflect them in its output. Such an ability is well-suited for and often optimized by preference learning. However, existing methods often directly sample multiple independent responses from the model when creating preference pairs. Such practice can introduce content variations irrelevant to whether the instruction is precisely followed (e.g., different expressions about the same semantic), interfering with the goal of teaching models to recognize the key differences that lead to improved instruction following. In light of this, we introduce SPaR, a self-play framework integrating tree-search self-refinement to yield valid and comparable preference pairs free from distractions. By playing against itself, an LLM employs a tree-search strategy to refine its previous responses with respect to the instruction while minimizing unnecessary variations. Our experiments show that a LLaMA3-8B model, trained over three iterations guided by SPaR, surpasses GPT-4-Turbo on the IFEval benchmark without losing general capabilities. Furthermore, SPaR demonstrates promising scalability, greatly enhancing models like GLM-4-9B and LLaMA3-70B. We also identify how inference scaling in tree search would impact model performance. Our code and data are publicly available at https://github.com/thu-coai/SPaR.

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/

Search is an ability foundational in many important tasks, and recent studies have shown that large language models (LLMs) struggle to perform search robustly. It is unknown whether this inability is due to a lack of data, insufficient model parameters, or fundamental limitations of the transformer architecture. In this work, we use the foundational graph connectivity problem as a testbed to generate effectively limitless high-coverage data to train small transformers and test whether they can learn to perform search. We find that, when given the right training distribution, the transformer is able to learn to search. We analyze the algorithm that the transformer has learned through a novel mechanistic interpretability technique that enables us to extract the computation graph from the trained model. We find that for each vertex in the input graph, transformers compute the set of vertices reachable from that vertex. Each layer then progressively expands these sets, allowing the model to search over a number of vertices exponential in the number of layers. However, we find that as the input graph size increases, the transformer has greater difficulty in learning the task. This difficulty is not resolved even as the number of parameters is increased, suggesting that increasing model scale will not lead to robust search abilities. We also find that performing search in-context (i.e., chain-of-thought) does not resolve this inability to learn to search on larger graphs.

Large-scale datasets have been pivotal to the advancements of deep learning models in recent years, but training on such large datasets inevitably incurs substantial storage and computational overhead. Meanwhile, real-world datasets often contain redundant and noisy data, imposing a negative impact on training efficiency and model performance. Data selection has shown promise in identifying the most representative samples from the entire dataset, which aims to minimize the performance gap with reduced training costs. Existing works typically rely on single-modality information to assign importance scores for individual samples, which may lead to inaccurate assessments, especially when dealing with noisy or corrupted samples. To address this limitation, we propose a novel CLIP-powered data selection framework that leverages multimodal information for more robust and generalizable sample selection. Specifically, our framework consists of three key modules—dataset adaptation, sample scoring, and selection optimization—that together harness extensive pre-trained multimodal knowledge to comprehensively assess sample influence and optimize the selection results through multi-objective optimization. Extensive experiments demonstrate that our approach consistently outperforms existing state-of-the-art baselines on various benchmark datasets. Notably, our method effectively removes noisy or damaged samples from the dataset, enabling it to achieve even higher performance with less data. This indicates that it is not only a way to accelerate training but can also improve overall data quality. The implementation is available at https://github.com/Jackbrocp/clip-powered-data-selection.

Continual learning, also known as lifelong learning or incremental learning, refers to the process by which a model learns from a stream of incoming data over time. A common problem in continual learning is the classification layer’s bias towards the most recent task. Traditionally, methods have relied on incorporating data from past tasks during training to mitigate this issue. However, the recent shift in continual learning to memory-free environments has rendered these approaches infeasible. In this study, we propose a solution focused on the testing phase. We first introduce a simple Out-of-Task Detection method, OTD, designed to accurately identify samples from past tasks during testing. Leveraging OTD, we then propose: (1) an Adaptive Retention mechanism for dynamically tuning the classifier layer on past task data; (2) an Adaptive Correction mechanism for revising predictions when the model classifies data from previous tasks into classes from the current task. We name our approach Adaptive Retention & Correction (ARC). While designed for memory-free environments, ARC also proves effective in memorybased settings. Extensive experiments show that our proposed method can be plugged in to virtually any existing continual learning approach without requiring any modifications to its training procedure. Specifically, when integrated with state-of-the-art approaches, ARC achieves an average performance increase of 2.7% and 2.6% on the CIFAR-100 and Imagenet-R datasets, respectively