ICML 2025 Papers

Personalized learning, especially data-based methods, has garnered widespread attention in recent years, aiming to meet individual student needs. However, many works rely on the implicit assumption that benchmarks are high-quality and well-annotated, which limits their practical applicability. In real-world scenarios, these benchmarks often exhibit long-tail distributions, significantly impacting model performance. To address this challenge, we propose a novel method called **N**eural-**C**ollapse-**A**dvanced personalized **L**earning (NCAL), designed to learn features that conform to the same simplex equiangular tight frame (ETF) structure. NCAL introduces Text-modality Collapse (TC) regularization to optimize the distribution of text embeddings within the large language model (LLM) representation space. Notably, NCAL is model-agnostic, making it compatible with various architectures and approaches, thereby ensuring broad applicability. Extensive experiments demonstrate that NCAL effectively enhances existing works, achieving new state-of-the-art performance. Additionally, NCAL mitigates class imbalance, significantly improving the model’s generalization ability.

We propose a *gradient-free* deep reinforcement learning algorithm to solve *high-dimensional*, finite-horizon stochastic control problems. Although the recently developed deep reinforcement learning framework has achieved great success in solving these problems, direct estimation of policy gradients from Monte Carlo sampling often suffers from high variance. To address this, we introduce the Momentum Consensus-Based Optimization (M-CBO) and Adaptive Momentum Consensus-Based Optimization (Adam-CBO) frameworks. These methods optimize policies using Monte Carlo estimates of the value function, rather than its gradients. Adjustable Gaussian noise supports efficient exploration, helping the algorithm converge to optimal policies in complex, nonconvex environments. Numerical results confirm the accuracy and scalability of our approach across various problem dimensions and show the potential for extension to mean-field control problems. Theoretically, we prove that M-CBO can converge to the optimal policy under some assumptions.

Communication between multiple language model (LM) agents has been shown to scale up the reasoning ability of LMs. While natural language has been the dominant medium for inter-LM communication, it is not obvious this should be the standard: not only does natural language communication incur high inference costs that scale quickly with the number of both agents and messages, but also the decoding process abstracts away too much rich information that could be otherwise accessed from the internal activations. In this work, we propose a simple technique whereby LMs communicate via *activations*; concretely, we pause an LM $B$'s computation at an intermediate layer, combine its current activation with another LM $A$'s intermediate activation via some function $f$, then pass $f$'s output into the next layer of $B$ and continue the forward pass till decoding is complete. This approach scales up LMs on new tasks with *zero* additional parameters and data, and saves a *substantial amount of compute* over natural language communication. We test our method with various functional forms $f$ on two experimental setups—multi-player coordination games and reasoning benchmarks—and find that it achieves up to $27$% improvement over natural language communication across datasets with $<$$1/4$ the compute, illustrating the superiority and robustness of activations as an alternative "language" for communication between LMs.

Molecular Dynamics (MD) simulations are essential for understanding the atomic-level behavior of molecular systems, giving insights into their transitions and interactions. However, classical MD techniques are limited by the trade-off between accuracy and efficiency, while recent deep learning-based improvements have mostly focused on single-domain molecules, lacking transferability to unfamiliar molecular systems. Therefore, we propose **Uni**fied **Sim**ulator (UniSim), which leverages cross-domain knowledge to enhance the understanding of atomic interactions. First, we employ a multi-head pretraining approach to learn a unified atomic representation model from a large and diverse set of molecular data. Then, based on the stochastic interpolant framework, we learn the state transition patterns over long timesteps from MD trajectories, and introduce a force guidance module for rapidly adapting to different chemical environments. Our experiments demonstrate that UniSim achieves highly competitive performance across small molecules, peptides, and proteins.

Pretrained language models (LMs) can generalize to implications of facts that they are finetuned on. For example, if finetuned on "John Doe lives in Tokyo," LMs correctly answer "What language do the people in John Doe's city speak?'' with "Japanese''. However, little is known about the mechanisms that enable this generalization or how they are learned during pretraining. We introduce extractive structures as a framework for describing how components in LMs (e.g., MLPs or attention heads) coordinate to enable this generalization. The structures consist of informative components that store training facts as weight changes, and upstream and downstream extractive components that query and process the stored information to produce the correct implication. We hypothesize that extractive structures are learned during pretraining when encountering implications of previously known facts. This yields two predictions: a data ordering effect where extractive structures can be learned only if facts precede their implications, and a weight grafting effect where extractive structures can be grafted to predict counterfactual implications. We empirically show these effects in the OLMo-7b, Llama 3-8b, Gemma 2-9b, and Qwen 2-7b models. Of independent interest, our results also indicate that fact learning can occur at both early and late layers, which lead to different forms of generalization.

Contradiction retrieval refers to identifying and extracting documents that explicitly disagree with or refute the content of a query, which is important to many downstream applications like fact checking and data cleaning. To retrieve contradiction argument to the query from large document corpora, existing methods such as similarity search and cross-encoder models exhibit different limitations. To address these challenges, we introduce a novel approach: SparseCL that leverages specially trained sentence embeddings designed to preserve subtle, contradictory nuances between sentences. Our method utilizes a combined metric of cosine similarity and a sparsity function to efficiently identify and retrieve documents that contradict a given query. This approach dramatically enhances the speed of contradiction detection by reducing the need for exhaustive document comparisons to simple vector calculations. We conduct contradiction retrieval experiments on Arguana, MSMARCO, and HotpotQA, where our method produces an average improvement of $11.0\%$ across different models. We also validate our method on downstream tasks like natural language inference and cleaning corrupted corpora. This paper outlines a promising direction for non-similarity-based information retrieval which is currently underexplored.

Off-policy deep reinforcement learning (RL) agents typically leverage replay buffers for reusing past experiences during learning. This can help sample efficiency when the collected data is informative and aligned with the learning objectives; when that is not the case, it has the effect of ``polluting'' the replay buffer with data that can exacerbate optimization challenges in addition to wasting environment interactions due to redundant sampling. We argue that sampling these uninformative and wasteful transitions can be avoided by addressing the **sunk cost fallacy** which, in the context of deep RL, is the tendency towards continuing an episode until termination. To address this, we propose the *learn to stop* (**LEAST**) mechanism which uses statistics based on $Q$-values and gradient to guide early episode termination which helps agents recognize when to terminate unproductive episodes early. We demonstrate that our method improves learning efficiency on a variety of RL algorithms, evaluated on both the MuJoCo and DeepMind Control Suite benchmarks.

We study the task of list-decodable linear regression using batches, recently introduced by Das et al. 2023.. In this setting, we are given $m$ batches with each batch containing $n$ points in $\mathbb R^d$. A batch is called clean if the points it contains are i.i.d. samples from an unknown linear regression distribution. For a parameter $\alpha \in (0, 1/2)$, an unknown $\alpha$-fraction of the batches are clean and no assumptions are made on the remaining batches. The goal is to output a small list of vectors at least one of which is close to the true regressor vector in $\ell_2$-norm. Das et al. 2023 gave an efficient algorithm for this task, under natural distributional assumptions, with the following guarantee. Under the assumption that the batch size satisfies $n \geq \tilde{\Omega}(\alpha^{-1})$ and the total number of batches is $m = \text{poly}(d, n, 1/\alpha)$, their algorithm runs in polynomial time and outputs a list of $O(1/\alpha^2)$ vectors at least one of which is $\tilde{O}(\alpha^{-1/2}/\sqrt{n})$ close to the target regressor. Here we design a new polynomial-time algorithm for this task with significantly stronger guarantees under the assumption that the low-degree moments of the covariates distribution are Sum-of-Squares (SoS) certifiably bounded. Specifically, for any constant $\delta>0$, as long as the batch size is $n \geq \Omega_{\delta}(\alpha^{-\delta})$ and the degree-$\Theta(1/\delta)$ moments of the covariates are SoS certifiably bounded, our algorithm uses $m = \text{poly}((dn)^{1/\delta}, 1/\alpha)$ batches, runs in polynomial-time, and outputs an $O(1/\alpha)$-sized list of vectors one of which is $O(\alpha^{-\delta/2}/\sqrt{n})$ close to the target. That is, our algorithm substantially improves both the minimum batch size and the final error guarantee, while achieving the optimal list size. Our approach leverages higher-order moment information by carefully combining the SoS paradigm interleaved with an iterative method and a novel list pruning procedure for this setting. In the process, we give an SoS proof of the Marcinkiewicz-Zygmund inequality that may be of broader applicability.

Modern science emerged from reasoning over repeatedly-observed planetary motions. We present Gravity-Bench-v1, an environment-based benchmark that challenges AI agents on tasks that parallel this historical development. Gravity-Bench-v1 evaluates agents on the discovery of physics concealed within a dynamic environment, using rigorous gravitational dynamics simulations. Gravity-Bench includes out-of-distribution cases, i.e. with physics that deviates from the real world, to evaluate true scientific generalization capabilities. Agents must plan to collect data within an experimental budget and must perform a dynamic form of data analysis and reasoning to solve tasks efficiently. Our benchmark admits an open-ended space of solutions. Reference solutions for each task are provided to calibrate AI performance against human expertise. Technically at an upper-undergraduate level, our benchmark proves challenging to baseline AI agents. Gravity-Bench-v1 and planned extensions should help map out AI progress towards scientific discovery capabilities.

Large language models (LLMs) are increasingly used for long-content generation (e.g., long Chain-of-Thought reasoning) where decoding efficiency becomes a critical bottleneck: Autoregressive decoding is inherently limited by its sequential token generation process, where each token must be generated before the next can be processed. This sequential dependency restricts the ability to fully leverage modern hardware’s parallel processing capabilities. Existing methods like speculative decoding and layer skipping offer potential speedups but have notable drawbacks: speculative decoding relies on an auxiliary “drafter” model, which can be challenging to acquire and increases memory overhead, while layer skipping may introduce discrepancies in the generated outputs due to the missing key-value cache at skipped layers. In this work, we propose AdaDecode, which accelerates LLM decoding without requiring auxiliary models or changes to the original model parameters, while ensuring output consistency. AdaDecode leverages the insight that many tokens—particularly simple or highly-predictable ones—can accurately be generated at intermediate layers, as further layers often do not significantly alter predictions once the model reaches a certain confidence. By adaptively generating tokens at intermediate layers when confidence is high, AdaDecode enables the next token’s computation to begin immediately. The remaining layer computations for early-predicted tokens are deferred and executed in parallel with subsequent tokens when needed, maximizing hardware utilization and reducing decoding latency. A final verification step ensures that early predictions match the results of standard autoregressive decoding, preserving output parity. Experiments across diverse generation tasks shows that AdaDecode consistently achieves superior decoding throughput compared to baselines with up to 1.73$\times$ speedup, while guaranteeing output parity with standard autoregressive decoding.

Estimating Individual Treatment Effects (ITE) from observational data is challenging due to covariate shift and counterfactual absence. While existing methods attempt to balance distributions globally, they often lack fine-grained sample-level alignment, especially in scenarios with significant individual heterogeneity. To address these issues, we reconsider counterfactual as a proxy to emulate balanced randomization. Furthermore, we derive a theoretical bound that links the expected ITE estimation error to both factual prediction errors and representation distances between factuals and counterfactuals. Building on this theoretical foundation, we propose FCCL, a novel method designed to effectively capture the nuances of potential outcomes under different treatments by (i) generating diffeomorphic counterfactuals that adhere to the data manifold while maintaining high semantic similarity to their factual counterparts, and (ii) mitigating distribution shift via sample-level alignment grounded in our derived generalization-error bound, which considers factual-counterfactual similarity and category consistency. Extensive evaluations on benchmark datasets demonstrate that FCCL outperforms 13 state-of-the-art methods, particularly in capturing individual-level heterogeneity and handling sparse boundary samples.

We consider the problem of novel view synthesis from unposed images in a single feed-forward. Our framework capitalizes on fast speed, scalability, and high-quality 3D reconstruction and view synthesis capabilities of 3DGS, where we further extend it to offer a practical solution that relaxes common assumptions such as dense image views, accurate camera poses, and substantial image overlaps. We achieve this through identifying and addressing unique challenges arising from the use of pixel-aligned 3DGS: misaligned 3D Gaussians across different views induce noisy or sparse gradients that destabilize training and hinder convergence, especially when above assumptions are not met. To mitigate this, we employ pre-trained monocular depth estimation and visual correspondence models to achieve coarse alignments of 3D Gaussians. We then introduce lightweight, learnable modules to refine depth and pose estimates from the coarse alignments, improving the quality of 3D reconstruction and novel view synthesis. Furthermore, the refined estimates are leveraged to estimate geometry confidence scores, which assess the reliability of 3D Gaussian centers and condition the prediction of Gaussian parameters accordingly. Extensive evaluations on large-scale real-world datasets demonstrate that PF3plat sets a new state-of-the-art across all benchmarks, supported by comprehensive ablation studies validating our design choices. We will make the code and weights publicly available.

Current approaches for training Process Reward Models (PRMs) often involve deconposing responses into multiple reasoning steps using rule-based techniques, such as using predefined placeholder tokens or setting the reasoning step's length to a fixed size. These approaches overlook the fact that certain words don't usually indicate true decision points. To address this, we propose AdaptiveStep, a method that divides reasoning steps based on the model's confidence in predicting the next word, offering more information on decision-making at each step, improving downstream tasks like reward model training. Moreover, our method requires no manual annotation. Experiments with AdaptiveStep-trained PRMs in mathematical reasoning and code generation show that the outcome PRM achieves state-of-the-art Best-of-N performance, surpassing greedy search strategy with token-level value-guided decoding, while also reducing construction costs by over 30% compared to existing open-source PRMs. We also provide a thorough analysis and case study on its performance, transferability, and generalization capabilities. We provide our code on https://github.com/Lux0926/ASPRM.

Deep reinforcement learning for continuous control has recently achieved impressive progress. However, existing methods often suffer from primacy bias—a tendency to overfit early experiences stored in the replay buffer—which limits an RL agent’s sample efficiency and generalizability. A common existing approach to mitigate this issue is periodically resetting the agent during training. Yet, even after multiple resets, RL agents could still be impacted by early experiences. In contrast, humans are less susceptible to such bias, partly due to *infantile amnesia*, where the formation of new neurons disrupts early memory traces, leading to the forgetting of initial experiences. Inspired by this dual processes of forgetting and growing in neuroscience, in this paper, we propose *Forget and Grow* (**FoG**), a new deep RL algorithm with two mechanisms introduced. First, *Experience Replay Decay (ER Decay)*—"forgetting early experience''—which balances memory by gradually reducing the influence of early experiences. Second, *Network Expansion*—"growing neural capacity''—which enhances agents' capability to exploit the patterns of existing data by dynamically adding new parameters during training. Empirical results on four major continuous control benchmarks with more than 40 tasks demonstrate the superior performance of **FoG** against SoTA existing deep RL algorithms, including BRO, SimBa and TD-MPC2.

Training with diverse teammates is the key for learning generalizable agents. Typical approaches aim to generate diverse teammates by utilizing techniques like randomization, designing regularization terms, or reducing policy compatibility, etc. However, such teammates lack semantic information, resulting in inefficient teammate generation and poor adaptability of the agents. To tackle these challenges, we propose Semantically Diverse Teammate Generation (SemDiv), a novel framework leveraging the capabilities of large language models (LLMs) to discover and learn diverse coordination behaviors at the semantic level. In each iteration, SemDiv first generates a novel coordination behavior described in natural language, then translates it into a reward function to train a teammate policy. Once the policy is verified to be meaningful, novel, and aligned with the behavior, the agents train a policy for coordination. Through this iterative process, SemDiv efficiently generates a diverse set of semantically grounded teammates, enabling agents to develop specialized policies, and select the most suitable ones through language-based reasoning to adapt to unseen teammates. Experiments show that SemDiv generates teammates covering a wide range of coordination behaviors, including those unreachable by baseline methods. Evaluation across four MARL environments, each with five unseen representative teammates, demonstrates SemDiv's superior coordination and adaptability. Our code is available at https://github.com/lilh76/SemDiv.

We investigate the Sobolev IPM problem for probability measures supported on a graph metric space. Sobolev IPM is an important instance of integral probability metrics (IPM), and is obtained by constraining a critic function within a unit ball defined by the Sobolev norm. In particular, it has been used to compare probability measures and is crucial for several theoretical works in machine learning. However, to our knowledge, there are no efficient algorithmic approaches to compute Sobolev IPM effectively, which hinders its practical applications. In this work, we establish a relation between Sobolev norm and weighted $L^p$-norm, and leverage it to propose a *novel regularization* for Sobolev IPM. By exploiting the graph structure, we demonstrate that the regularized Sobolev IPM provides a *closed-form* expression for fast computation. This advancement addresses long-standing computational challenges, and paves the way to apply Sobolev IPM for practical applications, even in large-scale settings. Additionally, the regularized Sobolev IPM is negative definite. Utilizing this property, we design positive-definite kernels upon the regularized Sobolev IPM, and provide preliminary evidences of their advantages for comparing probability measures on a given graph for document classification and topological data analysis.

Distribution matching (DM) is a versatile domain-invariant representation learning technique that has been applied to tasks such as fair classification, domain adaptation, and domain translation. Non-parametric DM methods struggle with scalability and adversarial DM approaches suffer from instability and mode collapse. While likelihood-based methods are a promising alternative, they often impose unnecessary biases through fixed priors or require explicit density models (e.g., flows) that can be challenging to train. We address this limitation by introducing a novel approach to training likelihood-based DM using expressive score-based prior distributions. Our key insight is that gradient-based DM training only requires the prior's score function---not its density---allowing us to train the prior via denoising score matching. This approach eliminates biases from fixed priors (e.g., in VAEs), enabling more effective use of geometry-preserving regularization, while avoiding the challenge of learning an explicit prior density model (e.g., a flow-based prior). Our method also demonstrates better stability and computational efficiency compared to other diffusion-based priors (e.g., LSGM). Furthermore, experiments demonstrate superior performance across multiple tasks, establishing our score-based method as a stable and effective approach to distribution matching. Source code available at https://github.com/inouye-lab/SAUB.

Recent research has shown an increasing interest in utilizing pre-trained large language models (LLMs) for a variety of time series applications. However, there are three main challenges when using LLMs as foundational models for time series forecasting: (1) Cross-domain generalization. (2) Cross-modality alignment. (3) Error accumulation in autoregressive frameworks. To address these challenges, we proposed **LangTime**, a **lan**guage-**g**uided unified model for **time** series forecasting that incorporates cross-domain pre-training with reinforcement learning-based fine-tuning. Specifically, LangTime constructs Temporal Comprehension Prompts (TCPs), which include dataset-wise and channel-wise instructions, to facilitate domain adaptation and condense time series into a single token, enabling LLMs to understand better and align temporal data. To improve autoregressive forecasting, we introduce TimePPO, a reinforcement learning-based fine-tuning algorithm. TimePPO mitigates error accumulation by leveraging a multidimensional rewards function tailored for time series and a repeat-based value estimation strategy. Extensive experiments demonstrate that LangTime achieves state-of-the-art cross-domain forecasting performance, while TimePPO fine-tuning effectively enhances the stability and accuracy of autoregressive forecasting.

In this paper, we present DiffusionVLA, a novel framework that integrates autoregressive reasoning with diffusion policies to address the limitations of existing methods: while autoregressive Vision-Language-Action (VLA) models lack precise and robust action generation, diffusion-based policies inherently lack reasoning capabilities. Central to our approach is autoregressive reasoning — a task decomposition and explanation process enabled by a pre-trained VLM — to guide diffusion-based action policies. To tightly couple reasoning with action generation, we introduce a reasoning injection module that directly embeds self-generated reasoning phrases into the policy learning process. The framework is simple, flexible, and efficient, enabling seamless deployment across diverse robotic platforms. We conduct extensive experiments using multiple real robots to validate the effectiveness of DiVLA. Our tests include a challenging factory sorting task, where DiVLA successfully categorizes objects, including those not seen during training. The reasoning injection module enhances interpretability, enabling explicit failure diagnosis by visualizing the model’s decision process. Additionally, we test DiVLA on a zero-shot bin-picking task, achieving \textbf{63.7\% accuracy on 102 previously unseen objects}. Our method demonstrates robustness to visual changes, such as distractors and new backgrounds, and easily adapts to new embodiments. Furthermore, DiVLA can follow novel instructions and retain conversational ability. Notably, DiVLA is data-efficient and fast at inference; our smallest DiVLA-2B runs 82Hz on a single A6000 GPU. Finally, we scale the model from 2B to 72B parameters, showcasing improved generalization capabilities with increased model size.