NEURIPS 2025

Despite the recent progresses, particularly in developing Language Models, there are fundamental challenges and unanswered questions about how such models can continually learn/memorize, self-improve, and find effective solutions. In this paper, we present a new learning paradigm, called Nested Learning (NL), that coherently represents a machine learning model with a set of nested, multi-level, and/or parallel optimization problems, each of which with its own context flow. Through the lenses of NL, existing deep learning methods learns from data through compressing their own context flow, and in-context learning naturally emerges in large models. NL suggests a philosophy to design more expressive learning algorithms with more levels, resulting in higher-order in-context learning and potentially unlocking effective continual learning capabilities. We advocate for NL by presenting three core contributions: (1) Expressive Optimizers: We show that known gradient-based optimizers, such as Adam, SGD with Momentum, etc., are in fact associative memory modules that aim to compress the gradients' information (by gradient descent). Building on this insight, we present other more expressive optimizers with deep memory and/or more powerful learning rules; (2) Self-Modifying Learning Module: Taking advantage of NL's insights on learning algorithms, we present a sequence model that learns how to modify itself by learning its own update algorithm; and (3) Continuum Memory System: We present a new formulation for memory system that generalizes the traditional viewpoint of long/short-term memory. Combining our self-modifying sequence model with the continuum memory system, we present a continual learning module, called Hope, showing promising results in language modeling, knowledge incorporation, and few-shot generalization tasks, continual learning, and long-context reasoning tasks.

Binary choices, as often used for reinforcement learning from human feedback (RLHF), convey only the direction of a preference. A person may choose apples over oranges and bananas over grapes, but which preference is stronger? Strength is crucial for decision-making under uncertainty and generalization of preference models, but hard to measure reliably. Metadata such as response times and inter-annotator agreement can serve as proxies for strength, but are often noisy and confounded. We propose ResponseRank to address the challenge of learning from noisy strength signals. Our method uses relative differences in proxy signals to rank responses to pairwise comparisons by their inferred preference strength. To control for systemic variation, we compare signals only locally within carefully constructed strata. This enables robust learning of utility differences consistent with strength-derived rankings while making minimal assumptions about the strength signal. Our contributions are threefold: (1) ResponseRank, a novel method that robustly learns preference strength by leveraging locally valid relative strength signals; (2) empirical evidence of improved sample efficiency and robustness across diverse tasks: synthetic preference learning (with simulated response times), language modeling (with annotator agreement), and RL control tasks (with simulated episode returns); and (3) the Pearson Distance Correlation (PDC), a novel metric that isolates cardinal utility learning from ordinal accuracy.

The balanced loss is a widely adopted objective for multi-class classification under class imbalance. By assigning equal importance to all classes, regardless of their frequency, it promotes fairness and ensures that minority classes are not overlooked. However, directly minimizing the balanced classification loss is typically intractable, which makes the design of effective surrogate losses a central question. This paper introduces and studies two advanced surrogate loss families: Generalized Logit-Adjusted (GLA) loss functions and Generalized Class-Aware weighted (GCA) losses. GLA losses generalize Logit-Adjusted losses, which shift logits based on class priors, to the broader general cross-entropy loss family. GCA loss functions extend the standard class-weighted losses, which scale losses inversely by class frequency, by incorporating class-dependent confidence margins and extending them to the general cross-entropy family. We present a comprehensive theoretical analysis of consistency for both loss families. We show that GLA losses are Bayes-consistent, but only $H$-consistent for complete (i.e., unbounded) hypothesis sets. Moreover, their $H$-consistency bounds depend inversely on the minimum class probability, scaling at least as $1/\mathsf p_{\min}$. In contrast, GCA losses are $H$-consistent for any hypothesis set that is bounded or complete, with $H$-consistency bounds that scale more favorably as $1/\sqrt{\mathsf p_{\min}}$, offering significantly stronger theoretical guarantees in imbalanced settings. We report the results of experiments demonstrating that, empirically, both the GCA losses with calibrated class-dependent confidence margins and GLA losses can greatly outperform straightforward class-weighted losses as well as the LA losses. GLA generally performs slightly better in common benchmarks, whereas GCA exhibits a slight edge in highly imbalanced settings.

Egocentric Referring Video Object Segmentation (Ego-RVOS) aims to segment the specific object actively involved in a human action, as described by a language query, within first-person videos. This task is critical for understanding egocentric human behavior. However, achieving such segmentation robustly is challenging due to ambiguities inherent in egocentric videos and biases present in training data. Consequently, existing methods often struggle, learning spurious correlations from skewed object-action pairings in datasets and fundamental visual confounding factors of the egocentric perspective, such as rapid motion and frequent occlusions. To address these limitations, we introduce Causal Ego-REferring Segmentation (CERES), a plug-in causal framework that adapts strong, pre-trained RVOS backbones to the egocentric domain. CERES implements dual-modal causal intervention: applying backdoor adjustment principles to counteract language representation biases learned from dataset statistics, and leveraging front-door adjustment concepts to address visual confounding by intelligently integrating semantic visual features with geometric depth information guided by causal principles, creating representations more robust to egocentric distortions. Extensive experiments demonstrate that CERES achieves state-of-the-art performance on Ego-RVOS benchmarks, highlighting the potential of applying causal reasoning to build more reliable models for broader egocentric video understanding.

Neural processes (NPs) are a class of models that learn stochastic processes directly from data and can be used for inference, sampling and conditional sampling. We introduce a new NP model based on flow matching, a generative modeling paradigm that has demonstrated strong performance on various data modalities. Following the NP training framework, the model provides amortized predictions of conditional distributions over any arbitrary points in the data. Compared to previous NP models, our model is simple to implement and can be used to sample from conditional distributions using an ODE solver, without requiring auxiliary conditioning methods. In addition, the model provides a controllable tradeoff between accuracy and running time via the number of steps in the ODE solver. We show that our model outperforms previous state-of-the-art neural process methods on various benchmarks including synthetic 1D Gaussian processes data, 2D images, and real-world weather data.

Speculative decoding improves LLM inference by generating and verifying multiple tokens in parallel, but existing systems suffer from suboptimal performance due to a mismatch between dynamic speculation and static runtime assumptions. We present Yggdrasil, a co-designed system that enables latency-optimal speculative decoding through context-aware tree drafting and compiler-friendly execution. Yggdrasil introduces an equal-growth tree structure for static graph compatibility, a latency-aware optimization objective for draft selection, and stage-based scheduling to reduce overhead. Yggdrasil supports unmodified LLMs and achieves up to $3.98\times$ speedup over state-of-the-art baselines across multiple hardware setups.

This paper presents JavisGPT, the first unified multimodal large language model (MLLM) for joint audio-video (JAV) comprehension and generation. JavisGPT has a concise encoder-LLM-decoder architecture, which has a SyncFusion module for spatio-temporal audio-video fusion and synchrony-aware learnable queries to bridge a pretrained JAV-DiT generator. This design enables temporally coherent video-audio understanding and generation from multimodal instructions. We design an effective three-stage training pipeline consisting of multimodal pretraining, audio-video fine-tuning, and large-scale instruction-tuning, to progressively build multimodal comprehension and generation from existing vision-language models. For instruction tuning, we construct JavisInst-Omni, a high-quality instruction dataset with over 200K GPT-4o-curated audio-video-text dialogues that cover diverse and multi-level comprehension and generation scenarios. On JAV comprehension and generation benchmarks, our experiments show that JavisGPT outperforms existing MLLMs, particularly in complex and temporally synchronized settings.

In machine learning, effective modeling requires a holistic consideration of how to encode inputs, make predictions (i.e., decoding), and train the model. However, in time-series forecasting, prior work has predominantly focused on encoder design, often treating prediction and training as separate or secondary concerns. In this paper, we propose TimePerceiver, a unified encoder-decoder forecasting framework that is tightly aligned with an effective training strategy. To be specific, we first generalize the forecasting task to include diverse temporal prediction objectives such as extrapolation, interpolation, and imputation. Since this generalization requires handling input and target segments that are arbitrarily positioned along the temporal axis, we design a novel encoder-decoder architecture that can flexibly perceive and adapt to these varying positions. For encoding, we introduce a set of latent bottleneck representations that can interact with all input segments to jointly capture temporal and cross-channel dependencies. For decoding, we leverage learnable queries corresponding to target timestamps to effectively retrieve relevant information. Extensive experiments demonstrate that our framework consistently and significantly outperforms prior state-of-the-art baselines across a wide range of benchmark datasets. The code is available at https://github.com/efficient-learning-lab/TimePerceiver.

In the big data era, the computer vision field benefits from large-scale datasets such as LAION-2B, LAION-400M, and ImageNet-21K, Kinetics, on which popular models like the ViT and ConvNeXt series have been pre-trained, acquiring substantial knowledge. However, numerous downstream tasks in specialized and data-limited scientific domains continue to pose significant challenges. In this paper, we propose a novel Cluster Attention Adapter (CLAdapter), which refines and adapts the rich representations learned from large-scale data to various data-limited downstream tasks. Specifically, CLAdapter introduces attention mechanisms and cluster centers to personalize the enhancement of transformed features through distribution correlation and transformation matrices. This enables models fine-tuned with CLAdapter to learn distinct representations tailored to different feature sets, facilitating the models' adaptation from rich pre-trained features to various downstream scenarios effectively. In addition, CLAdapter's unified interface design allows for seamless integration with multiple model architectures, including CNNs and Transformers, in both 2D and 3D contexts. Through extensive experiments on 10 datasets spanning domains such as generic, multimedia, biological, medical, industrial, agricultural, environmental, geographical, materials science, out-of-distribution (OOD), and 3D analysis, CLAdapter achieves state-of-the-art performance across diverse data-limited scientific domains, demonstrating its effectiveness in unleashing the potential of foundation vision models via adaptive transfer. Code is available at https://github.com/qklee-lz/CLAdapter.

Humans can efficiently extract knowledge and learn skills from the videos within only a few trials and errors. However, it poses a big challenge to replicate this learning process for autonomous agents, due to the complexity of visual input, the absence of action or reward signals, and the limitations of interaction steps. In this paper, we propose a novel, unsupervised, and sample-efficient framework to achieve imitation learning from videos (ILV), named Behavior Cloning from Videos via Latent Representations (BCV-LR). BCV-LR extracts action-related latent features from high-dimensional video inputs through self-supervised tasks, and then leverages a dynamics-based unsupervised objective to predict latent actions between consecutive frames. The pre-trained latent actions are fine-tuned and efficiently aligned to the real action space online (with collected interactions) for policy behavior cloning. The cloned policy in turn enriches the agent experience for further latent action finetuning, resulting in an iterative policy improvement that is highly sample-efficient. We conduct extensive experiments on a set of challenging visual tasks, including both discrete control and continuous control. BCV-LR enables effective (even expert-level on some tasks) policy performance with only a few interactions, surpassing state-of-the-art ILV baselines and reinforcement learning methods (provided with environmental rewards) in terms of sample efficiency across 24/28 tasks. To the best of our knowledge, this work for the first time demonstrates that videos can support extremely sample-efficient visual policy learning, without the need to access any other expert supervision.

Novel view synthesis (NVS) seeks to render photorealistic, 3D-consistent images of a scene from unseen camera poses given only a sparse set of posed views. Existing deterministic networks render observed regions quickly but blur unobserved areas, whereas stochastic diffusion-based methods hallucinate plausible content yet incur heavy training- and inference-time costs. In this paper, we propose a hybrid framework that unifies the strengths of both paradigms. A bidirectional transformer encodes multi-view image tokens and Plucker-ray embeddings, producing a shared latent representation. Two lightweight heads then act on this representation: (i) a feed-forward regression head that renders pixels where geometry is well constrained, and (ii) a masked autoregressive diffusion head that completes occluded or unseen regions. The entire model is trained end-to-end with joint photometric and diffusion losses, without handcrafted 3D inductive biases, enabling scalability across diverse scenes. Experiments demonstrate that our method attains state-of-the-art image quality while reducing rendering time by an order of magnitude compared with fully generative baselines.

We introduce a novel deep learning algorithm for computing convex conjugates of differentiable convex functions, a fundamental operation in convex analysis with various applications in different fields such as optimization, control theory, physics and economics. While traditional numerical methods suffer from the curse of dimensionality and become computationally intractable in high dimensions, more recent neural network-based approaches scale better, but have mostly been studied with the aim of solving optimal transport problems and require the solution of complicated optimization or max-min problems. Using an implicit Fenchel formulation of convex conjugation, our approach facilitates an efficient gradient-based framework for the minimization of approximation errors and, as a byproduct, also provides a posteriori error estimates for the approximation quality. Numerical experiments demonstrate our method's ability to deliver accurate results across different high-dimensional examples. Moreover, by employing symbolic regression with Kolmogorov--Arnold networks, it is able to obtain the exact convex conjugates of specific convex functions.

Transformers can implement both generalizable algorithms (e.g., induction heads) and simple positional shortcuts (e.g., memorizing fixed output positions). In this work, we study how the choice of pretraining data distribution steers a shallow transformer toward one behavior or the other. Focusing on a minimal trigger-output prediction task -- copying the token immediately following a special trigger upon its second occurrence -- we present a rigorous analysis of gradient-based training of a single-layer transformer. In both the infinite and finite sample regimes, we prove a transition in the learned mechanism: if input sequences exhibit sufficient diversity, measured by a low ``max-sum'' ratio of trigger-to-trigger distances, the trained model implements an induction head and generalizes to unseen contexts; by contrast, when this ratio is large, the model resorts to a positional shortcut and fails to generalize out-of-distribution (OOD). We also reveal a trade-off between the pretraining context length and OOD generalization, and derive the optimal pretraining distribution that minimizes computational cost per sample. Finally, we validate our theoretical predictions with controlled synthetic experiments, demonstrating that broadening context distributions robustly induces induction heads and enables OOD generalization. Our results shed light on the algorithmic biases of pretrained transformers and offer conceptual guidelines for data-driven control of their learned behaviors.

Understanding architectural differences in language models is challenging, especially at academic-scale pretraining (e.g., 1.3B parameters, 100B tokens), where results are often dominated by noise and randomness. To overcome this, we introduce controlled synthetic pretraining tasks that isolate and evaluate core model capabilities. Within this framework, we discover CANON LAYERS: lightweight architectural components -- named after the musical term "canon" -- that promote horizontal information flow across neighboring tokens. Canon layers compute weighted sums of nearby token representations and integrate seamlessly into Transformers, linear attention, state-space models, or any sequence architecture. We present 12 key results. This includes how Canon layers enhance reasoning depth (e.g., by $2\times$), reasoning breadth, knowledge manipulation, etc. They lift weak architectures like NoPE to match RoPE, and linear attention to rival SOTA linear models like Mamba2/GDN -- validated both through synthetic tasks and real-world academic-scale pretraining. This synthetic playground offers an economical, principled path to isolate core model capabilities often obscured at academic scales. Equipped with infinite high-quality data, it may even PREDICT how future architectures will behave as training pipelines improve -- e.g., through better data curation or RL-based post-training -- unlocking deeper reasoning and hierarchical inference.

Consistent pose-driven character animation has achieved remarkable progress in single-character scenarios. However, extending these advances to multi-character settings is non-trivial, especially when position swap is involved. Beyond mere scaling, the core challenge lies in enforcing correct Identity Correspondence (IC) between characters in reference and generated frames. To address this, we introduce EverybodyDance, a systematic solution targeting IC correctness in multi-character animation. EverybodyDance is built around the Identity Matching Graph (IMG), which models characters in the generated and reference frames as two node sets in a weighted complete bipartite graph. Edge weights, computed via our proposed Mask-Query Attention (MQA), quantify the affinity between each pair of characters. Our key insight is to formalize IC correctness as a graph structural metric and to optimize it during training. We also propose a series of targeted strategies tailored for multi-character animation, including identity-embedded guidance, a multi-scale matching strategy, and pre-classified sampling, which work synergistically. Finally, to evaluate IC performance, we curate the Identity Correspondence Evaluation benchmark, dedicated to multi-character IC correctness. Extensive experiments demonstrate that EverybodyDance substantially outperforms state-of-the-art baselines in both IC and visual fidelity.

The creation of photorealistic dynamic hair remains a major challenge in digital human modeling because of the complex motions, occlusions, and light scattering. Existing methods often resort to static capture and physics-based models that do not scale as they require manual parameter fine-tuning to handle the diversity of hairstyles and motions, and heavy computation to obtain high-quality appearance. In this paper, we present Dynamic Gaussian Hair (DGH), a novel framework that efficiently learns hair dynamics and appearance. We propose: (1) a coarse-to-fine model that learns temporally coherent hair motion dynamics across diverse hairstyles; (2) a strand-guided optimization module that learns a dynamic 3D Gaussian representation for hair appearance with support for differentiable rendering, enabling gradient-based learning of view-consistent appearance under motion. Unlike prior simulation-based pipelines, our approach is fully data-driven, scales with training data, and generalizes across various hairstyles and head motion sequences. Additionally, DGH can be seamlessly integrated into a 3D Gaussian avatar framework, enabling realistic, animatable hair for high-fidelity avatar representation. DGH achieves promising geometry and appearance results, providing a scalable, data-driven alternative to physics-based simulation and rendering.

The success of modern multimodal representation learning relies on internet-scale datasets. Due to the low quality of a large fraction of raw web data, data curation has become a critical step in the training pipeline. Filtering using a trained model (i.e., teacher-based filtering) has emerged as a successful solution, leveraging a pre-trained model to compute quality scores. To explain the empirical success of teacher-based filtering, we characterize the performance of filtered contrastive learning under the standard bimodal data generation model. Denoting $η\in(0,1]$ as the fraction of data with correctly matched modalities among $n$ paired samples, we utilize a linear contrastive learning setup to show a provable benefit of data filtering: $(i)$ the error without filtering is upper and lower bounded by $\frac{1}{η\sqrt{n}}$, and $(ii)$ the error with teacher-based filtering is upper bounded by $\frac{1}{\sqrt{ηn}}$ in the large $η$ regime, and by $\frac{1}{\sqrt{n}}$ in the small $η$ regime.

Contrastive Language-Image Pretraining (CLIP) achieves strong generalization in vision-language tasks by aligning images and texts in a shared embedding space. However, recent findings show that CLIP-like models still underutilize fine-grained semantic signals in text, and this issue becomes even more pronounced when dealing with long and detailed captions. This stems from CLIP's training objective, which optimizes only global image-text similarity and overlooks token-level supervision - limiting its ability to achieve fine-grained visual-text alignment. To address this, we propose SuperCLIP, a simple yet effective framework that augments contrastive learning with classification-based supervision. By adding only a lightweight linear layer to the vision encoder, SuperCLIP leverages token-level cues to enhance visual-textual alignment - with just a 0.077% increase in total FLOPs, and no need for additional annotated data. Experiments show that SuperCLIP consistently improves zero-shot classification, image-text retrieval, and purely visual tasks. These gains hold regardless of whether the model is trained on original web data or rich re-captioned data, demonstrating SuperCLIP's ability to recover textual supervision in both cases. Furthermore, SuperCLIP alleviates CLIP's small-batch performance drop through classification-based supervision that avoids reliance on large batch sizes. Code and models will be made open source.

We propose VASA-3D, an audio-driven, single-shot 3D head avatar generator. This research tackles two major challenges: capturing the subtle expression details present in real human faces, and reconstructing an intricate 3D head avatar from a single portrait image. To accurately model expression details, VASA-3D leverages the motion latent of VASA-1, a method that yields exceptional realism and vividness in 2D talking heads. A critical element of our work is translating this motion latent to 3D, which is accomplished by devising a 3D head model that is conditioned on the motion latent. Customization of this model to a single image is achieved through an optimization framework that employs numerous video frames of the reference head synthesized from the input image. The optimization takes various training losses robust to artifacts and limited pose coverage in the generated training data. Our experiment shows that VASA-3D produces realistic 3D talking heads that cannot be achieved by prior art, and it supports the online generation of 512x512 free-viewpoint videos at up to 75 FPS, facilitating more immersive engagements with lifelike 3D avatars.

A single biomedical image can be meaningfully segmented in multiple ways, depending on the desired application. For instance, a brain MRI can be segmented according to tissue types, vascular territories, broad anatomical regions, fine-grained anatomy, or pathology, etc. Existing automatic segmentation models typically either (1) support only a single protocol, the one they were trained on, or (2) require labor-intensive manual prompting to specify the desired segmentation. We introduce Pancakes, a framework that, given a new image from a previously unseen domain, automatically generates multi-label segmentation maps for multiple plausible protocols, while maintaining semantic consistency across related images. Pancakes introduces a new problem formulation that is not currently attainable by existing foundation models. In a series of experiments on seven held-out datasets, we demonstrate that our model can significantly outperform existing foundation models in producing several plausible whole-image segmentations, that are semantically coherent across images.