ICLR 2026 论文列表

Recent advances in 3D Gaussian Splatting (3DGS) have revolutionized scene reconstruction, opening new possibilities for 3D steganography by hiding 3D secrets within 3D covers. The key challenge in steganography is ensuring imperceptibility while maintaining high-fidelity reconstruction. However, existing methods often suffer from detectability risks and utilize only suboptimal 3DGS attributes, limiting their full potential. We propose a novel end-to-end key-secured 3D steganography framework (KeySS) that jointly optimizes a 3DGS model and a key-secured decoder for secret reconstruction. Our approach reveals that Gaussian attributes contribute unequally to secret hiding. The framework incorporates a key-controllable mechanism enabling multi-secret hiding and unauthorized access prevention, while systematically exploring optimal attribute update to balance fidelity and security. To rigorously evaluate steganographic imperceptibility beyond conventional 2D metrics, we introduce 3D-Sinkhorn distance analysis, which quantifies distributional differences between original and steganographic Gaussian parameters in the representation space. Extensive experiments show that our method achieves state-of-the-art performance in 3D reconstruction while ensuring high levels of steganographic security. The framework is highly efficient and readily extensible to multi-GPU training. Our code will be publicly available.

Speech-to-Speech (S2S) models have shown promising dialogue capabilities, but their ability to handle paralinguistic cues—such as emotion, tone, and speaker attributes—and to respond appropriately in both content and style remains underexplored. Progress is further hindered by the scarcity of high-quality and expressive demonstrations. To address this, we introduce a novel reinforcement learning (RL) framework for paralinguistic-aware S2S, ParaS2S, which evaluates and optimizes both content and speaking style directly at the waveform level. We first construct ParaS2SBench, a benchmark comprehensively evaluates S2S models’ output for content and style appropriateness from diverse and challenging input queries. It scores the fitness of input-output pairs and aligns well with human judgments, serving as an automatic judge for model outputs. With this scalable scoring feedback, we enable the model to explore and learn from diverse unlabeled speech via Group Relative Policy Optimization (GRPO). Experiments show that existing S2S models fail to respond appropriately to paralinguistic attributes, performing no better than pipeline-based baselines. Our RL-based strong baseline achieves a 11% relative improvement in response content and style’s appropriateness on ParaS2SBench over supervised fine-tuning (SFT), surpassing all prior models while requiring substantially fewer warm-up annotations than pure SFT.

Replication of experimental results has been a challenge faced by many scientific disciplines, including the field of machine learning. Recent work on the theory of machine learning has formalized replicability as the demand that an algorithm produce identical outcomes when executed twice on different samples from the same distribution. Provably replicable algorithms are especially interesting for reinforcement learning (RL), where algorithms are known to be unstable in practice. While replicable algorithms exist for tabular RL settings, extending these guarantees to more practical function approximation settings has remained an open problem. In this work, we make progress by developing replicable methods for linear function approximation in RL. We first introduce two efficient algorithms for replicable random design regression and uncentered covariance estimation, each of independent interest. We then leverage these tools to provide the first provably efficient replicable RL algorithms for linear Markov decision processes in both the generative model and episodic settings. Finally, we evaluate our algorithms experimentally and show how they can inspire more consistent neural policies.

While visual-language models have profoundly linked features between texts and images, the incorporation of 3D modality data, such as point clouds and 3D Gaussians, further enables pretraining for 3D-related tasks, e.g., cross-modal retrieval, zero-shot classification, and scene recognition. As challenges remain in extracting 3D modal features and bridging the gap between different modalities, we propose TIGaussian, a framework that harnesses 3D Gaussian Splatting (3DGS) characteristics to strengthen cross-modality alignment through multi-branch 3DGS tokenizer and modality-specific 3D feature alignment strategies. Specifically, our multi-branch 3DGS tokenizer decouples the intrinsic properties of 3DGS structures into compact latent representations, enabling more generalizable feature extraction. To further bridge the modality gap, we develop a bidirectional cross-modal alignment strategies: a multi-view feature fusion mechanism that leverages diffusion priors to resolve perspective ambiguity in image-3D alignment, while a text-3D projection module adaptively maps 3D features to text embedding space for better text-3D alignment. Extensive experiments on various datasets demonstrate the state-of-the-art performance of TIGaussian in multiple tasks. Code repository: https://github.com/RUiN-jiarun/TIGaussian.

Reinforcement finetuning (RFT) has garnered significant attention in recent years, particularly for enhancing large reasoning models such as OpenAI o1 and Deepseek R1. The appeal of RFT largely stems from its ability to refine model knowledge, better align outputs with user intent, and address challenges associated with limited finetuning data. Despite these advantages, the application of RFT in large Transformer-based generative agents remains relatively underexplored. Although these agents are designed to address multiple tasks through large-scale autoregressive pretraining and share many properties with large reasoning models, current adaptation strategies predominantly rely on supervised finetuning (SFT). In this work, we conduct a systematic investigation of several RFT techniques across a variety of finetuning parameter configurations and meta-reinforcement learning (meta-RL) environments, employing few-shot offline datasets. We provide a comprehensive analysis of RFT algorithm performance under diverse experimental conditions and, based on our empirical findings, introduce a lightweight enhancement to existing RFT methods. This enhancement consistently improves outcomes by combining the strengths of both SFT and RFT. Our findings provide valuable insights for advancing the effectiveness of RFT approaches and broadening their applicability to meta-RL tasks with large Transformer-based generative agents, motivating further research in broader domains.

Large Language Models (LLMs) have achieved impressive performance on complex tasks by generating human-like, step-by-step rationales, referred to as \textit{reasoning trajectory}, before arriving at final answers. However, the length of these reasoning trajectories often far exceeds that of the final answers, which incurs substantial inference costs even for relatively simple tasks. Advanced methods typically attempt to compress reasoning trajectory length through post-training, but they remain decoding-intensive and fail to inherently mitigate the efficiency challenge. In this work, we challenge the necessity of generating full reasoning trajectories and empirically demonstrate that LLMs can generate accurate answers using only fragmental reasoning paths, without relying on complete token-by-token sequences. To this end, we propose a novel \textbf{Latent Reasoning Tuning (LRT)} framework, which empowers LLMs to perform reasoning using implicit, compact, learnable representations instead of explicit textual trajectories. Technically, LRT replaces the costly autoregressive generation of reasoning steps with a single forward pass through a lightweight reasoning network, which generates latent vectors that encapsulate the necessary reasoning logic and condition the LLM to produce the final answer. Experiments on mathematical and out-of-domain benchmarks demonstrate that our LRT consistently outperforms relevant efficient reasoning methods. Moreover, by transforming explicit reasoning into latent reasoning, our approach surpasses the state-of-the-art Qwen3 hybrid reasoning framework.

Front-end development constitutes a substantial portion of software engineering, yet converting design mockups into production-ready *User Interface* (UI) code remains tedious and time-costly. While recent work has explored automating this process with *Multimodal Large Language Models* (MLLMs), existing approaches typically rely solely on design images. As a result, they must infer complex UI details from images alone, often leading to degraded results. In real-world development workflows, however, design mockups are usually delivered as Figma files—a widely used tool for front-end design—that embed rich multimodal information (e.g., metadata and assets) essential for generating high-quality UI. To bridge this gap, we introduce Figma2Code, a new task that generalizes *design-to-code* into a multimodal setting and aims to automate *design-to-code* in the wild. Specifically, we collect paired design images and their corresponding metadata files from the Figma community. We then apply a series of processing operations, including rule-based filtering, human and MLLM-based annotation and screening, and metadata refinement. This process yields 3,055 samples, from which designers curate a balanced dataset of 213 high-quality cases. Using this dataset, we benchmark ten state-of-the-art open-source and proprietary MLLMs. Our results show that while proprietary models achieve superior visual fidelity, they remain limited in layout responsiveness and code maintainability. Further experiments across modalities and ablation studies corroborate this limitation, partly due to models’ tendency to directly map primitive visual attributes from Figma metadata.

As large language models (LLMs) continue to scale up, their performance on various downstream tasks has significantly improved. However, evaluating their capabilities has become increasingly expensive, as performing inference on a large number of benchmark samples incurs high computational costs. In this paper, we revisit the model-item performance matrix and show that it exhibits sparsity, that representative items can be selected as anchors, and that the task of efficient benchmarking can be formulated as a sparse optimization problem. Based on these insights, we propose SparseEval, a method that, for the first time, adopts gradient descent to optimize anchor weights and employs an iterative refinement strategy for anchor selection. We utilize the representation capacity of MLP to handle sparse optimization and propose the Anchor Importance Score and Candidate Importance Score to evaluate the value of each item for task-aware refinement. Extensive experiments demonstrate the low estimation error and high Kendall’s~$\tau$ of our method across a variety of benchmarks, showcasing its superior robustness and practicality in real-world scenarios.

Attention-based multiple instance learning (MIL) has emerged as a powerful framework for whole slide image (WSI) diagnosis, leveraging attention to aggregate instance-level features into bag-level predictions. Despite this success, we find that such methods exhibit a new failure mode: unstable attention dynamics. Across four representative attention-based MIL methods and two public WSI datasets, we observe that attention distributions oscillate across epochs rather than converging to a consistent pattern, degrading performance. This instability adds to two previously reported challenges: overfitting and over-concentrated attention distribution. To simultaneously overcome these three limitations, we introduce attention-stabilized multiple instance learning (ASMIL), a novel unified framework. ASMIL uses an anchor model to stabilize attention, replaces softmax with a normalized sigmoid function in the anchor to prevent over-concentration, and applies token random dropping to mitigate overfitting. Extensive experiments demonstrate that ASMIL achieves up to a 6.49% F1 score improvement over state-of-the-art methods. Moreover, integrating the anchor model and normalized sigmoid into existing attention-based MIL methods consistently boosts their performance, with F1 score gains up to 10.73%. All code and data are publicly available at https://anonymous.4open.science/r/ASMIL-5018/.

Spectral bias, the tendency of neural networks to learn low frequencies first, can be both a blessing and a curse. While it enhances the generalization capabilities by suppressing high-frequency noise, it can be a limitation in scientific tasks that require capturing fine-scale structures. The delayed generalization phenomenon known as grokking is another barrier to rapid training of neural networks. Grokking has been hypothesized to arise as learning transitions from the NTK to the feature-rich regime. This paper explores the impact of preconditioned gradient descent (PGD), such as Gauss-Newton, on spectral bias and grokking phenomena. We demonstrate through theoretical and empirical results how PGD can mitigate issues associated with spectral bias. Additionally, building on the rich learning grokking hypothesis, we study how PGD can be used to reduce delays associated with grokking. Our conjecture is that PGD, without the impediment of spectral bias, enables uniform exploration of the parameter space in the NTK regime. Our experimental results confirm this prediction, providing strong evidence that grokking represents a transitional behavior between the lazy regime characterized by the NTK and the rich regime. These findings deepen our understanding of the interplay between optimization dynamics, spectral bias, and the phases of neural network learning.

Estimating causal quantities (CQs) typically requires large datasets, which can be expensive to obtain, especially when measuring individual outcomes is costly. This challenge highlights the importance of sample-efficient active learning strategies. To address the narrow focus of prior work on the conditional average treatment effect, we formalize the broader task of Actively estimating Causal Quantities (ActiveCQ) and propose a unified framework for this general problem. Built upon the insight that many CQs are integrals of regression functions, our framework models the regression function with a Gaussian Process. For the distribution component, we explore both a baseline using explicit density estimators and a more integrated method using conditional mean embeddings in a reproducing kernel Hilbert space. This latter approach offers key advantages: it bypasses explicit density estimation, operates within the same function space as the GP, and adaptively refines the distributional model after each update. Our framework enables the principled derivation of acquisition strategies from the CQ's posterior uncertainty; we instantiate this principle with two utility functions based on information gain and total variance reduction. A range of simulated and semi-synthetic experiments demonstrate that our principled framework significantly outperforms relevant baselines, achieving substantial gains in sample efficiency across a variety of CQs.

Universal multimodal retrieval (UMR), which aims to address complex retrieval tasks where both queries and candidates span diverse modalities, has been significantly advanced by the emergence of MLLMs. While state-of-the-art MLLM-based methods in the literature predominantly adopt contrastive learning principles, they often differ in their specific training recipes. Despite their success, the mechanisms underlying their retrieval capabilities remain largely unexplored, potentially resulting in suboptimal performance and limited generalization ability. To address these issues, we present a comprehensive study aimed at uncovering the key factors that drive effective embedding learning for UMR using MLLMs. We begin by implementing a general MLLM-based embedding learning pipeline, and systematically analyze the primary contributors to high-performing universal retrieval systems. Based on this, we explore various aspects of the details in embedding generation and training strategies, including progressive transition, hard negative mining and re-ranker distillation. Notably, our findings reveal that often-overlooked factors can have a substantial impact on model performance. Building on these discoveries, we introduce a unified framework termed U-MARVEL (Universal MultimodAl RetrieVal via Embedding Learning), which outperforms state-of-the-art competitors on the M-BEIR benchmark by a large margin in supervised settings, and also exhibits strong zero-shot performance on several tasks such as composed image retrieval and text-to-video retrieval. These results underscore the generalization potential of our framework across various embedding-based retrieval tasks.

Classifier-Free Guidance (CFG) improves diffusion sampling by encouraging conditional generations while discouraging unconditional ones. Existing preference alignment methods, however, focus only on positive preference pairs, limiting their ability to actively suppress undesirable outputs. Diffusion Negative Preference Optimization (Diff-NPO) approaches this limitation by introducing a separate negative model trained with inverted labels, allowing it to capture signals for suppressing undesirable generations. However, this design comes with two key drawbacks. First, maintaining two distinct models throughout training and inference substantially increases computational cost, making the approach less practical. Second, at inference time, Diff-NPO relies on weight merging between the positive and negative models, a process that dilutes the learned negative alignment and undermines its effectiveness. To overcome these issues, we introduce Diff-SNPO, a single-network framework that jointly learns from both positive and negative preferences. Our method employs a bounded preference objective to prevent winner-likelihood collapse, ensuring stable optimization. Diff-SNPO delivers strong alignment performance with significantly lower computational overhead, showing that explicit negative preference modeling can be simple, stable, and efficient within a unified diffusion framework. Code will be released.

In the era of large foundation models, the quality of embeddings has become a central determinant of downstream task performance and overall system capability. Yet widely used dense embeddings are often extremely high-dimensional (e.g., 4096), incurring substantial costs in storage, memory, and inference latency. To address these, Contrastive Sparse Representation (CSR) is recently proposed as a promising direction, mapping dense embeddings into high-dimensional but $k$-sparse vectors, in contrast to compact dense embeddings such as Matryoshka Representation Learning (MRL). Despite its promise, CSR suffers severe degradation in the ultra-sparse regime (e.g., $k \leq 4$), where over 80\% of neurons remain inactive, leaving much of its efficiency potential unrealized. In this paper, we introduce CSRv2, a principled training approach designed to make ultra-sparse embeddings viable. CSRv2 stabilizes sparsity learning through progressive $k$-annealing, enhances representational quality via supervised contrastive objectives, and ensures end-to-end adaptability with full backbone finetuning. CSRv2 reduces dead neurons from 80\% to 20\% and delivers a 14\% accuracy gain at $k=2$, bringing ultra-sparse embeddings on par with CSR at $k=8$ and MRL at 32 dimensions, all with only two active features. While maintaining comparable performance, CSRv2 delivers a 7$\times$ speedup over MRL, and yields up to 300$\times$ improvements in compute and memory efficiency relative to dense embeddings in e5-mistral-7b-instruct-based text representation. Extensive experiments across text (MTEB, multiple state-of-the-art LLM embeddings (Qwen and e5-Mistral-7B), SPLADEv3, GraphRAG) and vision (ImageNet-1k) demonstrate that CSRv2 makes ultra-sparse embeddings practical without compromising performance, where CSRv2 achieves 7\%/4\% improvement over CSR when $k=4$ and further increases this gap to 14\%/6\% when $k=2$ in text/vision representation. By making extreme sparsity viable, CSRv2 broadens the design space for large-scale, real-time, and edge-deployable AI systems where both embedding quality and efficiency are critical. Code is available at https://github.com/Y-Research-SBU/CSRv2.

Human-Object Interaction (HOI) detection with vision-language models (VLMs) has progressed rapidly, yet a trade-off persists between specialization and generalization. Two major challenges remain: (1) the sparsity of supervision, which hampers effective transfer of foundation models to HOI tasks, and (2) the absence of a generalizable architecture that can excel in both fully supervised and zero-shot scenarios. To address these issues, we propose \textbf{LINK}, \textbf{L}earning \textbf{IN}stance-level \textbf{K}nowledge. First, we introduce a HOI detection framework equipped with a {Human-Object Geometrical Encoder} and a {VLM Linking Decoder}. By decoupling from detector-specific features, our design ensures plug-and-play compatibility with arbitrary object detectors and consistent adaptability across diverse settings. Building on this foundation, we develop a {Progressive Learning Strategy} under a teacher-student paradigm, which delivers dense supervision over all potential human-object pairs. By contrasting subtle spatial and semantic differences between positive and negative instances, the model learns robust and transferable HOI representations. Extensive experiments on SWiG-HOI, HICO-DET, and V-COCO demonstrate state-of-the-art results, showing that our method achieves strong performance in both zero-shot and fully supervised settings while also exhibiting open-vocabulary capability.

Recent advances in reasoning domains with neural networks have primarily been enabled by a training recipe that optimizes Large Language Models, previously trained to predict the next-token in a sequence, with reinforcement learning algorithms. We introduce a framework to study the success of this paradigm, and we theoretically expose the optimization mechanisms by which reinforcement learning improves over next-token prediction in this setting. We study learning from mixture distributions of short and long “chain-of-thought” sequences encoding a single task. In particular, when the task consists of predicting the parity of $d$ bits and long sequences are rare, we show how reinforcement learning after next-token prediction enables autoregressive transformers to generalize, whereas mere next-token prediction requires extreme statistical or computational resources to do so. We further explain how reinforcement learning leverages increased test-time computation, manifested in longer responses, to facilitate this learning process. In a simplified setting, we theoretically prove that autoregressive linear models following this training recipe can efficiently learn to predict the parity of $d$ bits as long as the proportion of long demonstrations in the data mix is not exponentially small in the input dimension $d$. Finally, we demonstrate these same phenomena in other settings, including the post-training of Llama-series models on mixture variations of common mathematical reasoning benchmarks.

Reinforcement Learning with Verifiable Rewards (RLVR) for large language models (LLMs) has achieved remarkable progress in enhancing LLMs’ reasoning capabilities on tasks with clear correctness criteria, such as mathematical reasoning tasks. Several training metrics, such as entropy or response length, have been observed to correlate with different reasoning behaviors in reinforcement learning. Prior approaches incorporate such priors through reward or advantage shaping, which often relies on hand-crafted penalties and preferences (e.g., higher-is-better or lower-is-better). However, without careful hyper-parameter tuning, these directional priors can be overly biased and may lead to failure. To this end, we introduce ***C****onditional adv****AN****tage estimati****ON*** (***CANON***), amplifying the impact of the target metric without presuming its direction. Specifically, *CANON* regroups the sampled responses into two groups based on the higher or lower value of a target metric, measures which metric trend contributes to better performance through inter-group comparison, and identifies the better response within the same group. In summary, *CANON* based on entropy consistently outperforms prior methods across three LLMs on both math reasoning and high-complexity logic tasks. When applied to response length, *CANON* further improves token efficiency, yielding a more favorable Pareto frontier in the performance–cost trade-off.

Model merging has recently emerged as a lightweight alternative to ensembling, combining multiple fine-tuned models into a single set of parameters with no additional training overhead. Yet, existing merging methods fall short of matching the full accuracy of separately fine-tuned endpoints. We present MASS (MoErging through Adaptive Subspace Selection), a new approach that closes this gap by unifying multiple fine-tuned models while retaining near state-of-the-art performance across tasks. Building on the low-rank decomposition of per-task updates, MASS stores only the most salient singular components for each task and merges them into a shared model. At inference time, a non-parametric, data-free router identifies which subspace (or combination thereof) best explains an input's intermediate features and activates the corresponding task-specific block. This procedure is fully training-free and introduces only a two-pass inference overhead plus a ~2 storage factor compared to a single pretrained model, irrespective of the number of tasks. We evaluate MASS on CLIP-based image classification using ViT-B-16, ViT-B-32 and ViT-L-14 for benchmarks of 8, 14 and 20 tasks respectively, establishing a new state-of-the-art. Most notably, MASS recovers up to ~98% of the average accuracy of individual fine-tuned models, making it a practical alternative to ensembling at a fraction of the storage cost.

When a new release of a foundation model is published, practitioners typically need to repeat fine-tuning, even if the same task was already tackled in the previous version. A promising alternative is to reuse the parameter changes (i.e., task vectors) that capture how a model adapts to a specific task. However, these vectors often fail to transfer across different pre-trained models because their parameter spaces are misaligned. In this work, we show that successful transfer depends strongly on the gradient-sign structure of the new model. Based on this insight, we propose GradFix, which approximates the ideal sign structure and leverages it to transfer knowledge using only a handful of labeled samples. Notably, this requires no additional fine-tuning: we only compute a few target-model gradients without parameter updates and mask the source task vector accordingly. This yields an update that is locally aligned with the target loss landscape, effectively rebasing the task vector onto the new pre-training. We provide a theoretical guarantee that our method ensures first-order descent. Empirically, we demonstrate significant performance gains on vision and language benchmarks, consistently outperforming naive task vector addition and few-shot fine-tuning. We further show that transporting task vectors improves multi-task and multi-source model merging. Code is available at https://github.com/fillo-rinaldi/GradFix.