ICLR 2026 Papers

Understanding and predicting dynamics of the physical world can enhance a robot's ability to plan and interact effectively in complex environments. While recent video generation models have shown strong potential in modeling dynamic scenes, generating videos that are both temporally coherent and geometrically consistent across camera views remains a significant challenge. To address this, we propose a 4D video generation model that enforces multi-view 3D consistency of generated videos by supervising the model with cross-view pointmap alignment during training. Through this geometric supervision, the model learns a shared 3D scene representation, enabling it to generate spatio-temporally aligned future video sequences from novel viewpoints given a single RGB-D image per view, and without relying on camera poses as input. Compared to existing baselines, our method produces more visually stable and spatially aligned predictions across multiple simulated and real-world robotic datasets. We further show that the predicted 4D videos can be used to recover robot end-effector trajectories using an off-the-shelf 6DoF pose tracker, yielding robot manipulation policies that generalize well to novel camera viewpoints.

We study the problem of constructing $(1+\varepsilon)$-coresets for Euclidean $(k,z)$-clustering in the distributed setting, where $n$ data points are partitioned across $s$ sites. We focus on two prominent communication models: the coordinator model and the blackboard model. In the coordinator model, we design a protocol that achieves a $(1+\varepsilon)$-strong coreset with total communication complexity $\tilde{O}\left(sk + \frac{dk}{\min(\varepsilon^4,\varepsilon^{2+z})} + dk\log(n\Delta)\right)$ bits, improving upon prior work (Chen et al., NeurIPS 2016) by eliminating the need to communicate explicit point coordinates in-the-clear across all servers. In the blackboard model, we further reduce the communication complexity to $\tilde{O}\left(s\log(n\Delta) + dk\log(n\Delta) + \frac{dk}{\min(\varepsilon^4,\varepsilon^{2+z})}\right)$ bits, achieving better bounds than previous approaches while upgrading from constant-factor to $(1+\varepsilon)$-approximation guarantees. Our techniques combine new strategies for constant-factor approximation with efficient coreset constructions and compact encoding schemes, leading to optimal protocols that match both the communication costs of the best-known offline coreset constructions and existing lower bounds (Chen et al., NeurIPS 2016, Huang et. al., STOC 2024), up to polylogarithmic factors.

Algorithmic fairness is a socially crucial topic in real-world applications of AI. Among many notions of fairness, subgroup fairness is widely studied when multiple sensitive attributes (e.g., gender, race, age) are present. However, as the number of sensitive attributes grows, the number of subgroups increases accordingly, creating heavy computational burdens and data sparsity problem (subgroups with too small sizes). In this paper, we develop a novel learning algorithm for subgroup fairness which resolves these issues by focusing on subgroups with sufficient sample sizes as well as marginal fairness (fairness for each sensitive attribute). To this end, we formalize a notion of subgroup-subset fairness and introduce a corresponding distributional fairness measure called the supremum Integral Probability Metric (supIPM). Building on this formulation, we propose the Doubly Regressing Adversarial learning for subgroup Fairness (DRAF) algorithm, which reduces a surrogate fairness gap for supIPM with much less computation than directly reducing supIPM. Theoretically, we prove that the proposed surrogate fairness gap is an upper bound of supIPM. Empirically, we show that the DRAF algorithm outperforms baseline methods in benchmark datasets, specifically when the number of sensitive attributes is large so that many subgroups are very small.

Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; even at high Signal-to-Noise Ratios (SNRs) where speech is perfectly intelligible, their output token sequences can change drastically, increasing the learning burden for downstream LLMs. This instability stems from two flaws: a brittle single-path quantization architecture and a distant training signal indifferent to intermediate token stability. To address this, we introduce StableToken, a tokenizer that achieves stability through a consensus-driven mechanism. Its multi-branch architecture processes audio in parallel, and these representations are merged via a powerful bit-wise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks.

Expressive performance rendering (EPR) and automatic piano transcription (APT) are fundamental yet inverse tasks in music information retrieval: EPR generates expressive performances from symbolic scores, while APT recovers scores from performances. Despite their dual nature, prior work has addressed them independently. In this paper, we propose a unified framework that jointly models EPR and APT by disentangling note-level score content and global performance style representations from both paired and unpaired data. Our framework is built on a transformer-based sequence-to-sequence (Seq2Seq) architecture and is trained using only sequence-aligned data, without requiring fine-grained note-level alignment. To automate the rendering process while ensuring stylistic compatibility with the score, we introduce an independent diffusion-based performance style recommendation (PSR) module that generates style embeddings directly from score content. This modular component supports both style transfer and flexible rendering across a range of expressive styles. Experimental results from both objective and subjective evaluations demonstrate that our framework achieves competitive performance on EPR and APT tasks, while enabling effective content–style disentanglement, reliable style transfer, and stylistically appropriate rendering. Demos are available at https://wei-zeng98.github.io/joint-apt-epr/.

While reinforcement learning (RL) has proven highly effective for general reasoning in vision-language models, its application to tasks requiring deep understanding of information-rich images and structured output generation remains underexplored. Chart-to-code generation exemplifies this challenge, demanding complex reasoning over visual charts to produce structured code. Supervised fine-tuning (SFT) alone is often insufficient, highlighting the need for effective RL strategies tailored to structured outputs. In this paper, we systematically investigate the performance plateau of SFT through large-scale experiments and propose Multimodal Structured Reinforcement Learning (MSRL) for chart-to-code generation. We construct the largest training corpus to date, with 3 million chart-code pairs curated from real-world tables in arXiv papers, addressing the limitations of previous synthetic datasets. Despite achieving state-of-the-art performance, our experiments show that simply increasing SFT data eventually leads to diminishing improvements. To break this plateau, MSRL employs a multi-granularity reward system that integrates both textual and visual feedback. At the textual level, rule-based rewards validate fine-grained code details, while at the visual level, a model-based reward assesses the structural similarity between rendered code and ground-truth charts. We implement a two-stage curriculum training strategy, first optimizing the model with textual rewards and then incorporating visual signals for further enhancement. Experimental results demonstrate that MSRL substantially breaks the SFT plateau, improving high-level metrics by 6.2% and 9.9% on ChartMimic and ReachQA benchmarks, respectively. Notably, our method outperforms all existing approaches in the chart domain and achieves competitive results with advanced closed-source models.

Mathematical reasoning represents a critical frontier in advancing large language models (LLMs). While step-by-step approaches have emerged as the dominant paradigm for mathematical problem-solving in LLMs, the quality of reasoning steps in training data fundamentally constrains model performance. Recent studies has demonstrated that more detailed intermediate steps can enhance model performance, yet existing methods for step expansion either require more powerful external models or incur substantial computational costs. In this paper, we introduce MathFimer, a novel framework for mathematical reasoning step expansion inspired by the "Fill-in-the-middle" from code completion. By decomposing solution chains into prefix-suffix pairs and training models to reconstruct missing intermediate steps, we develop a specialized model, MathFimer-7B, on our carefully curated NuminaMath-Fim dataset. We then apply these models to enhance existing mathematical reasoning datasets by inserting detailed intermediate steps into their solution chains, creating MathFimer-expanded versions. Through comprehensive experiments on multiple mathematical reasoning datasets, including MathInstruct and MetaMathQA, we demonstrate that models trained on MathFimer-expanded data consistently outperform their counterparts trained on original data across various benchmarks such as GSM8K and MATH. Our approach offers a practical, scalable solution for enhancing mathematical reasoning capabilities in LLMs without relying on more powerful external models or expensive inference procedures.

Adapting large language models (LLMs) using smaller proxy models has been shown to improve training efficiency, where the LLMs remain frozen while the proxies are tuned on top. However, this approach typically requires access to the output probability distributions of LLMs, which are often inaccessible or unstable. To address this limitation, we propose KnowProxy, a knowledge-guided proxy framework in which the proxy is trained with textual knowledge rather than probability distributions. Specifically, we first elicit textual knowledge and reasoning from frozen LLMs through prompting, and then the proxy model learns to adapt this reasoning to target task distributions. We evaluate KnowProxy on diverse reasoning benchmarks with different fine-tuning scenarios. Comprehensive results show that KnowProxy achieves competitive or even better performance without direct access to probability distributions, thereby providing a scalable and versatile alternative to traditional fine-tuning.

Transformers struggle with tasks that require symbolic planning loops, variable updates, and conditional branching, especially under length extrapolation. We introduce discrete controller modules that insert a small set of program primitives (ASSIGN, ADD, COMPARE, BRANCH) into Transformer blocks via a Gumbel–Softmax selector over operations and a compact program state of registers, flags, and optional memory. We prove that the augmented model can simulate any bounded-step program by mapping each primitive step to one controller step, and we bound the deviation of relaxed execution from its discrete trace by $O(\tau+\kappa^{-1})$ (selection temperature $\tau$, comparison sharpness $\kappa$). Empirically, the controller-augmented Transformer achieves strong length generalization on algorithmic benchmarks (Sorting, Sum-of-List, BFS), improving longest-length accuracy by up to $20$–$40$ points over strong baselines, and yields consistent gains on symbolic QA (DROP) and program-synthesis-style tasks (RobustFill) with reduced compositionality drop-off. The learned execution is interpretable: operation traces align with ground truth, register roles are linearly decodable, and targeted knockouts cause localized accuracy losses. The approach adds only $\sim$5–7% FLOPs and can be applied sparsely (every $p$-th layer).

Biosignals offer valuable insights into the physiological states of the human body. Although biosignal modalities differ in functionality, signal fidelity, sensor comfort, and cost, they are often intercorrelated, reflecting the holistic and interconnected nature of human physiology. This opens up the possibility of performing the same tasks using alternative biosignal modalities, thereby improving the accessibility, usability, and adaptability of health monitoring systems. However, the limited availability of large labeled datasets presents challenges for training models tailored to specific tasks and modalities of interest. Unsupervised cross-modal knowledge transfer offers a promising solution by leveraging knowledge from an existing modality to support model training for a new modality. Existing methods are typically based on knowledge distillation, which requires running a teacher model alongside student model training, resulting in high computational and memory overhead. This challenge is further exacerbated by the recent development of foundation models that demonstrate superior performance and generalization across tasks at the cost of large model sizes. To this end, we explore a new framework for unsupervised cross-modal knowledge transfer of biosignals by training a lightweight bridge network to align the intermediate representations and enable information flow between foundation models and across modalities. Specifically, we introduce an efficient strategy for selecting alignment positions where the bridge should be constructed, along with a flexible prototype network as the bridge architecture. Extensive experiments across multiple biosignal modalities, tasks, and datasets show that BioX-Bridge reduces the number of trainable parameters by 88-99\% while maintaining or even improving transfer performance compared to state-of-the-art methods.

Sets are a fundamental data structure, and learning their vectorized representations is crucial for many computational problems. Existing methods typically focus on intra-set properties such as permutation invariance and cardinality independence. While effective at preserving basic intra-set semantics, these approaches may be insufficient in explicitly modeling inter-set correlations, which are critical for tasks requiring fine-grained comparisons between sets. In this work, we propose SRAL, a Set Representation Auxiliary Learning framework for capturing inter-set correlations that is compatible with various downstream tasks. SRAL conceptualizes sets as high-dimensional distributions and leverages the 2-Sliced-Wasserstein distance to derive their distributional discrepancies into set representation encoding. More importantly, we introduce a novel adversarial auxiliary learning scheme. Instead of manipulating the input data, our method perturbs the set encoding process itself and compels the model to be robust against worst-case perturbations through a min-max optimization. Our theoretical analysis shows that this objective, in expectation, directly optimizes for the set-wise Wasserstein distances, forcing the model to learn highly discriminative representations. Comprehensive evaluations across four downstream tasks examine SRAL’s performance relative to baseline methods, showing consistent effectiveness in both inter-set relation-sensitive retrieval and intra-set information-oriented processing tasks.

Simulation provides a cost-effective and flexible platform for data generation and policy learning to develop robotic systems. However, bridging the gap between simulation and real-world dynamics remains a significant challenge, especially in physical parameter identification. In this work, we introduce a real-to-sim-to-real engine that leverages the Gaussian Splat representations to build a differentiable engine, enabling object mass identification from real-world visual observations and robot control signals, while enabling grasping policy learning simultaneously. Through optimizing the mass of the manipulated object, our method automatically builds high-fidelity and physically plausible digital twins. Additionally, we propose a novel approach to train force-aware grasping policies from limited data by transferring feasible human demonstrations into simulated robot demonstrations. Through comprehensive experiments, we demonstrate that our engine achieves accurate and robust performance in mass identification across various object geometries and mass values. Those optimized mass values facilitate force-aware policy learning, achieving superior and high performance in object grasping, effectively reducing the sim-to-real gap. Our code is included in the Supplementary Material and will be open source to facilitate reproducibility. Anonymous project page is available at https://robot-drex-engine.github.io.

Diffusion models deliver state-of-the-art image quality. However, sampling is costly at inference time because it requires many model evaluations (number of function evaluations, NFEs). To reduce NFEs, classical ODE multistep methods have been adopted. Yet differences in the choice of prediction type (noise/data/velocity) and integration domain (half log-SNR/noise-to-signal ratio) lead to different outcomes. We introduce Dual-Solver, which generalizes multistep samplers by introducing learnable parameters that continuously (i) interpolate among prediction types, (ii) select the integration domain, and (iii) adjust the residual terms. It maintains the traditional predictor-corrector structure and guarantees second-order local accuracy. These parameters are learned with a classification-based objective using a frozen pretrained classifier (e.g., ViT or CLIP). On ImageNet class-conditional generation (DiT, GM-DiT) and text-to-image (SANA, PixArt-$\alpha$), Dual-Solver improves FID and CLIP scores in the low-NFE regime ($3\le$ NFE $\le 9$) across backbones.

Large language models (LLMs) exhibit strong generative capabilities but remain vulnerable to confabulations, fluent yet unreliable outputs that vary arbitrarily even under identical prompts. Leveraging a quantum tensor network–based pipeline, we propose a quantum physics-inspired uncertainty quantification framework that accounts for the aleatoric uncertainty in token sequence probability for semantic equivalence-based clustering of LLM generations. In turn, this offers a principled and interpretable scheme for hallucination detection. We further introduce an entropy-maximization strategy that prioritizes high-certainty, semantically coherent outputs and highlights entropy regions where LLM decisions are likely to be unreliable, offering practical guidelines for when human oversight is warranted. We evaluate the robustness of our scheme under different generation lengths and quantization levels, dimensions overlooked in prior studies, demonstrating that our approach remains reliable even in resource-constrained deployments. A total of 116 experiments on TriviaQA, NQ, SVAMP, and SQuAD across multiple architectures (Mistral-7B, Mistral-7B-instruct, Falcon-rw-1b, LLaMA-3.2-1b, LLaMA-2-13b-chat, LLaMA-2-7b-chat, LLaMA-2-13b and LLaMA-2-7b) show consistent improvements in AUROC and AURAC over state-of-the-art baselines.

Adversarial training is one of the most effective adversarial defenses, but it incurs a high computational cost. In this study, we present the first theoretical analysis suggesting that adversarially pretrained transformers can serve as universally robust foundation models, models that can robustly adapt to diverse downstream tasks with only lightweight tuning. Specifically, we demonstrate that single-layer linear transformers, after adversarial pretraining across a variety of classification tasks, can robustly generalize to unseen classification tasks through in-context learning from clean demonstrations (i.e., without requiring additional adversarial training or examples). This universal robustness stems from the model's ability to adaptively focus on robust features within given tasks. We also show the two open challenges for attaining robustness: accuracy-robustness trade-off and sample-hungry training. This study initiates the discussion on the utility of universally robust foundation models. While their training is expensive, the investment would prove worthwhile as downstream tasks can enjoy free adversarial robustness.

The choice of initial noise strongly affects quality and prompt alignment in video diffusion; different seeds for the same prompt can yield drastically different results. While recent methods use externally designed priors (e.g., frequency filtering or inter-frame smoothing), they often overlook internal model signals that indicate inherently preferable seeds. To address this, we propose ANSE (Active Noise Selection for Generation), a model-aware framework that selects high-quality seeds by quantifying attention-based uncertainty. At its core is BANSA (Bayesian Active Noise Selection via Attention), an acquisition function that measures entropy disagreement across multiple stochastic attention samples to estimate model confidence and consistency. For efficient inference-time deployment, we introduce a Bernoulli-masked approximation of BANSA that estimates scores from a single diffusion step and a subset of informative attention layers. Experiments across diverse text-to-video backbones demonstrate improved video quality and temporal coherence with marginal inference overhead, providing a principled and generalizable approach to noise selection in video diffusion.

It is important for Large Language Models (LLMs) to be aware of the boundary of their knowledge, i.e., the mechanism of identifying known and unknown queries. This type of awareness enables models to perform adaptive inference, such as invoking retrieval-augmented generation (RAG), engaging in slow and deep thinking, or abstaining from answering when appropriate. These mechanisms are beneficial to the development of efficient and trustworthy AI. In this work, we propose a method to detect knowledge boundaries via \textbf{\emph{Query-Level Uncertainty }}, which estimates if a model is capable of to answering a given query before generating any tokens. To this end, we propose a novel, training-free method called \textbf{\emph{Internal Confidence}}, which leverages self-evaluations across layers and tokens to provide a reliable signal of uncertainty. Empirical studies on both factual question answering and mathematical reasoning tasks demonstrate that our internal confidence can outperform several baselines. Furthermore, we showcase that our proposed method can be used for adaptive inference, such as efficient RAG and model cascading, thereby reducing inference costs while preserving overall performance.

High Dynamic Range Novel View Synthesis (HDR NVS) seeks to learn an HDR 3D model from Low Dynamic Range (LDR) training images captured under conventional imaging conditions. Current methods primarily focus on static scenes, implicitly assuming all scene elements remain stationary and non-living. However, real-world scenarios frequently feature dynamic elements, such as moving objects, varying lighting conditions, and other temporal events, thereby presenting a significantly more challenging scenario. To address this gap, we propose a more realistic problem named HDR Dynamic Novel View Synthesis (HDR DNVS), where the additional dimension ``Dynamic'' emphasizes the necessity of jointly modeling temporal radiance variations alongside sophisticated 3D translation between LDR and HDR. To tackle this complex, intertwined challenge, we introduce HDR-4DGS, a Gaussian Splatting-based architecture featured with an innovative dynamic tone-mapping module that explicitly connects HDR and LDR domains, maintaining temporal radiance coherence by dynamically adapting tone-mapping functions according to the evolving radiance distributions across the temporal dimension. As a result, HDR-4DGS achieves both temporal radiance consistency and spatially accurate color translation, enabling photorealistic HDR renderings from arbitrary viewpoints and time instances. Extensive experiments demonstrate that HDR-4DGS surpasses existing state-of-the-art methods in both quantitative performance and visual fidelity. Source code is available at \url{https://github.com/prinasi/HDR-4DGS}.

We propose an alternative to sparse autoencoders (SAEs) as a simple and effective unsupervised method for extracting interpretable concepts from neural networks. The core idea is to cluster differences in activations, which we formally justify within a discriminant analysis framework. To enhance the diversity of extracted concepts, we refine the approach by weighting the clustering using the skewness of activations. The method aligns with Deleuze's modern view of concepts as differences. We evaluate the approach across five models and three modalities (vision, language, and audio), measuring concept quality, diversity, and consistency. Our results show that the proposed method achieves concept quality surpassing prior unsupervised SAE variants while approaching supervised baselines, and that the extracted concepts enable steering of a model’s inner representations, demonstrating their causal influence on downstream behavior.