ICML 2026 论文列表

Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen2.5 achieve significant gains even with spurious rewards. We investigate this phenomenon and identify ``Perplexity Paradox'': spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting model is bypassing reasoning in favor of memorization. Using a suite of mechanistic interpretability tools, including Path Patching and Logit Lens, we identify a previously unknown Anchor–Adapter circuit. This circuit enables model to bypass reasoning and directly retrieve memorized solutions under spurious RLVR. We localize a Functional Anchor in middle layers (L18–20) that triggers retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering, i.e., artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models.

While algorithmic stability is a central tool for understanding generalization of learning algorithms, existing high-probability guarantees typically rely on uniform boundedness or sub-Gaussian/sub-Weibull tail assumptions, which can be overly restrictive for modern settings with heavy-tailed or unbounded losses. We develop a stability-based framework that requires only a finite $L_p$ moment condition. Our first contribution is sharp concentration inequalities for functions of independent random variables under $L_p$ constraints, extending McDiarmid's bounded-differences techniques beyond the classical regime. Leveraging these results, we derive sharp high-probability generalization bounds across a range of learning paradigms, including empirical risk minimization, transductive regression, and meta-learning. These guarantees show that $L_p$ stability suffices for robust generalization even when boundedness fails, substantially weakening the standard assumptions in the stability literature.

Sparse Mixture-of-Experts (MoE) architectures scale model capacity efficiently but suffer from massive static parameter footprints, creating significant deployment burdens on memory-constrained hardware. Existing post-training pruning methods often rely on scalar statistics, ignoring the representational geometry of expert feature spaces. This leads to sub-optimal resource allocation across layers and the retention of redundant experts. To address this, we propose a Rank-aware Geometric Expert Pruning (RaGEP) framework to compress MoE models by analyzing the geometric properties of expert activations. First, in the inter-layer allocation stage, we introduce a Rank-aware budget allocation mechanism that adaptively assigns expert budgets based on the effective rank of layer-wise representations. Second, in the intra-layer selection stage, we propose a Spectral-Salience Pruning metric that harmonizes subspace orthogonality and activation magnitude to identify high-energy orthogonal experts. Extensive experiments across MoE models of different scales show that our method consistently outperforms state-of-the-art baselines on a diverse set of zero-shot tasks, while reducing model size and inference cost.

Video Diffusion Models (VDMs) is constrained by immense computational costs. While offline calibration-based acceleration suffers from calibration data dependency, prohibitive calibration duration, and susceptibility to distribution shifts, offline calibration-free methods eliminate these hurdles. However, since they rely on instantaneous zero-order approximations where the mapping between input and output differences varies in real-time, they are susceptible to observational noise and ignore the intrinsic momentum within the diffusion trajectory. In this paper, we propose NaviCache, a plug-and-play test-time self-calibration method re-conceptualizing feature evolution as an Inertial Navigation System (INS) problem. NaviCache bridges the fundamental domain gap and the non-stationary nature of diffusion by modeling the relative coupling between input and output variations. We introduce a dual-state estimation architecture that adaptively tracks the feature change ratio and its latent drift, initialized via a specialized Initial Alignment phase. By integrating a time-dependent noise schedule with an uncertainty-aware Measurement Update mechanism, NaviCache provides a theoretically grounded mechanism for error-bounded block skipping. Extensive experiments on the HunyuanVideo, Wan, and Open-Sora series demonstrate that NaviCache exhibits more accurate error judgment for block skipping and achieves outstanding comprehensive performance.

While explicit Chain-of-Thought (CoT) equips Large Language Models (LLMs) with strong reasoning capabilities, it constrains the model's thoughts to a discrete vocabulary space. Recently, reasoning in continuous latent space has emerged as a promising alternative, but current paradigms suffer from feature collapse and instability due to distribution mismatch when recurrently reusing hidden states, or alignment issues when relying on assistant models. To address this, we propose $\textbf{Latent Thoughts Tuning (\textit{LT-Tuning})}$, a post-training framework that redefines how latent thoughts are constructed and deployed. Instead of relying solely on raw hidden states, our method introduces a $\textbf{Context-Prediction-Fusion}$ mechanism that jointly leverages contextual hidden states and predictive semantic guidance from the vocabulary embedding space. Combined with a progressive three-stage curriculum learning pipeline, $\textit{LT-Tuning}$ also enables dynamic switching between latent and explicit thinking modes. Experiments demonstrate that our method outperforms existing latent reasoning baselines, effectively mitigating feature collapse and achieving robust reasoning accuracy.

Reconstructing dynamic 3D scenes from blurry monocular videos is challenging because motion-induced blur entangles object motion and geometry, hindering geometric consistency. We present Kinematics-GS, a kinematics-aware framework that models blur as motion-aligned deformation and introduces a kinematic prior to reparameterize Gaussian shapes along motion trajectories, thereby mitigating degenerate shape collapse without auxiliary motion supervision. To stabilize optimization, we decompose scenes into dynamic and static components using temporal deformation variance and employ a coarse-to-fine deformation strategy to capture both global motion and fine-grained details. We also introduce a challenging real-world dataset of deformable and elastic objects exhibiting non-rigid motion with spatially non-uniform motion blur that obscures geometric cues. Extensive experiments on real-world benchmarks with realistic motion blur demonstrate that Kinematics-GS outperforms prior methods by a clear margin in monocular dynamic scene reconstruction, highlighting its effectiveness in handling complex and non-rigid motion scenarios.

We study differentially private (DP) regression in settings where each data sample includes public, non-sensitive features—common in applications like recommendation or advertising systems. While such label DP or DP with semi-sensitive features settings have been primarily explored in the context of classification, effective approaches for regression remain underexplored. We introduce $\textsf{Cond-DP}$, a conditioned variant of $\textsf{DPSGD}$ that leverages the structure of public feature matrices to improve optimization under privacy constraints. Motivated by the observation that these public features often exhibit rapidly decaying spectra, $\textsf{Cond-DP}$ incorporates a data-driven conditioning matrix to reshape the optimization landscape and accelerate convergence. We provide convergence guarantees for convex, strongly convex and non-convex settings, and recover standard $\textsf{DPSGD}$ as a special case when the conditioning matrix is the identity. We show how to construct an effective conditioning matrix for $\textsf{Cond-DP}$ directly from public features, enabling faster convergence than $\textsf{DPSGD}$ in private linear regression, without incurring additional privacy cost. Empirically, $\textsf{Cond-DP}$ with this conditioning matrix consistently outperforms state-of-the-art baselines across a wide range of datasets and model architectures under label DP, demonstrating strong and robust performance in practice.

Continuous Variational Autoencoders (VAEs) serve as the fundamental continuous tokenizer for modern neural audio generation systems, enabling high-fidelity reconstruction while providing a compact, smooth latent space for downstream generative priors. However, continuous VAEs face a fundamental conflict when balancing *compression rate*, *reconstruction fidelity*, and *latent space topology*—a challenge we formalize as the **Rate-Distortion-Regularity Trilemma**. This trilemma stems from a critical *topological mismatch*: the prevailing isotropic Gaussian prior in standard VAEs imposes a *flat* latent geometry that fails to accommodate audio's *hierarchical* nature, where low-frequency components are structured and compressible while high-frequency components are stochastic and incompressible, leading to *disordered information packing* where crucial semantic features are randomly interleaved with high-entropy noise. To resolve this challenge, we propose **Structured Topology-Aware Regularization (STAR)**, a general training strategy that reshapes latent space geometry by imposing a growth-based constraint field, routing structural and textural information into channel subspaces with matching capacities. STAR is applicable to any VAE architecture and effectively resolves the trilemma, as demonstrated in CNN-based VAEs. To fully exploit STAR's potential, we present **STAR-VAE**, combining STAR with a hybrid CNN-Mamba architecture that synergizes local feature extraction with linear-complexity global context modeling, achieving state-of-the-art performance. We further propose **STAR-Gen**, an LLM-based Flow Matching framework that leverages STAR-VAE's structured latent space for high-fidelity generation without suffering from vector quantization artifacts. Empirical results demonstrate that STAR-VAE successfully resolves the trilemma, achieving state-of-the-art reconstruction fidelity and enhanced semantic information preservation across diverse audio domains. The structured latent space improves both traditional diffusion models and our **STAR-Gen** paradigm, achieving state-of-the-art performance in text-to-audio generation. The project page is available at https://STAR-VAE.github.io.

Open-Ended Deep Research (OEDR) pushes LLM agents beyond short-form QA toward long-horizon workflows that iteratively search, connect, and synthesize evidence into structured reports. However, existing OEDR agents largely follow either linear "search-then-generate" accumulation or outline-centric planning. The former suffers from lost-in-the-middle failures as evidence grows, while the latter relies on the LLM to implicitly infer knowledge gaps from the outline alone, providing weak supervision for identifying missing relations and triggering targeted exploration. We present DualGraph memory, an architecture that separates what the agent knows from how it writes. DualGraph maintains two co-evolving graphs: an Outline Graph (OG), and a Knowledge Graph (KG), a semantic memory that stores fine-grained knowledge units, including core entities, concepts, and their relations. By analyzing the KG topology together with structural signals from the OG, DualGraph generates targeted search queries, enabling more efficient and comprehensive iterative knowledge-driven exploration and refinement. Across four established OEDR benchmarks, DualGraph consistently outperforms state-of-the-art baselines in report depth, breadth, and factual grounding; for example, it reaches a 53.08 RACE score on DeepResearch Bench with GPT-5. Moreover, ablation studies confirm the central role of the dual-graph design.

Instruction tuning for speech language models (SLMs) is substantially more challenging than for text-based large language models (LLMs), as it requires learning a new modality and a wide range of speech-specific instructions in addition to those supported by text LLMs. Existing SLM training approaches largely replicate the text LLM training paradigm by synthesizing large-scale speech pre-training and instruction-tuning datasets. However, this strategy is difficult to scale, since speech sequences are significantly longer than text sequences. In this paper, we propose SpeechCombine, an instruction-following speech language model trained **without any instruction tuning**, using only a single round of speech pre-training on as little as 30k hours of speech data. Starting from a text LLM base model, we perform continuous pre-training on speech utterances to obtain a speech-adapted model, and then directly combine its weights with the weight difference between the instruction-tuned and base versions of the text LLM. Our results show that this simple combination strategy not only preserves the knowledge and capabilities of the original text LLM, but also effectively transfers them to the speech domain. These findings suggest a new direction for SLM training that avoids reliance on massive volumes of speech data.

Recently, small models with latent recursion have obtained promising results on complex reasoning tasks. These results are typically explained by the theory that such recursion increases a network’s depth, allowing it to compactly emulate the capacity of larger models. However, the performance of recursively added layers remains behind the capabilities of one‑pass models with the same feed-forward depth. This means that in the looped version, not every recursive step effectively contributes to depth. This raises the question: when and why does latent reasoning improve performance, and when does it result in dead compute? In our work, we analyze the algorithms that latent reasoning provides answer to this question. We show that latent reasoning can be formalized as a classifier‑free guidance and policy improvement algorithm. Building on these insights, we propose to use a training schemes from RL and diffusion methods for latent reasoning modles. Using the Tiny Recursive Model as our testbed, we show that with our modifications we can avoid dead compute steps and reduce the total number of forward passes by 18× while maintaining performance. Broadly speaking, we show how a policy improvement perspective on recursive steps can explain model behavior and provide insights for further improvements.

Predicting outputs that are located in non-Euclidean spaces, such as probability distributions, networks, and symmetric positive-definite matrices, is becoming increasingly important in modern data analysis, particularly when inputs are high-dimensional. We propose DeSI (Deep Single-Index Fréchet Regression), a semiparametric framework for regression with metric space-valued outputs and multivariate inputs that assumes a single-index structure for the conditional Fréchet mean. DeSI estimates an interpretable index direction, which quantifies the relative importance of inputs, using a deep neural network, and performs Fréchet regression along the resulting one-dimensional index in the target metric space. This structure mitigates the curse of dimensionality while retaining interpretability, which stands in contrast to standard deep neural networks. We establish theoretical guarantees for DeSI, including uniform approximation and convergence rates, and demonstrate its strong predictive performance through simulations on distributions, networks, and symmetric positive-definite matrices, as well as an application to compositional mood data from New Jersey.

Understanding how anatomical shapes evolve in response to developmental covariates—and quantifying their spatially varying uncertainties—is critical in healthcare research. Existing approaches typically rely on global time-warping formulations that ignore spatially heterogeneous dynamics. We introduce \texttt{PRISM}, a novel framework that bridges implicit neural representations with uncertainty-aware statistical shape analysis. \texttt{PRISM} models the conditional distribution of shapes given covariates, providing spatially continuous estimates of both the population mean and covariate-dependent uncertainty at arbitrary locations. A key theoretical contribution is a closed-form Fisher Information metric that enables efficient, analytically tractable local temporal uncertainty quantification via automatic differentiation. Experiments on three synthetic datasets and one clinical dataset demonstrate \texttt{PRISM}'s strong performance across diverse tasks—from modeling shape evolution to anomaly detection—within a unified framework, while providing interpretable and clinically meaningful uncertainty estimates.

Recent works propose complex multi-modal models that handle both time series and language, ultimately claiming high performance on complex tasks like time series reasoning and cross-modal question answering. However, they skip foundational evaluations that such complex models should have mastered. So we ask a simple question:$\textit{How well can recent models describe structural properties of time series?}$ To answer this, we propose that successful models should be able to $\textit{recognize}$, $\textit{differentiate}$, and $\textit{generate}$ descriptions of univariate time series. We then create $\textbf{BEDTime}$, a benchmark to assess these novel tasks, that comprises $\textbf{five datasets}$ reformatted across $\textbf{three modalities}$. In evaluating $\textbf{17 state-of-the-art models}$, we find that (1) surprisingly, dedicated time series-language models fall short, despite being designed for similar tasks, (2) vision language models are quite capable, (3) language only methods perform worst, despite many lauding their potential, and (4) all approaches are clearly fragile to a range of real world robustness tests, indicating directions for future work. Together, our findings critique prior works' claims and provide avenues for advancing multi-modal time series modeling.

Reliable watermarking of panoramic imagery is fundamentally challenged by arbitrary 3D rotations. As panoramas are defined on the sphere, they naturally transform under the action of $SO(3)$, rendering conventional planar representations and augmentation-based robustness strategies inadequate and devoid of theoretical guarantees. To address this, we formulate panoramas as spherical signals and leverage $SO(3)$ representation theory to derive provably rotation-invariant descriptors. While spherical harmonic coefficients transform equivariantly under rotations, the natural invariant constructions are typically limited to zeroth-order statistics which eliminate directional information and severely constrain embedding capacity. In this work, we introduce a principled third-order invariant construction by coupling higher-order $SO(3)$ irreducible representations via tensor products and projecting onto the trivial representation. This yields a spherical invariant bispectrum that preserves phase information while remaining strictly rotation-invariant. Leveraging this property, we embed watermarks into higher-order spherical harmonic coefficients and recover them from invariant bispectral scalars, enabling reliable extraction under arbitrary 3D rotations. We provide a theoretical proof of $SO(3)$ invariance for it and demonstrate experimentally its near-perfect robustness to continuous rotations while maintaining high visual fidelity.

Cosine similarity estimation is a core primitive in coarse-to-fine retrieval pipelines, where early-stage candidate selection relies on approximate similarity estimates whose errors are amplified downstream. Widely used sign-based sketches arising from extreme quantization of random projections exhibit a structural variance peak near $\theta$ $\approx$ 90$^\circ$, the near-background region where candidate selection is most difficult. We propose QA-Cos, a query-aware decoder-side estimator that departs from the Hamming-agreement paradigm, treating sign bits as probabilistic observations rather than deterministic votes. Across simulations and BEIR benchmarks, QA-Cos reduces estimation error by up to $\sim$15--20\% in the near-orthogonal region and translates these gains into improved candidate selection in two-stage ANN pipelines, improving Hit@K by up to $\sim$30 percentage points at fixed budgets and reducing candidates by up to $\sim$45--50\% at fixed recall. In a native hnswlib scorer-replacement experiment, the same decoder improves the quality of the final HNSW frontier, and in a focused storage-aware setting a practical gated variant reduces end-to-end wall-clock latency at matched recall.

How can we train agents to navigate uncertainty over long horizons? In this work, we propose ∆Belief-RL, which leverages a language model's own intrinsic beliefs to reward intermediate progress. Our method utilizes the change in the probability an agent assigns to the target solution for credit assignment. By training on synthetic interaction data, ∆Belief-RL teaches information-seeking capabilities that consistently outperform purely outcome-based rewards for RL, with improvements generalizing to out-of-distribution applications ranging from customer service to personalization. Notably, the performance continues to improve as we scale test-time interactions beyond the training horizon, with interaction-efficiency increasing even on Pass@k metrics. Overall, our work introduces a scalable training strategy for navigating uncertainty over a long-horizon, by enabling credit assignment to intermediate actions via intrinsic ∆Belief rewards.

Time series are ubiquitous in real-world scenarios and crucial for applications ranging from energy management to traffic control. Consequently, the ability to reason over time series is a fundamental skill for generalist models to solve complex problems. However, current benchmarks for generalist models largely overlook this dimension. To bridge this gap, we introduce TSRBench, a comprehensive multi-modal benchmark designed to stress-test the full spectrum of time series reasoning capabilities. TSRBench features: i) a diverse set of 4125 problems from 14 domains, and is categorized into 4 major dimensions: Perception, Reasoning, Prediction, and Decision-Making. ii) 15 tasks from the 4 dimensions evaluating essential reasoning capabilities (e.g., numerical reasoning). Through extensive experiments, we evaluate over 30 leading proprietary and open-source LLMs, VLMs, and TSLLMs within TSRBench. Our findings reveal that: i) scaling laws hold for perception and reasoning but break down for prediction; ii) strong reasoning does not guarantee accurate context-aware forecasting, indicating a decoupling between semantic understanding and numerical prediction; and iii) despite the complementary nature of textual and visual forms of time series as inputs, current multimodal models fail to effectively fuse them for reciprocal performance gains. TSRBench provides a standardized evaluation platform that not only highlights existing challenges but also offers valuable insights to advance generalist models. Our code and dataset are available at https://tsrbench.github.io/.

Sparse Mixture-of-Experts (MoE) models scale Transformers efficiently but suffer from expert overlap, where different experts process similar tokens and learn redundant functions, resulting in ambiguous routing and underutilized capacity. While architectural solutions like DeepSeek-style shared experts promote specialization, they require substantial structural modifications and rely solely on intra-layer signals. We propose two plug-and-play auxiliary losses that enhance MoE specialization and routing efficiency without modifying routers or model architectures. First, an intra-layer specialization loss penalizes cosine similarity between experts' SwiGLU activations on identical tokens, encouraging experts to specialize in complementary functions. Second, a cross-layer dependency loss maximizes joint Top-$k$ routing probabilities across adjacent layers, establishing coherent expert pathways through network depth while reinforcing intra-layer specialization. Both losses are orthogonal to the standard load-balancing loss and compatible with shared-expert and vanilla Top-$k$ MoE architectures. We implement both losses as a drop-in Megatron-LM module. Extensive experiments across pre-training, fine-tuning, and zero-shot benchmarks demonstrate consistent task gains, higher expert specialization, and lower-entropy routing; together, these improvements translate into faster inference via more stable expert pathways.