ICLR 2026 论文列表

Cover songs constitute a vital aspect of musical culture, preserving the core melody of an original composition while reinterpreting it to infuse novel emotional depth and thematic emphasis. Although prior research has explored the reinterpretation of instrumental music through melody-conditioned text-to-music models, the task of cover song generation remains largely unaddressed. In this work, we reformulate our cover song generation as a conditional generation, which simultaneously generates new vocals and accompaniment conditioned on the original vocal melody and text prompts. To this end, we present SongEcho, which leverages Instance-Adaptive Element-wise Linear Modulation (IA-EiLM), a framework that incorporates controllable generation by improving both conditioning injection mechanism and conditional representation. To enhance the conditioning injection mechanism, we extend Feature-wise Linear Modulation (FiLM) to an Element-wise Linear Modulation (EiLM), to facilitate precise temporal alignment in melody control. For conditional representations, we propose Instance-Adaptive Condition Refinement (IACR), which refines conditioning features by interacting with the hidden states of the generative model, yielding instance-adaptive conditioning. Additionally, to address the scarcity of large-scale, open-source full-song datasets, we construct Suno70k, a high-quality AI song dataset enriched with comprehensive annotations. Experimental results across multiple datasets demonstrate that our approach generates superior cover songs compared to existing methods, while requiring fewer than 30% of the trainable parameters. The code, dataset, and demos are available at [https://github.com/lsfhuihuiff/SongEcho_ICLR2026](https://github.com/lsfhuihuiff/SongEcho_ICLR2026).

Diffusion models have recently demonstrated significant robustness in medical image segmentation, effectively accommodating variations across different imaging styles. However, their applications remain limited due to: (i) current successes being primarily confined to 2D segmentation tasks—we observe that diffusion models tend to collapse at the early stage when applied to 3D medical tasks; and (ii) the inherently isolated iteration along timesteps during training and inference. To tackle these limitations, we propose a novel framework named Cross-Timestep, which incorporates two key innovations: an Adaptive Priori Decoding Strategy (APDS) and a trans-temporal memory LSTM (tLSTM) mechanism. (i) The APDS provides prior guidance during the diffusion process by employing a Priori Decoder(PD) that focuses solely on the conditional branch, successfully stabilizing the reverse diffusion process. (ii) The tLSTM integrates convolution and linear layers into the LSTM gating structure, and enhances the memory cell mechanism to retain temporal state, explicitly preserving and propagating continuous temporal states across timesteps. Experimental results demonstrate that Cross-Timestep performs favorably on heterogeneous 3D medical datasets. Three experiments further analyze the collapse phenomenon in 3D medical diffusion models and validate that APDS effectively prevents initial-stage collapse without excessively constraining the model, while tLSTM facilitates the performance and scalability of diffusion models.

We introduce NeuralOS, a neural framework that simulates graphical user interfaces (GUIs) of operating systems by directly predicting screen frames in response to user inputs such as mouse movements, clicks, and keyboard events. NeuralOS combines a recurrent neural network (RNN), which tracks the computer state, with a diffusion-based neural renderer that generates screen images. The model is trained on a dataset of Ubuntu XFCE recordings, which include both randomly generated interactions and realistic interactions produced by AI agents. Experiments show that NeuralOS successfully renders realistic GUI sequences, accurately captures mouse interactions, and reliably predicts state transitions like application launches. Beyond reproducing existing systems, NeuralOS shows that synthesized training data can teach the model to simulate applications that were never installed, as illustrated by a Doom application, and suggests a path toward learning user interfaces purely from synthetic demonstrations.

Repeated Sampling (RS) is a simple inference-time algorithm that has been shown to improve model performance on complex tasks. Although it is an effective way of scaling inference time, it often struggles to generate diverse solution candidates, frequently relying on the same underlying approach to solve the problem and thus producing redundant samples. To address this limitation, we propose a new inference algorithm, GuidedSampling, which decouples the exploration and generation phases during inference, increasing diversity of generated candidate solutions. The exploration phase identifies multiple concepts that can be utilized to solve the problem, while the generation phase applies a specific concept to provide final solution candidates. We first define the theoretical bounds of GuidedSampling and then empirically demonstrate that it improves the performance of base model at pass@50 by on an average $\sim21.6$% across various benchmarks compared to RS. Furthermore, models trained on trajectories of GuidedSampling exhibit substantial performance improvements at pass@5 by on an average $\sim9.7$%, compared to models trained on traditional RS. Additionally, models trained with GuidedSampling increases the average number of concepts per instance ($1.67 \to 3.03$), yielding a diverse set of candidates than traditional RS.

We propose BioTamperNet, a novel framework for detecting duplicated regions in tampered biomedical images, leveraging affinity-guided attention inspired by State Space Model (SSM) approximations. Existing forensic models, primarily trained on natural images, often underperform on biomedical data where subtle manipulations can compromise experimental validity. To address this, BioTamperNet introduces an affinity-guided self-attention module to capture intra-image similarities and an affinity-guided cross-attention module to model cross-image correspondences. Our design integrates lightweight SSM-inspired linear attention mechanisms to enable efficient, fine-grained localization. Trained end-to-end, BioTamperNet simultaneously identifies tampered regions and their source counterparts. Extensive experiments on the benchmark bio-forensic datasets demonstrate significant improvements over competitive baselines in accurately detecting duplicated regions. All source code and dataset will be publicly available.

We study test-time scaling, where a model improves its answer through multi-round self-reflection at inference. We introduce In-Context Policy Optimization (ICPO), in which an agent optimizes its response in context using self-assessed or externally observed rewards without modifying its parameters. To explain this ICPO process, we theoretically show that with sufficient pretraining under a novel Fisher-weighted logit-matching objective, a single-layer linear self-attention model can provably imitate policy-optimization algorithm for linear bandits. Building on this theory, we propose Minimum-Entropy ICPO (ME-ICPO), a practical algorithm that iteratively uses its response and self-assessed reward to refine its response in-context at inference time. By selecting the responses and their rewards with minimum entropy, ME-ICPO ensures the robustness of the self-assessed rewards via majority voting. Across standard mathematical reasoning tasks, ME-ICPO attains competitive, top-tier performance while keeping inference costs affordable compared with other inference-time algorithms. Overall, ICPO provides a principled understanding of self-reflection in LLMs and yields practical benefits for test-time scaling for mathematical reasoning.

Identifying latent variables and their induced causal structure is fundamental in various scientific fields. Existing approaches often rely on restrictive structural assumptions (e.g., purity) and may become invalid when these assumptions are violated. We introduce Conditional Independent Component Analysis (CICA), a new tool that extracts components that are conditionally independent given latent variables. Under mild conditions, CICA can be optimized using a tractable proxy such as rank-deficiency constraints. Building on CICA, we establish an identifiability theory for linear non-Gaussian acyclic models with latent variables: solving CICA and then applying an appropriate row permutation to the sparsest CICA solution enables recovery of the causal structure. Accordingly, we propose an estimation method based on the identifiability theory and substantiate the algorithm with experiments on both synthetic and real-world datasets.

Classifier-free guidance (CFG) has helped diffusion models achieve great conditional generation in various fields. Recently, more diffusion guidance methods have emerged with improved generation quality and human preference. However, can these emerging diffusion guidance methods really achieve solid and significant improvements? In this paper, we rethink recent progress on diffusion guidance. Our work mainly consists of four contributions. First, we reveal a critical evaluation pitfall that common human preference models exhibit a strong bias towards large guidance scales. Simply increasing the CFG scale can easily improve quantitative evaluation scores due to strong semantic alignment, even if image quality is severely damaged (e.g., oversaturation and artifacts). Second, we introduce a novel guidance-aware evaluation (GA-Eval) framework that employs effective guidance scale calibration to enable fair comparison between current guidance methods and CFG by identifying the effects orthogonal and parallel to CFG effects. Third, motivated by the evaluation pitfall, we design Transcendent Diffusion Guidance (TDG) method that can significantly improve human preference scores in the conventional evaluation framework but actually does not work in practice. Fourth, in extensive experiments, we empirically evaluate recent eight diffusion guidance methods within the conventional evaluation framework and the proposed GA-Eval framework. Notably, simply increasing the CFG scales can compete with most studied diffusion guidance methods, while all methods suffer severely from winning rate degradation over standard CFG. Our work would strongly motivate the community to rethink the evaluation paradigm and future directions of this field.

A/B testing has become a gold standard for modern technological companies to conduct policy evaluation. Yet, its application to time series experiments, where treatments are sequentially assigned over time, remains challenging. Existing designs suffer from two limitations: (i) they do not fully leverage the entire history for treatment allocation; (ii) they rely on strong assumptions to approximate the objective function (e.g., the mean squared error of the estimated treatment effect) for optimizing the design. We first establish an impossibility theorem showing that failure to condition on the full history leads to suboptimal designs, due to the dynamic dependencies in time series experiments. To address both limitations simultaneously, we next propose a transformer reinforcement learning (RL) approach which leverages transformers to condition treatment allocation on the entire history and employs RL to directly optimize the MSE without relying on restrictive assumptions. Empirical evaluations on synthetic data, a publicly available dispatch simulator, and a real-world ridesharing dataset demonstrate that our proposal consistently outperforms existing designs.

Dynamic programming (DP) is central to combinatorial optimization, optimal control, and reinforcement learning, yet its perceived sequentiality has long hindered scalability. We introduce a general-purpose GPU framework that reformulates broad classes of forward DP recursions as batched min--plus matrix--vector products over layered DAGs, collapsing actions into masked state-to-state transitions that map directly to GPU kernels. This approach removes a major bottleneck in scenario-based stochastic programming (SP), where the use of DP has traditionally restricted the number of scenarios due to excessive computational cost. Our framework exposes massive parallelism across scenarios, transition layers, and, when applicable, route or action options, via self-designed GPU kernels that implement Bellman updates with warp-/block-level reductions and numerically safe masking. In a single GPU pass, these kernels can process over $10^6$ uncertainty realizations, far beyond the capacity of prior scenario-based methods. We demonstrate the approach in two canonical SP applications: (i) a vectorized split operator for the capacitated vehicle routing problem with stochastic demand, exploiting **2D** parallelism (scenarios $\times$ transitions); and (ii) a forward inventory reinsertion DP under an order-up-to policy, exploiting **3D** parallelism (scenarios $\times$ inventory transitions $\times$ route options). Across benchmarks, the implementation scales nearly linearly in the number of scenarios and achieves one to three orders of magnitude speedups over multithreaded CPU baselines, yielding tighter SAA estimates and consistently stronger first-stage decisions under identical wall-clock budgets. Viewed as hardware-aware software primitives, our min--plus DP kernels offer a drop-in path to scalable, GPU-accelerated stochastic discrete optimization.

Diffusion transformers (DiT) have already achieved appealing synthesis and scaling properties in content recreation, \emph{e.g.,} image and video generation. However, scaling laws of DiT are less explored, which usually offer precise predictions regarding optimal model size and data requirements given a specific compute budget. Therefore, experiments across a broad range of compute budgets, from 1e17 to 6e18 FLOPs are conducted to confirm the existence of scaling laws in DiT \emph{for the first time}. Concretely, the loss of pretraining DiT also follows a power-law relationship with the involved compute. Based on the scaling law, we can not only determine the optimal model size and required data but also accurately predict the text-to-image generation loss given a model with 1B parameters and a compute budget of 1.5e21 FLOPs. Additionally, we also demonstrate that the trend of pretraining loss matches the generation performances (e.g., FID), even across various datasets, which complements the mapping from compute to synthesis quality and thus provides a predictable benchmark that assesses model performance and data quality at a reduced cost.

This paper analyzes Cross-Entropy (CE) loss in knowledge distillation (KD) for recommender systems. KD for recommender systems targets at distilling rankings, especially among items most likely to be preferred, and can only be computed on a small subset of items. Considering these features, we reveal the connection between CE loss and NDCG in the field of KD. We prove that when performing KD on an item subset, minimizing CE loss maximizes the lower bound of NDCG, only if an assumption of closure is satisfied. It requires that the item subset consists of the student's top items. However, this contradicts our goal of distilling rankings on the teacher's top items. We empirically demonstrate the vast gap between these two kinds of top items. To bridge the gap between our goal and theoretical support, we propose Rejuvenated Cross-Entropy for Knowledge Distillation (RCE-KD). It splits the top items given by the teacher into two subsets based on whether or not it is ranked high by the student. For the subset that defies the condition, a sampling strategy is devised to use teacher-student collaboration to approximate our assumption of closure. We also combine the losses on the two subsets adaptively. Extensive experiments demonstrate the effectiveness of our method. Our code is available at https://anonymous.4open.science/r/RCE-KD.

Geometric graph neural networks (Geo-GNNs) have revolutionized molecular dynamics (MD) simulations by providing accurate and fast energy and force predictions. However, minor prediction errors could still destabilize MD trajectories in real MD simulations due to the limited coverage of molecular conformations in training datasets. Existing methods that focus on in-distribution predictions often fail to address extrapolation to unseen conformations, undermining the simulation stability. To tackle this, we propose Geometric Graph Neural Diffusion (GGND), a novel framework that can capture geometrically invariant topological features, thereby alleviating error accumulation and ensuring stable MD simulations. The core of our framework is that it iteratively refines atomic representations, enabling instantaneous information flow between arbitrary atomic pairs while maintaining equivariance. Our proposed GGND is a plug-and-play module that can seamlessly integrate with existing local equivariant message-passing frameworks, enhancing their predictive performance and simulation stability. We conducted sets of experiments on the 3BPA and SAMD23 benchmark datasets, which encompass diverse molecular conformations across varied temperatures. We also ran real MD simulations to evaluate the stability. GGND outperforms baseline models in both accuracy and stability under significant topological shifts, advancing stable molecular modeling for real-world applications.

Auto-regressive large language models (LLMs) exhibit a non-trivial capacity to "anticipate'' long-range future tokens despite being trained to predict only one token at a time. Nevertheless, how to systematically profile, enhance and leverage such capacity to practically improve LLM reasoning performance remains unclear. In this paper, we propose **Next Token-Bag Exploitation (Next-ToBE)** to tackle this challenge. Next-ToBE quantifies LLM’s anticipatory capacity by measuring how well tokens in the future window are pre-captured by the model’s current softmax probabilities. This capacity is strongly correlated with LLM generative quality but often suppressed by the rigid one-hot objective in next-token prediction. To address this, we replace the {one-hot target vector} in next-token prediction with a soft target distribution spanning additional future tokens. Specifically, the immediate next token retains the highest importance, while more distant ``look-ahead tokens'' are also included to enrich supervision, with their importance dynamically determined by temporal and semantic relevance patterns to inject forward-looking pressure. Besides, the fitting process emphasizes the model’s intrinsic anticipatory tendency, thus preserving the confidence and fidelity of the pre-trained model to improve training stability. Overall, Next-ToBE not only effectively activates LLM anticipatory capacity through fine-tuning, yielding notable gains in reasoning performance with higher memory and computational efficiency against the MTP baselines, but also shows great potential in pretraining setting by successfully cultivating this capacity from scratch. These highlight its value as an effective strategy to extend the prediction horizon of LLMs, enabling them to see further, and reason better.

Group-based reinforcement learning (RL), such as GRPO, has advanced the capabilities of large language models on long-horizon agentic tasks. To enable more fine-grained policy updates, recent research has increasingly shifted toward stepwise group-based policy optimization, which treats each step in a rollout trajectory independently while using a memory module to retain historical context. However, we find a key issue in estimating stepwise relative advantages, namely context inconsistency, where steps within the same group may differ in their historical contexts. Empirically, we reveal that this issue can lead to severely biased advantage estimation, thereby degrading policy optimization significantly. To address the issue, in this paper, we propose Hierarchical-of-Groups Policy Optimization (HGPO) for long-horizon agentic tasks. Specifically, within a group of rollout trajectories, HGPO assigns each step to multiple hierarchical groups according to the consistency of historic contexts. Then, for each step, HGPO computes distinct advantages within each group and aggregates them with an adaptive weighting scheme. In this way, HGPO can achieve a favorable bias-variance trade-off in stepwise advantage estimation, without extra models or rollouts. Evaluations on two challenging agentic tasks, ALFWorld and WebShop with Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct, show that HGPO significantly outperforms existing agentic RL methods under the same computational constraints.

Generative model ecosystems increasingly operate as competitive multi-platform markets, where platforms strategically select models from a shared pool and users with heterogeneous preferences choose among them. Understanding how platforms interact, when market equilibria exist, how outcomes are shaped by model-provider, platforms, and user behavior, and how social welfare is affected is critical for fostering beneficial market environment. In this paper, we formalize a three-layer *model-platfrom-user* market game and identify conditions for the existence of pure Nash equilibrium. Our analysis shows that market structure, whether platforms converge on similar models or differentiate by selecting distinct ones, depends not only on models’ global average performance but also on their localized attraction to user groups. We further examine welfare outcomes and show that expanding the model pool does not necessarily increase user welfare or market diversity. Finally, we design and evaluate best-response training schemes that allow model-provider to strategically introduce new models into competitive markets.

Object-level manipulation—relocating or reorienting objects in images or videos while preserving scene realism—is central to film post-production, AR, and creative editing. Yet existing methods struggle to jointly achieve three core goals: background preservation, geometric consistency under viewpoint shifts, and user-controllable transformations. Geometry-based approaches offer precise control but require explicit 3D reconstruction and generalize poorly; diffusion-based methods generalize better but lack fine-grained geometric control. We present **Ctrl&Shift**, an end-to-end diffusion framework to achieve geometry-consistent object manipulation without explicit 3D representations. Our key insight is to decompose manipulation into two stages—object removal and reference-guided inpainting under explicit camera pose control—and encode both within a unified diffusion process. To enable precise, disentangled control, we design a multi-task, multi-stage training strategy that separates background, identity, and pose signals across tasks. To improve generalization, we introduce a scalable real-world dataset construction pipeline that generates paired image and video samples with estimated relative camera poses. Extensive experiments demonstrate that **Ctrl&Shift** achieves state-of-the-art results in fidelity, viewpoint consistency, and controllability. *To our knowledge, this is the first framework to unify fine-grained geometric control and real-world generalization for object manipulation—without relying on any explicit 3D modeling.*

Reasoning Large Language Models (RLLMs) excelling in complex tasks present unique challenges for digital watermarking, as existing methods often disrupt logical coherence or incur high computational costs. Token-based watermarking techniques can corrupt the reasoning flow by applying pseudo-random biases, while semantic-aware approaches improve quality but introduce significant latency or require auxiliary models. This paper introduces ReasonMark, a novel watermarking framework specifically designed for reasoning-intensive LLMs. Our approach decouples generation into an undisturbed Thinking Phase and a watermarked Answering Phase. We propose a Criticality Score to identify semantically pivotal tokens from the reasoning trace, which are distilled into a Principal Semantic Vector (PSV). The PSV then guides a semantically-adaptive mechanism that modulates watermark strength based on token-PSV alignment, ensuring robustness without compromising logical integrity. Extensive experiments show ReasonMark surpasses state-of-the-art methods by reducing text Perplexity by 0.35, increasing translation BLEU score by 0.164, and raising mathematical accuracy by 0.67 points. These advancements are achieved alongside a 0.34% higher watermark detection AUC and stronger robustness to attacks, all with a negligible increase in latency. This work enables the traceable and trustworthy deployment of reasoning LLMs in real-world applications.

Vision–language models (VLMs) extend large language models (LLMs) with vision encoders, enabling text generation conditioned on both images and text. However, this multimodal integration expands the attack surface by exposing the model to image-based jailbreaks crafted to induce harmful responses. Existing gradient-based jailbreak methods transfer poorly, as adversarial patterns overfit to a single white-box surrogate and fail to generalise to black-box models. In this work, we propose **U**niversa**l** and **tra**nsferable jail**break** (**UltraBreak**), a framework that constrains adversarial patterns through transformations and regularisation in the vision space, while relaxing textual targets through semantic-based objectives. By defining its loss in the textual embedding space of the target LLM, UltraBreak discovers universal adversarial patterns that generalise across diverse jailbreak objectives. This combination of vision-level regularisation and semantically guided textual supervision mitigates surrogate overfitting and enables strong transferability across both models and attack targets. Extensive experiments show that UltraBreak consistently outperforms prior jailbreak methods. Further analysis reveals why earlier approaches fail to transfer, highlighting that smoothing the loss landscape via semantic objectives is crucial for enabling universal and transferable jailbreaks.