ICLR 2026 Papers

Diffusion models are widely used for generative tasks across domains. While pre-trained diffusion models effectively capture the training data distribution, it is often desirable to shape these distributions using reward functions to align with downstream applications. Policy gradient methods, such as Proximal Policy Optimization (PPO), are widely used in the context of autoregressive generation. However, the marginal likelihoods required for such methods are intractable for diffusion models, leading to alternative proposals and relaxations. In this context, we unify variants of Rejection sAmpling based Fine-Tuning (RAFT) as GRAFT, and show that this induces same solution as PPO objective with reshaped rewards. We then introduce P-GRAFT to shape distributions at intermediate noise levels and demonstrate empirically that this can lead to more effective fine-tuning. We mathematically explain this via a bias-variance tradeoff. Motivated by this, we propose inverse noise correction to improve flow models without leveraging explicit rewards. We empirically evaluate our methods on text-to-image(T2I) generation, layout generation, molecule generation and unconditional image generation. Notably, our framework, applied to Stable Diffusion 2, improves over policy gradient methods on popular T2I benchmarks in terms of VQAScore and shows an 8.81% relative improvement over the base model. For unconditional image generation, inverse noise correction improves FID of generated images at lower FLOPs/image.

Large multimodal models (LMMs) have achieved impressive performance on various vision-language tasks, but their substantial computational and memory costs hinder their practical deployment. Existing compression methods often decouple low-rank decomposition and quantization, leading to compounded reconstruction errors, especially in multimodal architectures with cross-modal redundancy. To address this issue, we propose LLaVA-FA, a novel efficient LMM that performs joint low-rank plus quantization approximation in the frequency domain. By leveraging the de-correlation and conjugate symmetry properties of Fourier transform, LLaVA-FA achieves more compact and accurate weight representations. Furthermore, we introduce PolarQuant, a polar-coordinate quantization method tailored for complex matrices, and an optional diagonal calibration (ODC) scheme that eliminates the need for large-scale calibration data. Extensive experimental results demonstrate that our proposed LLaVA-FA outperforms existing efficient multimodal models across multiple benchmarks while maintaining minimal activated parameters and low computational costs, validating its effectiveness as a powerful solution for compressing LMMs.

Converting pretrained attention modules such as *grouped-query attention* (GQA) into *multi-head latent attention* (MLA) can improve expressivity without increasing KV-cache cost, making it attractive for efficient inference. However, existing conversion methods typically apply naïve singular value decomposition (SVD). They focus on minimizing the difference between weight matrices rather than how those weights affect input activations, ignore the covariance structure of activations, and enforce a uniform rank across layers—causing activation drift and degraded attention fidelity. To address these issues, we propose **CARE** (**C**ovariance-**A**ware, **R**ank-**E**nhanced), a MLA conversion pipeline under a fixed KV width. CARE introduces three key steps: (i). **Activation-preserving factorization** — align the approximation with the actual input activations rather than just the weights. (ii). **Adjusted-rank allocation** — distribute a fixed KV budget across layers by giving more capacity to layers that need it most. (iii). **KV-parity mapping** — reparameterize the converted \(K\) and \(V\) to fit the MLA format while keeping the KV-cache size unchanged. Under a matched KV-cache budget, our method consistently outperforms a uniform-rank SVD baseline on Llama-3-8B, delivering up to **331%** relative gains in one-shot evaluation (higher accuracy, lower perplexity). With a brief post-SVD “healing” fine-tune, we fully recover the original model’s accuracy.

Image identity customization aims to synthesize realistic and diverse portraits of a specified identity, given a reference image and a text prompt. This task presents two key challenges: (1) generating realistic portraits that preserve fine-grained facial details of the reference identity, and (2) maintaining identity consistency while achieving strong alignment with the text prompt. Our findings suggest that existing single-stream methods fail to capture and guide fine-grained identity details. To address these challenges, we introduce \textit{OmniPortrait}, a novel diffusion-based framework for fine-grained identity fidelity and high editability in portrait synthesis. Our core idea is pivotal optimization, which leverages dual-stream identity guidance in a coarse-to-fine manner. First, a Pivot ID Encoder is proposed and trained with a face localization loss while avoiding the degradation of editability typically caused by fine-tuning the denoiser. Although this encoder primarily guides coarse-level identity synthesis, it provides a good initialization that serves as the identity pivot for optimization during inference. Second, we propose Reference-Based Guidance, which performs on-the-fly feature matching and optimization over diffusion intermediate features conditioned on the identity pivot. In addition, our approach is able to generalize naturally to multi-identity customized image generation scenarios. Extensive experiments demonstrate significant improvements in both identity preservation and text alignment, establishing a new benchmark for image identity customization.

Video Query Sound Separation (VQSS) aims to isolate target sounds conditioned on visual queries while suppressing off-screen interference—a task central to audiovisual understanding. However, existing methods often fail under conditions of homogeneous interference and overlapping soundtracks, due to limited temporal modeling and weak audiovisual alignment. We propose \textbf{AlignSep}, the first generative VQSS model based on flow matching, designed to address common issues such as spectral holes and incomplete separation. To better capture cross-modal correspondence, we introduce a series of temporal consistency mechanisms that guide the vector field estimator toward learning robust audiovisual alignment, enabling accurate and resilient separation in complex scenes. As a \textit{multi-conditioned generation} task, VQSS presents unique challenges that differ fundamentally from traditional flow matching setups. We provide an in-depth analysis of these differences and their implications for generative modeling. To systematically evaluate performance under realistic and difficult conditions, we further construct \textbf{VGGSound-Hard}, a challenging benchmark composed entirely of separation cases with homogeneous interference and strong reliance on temporal visual cues. Extensive experiments across multiple benchmarks demonstrate that AlignSep achieves state-of-the-art performance both quantitatively and perceptually, validating its practical value for real-world applications. More results and audio examples are available at: \url{https://AlignSep.github.io}.

Logic synthesis (LS), which aims to generate a *compact* logic circuit graph with minimized size while *accurately* satisfying a given functionality, plays an important role in chip design. However, existing LS methods struggle to balance circuit structure compactness and functional accuracy, often leading to suboptimal generation. To address this problem, we propose a novel *Circuit Program Evolution* framework, namely CircuitEvo, which iteratively leverages large language models (LLMs) to evolve circuit programs towards improved compactness while preserving functional accuracy. Specifically, CircuitEvo models the circuit graph as a structured program and leverages the strong generative capabilities of LLMs — guided by domain-specific evolutionary prompt strategies — to generate promising circuit candidates in each iteration. Moreover, a structure-aware circuit optimization module is introduced to correct functional discrepancies by appending necessary substructures to the generated circuits. To the best of our knowledge, CircuitEvo is *the first* LLM-based LS approach that can iteratively improve a circuit's compactness while ensuring functional accuracy. Experiments on several widely used benchmarks demonstrate that CircuitEvo can efficiently generate accurate circuits with up to 16 input number and 69 output number. Moreover, our method significantly outperforms state-of-the-art methods in terms of circuit size, achieving an average improvement of 6.74%.

The role of reasoning in Audio Large Language Models remains widely underexplored, as introducing a reasoning process often degrades rather than improves performance during inference, a phenomenon we term test-time inverse scaling, where longer reasoning chains yield progressively worse results. We demonstrate that this stems not from fundamental limitations of reasoning itself, but from inadequate training: models without proper guidance for the reasoning process produce hallucinatory, inconsistent reasoning that accumulates errors over longer chains. To address these challenges, we introduce CESAR (Consistent, Effective, and Scalable Audio Reasoners), shifting from outcome verification to rewarding the reasoning process. Our online reinforcement learning framework employs Group Relative Policy Optimization with a multi-faceted reward suite that incentivizes not only correctness and format but also consistency, structured analytical patterns, causal reasoning, domain-knowledge integration, and calibrated reasoning depth. CESAR resolves test-time inverse scaling, transforming reasoning from detriments into gains while revealing model-specific "reasoning sweet spots", where performance peaks during test-time scaling. We achieve state-of-the-art results on MMAU Test-mini, substantially outperforming Gemini 2.5 Pro and GPT-4o Audio, and near-human-level performance on MMSU reasoning tasks. Through AI-as-judge evaluations and qualitative comparisons, we provide both quantitative and qualitative validation of our improved reasoning quality. Importantly, enhanced reasoning creates synergistic effects, simultaneously improving multimodal reasoning and perception capabilities. Overall, CESAR establishes a principled method for developing robust and scalable reasoning in Audio LLMs.

Contrastive learning has emerged as a powerful framework for learning generalizable representations, yet its theoretical understanding remains limited, particularly under imbalanced data distributions that are prevalent in real-world applications. Such an imbalance can degrade representation quality and induce biased model behavior, yet a rigorous characterization of these effects is lacking. In this work, we develop a theoretical framework to analyze the training dynamics of contrastive learning with Transformer-based encoders under imbalanced data. Our results reveal that neuron weights evolve through three distinct stages of training, with different dynamics for majority features, minority features, and noise. We further show that minority features reduce representational capacity, increase the need for more complex architectures, and hinder the separation of ground-truth features from noise. Inspired by these neuron-level behaviors, we show that pruning restores performance degraded by imbalance and enhances feature separation, offering both conceptual insights and practical guidance. Major theoretical findings are validated through numerical experiments.

Credal networks extend Bayesian networks by incorporating imprecise probabilities through convex sets of probability distributions known as credal sets. MAP inference in credal networks, which seeks the most probable variable assignment given evidence, becomes inherently more difficult than in Bayesian networks because it involves computations over a complex joint credal set. In this paper, we introduce two tasks called \emph{maximax} and \emph{maximin} MAP, and develop depth-first branch-and-bound search algorithms for solving them \emph{exactly}. The algorithms exploit problem decomposition by exploring an AND/OR search space and use a partitioning-based heuristic function enhanced with a cost-shifting scheme to effectively guide the search. Our experimental results obtained on both random and realistic credal networks clearly demonstrate the effectiveness of the proposed algorithms as they scale to large and complex problem instances.

We introduce $\pi^3$, a feed-forward neural network that offers a novel approach to visual geometry reconstruction, breaking the reliance on a conventional fixed reference view. Previous methods often anchor their reconstructions to a designated viewpoint, an inductive bias that can lead to instability and failures if the reference is suboptimal. In contrast, $\pi^3$ employs a fully permutation-equivariant architecture to predict affine-invariant camera poses and scale-invariant local point maps without any reference frames. This design not only makes our model inherently robust to input ordering, but also leads to higher accuracy and performance. These advantages enable our simple and bias-free approach to achieve state-of-the-art performance on a wide range of tasks, including camera pose estimation, monocular/video depth estimation, and dense point map reconstruction. Code and models will be publicly available.

Recent benchmarks for Large Language Model (LLM) agents primarily focus on evaluating reasoning, planning, and execution capabilities, while another critical component—memory, encompassing how agents memorize, update, and retrieve long-term information—is under-evaluated due to the lack of benchmarks. We term agents with memory mechanisms as memory agents. In this paper, based on classic theories from memory science and cognitive science, we identify four core competencies essential for memory agents: accurate retrieval, test-time learning, long-range understanding, and selective forgetting. Existing benchmarks either rely on limited context lengths or are tailored for static, long-context settings like book-based QA, which do not reflect the interactive, multi-turn nature of memory agents that incrementally accumulate information. Moreover, no existing benchmarks cover all four competencies. We introduce MemoryAgentBench, a new benchmark specifically designed for memory agents. Our benchmark transforms existing long-context datasets and incorporates newly constructed datasets into a multi-turn format, effectively simulating the incremental information processing characteristic of memory agents. By carefully selecting and curating datasets, our benchmark provides comprehensive coverage of the four core memory competencies outlined above, thereby offering a systematic and challenging testbed for assessing memory quality. We evaluate a diverse set of memory agents, ranging from simple context-based and retrieval-augmented generation (RAG) systems to advanced agents with external memory modules and tool integration. Empirical results reveal that current methods fall short of mastering all four competencies, underscoring the need for further research into comprehensive memory mechanisms for LLM agents.

Multimodal Large Language Models (MLLMs) encode images into visual tokens, aligning visual and textual signals within a shared latent space to facilitate cross-modal representation learning. The CLIP model is a widely adopted foundational vision language model whose vision encoder has played a critical role in the development of MLLMs such as LLaVA. However, the CLIP vision encoder suffers from notable limitations including being constrained to only handling fixed input resolutions and a failure to produce separated embeddings for dissimilar images. Replacing the vision encoder of an existing model typically incurs substantial computational costs because such a change often necessitates retraining the entire model pipeline. In this work, we identify two factors which underlie the limitations of the CLIP vision encoder: mesoscopic bias and interpolation bias. To address these issues, we propose QLIP, a drop-in replacement for CLIP that can be seamlessly integrated with existing MLLMs with only a few lines of code and can enhance both coarse-grained and fine-grained visual understanding, without re-training. QLIP is designed around an image quadtree which replaces the standard uniform grid patches with a novel content aware patchification. Our experimental results demonstrate that QLIP improves the general visual question answering accuracy of the LLaVA-1.5 model series across various model sizes—without requiring retraining or fine-tuning of the full MLLM. Notably, QLIP boosts detailed understanding performance on the challenging $V^*$ benchmark by up to 13.6%.

Sparse autoencoders (SAEs) have recently emerged as a powerful tool for interpreting the features learned by large language models (LLMs). By reconstructing features with sparsely activated networks, SAEs aim to recover complex superposed polysemantic features into interpretable monosemantic ones. Despite their wide applications, it remains unclear under what conditions SAEs can fully recover the ground truth monosemantic features from the superposed polysemantic ones. In this paper, we provide the first theoretical analysis with a closed-form solution for SAEs, revealing that they generally fail to fully recover the ground truth monosemantic features unless the ground truth features are extremely sparse. To improve the feature recovery of SAEs in general cases, we propose a reweighting strategy targeting at enhancing the reconstruction of the ground truth monosemantic features instead of the observed polysemantic ones. We further establish a theoretical weight selection principle for our proposed weighted SAE (WSAE). Experiments across multiple settings validate our theoretical findings and demonstrate that our WSAE significantly improves feature monosemanticity and interpretability.

Segmentation in pathology image requires handling out-of-domain tissue morphologies and new pathologies beyond training distributions, where traditional closed-set segmentation approaches fail to generalize. Reasoning segmentation enables zero-shot generalization via prompting with text queries. However, existing reasoning segmentation models face three barriers when applied to pathology: (1) the vision encoder lack pathology-specific knowledge and robustness to staining variations, (2) the large language model (LLM) backbone for reasoning fails to identify whether it has gathered sufficient semantic context to trigger the segmentation output, and (3) no reasoning segmentation benchmarks and datasets exist for pathology analysis. Consequently, we introduce PathChat-SegR1, a reasoning segmentation model built upon pathology-specific vision encoders trained with a novel stain-invariant self-distillation for robust pathology image representations. Moreover, we propose Segmentation-Optimized GRPO (SO-GRPO), a reinforcement learning method specifically for reasoning segmentation that learns to determine optimal segmentation timing based on accumulated reasoning context. Finally, we construct a pathology-specific reasoning segmentation benchmark of 118,667 triplets of pathology image, ground-truth mask, query, and reasoning chain including both public and private pathology images. Zero-shot evaluation on pathology images with out-of-domain morphologies/pathologies shows 61\% improvement over state-of-the-art segmentation models.

Conditional image generation augments text-to-image synthesis with structural, spatial, or stylistic priors and is used in many domains. However, current methods struggle to harmonize guidance from both sources when conflicts arise: 1) input-level conflict, where the semantics of the conditioning image contradict the text prompt, and 2) model-bias conflict, where learned generative biases hinder alignment even when the condition and text are compatible. These scenarios demand nuanced, case-by-case trade-offs that standard supervised fine-tuning struggles to deliver. Preference-based optimization techniques, such as Direct Preference Optimization (DPO), offer a promising solution but remain limited: naive DPO suffers from gradient entanglement between text and condition signals and lacks disentangled, conflict-aware training data for multi-constraint tasks. To overcome these issues, we propose a self-driven, bidirectionally decoupled DPO framework (BideDPO). At its core, our method constructs two disentangled preference pairs for each sample—one for the condition and one for the text—to mitigate gradient entanglement. The influence of these pairs is then managed by an Adaptive Loss Balancing strategy for balanced optimization. To generate these pairs, we introduce an automated data pipeline that iteratively samples from the model and uses vision-language model checks to create disentangled, conflict-aware data. Finally, this entire process is embedded within an iterative optimization strategy that progressively refines both the model and the data. We construct a DualAlign benchmark to evaluate a model’s ability to resolve conflicts between text and condition, and experiments on commonly used modalities show that BideDPO delivers substantial gains in both text success rate (e.g., +35\%) and condition adherence. We also validated the robustness of our approach on the widely used COCO dataset. All models, code, and benchmarks will be released to support future work.

3D single object tracking (SOT) remains a highly challenging task due to the inherent crux in learning representations from point clouds to effectively capture both spatial shape features and temporal motion features. Most existing methods employ a category-specific optimization paradigm, training the tracking model individually for each object category to enhance tracking performance, albeit at the expense of generalizability across different categories. In this work, we propose a robust category-unified 3D SOT model, referred to as SpatioTemporal Point Relation Parsing model (*PointRePar*), which is capable of joint training across multiple categories while excelling in unified feature learning for both spatial shapes and temporal motions. Specifically, the proposed *PointRePar* captures and parses the latent point relations across both spatial and temporal domains to learn superior shape and motion characteristics for robust tracking. On the one hand, it models the multi-scale spatial point relations using a Mamba-based U-Net architecture with adaptive point-wise feature refinement. On the other hand, it captures both the point-level and box-level temporal relations to exploit the latent motion features. Extensive experiments across three benchmarks demonstrate that our *PointRePar* not only outperforms the existing category-unified 3D SOT methods significantly, but also compares favorably against the state-of-the-art category-specific methods. Codes will be released.

Most existing benchmarks for egocentric vision understanding focus primarily on daytime scenarios, overlooking the low-light conditions that are inevitable in real-world applications. To investigate this gap, we present EgoNight, the first comprehensive benchmark for nighttime egocentric vision, with visual question answering (VQA) as the core task. A key feature of EgoNight is the introduction of day–night aligned videos, which enhance night annotation quality using the daytime data and reveal clear performance gaps between lighting conditions. To achieve this, we collect both synthetic videos rendered by Blender and real-world recordings, ensuring that scenes and actions are visually and temporally aligned. Leveraging these paired videos, we construct EgoNight-VQA, supported by a novel day-augmented night auto-labeling engine and refinement through extensive human verification. Each QA pair is double-checked by annotators for reliability. In total, EgoNight-VQA contains 3658 QA pairs across 90 videos, spanning 12 diverse QA types, with more than 300 hours of human work. Evaluations of the state-of-the-art multimodal large language models (MLLMs) reveal substantial performance drops when transferring from day to night, underscoring the challenges of reasoning under low-light conditions. Beyond VQA, EgoNight also introduces two auxiliary tasks, day–night correspondence retrieval and egocentric depth estimation at night, that further explore the boundaries of existing models. We believe EgoNight-VQA provides a strong foundation for advancing application-driven egocentric vision research and for developing models that generalize across illumination domains.

Content Warning: This paper contains examples with harmful content, and the constructed dataset includes samples that may be considered offensive. As large language models (LLMs) continue to advance in capabilities, ensuring their safety against jailbreak attacks remains a critical challenge. In this paper, we introduce a novel safety alignment approach called Answer-Then-Check, which enhances LLM robustness against malicious prompts by applying thinking ability to mitigate jailbreaking problems before producing a final answer to the user. Our method enables models to answer the question directly in their thoughts and then critically evaluate its safety before deciding whether to provide it. To implement this approach, we construct the Reasoned Safety Alignment (ReSA) dataset, comprising 80K examples that teach models to reason through direct responses and then analyze their safety. Experimental results demonstrate that our approach achieves the Pareto frontier with superior safety capability while decreasing over-refusal rates on over-refusal benchmarks. Notably, the model fine-tuned with ReSA maintains general reasoning capabilities on benchmarks like MMLU, MATH500, and HumanEval. Besides, our method equips models with the ability to perform safe completion. Unlike post-hoc detection methods that can only reject harmful queries, our model can provide helpful and safe alternative responses for sensitive topics (e.g., self-harm). Our results show that inference-time strategies alone are insufficient, highlighting the necessity of safety training, and we find even 500 samples can yield performance comparable to the entire dataset, suggesting a promising path for data-efficient safety alignment.

Large Audio Language Models (LALMs) integrate the audio modality directly into the model, rather than converting speech into text and inputting text to Large Language Models (LLMs). While jailbreak attacks on LLMs have been extensively studied, the security of LALMs with audio modalities remains largely unexplored. Currently, there is a lack of an adversarial audio dataset and a unified framework specifically designed to evaluate and compare attacks and LALMs. In this paper, we present JALMBench, a comprehensive benchmark to assess the safety of LALMs against jailbreak attacks. JALMBench includes a dataset containing 11,316 text samples and 245,355 audio samples (>1,000 hours). It supports 12 mainstream LALMs, 4 text-transferred and 4 audio-originated attack methods, and 5 defense methods. Using JALMBench, we provide an in-depth analysis of attack efficiency, topic sensitivity, voice diversity, and architecture. Additionally, we explore mitigation strategies for the attacks at both the prompt level and the response level. We find that LALM safety is strongly influenced by modality and architectural choices, demonstrating that text-based safety alignment can partially transfer to audio inputs and that interleaving audio-text strategy enables more robust cross-modal generalization of safety. Moreover, current general moderation for jailbreak only slightly improves security, and calls on the community to explore more defense methods for LALMs. Our work is the first to systematically uncover these design principles, providing a roadmap for building resilient multimodal language models.