ICML 2026 Papers

KV-cache retrieval is essential for long-context LLM inference, yet existing methods struggle with distribution drift and high latency at scale. We introduce **ParisKV**, a drift-robust, GPU-native KV-cache retrieval framework based on collision-based candidate selection, followed by a quantized inner-product reranking estimator. For million-token contexts, ParisKV supports CPU-offloaded KV caches via Unified Virtual Addressing (UVA), enabling on-demand top-$k$ fetching with minimal overhead. ParisKV matches or outperforms full attention quality on both **long-input** and **long-generation** benchmarks. It achieves state-of-the-art long-context decoding efficiency: it matches or exceeds full-attention speed even at batch size 1 for long contexts, delivers up to **2.8$\times$** higher throughput within full attention’s runnable range, and scales to **million-token** contexts where full attention runs out of memory. At million-token scale, ParisKV reduces decode latency by **17$\times$** and **44$\times$** compared to MagicPIG and PQCache, respectively—two state-of-the-art KV-cache top-$k$ retrieval baselines; code is available at https://github.com/amy-77/ParisKV/tree/main.

In everyday photography, aesthetically appealing moments are often captured with structural flaws (e.g., composition, camera viewpoint, or pose) that existing retouching and portrait enhancement methods cannot fix. We formulate Aesthetic Photo Reconstruction (APR) as improving a photo’s aesthetic quality via structural reconstruction while preserving subject identity and scene semantics. Although recent advances in image editing models make APR feasible, they often lack aesthetic understanding, yielding edits that are semantically plausible yet aesthetically weak. To address this, we propose AesFormer, a two-stage framework that decouples aesthetic planning from image editing. In Stage 1, an aesthetic action model (AesThinker) analyzes the input along seven progressive photographic dimensions and outputs executable editing actions; we further apply GRPO-A to encourage broad exploration over diverse action plans beyond SFT. In Stage 2, an action-conditioned editor (AesEditor) performs structural edits guided by these actions. To support APR, we build a video-based corpus-mining pipeline (VCMP) and construct AesRecon, a benchmark of 9,071 strictly aligned (poor, good) image pairs. Experiments show that AesFormer substantially improves APR performance and is competitive with Nano Banana Pro. Code is available at: https://github.com/PKU-ICST-MIPL/AesFormer_ICML2026.

Prior probing-based analyses show that individual layers specialize in distinct linguistic and semantic functions, while a complementary line of work observes that deeper layers enter an alignment-dominated regime where residual transformations become increasingly constrained relative to the evolving hidden-state geometry. These findings appear to conflict: layer-wise functional specialization versus apparent representational stagnation in deeper layers. This raises a key question: Is the observed representational stagnation in deeper layers attributable to suboptimal acquisition or encoding of novel features, or are these features learned appropriately but exhibit minimal marginal contribution to the model’s predictive output? We address this by separating model behavior into (i) representational capacity: the richness and spectral diversity of the encoded features, and (ii) accessibility: the extent to which these features are aligned with, and exploited by, the output-relevant subspace. Analyzing hidden-state second-moment geometry across depth shows that intermediate layers maintain a broad representational span and rich spectral diversity, yet project only weakly onto the output-relevant subspace, indicating intact representational capacity but constrained accessibility within task-relevant dimensions. To probe their functional relevance, we add a controlled diagnostic pathway that routes intermediate-layer embeddings to the terminal readout. Analysis of the induced behavioral changes indicates that the apparent stagnation is better accounted for by geometric limitations on accessibility, rather than by a deficit in representational capacity. The source code is available here.

Vision language models (VLMs) extend the reasoning capabilities of large language models (LLMs) to cross-modal settings, yet remain highly vulnerable to multimodal jailbreak attacks. Existing defenses predominantly rely on safety fine-tuning or \textit{aggressive} token manipulations, incurring substantial training costs or significantly degrading utility. Recent research shows that LLMs inherently recognize unsafe content in text, and the incorporation of visual inputs in VLMs frequently dilutes risk-related signals. Motivated by this, we propose Risk Awareness Injection (RAI), a \textit{lightweight} and training-free framework for safety calibration that restores LLM-like risk recognition by amplifying unsafe signals in VLMs. Specifically, RAI constructs an Unsafe Prototype Subspace from language embeddings and performs targeted modulation on selected high-risk visual tokens, explicitly activating safety-critical signals within the cross-modal feature space. This modulation restores the model’s LLM-like ability to detect unsafe content from visual inputs, while preserving the semantic integrity of original tokens for cross-modal reasoning. Extensive experiments across multiple jailbreak and utility benchmarks demonstrate that RAI substantially reduces attack success rate without compromising task performance.

Large Vision-Language Models (LVLMs) represent a significant leap towards empathetic agents, demonstrating remarkable capabilities in emotion understanding. However, the internal mechanisms governing how LVLMs translate abstract visual stimuli into coherent emotional narratives remain largely unexplored, primarily due to the scarcity of visual counterfactuals and the diffuse nature of emotional expression. In this paper, we bridge this gap by introducing a steering-vector-based causal attribution framework tailored for descriptive emotional reasoning. To this end, we construct a specialized dataset to demystify the emotional circuits underlying the three-stage '' Adapt-Aggregate-Execute'' mechanism. Crucially, we discover a functional decoupling: visual emotional cues are aggregated in middle layers via $\textit{sentiment-specific}$ attention heads, but are subsequently translated into narrative generation in deep layers through $\textit{emotion-general}$ pathways. Guided by these insights, we regulate the emotional information routing to strengthen attention flow and amplify the semantic activation to consolidate expression. Extensive experiments on the comprehensive MER-UniBench demonstrate that our methods significantly improve performance via inference-time intervention, effectively mitigating emotional hallucinations and corroborating the causal fidelity of the discovered circuits.

Texture recognition is often framed as matching an image to a static training-set dictionary or codebook. In practice, this assumption is brittle: label-preserving transformations (illumination, scale, compression, blur) can shift test features away from the fixed training dictionary, producing a training-set codebook misalignment that limits accuracy. We propose STD-Former (Simple Texture Dictionary Transformer), a lightweight framework for image-conditioned texture dictionary encoding. Instead of comparing against a static codebook, STD-Former extracts a compact set of Intrinsic Textons (dictionary atoms / codewords) from the input image itself, yielding self-aligned representations at inference. Our design is intentionally simple and uses a decoupled two-stage recipe. In Stage 1, a Texture Dictionary Extractor (TDE) is pre-trained with a self-supervised Texton Coverage Loss that encourages the learned textons to collectively cover the image patch feature manifold. In Stage 2, a classifier is trained on the encoded dictionary representation; optionally, we add a Sparse Topological Loss derived from 0D persistent homology, which is equivalent to supervising only the (B-1) edges of a minimum spanning tree (MST) in each batch, providing efficient structure regularization. Across six standard texture benchmarks, STD-Former and STD-Former+ achieve new state-of-the-art results.

Hyperspectral image change detection (HSI-CD) suffers from severe annotation scarcity and complex change patterns, which fundamentally limit the effectiveness of directly fine-tuning pre-trained foundation models. Although semi-supervised learning provides a promising direction, existing approaches mainly rely on confidence-based pseudo-label selection, leading to limited data diversity or severe error propagation. In this paper, we propose Influence-Aware Semi-supervised Fine-tuning (IA-SFT), a novel framework that evaluates the influence of pseudo-labels on model decision behavior to identify truly valuable supervision signals. Instead of confidence-based selection, IA-SFT evaluates each low-confidence pseudo-label by measuring its impact on labeled data, enabling reliable filtering of high-value pseudo-labels with minimal noise. To further adapt foundation models to HSI-CD, we design an Adaptive Fusion Change Decoder (AFCD) that jointly models global semantic consistency and local change details. Extensive experiments on three benchmark datasets demonstrate that IA-SFT consistently improves pseudo-label quality and detection performance, achieving superior accuracy compared to state-of-the-art methods. Additional analyses validate the transferability of IA-SFT when integrated into different frameworks in a plug-and-play manner. Code will be released.

Robotic manipulation benefits from foundation models that describe goals, but today's agents still lack a principled way to learn from their own mistakes. We ask whether natural language can serve as feedback, an error-reasoning signal that helps embodied agents diagnose what went wrong and correct course. We introduce LAGEA (Language Guided Embodied Agents), a framework that turns episodic, schema-constrained reflections from a vision language model (VLM) into temporally grounded guidance for reinforcement learning. LAGEA summarizes each attempt in concise language, localizes the decisive moments in the trajectory, aligns feedback with visual state in a shared representation, and converts goal progress and feedback agreement into bounded, step-wise shaping rewardswhose influence is modulated by an adaptive, failure-aware coefficient. This design yields dense signals early when exploration needs direction and gracefully recedes as competence grows. On the Meta-World MT10 and Robotic Fetch embodied manipulation benchmarks, LAGEA improves average success over the state-of-the-art (SOTA) methods by 9.0% on random goals, 5.3% on fixed goals, and 17% on fetch tasks, while converging faster. These results support our hypothesis: language, when structured and grounded in time, is an effective mechanism for teaching robots to self-reflect on mistakes and make better choices. Code will be released soon.

Text-to-video (T2V) generation aims to synthesize videos with high visual quality and temporal consistency that are semantically aligned with input text. Reward-based post-training has emerged as a promising direction to improve the quality and semantic alignment of generated videos. However, recent methods either rely on large-scale human preference annotations or operate on misaligned embeddings from pre-trained vision-language models, leading to limited scalability or suboptimal supervision. We present PISCES, an annotation-free post-training algorithm that addresses these limitations via a novel Dual Optimal Transport (OT)-aligned Rewards module. To align reward signals with human judgment, PISCES uses OT to bridge text and video embeddings at both distributional and discrete token levels, enabling reward supervision to fulfill two objectives: (i) a Distributional OT-aligned Quality Reward that captures overall visual quality and temporal coherence; and (ii) a Discrete Token-level OT-aligned Semantic Reward that enforces semantic, spatio-temporal correspondence between text and video tokens. To our knowledge, PISCES is the first to improve annotation-free reward supervision in generative post-training through the lens of OT. Experiments on both short- and long-video generation show that PISCES outperforms both annotation-based and annotation-free methods on VBench across Quality and Semantic scores, with human preference studies further validating its effectiveness. We show that the Dual OT-aligned Rewards module is compatible with multiple optimization paradigms, including direct backpropagation and reinforcement learning fine-tuning.

With the rise of large-scale foundation models, efficiently adapting them to downstream tasks remains a central challenge. Linear probing, which freezes the backbone and trains a lightweight head, is computationally efficient but often restricted to last-layer representations. We show that task-relevant information is distributed across the network hierarchy rather than encoded solely in the last layers. To leverage this distribution of information, we apply an attentive probing mechanism that dynamically fuses representations from all layers of a Vision Transformer. This attentive layer fusion (ALF) learns to identify the most relevant layers for a target task and combines low-level structural cues with high-level semantic abstractions. Across 20 diverse datasets and multiple pretrained foundation models, ALF achieves consistent, substantial gains over standard linear probes. Attention heatmaps further reveal that tasks different from the pre-training domain benefit most from intermediate representations. Overall, our findings underscore the value of intermediate layers and demonstrate a principled, task-aware approach for unlocking their potential for probing-based adaptation.

Recent advances in large image editing models have shifted the paradigm from text-driven instructions to vision-prompt editing, where user intent is inferred directly from visual inputs such as marks, arrows, and visual–text prompts. While this paradigm greatly expands usability, it also introduces a critical and underexplored safety risk: the attack surface itself becomes visual. In this work, we propose Vision-Centric Jailbreak Attack (VJA), the first visual-to-visual jailbreak attack that conveys malicious instructions purely through visual inputs. To systematically study this emerging threat, we introduce IESBench, a safety-oriented benchmark for image editing models. Extensive experiments on IESBench demonstrate that VJA effectively compromises state-of-the-art commercial models, achieving attack success rates of up to 80.9% on Nano Banana Pro and 70.1% on GPT-Image-1.5. To mitigate this vulnerability, we propose a training-free defense based on introspective multimodal reasoning, which substantially improves the safety of poorly aligned models to a level comparable with commercial systems, without auxiliary guard models and with negligible computational overhead. Our findings expose new vulnerabilities, provide both a benchmark and practical defense to advance safe and trustworthy modern image editing systems.

Large Multimodal Models (LMMs) have recently demonstrated their proficiency in holistic visual comprehension. However, most of them struggle to tackle region-level perception guided by visual prompts, especially for cases where multiple regions are referred simultaneously, or scenarios where global contexts are necessary for precise visual referring. We introduce Contextual Latent Steering (CSteer), a training-free approach for guiding general LMMs to refer multiple regions contextually, without expensive fine-tuning or architectural modifications. CSteer starts with pre-computing contextual vectors that implicitly represent visual referring behaviors, such as differentiation among regions and attention to global contexts, followed by representation editing during inference time. Experimental results on multiple datasets indicate that general LMMs with CSteer outperform referring LMMs in most cases, suggesting a promising solution in training-free, and setting new state-of-the-art for this field. Code is available at https://github.com/xing0047/csteer.git.

While multimodal large language models (MLLMs) have made substantial progress in single-image spatial reasoning, multi-image spatial reasoning, which requires integration of information from multiple viewpoints, remains challenging. Cognitive studies suggest that humans address such tasks through two mechanisms: *cross-view correspondence*, which identifies regions across different views that correspond to the same physical locations, and *stepwise viewpoint transformation*, which composes relative viewpoint changes sequentially. However, existing studies incorporate these mechanisms only partially and often implicitly, without explicit supervision for both. We propose Human-Aware Training for Cross-view correspondence and viewpoint cHange (HATCH), a training framework with two complementary objectives: (1) Patch-Level Spatial Alignment, which encourages patch representations to align across views for spatially corresponding regions, and (2) Action-then-Answer Reasoning, which requires the model to generate explicit viewpoint transition actions before predicting the final answer. Experiments on three benchmarks demonstrate that HATCH consistently outperforms baselines of comparable size by a clear margin and achieves competitive results against much larger models, while preserving single-image reasoning capabilities.

Recently, the physics reasoning capabilities of (M)LLMs have attracted growing attention. However, existing physics benchmarks lack systematic coverage of recent physics Olympiads and direct comparison with human contestants. We present **HiPhO**, the first benchmark dedicated to high school physics Olympiads with human-aligned evaluation. HiPhO highlights three key innovations. **(1) Comprehensive data:** it compiles 13 latest Olympiads from 2024--2025, covering international and regional competitions and spanning mixed modalities from text-only to diagram-based problems. **(2) Professional evaluation:** it adopts official rubrics for fine-grained answer- and step-level grading aligned with human examiners. **(3) Human-level comparison:** it assigns gold, silver, and bronze medals to models based on official medal scores, enabling direct comparison with human contestants. Evaluating 30 (M)LLMs across 13 exams, we find that most open-source MLLMs remain at or below the bronze level, open-source LLMs demonstrate notable progress with multiple gold medals, and closed-source MLLMs achieve 6-13 gold medals, while most models still fall well short of full marks. These results underscore the substantial gap between current (M)LLMs and top human contestants, as well as the room for further improvement. The dataset and leaderboard are available at https://huggingface.co/datasets/SciYu/HiPhO and https://phyarena.github.io, respectively.

Multi-label recognition with frozen Vision-Language Models (VLMs) is brittle under distribution shift: standard zero-shot inference scores labels independently, ignoring co-occurrence structure and producing incoherent label sets where dominant concepts suppress weaker but compatible labels. We introduce Bayesian Conditional Priors (BCP) Estimation, a gradient-free test-time adaptation method that injects label dependency without tuning the backbone. BCP views zero-shot logits as a proxy for marginal posteriors under a fixed image-text likelihood and attributes shift-induced errors mainly to a mismatched label prior. For each test image, it selects a high-confidence anchor label and applies an anchor-conditioned Bayesian refinement. This update is closed-form in logit space and admits a pointwise mutual information (PMI) interpretation, explicitly promoting compatible labels and suppressing incompatible ones. BCP operates without target annotations by estimating anchor-conditioned priors online from the unlabeled test stream via lightweight second-order co-occurrence statistics, adding negligible overhead beyond a single forward pass. Across standard multi-label benchmarks and multiple CLIP backbones, BCP consistently outperforms strong TTA baselines, e.g., improving RN50 average mAP from 57.31 to 69.22 and ViT-B/16 from 62.61 to 71.79.

Long-horizon agentic reasoning necessitates effectively compressing growing interaction histories into a limited context window. Most existing memory systems serialize history as text, where token-level cost is uniform and scales linearly with length, often spending scarce budget on low-value details. To this end, we introduce **MemOCR**, a multimodal memory agent that improves long-horizon reasoning under tight context budgets by allocating memory space with adaptive information density through visual layout. Concretely, MemOCR maintains a structured rich-text memory (e.g., headings, highlights) and renders it into an image that the agent consults for memory access, visually prioritizing crucial evidence while aggressively compressing auxiliary details. To ensure robustness across varying memory budgets, we train MemOCR with reinforcement learning under budget-aware objectives that expose the agent to diverse compression levels. Across long-context multi-hop and single-hop question-answering benchmarks, MemOCR outperforms strong text-based baselines and achieves more effective context utilization under extreme budgets.

Large multimodal models (LMMs) have shown strong capabilities in audio-visual understanding by jointly reasoning over visual, auditory, and linguistic inputs. However, processing long-form audio-visual content often requires a large number of tokens, leading to substantial computational and memory overhead during inference. Existing efficiency-oriented methods typically apply uniform compression or pruning strategies, which overlook modality-specific characteristics and instruction-dependent reasoning behaviors in multimodal models. In this work, we present PRIM, an inference framework for efficient multimodal reasoning that systematically compresses audio-visual representations based on attention dynamics and instruction relevance. Motivated by an attention-based analysis revealing modality imbalance and layer-wise redundancy in LMMs, PRIM introduces a cooperative compression pipeline that spans both multimodal encoders and the language model. Specifically, PRIM performs early text-conditioned audio-visual fusion to externalize cross-modal interactions, applies attention-guided and frequency-aware strategies to remove redundant audio and video tokens, and further adapts token retention inside the language model according to task demands. Extensive experiments on multiple audio-visual benchmark datasets demonstrate that PRIM consistently achieves stable and superior efficiency--accuracy trade-offs across diverse tasks and datasets. These results demonstrate that PRIM, a multimodal cooperative compression approach, provides an effective pathway toward scalable and efficient audio-visual reasoning.

Vision Transformers (ViTs) have achieved remarkable success across various vision tasks, yet their deployment is often hindered by prohibitive computational costs. While structured weight pruning and token compression have emerged as promising solutions, they suffer from prolonged retraining and inter-layer dependencies that complicate optimization, respectively. We propose ToaSt, a decoupled framework applying specialized strategies to distinct ViT components. We apply coupled head-wise structured pruning to Multi-Head Self-Attention modules, leveraging attention operation characteristics to enhance robustness. For Feed-Forward Networks (over 60\% of FLOPs), we introduce Token Channel Selection (TCS), a training-free method that filters redundant noise channels at inference time. Extensive evaluations across nine diverse models, including DeiT, ViT-MAE, and Swin Transformer, demonstrate that ToaSt achieves superior trade-offs between accuracy and efficiency, consistently outperforming existing baselines. On ViT-MAE-Huge, ToaSt achieves 88.52\% accuracy (+1.64\%p) with 39.4\% FLOPs reduction. ToaSt also transfers effectively to diverse downstream tasks (COCO detection, ADE20K segmentation, CIFAR-100 classification), achieving 52.2 versus 51.9 mAP on COCO. Code: \href{https://github.com/SHANNonLab-HUFS/ToaSt}{github.com/SHANNonLab-HUFS/ToaSt}.

Scaling laws for dense LLMs under infinite data are well explored, but how sparsity interacts with limited data is not. In this work, we study sparse training in data-constrained regimes where limited unique tokens require multi-epoch training. Our experiments span models up to 1.92B parameters in the fitting set, sparsity up to 93.75\%, unique data budgets up to 2.6B tokens, and total training tokens up to 41.6B over 16 epochs; we further validate extrapolation on held-out dense-equivalent models up to 7.68B parameters. We find that: **1. Sparse scaling in data-limited settings:** We introduce a scaling law that models loss as a function of active parameters, unique tokens, data repetition, and sparsity, accurately predicting performance across compute and data budgets. **2. Delayed data saturation:** sparse training postpones diminishing returns from repeated data, making multi-epoch training more effective. **3. Resource trade-offs:** With fixed data, loss-optimal sparsity is moderate (~ 50%), while compute-optimal sparsity is higher and grows with data scale. Overall, sparsity is not just a tool for efficiency, but a mechanism for improving scaling trade-offs under data scarcity. Our code is available at: https://github.com/boqian333/sparse-dc-scaling.