ICLR 2026 论文列表

Foundation Vision-Language Models (VLMs) excel across benchmarks yet remain vulnerable to adversarial attacks. While adversarial fine-tuning improves robustness, attaining a desirable clean–robust performance trade-off typically requires costly hyperparameter searches with multiple retraining runs. A promising alternative is to merge task vectors (i.e., parameter displacements from pre-trained models) to balance accuracy and robustness without retraining. However, we find that naive task-vector merging produces a near-linear trade-off, as it equally weights all coordinates and fails to distinguish weights that aid both objectives from those that create conflicts. To overcome this limitation, we propose a prediction stability-aware merging framework that composes task vectors from off-the-shelf naturally and robustly fine-tuned VLMs. Our key insight is that prediction stability serves as a proxy for cross-objective compatibility, enabling us to favor perturbation-invariant parameters while attenuating those with high cross-objective impact. Specifically, we estimate per-parameter stability from gradients under both objectives, building complementary masks that retain jointly stable coordinates while suppressing counterpart-sensitive ones. We further refine these masks along adversarial parameter trajectories, with steps weighted by a prediction-sensitivity index. Our theoretical analysis shows that the masks provably contract first-order cross-objective interference, and the prediction criticality index tracks curvature, biasing the merge toward flatter minima and better generalization. Extensive experiments across benchmarks and scenarios demonstrate our method consistently achieves superior clean–robust trade-offs over prior approaches, with the learned balance transferring effectively to downstream tasks.

Large language models (LLMs) often experience language confusion, which is the unintended mixing of languages during text generation. Current solutions to this problem either necessitate model retraining or cannot differentiate between harmful confusion and acceptable code-switching. This paper introduces the \textbf{Language Confusion Gate} (LCG), a lightweight, plug-in solution that filters tokens during decoding without altering the base LLM. The LCG is trained using norm-adjusted self-distillation to predict appropriate language families and apply masking only when needed. Our method is based on the findings that language confusion is infrequent, correct-language tokens are usually among the top predictions, and output token embedding norms are larger for high-resource languages, which biases sampling. When evaluated across various models, including Qwen3, GPT-OSS, Gemma3, Llama3.1, LCG decreases language confusion significantly—often by an order of magnitude—without negatively impacting task performance.

Modern AI algorithms increasingly adopt fused kernels for performance, but implementing them remains complex due to the lack of fine-grained control in existing compilers like Triton. We introduce TileLang, a controllable programming system for fused neural kernels. TileLang provides explicit tile-level primitives for memory placement, data movement, and parallel scheduling. To guide developers in hardware-aware programming, the TileLang introduces two key techniques: tile inference which models tile programs as fused graphs and automatically deduces tile configuration from partial annotations; and tile recommendation that suggests efficient tile configurations based on hardware profiles and heuristics. TileLang makes it easy to express a wide range of fused attention kernels in under 80 lines of Python code, reducing code size by up to 90% compared to manual implementations. Evaluations show that TileLang achieves up to 5x speedup over Triton on NVIDIA H100 and up to 6 on AMD GPUs, demonstrating its ability to bridge programmability and performance.

Federated Learning (FL) offers a privacy-preserving paradigm for distributed model training by enabling clients to collaboratively learn a shared model without exchanging their raw data. However, the communication overhead associated with exchanging model updates remains a critical challenge, particularly for devices with limited bandwidth and battery resources. Existing communication compression methods largely rely on simple heuristics based on magnitude or randomness. For example, Top-k drops the elements with small magnitude, while low-rank methods such as ATOMO and PowerSGD truncate singular values with small magnitude. However, these rules do not account for the discrepancy between the compressed and the original outputs, which can lead to the loss of important information. To address this issue, we propose a novel discrepancy-aware communication compression method that enhances performance under severely constrained communication conditions. Each client uses a small subset of its local data as calibration data to directly measure the output discrepancy induced by dropping candidate compression units and uses it as a compression metric to guide the selection. By integrating this strategy, we can enhance existing mainstream compression schemes, enabling more efficient communication. Empirical results across multiple datasets and models show that our method achieves a significant improvement in accuracy under stringent communication constraints, notably an $18.9\\%$ relative accuracy improvement at a compression ratio of $0.1$, validating its efficacy for scalable and communication-efficient FL. Our code is available at https://github.com/wzy1026wzy/Discrepancy-aware-Compression-for-FL.

Linking genomic DNA to quantitative, context-specific expression remains a central challenge in computational biology. Current foundation models capture either tissue context or sequence features, but not both. Cross-omics systems, in turn, often overlook critical mechanisms such as alternative splicing and isoform reuse. We present CDBridge, a post-training strategy that unifies pretrained DNA and protein models into a context-aware framework without full retraining. CDBridge operates in two stages: (a) Seq-context learning, where a splicing-inspired token merge compresses long genomic regions into isoform-aware representations, and (b) Env-context learning, where a conditional decoder injects tissue embeddings to model expression under diverse biological contexts. To benchmark this setting, we introduce GTEx-Benchmark, derived from GTEx and Ensembl, which requires models to capture long-range exon dependencies, resolve isoform reuse, and predict tissue-specific expression levels. Across qualitative and quantitative tasks, CDBridge consistently outperforms prior methods that ignore central dogma constraints or context dependence, offering a scalable and biologically faithful solution for DNA-to-expression modeling.

Segment Anything Model 2 (SAM2) shows excellent performance in video object segmentation tasks; however, the heavy computational burden hinders its application in real-time video processing. Although there have been efforts to improve the efficiency of SAM2, most of them focus on retraining a lightweight backbone, with little exploration into post-training acceleration. In this paper, we observe that SAM2 exhibits sparse perception pattern as biological vision, which provides opportunities for eliminating redundant computation and acceleration: i) In mask decoder, the attention primarily focuses on the foreground objects, whereas the image encoder in the earlier stage exhibits a broad attention span, which results in unnecessary computation to background regions. ii) In memory bank, only a small subset of tokens in each frame contribute significantly to memory attention, and the salient regions exhibit temporal consistency, making full-token computation redundant. With these insights, we propose Efficient-SAM2, which promotes SAM2 to adaptively focus on object regions while eliminating task-irrelevant computations, thereby significantly improving inference efficiency. Specifically, for image encoder, we propose object-aware Sparse Window Routing (SWR), a window-level computation allocation mechanism that leverages the consistency and saliency cues from the previous-frame decoder to route background regions into a lightweight shortcut branch. Moreover, for memory attention, we propose object-aware Sparse Memory Retrieval (SMR), which allows only the salient memory tokens in each frame to participate in computation, with the saliency pattern reused from their first recollection. With negligible additional parameters and minimal training overhead, Efficient-SAM2 delivers 1.68$\times$ speedup on SAM2.1-L model with only 1.0\% accuracy drop on SA-V test set, where SWR and SMR provide 1.83$\times$ and 1.78$\times$ speedups, respectively.

Large Reasoning Models (LRMs) have achieved impressive performance on complex reasoning tasks by generating detailed chain-of-thought (CoT) explanations. However, these responses are often excessively long, containing redundant reasoning steps that inflate inference cost and reduce usability. Controlling the length of generated reasoning without sacrificing accuracy remains an open challenge. Through a systematic empirical analysis, we reveal a consistent positive correlation between model entropy and response length at different reasoning stages across diverse LRMs: the thinking phase exhibits higher entropy, reflecting exploratory behavior of longer responses, while the final answer phase shows lower entropy, indicating a more deterministic solution.This observation suggests that entropy at different reasoning stages can serve as a control knob for balancing conciseness and performance. Based on this insight, this paper introduces Phase Entropy Aware Reward (PEAR), a reward mechanism that incorporating phase-dependent entropy into the reward design. Instead of treating all tokens uniformly, PEAR penalize excessive entropy during the thinking phase and allowing moderate exploration at the final answer phase, which encourages models to generate concise reasoning traces that retain sufficient flexibility to solve the task correctly. This enables adaptive control of response length without relying on explicit length targets or rigid truncation rules. Extensive experiments across four benchmarks demonstrate that PEAR consistently reduces response length while sustaining competitive accuracy across model scales. In addition, PEAR demonstrates strong out-of-distribution (OOD) robustness beyond the training distribution.

The rapid growth of large language models (LLMs) has heightened the importance of post-training quantization (PTQ) for reducing memory and computation costs. Among PTQ methods, GPTQ has gained significant attention for its efficiency, enabling billion-scale LLMs to be quantized within a few GPU hours. However, GPTQ's assumption of layer-wise independence leads to severe accuracy drops in low-bit regimes. Recently, BoA improved upon GPTQ by incorporating inter-layer dependencies within attention modules, but its reliance on sequential quantization across all out-channels makes it substantially less efficient. In this paper, we propose TurboBoA, a new backpropagation-free PTQ algorithm that preserves the accuracy benefits of BoA while significantly accelerating the process. The proposed TurboBoA introduces three key innovations: (i) joint quantization of multiple out-channels with a closed-form error compensation rule, which reduces sequential bottlenecks and yields more than a three-fold speedup; (ii) a correction mechanism for errors propagated from preceding quantized layers; and (iii) adaptive grid computation with coordinate descent refinement to maintain alignment during iterative updates. Extensive experiments demonstrate that TurboBoA delivers substantial acceleration over BoA while consistently improving accuracy. When combined with outlier suppression techniques, it achieves state-of-the-art results in both weight-only and weight-activation quantization. The code will be available at https://github.com/SamsungLabs/TurboBoA.

We present Spatial Region 3D (SR-3D) aware vision-language model that connects single-view 2D images and multi-view 3D data through a shared visual token space. SR-3D supports flexible region prompting, allowing users to annotate regions with bounding boxes, segmentation masks on any frame, or directly in 3D, without the need for exhaustive multi-frame labeling. We achieve this by enriching 2D visual features with 3D positional embeddings, which allows the 3D model to draw upon strong 2D priors for more accurate spatial reasoning across frames, even when objects of interest do not co-occur within the same view. Extensive experiments on both general 2D vision language and specialized 3D spatial benchmarks demonstrate that SR-3D achieves state-of-the-art performance, underscoring its effectiveness for unifying 2D and 3D representation space on scene understanding. Moreover, we observe applicability to in-the-wild videos without sensory 3D inputs or ground-truth 3D annotations, where SR-3D accurately infers spatial relationships and metric measurements. We show more qualitative results anonymously at https://sr3d-iclr.github.io/.

Domain-specific embedding models have shown promise for applications that require specialized semantic understanding, such as coding agents and financial retrieval systems, often achieving higher performance gains than general models. However, state-of-the-art embedding models are typically based on LLMs, which contain billions of parameters, making deployment challenging in resource-constrained environments. Model compression through pruning offers a promising solution, but existing pruning methods treat all parameters uniformly, failing to distinguish between general semantic representations and domain-specific patterns, leading to suboptimal pruning decisions. Thus, we propose GAPrune, a pruning framework that addresses this challenge by considering both domain importance and preserving general linguistic foundation. Our method uses Fisher Information to measure importance and general-domain gradient alignment to assess parameter behavior, then combines these signals using our Domain Alignment Importance (DAI) scoring. Lower DAI scores indicate that the parameter is either less important for the domain task or creates conflicts between domain and general objectives. Experiments on two domain benchmarks, FinMTEB and ChemTEB, show that GAPrune maintains performance within 2.5\% of dense models in one-shot pruning at 50\% sparsity, while outperforming all baselines. With retraining in 100 steps, GAPrune achieves +4.51\% improvement on FinMTEB and +1.73\% on ChemTEB, demonstrating that our pruning strategy not only preserves but enhances domain-specific capabilities. Our findings demonstrate that principled pruning strategies can achieve model compression and enhanced domain specialization, providing the research community with a new approach for development.

Vision-based reinforcement learning can benefit from object-centric scene representation, which factorizes the visual observation into individual objects and their attributes, such as color, shape, size, and position. While such object-centric representations can extract components that generalize well for various multi-object manipulation tasks, they are prone to issues with occlusions and 3D ambiguity of object properties due to their reliance on single-view 2D image features. Furthermore, the entanglement between object configurations and camera poses complicates the object-centric disentanglement in 3D, leading to poor 3D reasoning by the agent in vision-based reinforcement learning applications. To address the lack of 3D awareness and the object-camera entanglement problem, we propose an enhanced 3D object-centric representation that utilizes multi-view 3D features and enforces more explicit 3D-aware disentanglement. The enhancement is based on the integration of the recent success of multi-view Transformer and the prototypical representation learning among the object-centric representations. The representation, therefore, can stably identify proxies of 3D positions of individual objects along with their semantic and physical properties, exhibiting excellent interpretability and controllability. Then, our proposed block transformer policy effectively performs novel tasks by assembling desired properties adaptive to the new goal states, even when provided with unseen viewpoints at test time. We demonstrate that our 3D-aware block representation is scalable to compose diverse novel scenes and enjoys superior performance in out-of-distribution tasks with multi-object manipulations under both seen and unseen viewpoints compared to existing methods.

Different federated optimization algorithms typically employ distinct client-selection strategies: some methods communicate only with a randomly sampled subset of clients at each round, while others need to periodically communicate with all clients or use a hybrid scheme that combines both strategies. However, existing metrics for comparing optimization methods typically do not distinguish between these strategies, which often incur different communication costs in practice. To address this disparity, we introduce a simple and natural model of federated optimization that quantifies communication and local computation complexities. This new model allows for several commonly used client-selection strategies and explicitly associates each with a distinct cost. Within this setting, we propose a new algorithm that achieves the best-known communication and local complexities among existing federated optimization methods for non-convex optimization. This algorithm is based on the inexact composite gradient method with a carefully constructed gradient estimator and a special procedure for solving the auxiliary subproblem at each iteration. The gradient estimator is based on SAGA, a popular variance-reduced gradient estimator. We first derive a new variance bound for it, showing that SAGA can exploit functional similarity. We then introduce the Recursive-Gradient technique as a general way to potentially improve the error bound of a given conditionally unbiased gradient estimator, including both SAGA and SVRG. By applying this technique to SAGA, we obtain a new estimator, RG-SAGA, which has an improved error bound compared to the original one.

The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100\% of full-KV performance when only keeping 256 tokens, and achieves 2.56$\times$ speedup using just 18.9\% of the cache on AIME24.

We study the implicit bias of sharpness-aware minimization (SAM) when training $L$-layer linear diagonal networks on linearly separable binary classification. For linear models ($L=1$), both $\ell_\infty$- and $\ell_2$-SAM recover the $\ell_2$ max-margin classifier, matching gradient descent (GD). However, for depth $L = 2$, the behavior changes drastically—even on a single-example dataset where we can analyze the dynamics. For $\ell_\infty$-SAM, the limit direction depends critically on initialization and can converge to $0$ or to any standard basis vector; this is in stark contrast to GD, whose limit aligns with the basis vector of the dominant coordinate in the data. For $\ell_2$-SAM, we uncover a phenomenon we call *sequential feature discovery*, in which the predictor initially relies on minor coordinates and gradually shifts to larger ones as training proceeds or initialization grows. Our theoretical analysis attributes this phenomenon to $\ell_2$-SAM’s gradient normalization factor applied in its perturbation, which amplifies minor coordinates early and allows major ones to dominate later. Synthetic and real-data experiments corroborate our findings.

Submissions are rising fast, and venues use different rules, data formats, and update times. As a result, signals of progress get split across places, and key moments (rebuttal, discussion, final decision) are easy to miss, making analysis hard. We present Paper Copilot, a system and scalable peer-review archive that pulls data from official sites, OpenReview, and opt-in forms into a single, standardized, versioned record with timestamps. This lets us track trends over time and compare venues, institutions, and countries in a consistent way. Using the archive for ICLR 2024/2025, we see larger score changes after rebuttal for higher-tier papers, reviewer agreement that dips during active discussion and tightens by the end, and in 2025 a sharper, mean-score–driven assignment of tiers with lower decision uncertainty than expected at that scale. We also state simple rules for ethics—clear sourcing and consent, privacy protection, and limits on use for closed venues. Together, we provide a clear, reusable base for tracking AI/ML progress, and, with this data, enable validation, benchmarking, and otherwise hard-to-run studies.

Large Language Models (LLMs) have achieved remarkable success in various NLP tasks, but their enormous memory footprints pose significant challenges for deployment on consumer-grade GPUs. Prior solutions, such as PowerInfer, combine offloading and sparse activation to reduce memory and computational overhead, but suffer from static neuron partitioning, leading to suboptimal GPU utilization and increased latency. In this work, we present DynamicInfer, a runtime neuron offloading framework that dynamically adapts neuron scheduling based on input-dependent activation patterns. DynamicInfer introduces (1) a hierarchical neural caching strategies, (2) a load-aware neuron activation mechanism tailored to heterogeneous hardware, and (3) an activation-aware prefetching pipeline that overlaps data transfer with computation. Extensive experiments on ReluLLaMA and Prosparse models across multiple hardware platforms demonstrate that DynamicInfer achieves up to 253\% speedup over llama.cpp and 59\% over PowerInfer, while retaining model accuracy. Our approach offers a practical and scalable solution for high-performance LLM inference on resource-constrained devices.

Reasoning language models have demonstrated remarkable performance on many challenging tasks in math, science, and coding. Choosing the right reasoning model for practical deployment involves a performance and cost tradeoff at two key levels: model size and reasoning budget, where larger models and higher reasoning budget lead to better performance but with increased cost and latency. In this work, we tackle this tradeoff from the angle of model configuration routing for different queries, and present RADAR (Reasoning–Ability and Difficulty-Aware Routing), a lightweight, interpretable, and scalable routing framework. Inspired by psychometrics, RADAR learns an item response model from model responses with different budgets to different queries, with interpretable parameters including query difficulties and model-budget abilities. RADAR then routes queries with higher difficulty to model-budget pairs with higher ability, and vice versa. We conduct extensive experiments on 8 widely used challenging reasoning benchmarks, demonstrating the superior performance of RADAR compared to state-of-the-art model routing methods. RADAR also exhibits query generalization capabilities, showing strong performance on out-of-distribution queries in all benchmarks. RADAR is also scalable and can efficiently integrate additional models, by dynamically selecting a small set of evaluation queries to estimate their abilities.

Model merging aims to integrate multiple independently fine-tuned expert models into a single model while preserving the knowledge of all experts. However, existing approaches mainly address parameter conflicts at the merging stage and overlook the role of the fine-tuning process, which often leads to significant post-merge performance degradation. To address this limitation, we propose a novel merging-aware optimizer (abbreviated as MergOPT) that injects principled merge-induced parameter shifts into the weight update steps so that the fine-tuned model exhibits a more stable loss landscape under subsequent merging operations. Specifically, we first formulate model merging as a distributionally robust optimization problem in the weight space: the parameters of other experts to be merged are viewed as adversarial merge-offsets, and fine-tuning adapts to the worst-case merging scenario. Building on this formulation, we analyze the distribution of parameter updates and the effects of merging hyperparameters, from which we derive a merging-guided feasible region for weight shifts. Finally, extensive experiments across four large language models (LLMs) and one vision model show that our approach consistently outperforms standard fine-tuning, yielding an average relative gain of 3.5\% and a maximum gain of 9.5\% across four merging strategies when merging seven experts.

Speculative decoding (SD) has emerged as a promising technique to accelerate LLM inference by employing a small, efficient draft model to propose draft tokens in advance, and subsequently validating them in parallel with the large target model. However, the existing SD methods still remain fundamentally constrained by their serialized execution, which inevitably causes mutual waiting bubbles between the draft and target models. To address this critical challenge, we draw inspiration from sophisticated branch prediction mechanisms in modern processors and propose a novel framework, \textbf{SpecBranch}, to fully unlock branch parallelism in SD. Specifically, we first conduct an in-depth analysis of the potential of branch parallelism in SD, and recognize that the key challenge lies in the intricate trade-offs between parallelization and token rollback. Based on this analysis, we introduce parallel speculative branches to preemptively hedge against likely rejections. Meanwhile, to significantly enhance parallelism, we jointly orchestrate adaptive draft lengths with a hybrid combination of the implicit draft model confidence and explicit reusing of target model features. Extensive experiments conducted across various models and benchmarks show that \textbf{SpecBranch} achieves impressive speedups of over \textbf{1.8}$\times \sim$ \textbf{4.5}$\times$ against the standard auto-regressive decoding and reduces rollback tokens by \textbf{50}\% for poorly aligned models, while maintaining an identical sampling distribution. Our code is available at \url{https://github.com/Sylvan820/Specbranch}.