ICLR 2026 Papers

Rotary Positional Embedding (RoPE) is a widely adopted technique for encoding position in language models, which, while effective, causes performance breakdown when input length exceeds training length. Prior analyses assert (rightly) that long inputs cause channels to rotate "out of distribution," but it is not clear how extra rotation relates to or causes pathological behavior. Through empirical analysis we advance a unified geometric understanding of attention behavior with RoPE. We find that attention induces tight clustering of separated key and query latent point clouds, allowing for creation of sink tokens: placeholders that allow attention heads to avoid token mixing when not required. RoPE applied to longer inputs damages this key/query cluster separation, producing pathological behavior by inhibiting sink token functionality. From this geometric perspective, we propose RoPE-ID (In Distribution), a straightforward modification that allows attention layers to generalize to longer inputs out of the box: apply RoPE with high frequency to a subset of channels. We demonstrate the effectiveness of RoPE-ID for extended inputs using 1B and 3B parameter Transformers on the LongBench and RULER information retrieval benchmarks.

Concerns have been raised that knowledge distillation may transfer test-set knowledge from a contaminated teacher to a clean student—a “data laundering” effect that potentially threatens evaluation integrity. In this paper, we assess the severity of this phenomenon. If these concerns regarding data laundering are minor, then distillation could be used to mitigate risks of direct data exposure. Across eight classification benchmarks, we find that substantial laundering is the exception rather than the rule: unlike the large performance gains from direct contamination, any accuracy inflation from laundering is consistently smaller and statistically insignificant in all but two cases. More broadly, we find that the two phenomena are weakly correlated, suggesting that laundering is not simply a diluted form of contamination but a distinct effect that arises primarily when benchmarks exhibit large train–test distribution gaps. Motivated by this, we conduct controlled experiments that systematically enlarge the train–test distance on two benchmarks where laundering was initially negligible, and observe that laundering becomes more significant as the gap widens. Taken together, our results indicate that knowledge distillation, despite rare benchmark-specific residues, can be expected to function as an effective decontamination technique that largely mitigates test-data leakage.

Transformer-based large language models (LLMs) have achieved remarkable success, yet their standard attention mechanism incurs quadratic computation and memory costs with respect to sequence length, posing a major bottleneck for long-context training. Prior work tackles this challenge along two directions: (1) kernel-level optimizations, which accelerate dense and sparse attention operators; and (2) module-level strategies, often referred to as distributed attention or context parallel training, which scale attention across multiple devices. However, systematic evaluation still remains limited: operator-level comparisons are often incomplete, while context parallel strategies are typically framework-specific, with unclear performance analysis across contexts. To address these gaps, we propose a unified benchmark that integrates representative attention kernels and context parallel mechanisms with a modular and extensible interface for evaluation. The benchmark evaluates methods along two critical dimensions: (1) attention mask patterns, which strongly affect efficiency, scalability, and usability, and (2) sequence length and distributed scale, which determine performance under extreme long-context training. Through comprehensive experiments on the cluster of up to 96 GPUs, our benchmark enables reproducible comparisons, highlights method-specific trade-offs, and provides practical guidance for designing and deploying attention mechanisms in long-context LLM training.

Hands play a central role in daily life, yet modeling natural hand motions remains underexplored. Existing methods that tackle text-to-hand-motion generation or hand animation captioning rely on studio-captured datasets with limited actions and contexts, making them costly to scale to “in-the-wild” settings. Further, contemporary models and their training schemes struggle to capture animation fidelity with text–motion alignment. To address this, we (1) introduce ‘3D Hands in the Wild’ (3D-HIW), a dataset of 32K 3D hand-motion sequences and aligned text, and (2) propose CLUTCH, an LLM-based hand animation system with two critical innovations: (a) SHIFT, a novel VQ-VAE architecture to tokenize hand motion, and (b) a geometric refinement stage to finetune the LLM. To build 3D- HIW, we propose a data annotation pipeline that combines vision–language models (VLMs) and state-of-the-art 3D hand trackers, and apply it to a large corpus of egocentric action videos covering a wide range of scenarios. To fully capture motion in-the-wild, CLUTCH employs SHIFT, a part–modality decomposed VQ- VAE, which improves generalization and reconstruction fidelity. Finally, to improve animation quality, we introduce a geometric refinement stage, where CLUTCH is co-supervised with a reconstruction loss applied directly to decoded hand motion parameters. Experiments demonstrate state-of-the-art performance on text-to- motion and motion-to-text tasks, establishing the first benchmark for scalable in-the-wild hand motion modelling. Code, data and models will be released.

Facts change over time, making it essential for Large Language Models (LLMs) to handle time-sensitive factual knowledge accurately and reliably. Although factual Time-Sensitive Question-Answering (TSQA) tasks have been widely developed, existing benchmarks often face manual bottlenecks that limit scalable and comprehensive TSQA evaluation. To address this issue, we propose TDBench, a new benchmark that systematically constructs TSQA pairs by harnessing temporal databases and database techniques, such as temporal functional dependencies, temporal SQL, and temporal joins. We also introduce a new evaluation metric called time accuracy, which assesses the validity of time references in model explanations alongside traditional answer accuracy for a more fine-grained TSQA evaluation. Extensive experiments on contemporary LLMs show how TDBench enables scalable and comprehensive TSQA evaluation while reducing the reliance on human labor, complementing current TSQA evaluation approaches that largely center on Wikipedia/Wikidata by enabling LLM evaluation on application-specific data.

Model merging has emerged as an efficient and flexible paradigm for multi-task learning, with numerous methods being proposed in recent years. However, these state-of-the-art techniques are typically evaluated on benchmark suites that are highly favorable to model merging, and their robustness in more realistic settings remains largely unexplored. In this work, we first investigate the vulnerabilities of model-merging methods and pinpoint the source-model characteristics that critically underlie them. Specifically, we identify two factors that are particularly harmful to the merging process: (1) disparities in task vector norms, and (2) the low confidence of the source models. To address this issue, we propose **DisTaC** (**Dis**tillation for **Ta**sk vector **C**onditioning), a novel method that pre-conditions these problematic task vectors before the merge. DisTaC leverages knowledge distillation to adjust a task vector's norm and increase source-model confidence while preserving its essential task-specific knowledge. Our extensive experiments demonstrate that by pre-conditioning task vectors with DisTaC, state-of-the-art merging techniques can successfully integrate models that exhibit these harmful traits, where they would otherwise fail, and achieve significant performance gains.

Training large neural networks on large-scale datasets requires substantial computational resources, particularly for dense prediction tasks such as object detection. Although dataset distillation (DD) has been proposed to alleviate these demands by synthesizing compact datasets from larger ones, most existing work focuses solely on image classification, leaving the more complex detection setting largely unexplored. In this paper, we introduce OD$^3$, a novel optimization-free data distillation framework specifically designed for object detection. Our approach involves two stages: first, a candidate selection process in which object instances are iteratively placed in synthesized images based on their suitable locations, and second, a candidate screening process using a pre-trained observer model to remove low-confidence objects. We perform our data synthesis framework on MS COCO and PASCAL VOC, two popular detection datasets, with compression ratios ranging from 0.25% to 5%. Compared to the prior solely existing dataset distillation method on detection and conventional core set selection methods, OD$^3$ delivers superior accuracy, establishes new state-of-the-art results, surpassing prior best method by more than 14% on COCO mAP$_{50}$ at a compression ratio of 1.0%. Code is available at https://github.com/VILA-Lab/OD3.

We introduce NerVE, a unified eigenspectral framework for understanding how feed-forward networks (FFNs) in large language models (LLMs) organize and regulate information flow in high-dimensional latent space. Despite FFNs dominating the parameter budget, their high-dimensional dynamics remain poorly understood. NerVE addresses this gap through lightweight, memory-efficient tracking of eigenspectrum dynamics via four complementary metrics: Spectral Entropy (dispersion), Participation Ratio (effective dimensionality), Eigenvalue Early Enrichment (top-heaviness), and Jensen-Shannon divergence (distributional shifts). Our {\em key insight} is that FFN nonlinearities reinject and reshape variance across eigenmodes, fundamentally governing latent dimension utilization. We validate NerVE across model scales and diverse architectural configurations that each uniquely shape FFN dynamics: normalization strategies (PreLN, PostLN, MixLN, Norm-Free) controlling variance flow; FFN weight geometries constraining latent space; positional encoding and activation functions modulating information propagation. Across these settings, NerVE consistently recovers stable spectral signatures that correlate with model's generalization ability and respond predictably to design choices, providing actionable insights for architectural optimization beyond trial-and-error.

Cybersecurity operations demand assistant LLMs that support diverse workflows without exposing sensitive data. Existing solutions either rely on proprietary APIs with privacy risks or on open models lacking domain adaptation. To bridge this gap, we curate 11.8B tokens of cybersecurity-focused continual pretraining data via large-scale web filtering and manual collection of high-quality resources, spanning 28.6K documents across frameworks, offensive techniques, and security tools. Building on this, we design an agentic augmentation pipeline that simulates expert workflows to generate 266K multi-turn cybersecurity samples for supervised fine-tuning. Combined with general open-source LLM data, these resources enable the training of RedSage, an open-source, locally deployable cybersecurity assistant with domain-aware pretraining and post-training. To rigorously evaluate the models, we introduce RedSage-Bench, a benchmark with 30K multiple-choice and 240 open-ended Q\&A items covering cybersecurity knowledge, skills, and tool expertise. RedSage is further evaluated on established cybersecurity benchmarks (e.g., CTI-Bench, CyberMetric, SECURE) and general LLM benchmarks to assess broader generalization. At the 8B scale, RedSage achieves consistently better results, surpassing the baseline models by up to +5.59 points on cybersecurity benchmarks and +5.05 points on Open LLM Leaderboard tasks. These findings demonstrate that domain-aware agentic augmentation and pre/post-training can not only enhance cybersecurity-specific expertise but also help to improve general reasoning and instruction-following. All models, datasets, and code will be released to advance reproducibility and open cybersecurity LLM research.

Articulated objects are ubiquitous and important in robotics, AR/VR, and digital twins. Most self-supervised methods for articulated object modeling reconstruct discrete interaction states and relate them via cross-state geometric consistency, yielding representational fragmentation and drift that hinder smooth control of articulated configurations. We introduce PD$^{2}$GS, a novel framework that learns a shared canonical Gaussian field and models the arbitrary interaction state as its continuous deformation, jointly encoding geometry and kinematics. By associating each interaction state with a latent code and refining part boundaries using generic vision priors, PD$^{2}$GS enables accurate and reliable part-level decoupling while enforcing mutual exclusivity between parts and preserving scene-level coherence. This unified formulation supports part-aware reconstruction, fine-grained continuous control, and accurate kinematic modeling, all without manual supervision. To assess realism and generalization, we release RS-Art, a real-to-sim RGB-D dataset aligned with reverse-engineered 3D models, supporting real-world evaluation. Extensive experiments demonstrate that PD$^{2}$GS surpasses prior methods in geometric and kinematic accuracy, and in consistency under continuous control, both on synthetic and real data.

Videos captured in the wild often suffer from rain streaks, blur, and noise. In addition, even slight changes in camera pose can amplify cross-frame mismatches and temporal artifacts. Existing methods rely on optical flow or heuristic alignment, which are computationally expensive and less robust. To address these challenges, Lie groups provide a principled way to represent continuous geometric transformations, making them well-suited for enforcing spatial and temporal consistency in video modeling. Building on this insight, we propose DeLiVR, an efficient video deraining method that injects spatiotemporal Lie-group differential biases directly into attention scores of the network. Specifically, the method introduces two complementary components. First, a rotation-bounded Lie relative bias predicts the in-plane  angle of each frame using a compact prediction module, which normalized coordinates are rotated and compared with base coordinates to achieve geometry-consistent alignment before feature aggregation. Second, a differential group displacement computes angular differences between adjacent frames to estimate a velocity. These biases are combined with temporal decay and a banded attention mask to emphasize short-range reliable relations while suppressing long-range noise. DeLiVR achieves sharper details, fewer rain remnants, and stronger temporal coherence on both synthetic and real rainy benchmarks. The code is publicly available at https://github.com/Shuning0312/ICLR-DeLiVR.

Recent large language models (LLMs) have increasingly adopted the Mixture-of-Experts (MoE) architecture for efficiency. MoE-based LLMs heavily depend on a superficial safety mechanism in which harmful inputs are routed safety-critical experts. However, our analysis reveals that routing decisions for harmful inputs drift significantly after fine-tuning, exposing a critical vulnerability to harmful fine-tuning (HFT) attacks. Existing defenses, primarily designed for monolithic LLMs, are less effective for MoE LLMs as they fail to prevent drift in harmful input routing. To address this limitation, we propose SafeMoE, a safe fine-tuning method tailored to MoE LLMs. SafeMoE directly mitigates routing drift by penalizing the gap between the routing weights of a fine-tuned model and those of the initial safety-aligned model, thereby preserving the safety-aligned routing of harmful inputs to safety-critical experts. Experiments on open-source MoE LLMs ranging from 7B to 141B parameters demonstrate that SafeMoE effectively mitigates HFT attacks, reducing the harmfulness score of OLMoE from 62.0 to 5.0, for example, while maintaining task utility within 1% degradation and incurring only 2\% overhead. It significantly outperforms state-of-the-art defense methods for safeguarding LLM fine-tuning and remains effective in recent large-scale MoE LLMs such as gpt-oss and Llama 4. Our implementation is available at https://anonymous.4open.science/r/SafeMoE.

Evaluating progress in large language models (LLMs) is often constrained by the challenge of verifying responses, limiting assessments to tasks like mathematics, programming, and short-form question-answering. However, many real-world applications require evaluating LLMs in processing professional documents, synthesizing information, and generating comprehensive reports in response to user queries. We introduce ProfBench: a set of over 7000 response-criterion pairs as evaluated by human-experts with professional knowledge across Physics PhD, Chemistry PhD, Finance MBA and Consulting MBA. We build robust and affordable LLM-Judges to evaluate ProfBench rubrics, by mitigating self-enhancement bias and reducing the cost of evaluation by 2-3 orders of magnitude, to make it fair and accessible to the broader community. Our findings reveal that ProfBench poses significant challenges even for state-of-the-art LLMs, with top-performing models like GPT-5-high achieving only 65.9\% overall performance. Furthermore, we identify notable performance disparities between proprietary and open-weight models and provide insights into the role that extended thinking plays in addressing complex, professional-domain tasks.

Deep neural networks often fail when deployed in real-world contexts due to distribution shift, a critical barrier to building safe and reliable systems. An emerging approach to address this problem relies on _disagreement discrepancy_—a measure of how the disagreement between two models changes under a shifting distribution. The process of maximizing this measure has seen applications in bounding error under shifts, testing for harmful shifts, and training more robust models. However, this optimization involves the non-differentiable zero-one loss, necessitating the use of practical surrogate losses. We prove that existing surrogates for disagreement discrepancy are not Bayes consistent, revealing a fundamental flaw: maximizing these surrogates can fail to maximize the true disagreement discrepancy. To address this, we introduce new theoretical results providing both upper and lower bounds on the optimality gap for such surrogates. Guided by this theory, we propose a novel disagreement loss that, when paired with cross-entropy, yields a provably consistent surrogate for disagreement discrepancy. Empirical evaluations across diverse benchmarks demonstrate that our method provides more accurate and robust estimates of disagreement discrepancy than existing approaches, particularly under challenging adversarial conditions.

We develop new classifiers under group fairness in the attribute-aware setting for binary classification with multiple group fairness constraints (e.g., demographic parity (DP), equalized odds (EO), and predictive parity (PP)). We propose a novel approach based on directly intervening on the operating characteristics of a pre-trained base classifier, by: (i) identifying optimal operating characteristics using the base classifier's group-wise ROC convex hulls; (ii) post-processing the base classifier to match those targets. As practical post-processors, we consider randomizing a mixture of group-wise thresholding rules subject to minimizing the expected number of interventions. We further extend our approach to handle multiple protected attributes and multiple linear fractional constraints. On standard datasets (COMPAS and ACSIncome), our method simultaneously satisfies approximate DP, EO, and PP with few interventions and a nearly optimal drop in accuracy; and compare favorably to previous methods.

Integrating multimodal screening data is challenging because biological signals only partially overlap and cell-level pairing is frequently unavailable. Existing approaches either require pairing or fail to capture both shared and modality-specific information in an end-to-end manner. We present PETRI, an early-fusion transformer that learns a unified cell embedding from unpaired cellular images and gene expression profiles. PETRI groups cells by shared experimental context into multimodal “documents” and performs masked joint reconstruction with cross-modal attention, permitting information sharing while preserving modality-specific capacity. The resulting latent space supports construction of perturbation-level profiles by simple averaging across modalities. Applying sparse autoencoders to the embeddings reveals learned concepts that are biologically meaningful, multimodal, and retain perturbation-specific effects. To support further machine learning research, we release a blinded, matched optical pooled screen (OPS) and Perturb-seq dataset in HepG2 cells.

Deploying semantic segmentation in driving and robotics requires both real-time inference and robustness to domain shifts, formalized as Real-Time Domain-Generalized Semantic Segmentation (RT-DGSS), which has not been fully addressed. Existing methods often treat real-time(RT) inference and domain generalization (DG) separately, with DG improving robustness but lacking real-time performance, and real-time models being brittle under distribution shifts. To address the RT-DGSS problem, we propose QPrompt-R1, a real-time Query-Prompt architecture built on a ViT backbone. QPrompt-R1 injects a small set of learnable queries only at the final transformer block, performing a single query–image alignment step and eliminating decoder overhead. To further enhance alignment without test-time cost, we introduce a Group Relative Query Alignment (GRQA) objective, which uses group-relative supervision within each group to align queries with features, improving domain generalization through group-relative rewards. QPrompt-R1 achieves 54 FPS, delivering strong performance in synthetic-to-real transfer, real-to-real generalization, and robustness under adverse conditions. Additionally, GRQA is plug-and-play, improving state-of-the-art DGSS methods like REIN (+1.2) and SoMA (+0.5) without inference-time overhead.

We study the challenge of achieving theoretically grounded feature recovery using Sparse Autoencoders (SAEs) for the interpretation of Large Language Models. Existing SAE training algorithms often lack rigorous mathematical guarantees and suffer from practical limitations such as hyperparameter sensitivity and instability. We rethink this problem from the perspective of neuron activation frequencies, and through controlled experiments, we identify a striking phenomenon we term neuron resonance: neurons reliably learn monosemantic features when their activation frequency matches the feature's occurrence frequency in the data. Building on this finding, we introduce a new SAE training algorithm based on ``bias adaptation'', a technique that adaptively adjusts neural network bias parameters to ensure appropriate activation sparsity. We theoretically prove that this algorithm correctly recovers all monosemantic features when input data is sampled from our proposed statistical model. Furthermore, we develop an improved empirical variant, Group Bias Adaptation (GBA), and demonstrate its superior performance against benchmark methods when applied to LLMs with up to 2 billion parameters. This work represents a foundational step in demystifying SAE training by providing the first SAE algorithm with theoretical recovery guarantees and practical effectiveness for LLM interpretation.

The effectiveness of collective decision-making is often challenged by the bounded rationality and inherent stochasticity of individual agents. We investigate this by analyzing how to aggregate decisions from $n$ experts, each receiving a private signal about an unknown state. Assuming signals are conditionally independent and identically distributed, we depart from the fully rational paradigm and model expert behavior using quantal response—a stochastic choice model capturing bounded rationality. Within a minimax regret framework, we show that majority voting is the optimal robust aggregator when individual rationality falls below a certain threshold. Interestingly, such groups can outperform perfectly rational agents, as their decision randomness encodes weak but informative signals lost in deterministic behavior. We validate these findings using large language models (LLMs), which naturally exhibit quantal response via their temperature parameter. Aggregating moderately stochastic LLM outputs significantly improves accuracy on complex reasoning tasks, highlighting bounded rationality not as a limitation, but as a potential strength in collective intelligence.