ICLR 2026 论文列表

Precisely evaluating video understanding models remains challenging: commonly used metrics such as BLEU, ROUGE, and BERTScore fail to capture the fineness of human judgment, while obtaining such judgments through manual evaluation is costly. Recent work has explored using large language models (LLMs) or multimodal LLMs (MLLMs) as evaluators, but their extension to video understanding remains relatively unexplored. In this work, we introduce VideoJudge, a 3B and 7B-sized MLLM judge specialized to evaluate outputs from video understanding models (\textit{i.e.}, text responses conditioned on videos). To train VideoJudge, our recipe builds on the interplay between a generator and an evaluator: the generator is prompted to produce responses conditioned on a target rating, and responses not matching the evaluator's rating are discarded. Across three out of four meta-evaluation benchmarks, VideoJudge-7B outperforms larger MLLM judge baselines such as Qwen2.5-VL-32B and Qwen2.5-VL-72B. Notably, we find that LLM judges (Qwen3) models perform worse than MLLM judges (Qwen2.5-VL) and that long chain-of-thought reasoning does not improve performance, indicating that providing video inputs is crucial for evaluation of video understanding tasks.

With the rapid development of language models, the number of small language models (SLMs) has grown significantly. Although they do not achieve state-of-the-art accuracy, they are more efficient and often excel at specific tasks. This raises a natural question: can multiple SLMs be orchestrated into a system where each contributes effectively, achieving higher accuracy than any individual model? Existing orchestration methods have primarily targeted frontier models (e.g., GPT-4) and perform suboptimally when applied to SLMs. To address this gap, we propose a three-stage approach for orchestrating SLMs. First, we introduce SLM-MUX, a multi-model architecture that effectively coordinates multiple SLMs. Building on this, we develop two optimization strategies: (i) a model selection search that identifies the most complementary SLMs from a given pool, and (ii) test-time scaling tailored to SLM-MUX. Our approach delivers strong results: Compared to existing orchestration methods, our approach achieves up to 13.4% improvement on MATH, 8.8% on GPQA, and 7.0% on GSM8K. With just two SLMs, SLM-MUX outperforms Qwen 2.5 72B on GPQA and GSM8K, and matches its performance on MATH. We further provide theoretical analyses to substantiate the advantages of our method. Additional experiments show that the core principle of SLM-MUX extends to open-ended generation tasks (e.g., HumanEval) and benefits other model classes, including frontier LLMs and domain-specific fine-tuned SLMs. In summary, we demonstrate that SLMs can be effectively orchestrated into more accurate and efficient systems through the proposed approach.

We present a target-aware video diffusion model that generates videos from an input image, in which an actor interacts with a specified target while performing a desired action. The target is defined by a segmentation mask, and the action is described through a text prompt. Our key motivation is to incorporate target awareness into video generation, enabling actors to perform directed actions on designated objects. This enables video diffusion models to act as motion planners, producing plausible predictions of human–object interactions by leveraging the priors of large-scale video generative models. We build our target-aware model by extending a baseline model to incorporate the target mask as an additional input. To enforce target awareness, we introduce a special token that encodes the target's spatial information within the text prompt. We then fine-tune the model with our curated dataset using an additional cross-attention loss that aligns the cross-attention maps associated with this token with the input target mask. To further improve performance, we selectively apply this loss to the most semantically relevant attention regions and transformer blocks. Experimental results show that our target-aware model outperforms existing solutions in generating videos where actors interact accurately with the specified targets. We further demonstrate its efficacy in two downstream applications: zero-shot 3D HOI motion synthesis with physical plausibility and long-term video content creation.

Vision-Language Models (VLMs) are increasingly deployed as autonomous agents to navigate mobile Graphical User Interfaces (GUIs). However, their operation within dynamic on-device ecosystems, which include notifications, pop-ups, and inter-app interactions, exposes them to a unique and underexplored threat vector: environmental injection. Unlike traditional prompt-based attacks that manipulate textual instructions, environmental injection contaminates the agent's visual perception by inserting adversarial UI elements, such as deceptive overlays or spoofed notifications, directly into the GUI. This bypasses textual safeguards and can derail agent execution, leading to privacy leakage, financial loss, or irreversible device compromise. To systematically evaluate this threat, we introduce GhostEI-Bench, the first benchmark dedicated to assessing mobile agents under environmental injection attacks within dynamic, executable environments. Moving beyond static image-based assessments, our benchmark injects adversarial events into realistic application workflows inside fully operational Android emulators, assessing agent performance across a range of critical risk scenarios. We also introduce a novel evaluation protocol where a judge LLM performs fine-grained failure analysis by reviewing the agent's action trajectory alongside the corresponding sequence of screenshots. This protocol identifies the precise point of failure, whether in perception, recognition, or reasoning. Our comprehensive evaluation of state-of-the-art agents reveals their profound vulnerability to deceptive environmental cues. The results demonstrate that current models systematically fail to perceive and reason about manipulated UIs. GhostEI-Bench provides an essential framework for quantifying and mitigating this emerging threat, paving the way for the development of more robust and secure embodied agents.

Compression methods, including quantization, distillation, and pruning, improve the computational efficiency of large reasoning models (LRMs). However, existing studies either fail to sufficiently compare all three compression methods on LRMs or lack in-depth interpretation analysis. In this paper, we investigate how the reasoning capabilities of LRMs are compromised during compression, through performance benchmarking and mechanistic interpretation. To uncover the effects of compression on reasoning performance, we benchmark quantized, distilled, and pruned DeepSeek-R1 models on four reasoning datasets (AIME 2024, FOLIO, Temporal Sequences, and MuSiQue). To precisely locate compression effects on model weights, we adapt difference of means and attribution patching techniques, focusing on the activation of every linear component in compressed LRMs, to interpret fine-grained causal relationships between weights and various reasoning capabilities. This fine-grained interpretation addresses a fundamental question of compression: which weights are the most important for reasoning? Overall, we find dynamically quantized 2.51-bit R1 reaches close-to-R1 performance. With empirical verification, we present three main findings that generalize across both Llama and Qwen: (1) Weight count has a greater impact on LRMs' knowledge memorization than reasoning, highlighting the risks of pruning and distillation; (2) The MLP up projection in the final layer of distilled LRMs is one of the most important components, offering a new perspective on locating critical weights - a fundamental problem in model compression; and (3) Current quantization methods overly compress the final-layer modules and MLP gate projections, so protecting just 2% of all weights that are excessively compressed can raise average accuracy by 6.57%, greatly surpassing the state-of-the-art.

Spatio-temporal reasoning is a remarkable capability of Vision Language Models (VLMs), but the underlying mechanisms of such abilities remain largely opaque. We postulate that visual/geometrical and textual representations of spatial structure must be combined at some point in VLM computations. We search for such confluence, and ask whether the identified representation can causally explain aspects of input-output model behavior through a linear model. We show empirically that VLMs encode object locations by linearly binding \textit{spatial IDs to textual activations, then perform reasoning via language tokens. Through rigorous causal interventions we demonstrate that these IDs, which are ubiquitous across the model, can systematically mediate model beliefs at intermediate VLM layers. Additionally, we find that spatial IDs serve as a diagnostic tool for identifying limitations and bottlenecks in existing VLMs. We extend our analysis to video VLMs and identify an analogous linear temporal ID mechanism. By characterizing our proposed spatiotemporal ID mechanism, we elucidate a previously underexplored internal reasoning process in VLMs, toward improved interpretability and the principled design of more aligned and capable models.

For real-world BCI applications, lightweight Electroencephalography (EEG) systems offer the best cost–deployment balance. However, such spatial sparsity of EEG limits spatial fidelity, hurting learning and introducing bias. EEG spatial super-resolution methods aim to recover high-density EEG signals from sparse measurements, yet is often hindered by distribution shift and signal distortion and thus reducing fidelity and usability for EEG analysis and visualization. To overcome these challenges, we introduce SRGDiff, a step-aware residual-guided diffusion model that formulates EEG spatial super-resolution as dynamic conditional generation. Our key idea is to learn a dynamic residual condition from the low-density input that predicts the step-wise temporal and spatial details to add and uses the evolving cue to steer the denoising process toward high density reconstructions. At each denoising step, the proposed residual condition is additively fused with the previous denoiser feature maps, then a step-dependent affine modulation scales and shifts the activation to produce the current features. This iterative procedure dynamically extracts step-wise temporal rhythms and spatial-topographic cues to steer high-density recovery and maintain a fidelity–consistency balance. We adopt a comprehensive evaluation protocol spanning signal-, feature-, and downstream-level metrics across SEED, SEED-IV, and Localize-MI and multiple upsampling scales. SRGDiff achieves consistent gains of up to 40\% over strong baselines, proving its superiority in the task of EEG spatial super-resolution. Moreover, topographic visualizations comparison and substantial EEG-FID gains jointly indicate that our SR EEG mitigates the spatial–spectral shift between low- and high-density recordings. Our code is available at https://anonymous.4open.science/r/DDPM-VAE-6F6B.

Graph Domain Adaptation (GDA) aims to bridge distribution shifts between domains by transferring knowledge from well-labeled source graphs to given unlabeled target graphs. One promising recent approach addresses graph transfer by discretizing the adaptation process, typically through the construction of intermediate graphs or stepwise alignment procedures. However, such discrete strategies often fail in real-world scenarios, where graph structures evolve continuously and nonlinearly, making it difficult for fixed-step alignment to approximate the actual transformation process. To address these limitations, we propose \textbf{DiffGDA}, a \textbf{Diff}usion-based \textbf{GDA} method that models the domain adaptation process as a continuous-time generative process. We formulate the evolution from source to target graphs using stochastic differential equations (SDEs), enabling the joint modeling of structural and semantic transitions. To guide this evolution, a domain-aware network is introduced to steer the generative process toward the target domain, encouraging the diffusion trajectory to follow an optimal adaptation path. We theoretically show that the diffusion process converges to the optimal solution bridging the source and target domains in the latent space. Extensive experiments on 14 graph transfer tasks across 8 real-world datasets demonstrate DiffGDA consistently outperforms state-of-the-art baselines.

Although Large Reasoning Models (LRMs) have progressed in solving complex problems, their chain-of-thought (CoT) reasoning often contains harmful content that can persist even when the final responses appear safe. We show that this issue still remains in existing methods which overlook the unique significance of safe reasoning, undermining their trustworthiness and posing potential risks in applications if unsafe reasoning is accessible for and exploited by malicious users. We therefore shift our focus to aligning the safety of reasoning itself in this paper and explore process supervision as the solution. However, simply rewarding safe reasoning proves inadequate due to low rollout diversity and limited training signals. To tackle this challenge, we first delve into the characteristics of safe reasoning and uncover several critical insights that 1) safe reasoning is often consolidated by a few critical steps of _safety triggers_; 2) _compliance cues_ strongly correlate with unsafe continuations; and 3) corrective interventions reliably steer unsafe trajectories towards safer traces. Motivated by these, we propose **Intervened Preference Optimization (IPO)**, an alignment method that enforces safe reasoning by substituting compliance steps with safety triggers and constructing pairs for preference learning with strong signals. Experiments on jailbreak and adversarial safety benchmarks demonstrate that IPO remarkably improves overall safety regarding both reasoning and responses, outperforming SFT-based and RL-based baselines with a relative reduction of over 30\% in harmfulness, while preserving excellent performance across diverse reasoning tasks. The results highlight the importance of explicit alignment for reasoning and provide a practical path to safer LRMs.

Although continuous-time consistency models (e.g., sCM, MeanFlow) are theoretically principled and empirically powerful for fast academic-scale diffusion, its applicability to large-scale text-to-image and video tasks remains unclear due to infrastructure challenges in Jacobian-vector product (JVP) computation and the limitations of evaluation benchmarks like FID. This work represents the first effort to scale up continuous-time consistency to general application-level image and video diffusion models, and to make JVP-based distillation effective at large scale. We first develop a parallelism-compatible FlashAttention-2 JVP kernel, enabling sCM training on models with over 10 billion parameters and high-dimensional video tasks. Our investigation reveals fundamental quality limitations of sCM in fine-detail generation, which we attribute to error accumulation and the “mode-covering” nature of its forward-divergence objective. To remedy this, we propose the score-regularized continuous-time consistency model (rCM), which incorporates score distillation as a long-skip regularizer. This integration complements sCM with the “mode-seeking” reverse divergence, effectively improving visual quality while maintaining high generation diversity. Validated on large-scale models (Cosmos-Predict2, Wan2.1) up to 14B parameters and 5-second videos, rCM generally matches the state-of-the-art distillation method DMD2 on quality metrics while mitigating mode collapse and offering notable advantages in diversity, all without GAN tuning or extensive hyperparameter searches. The distilled models generate high-fidelity samples in only $1\sim4$ steps, accelerating diffusion sampling by $15\times\sim50\times$. These results position rCM as a practical and theoretically grounded framework for advancing large-scale diffusion distillation. Code is available at https://github.com/NVlabs/rcm.

Large Reasoning Models (LRMs) have achieved impressive performance on challenging tasks, yet their deep reasoning often incurs substantial computational costs. To achieve efficient reasoning, existing reinforcement learning methods still struggle to construct short reasoning path during the rollout stage, limiting effective learning. Inspired by Evidence Accumulation Models, we find that LRMs have accumulated sufficient information early in reasoning, making further reasoning steps redundant. Based on this insight, we propose Just-Enough Thinking (JET), which trains models to proactively terminate unnecessary reasoning. JET performs trajectory truncation during rollout to expose the model to short, distributionally consistent reasoning paths. Besides, it uses a quality-controlled length reward to better encourage concise reasoning while maintaining correctness. Extensive experiments demonstrate that JET significantly improves reasoning efficiency without sacrificing accuracy. In particular, JET delivers a 4.6% accuracy improvement while reducing the output length by 46.3% on the Olympiad benchmark using DeepSeek-R1-Distill-Qwen-1.5B. Our code is available in the GitHub.

Foundation models are reshaping EEG analysis, yet an important problem of EEG tokenization remains a challenge. This paper presents TFM-Tokenizer, a novel tokenization framework that learns a vocabulary of time-frequency motifs from *single-channel* EEG signals and encodes them into discrete tokens. We propose a dual-path architecture with time–frequency masking to capture robust motif representations, and it is model-agnostic, supporting both lightweight transformers and existing foundation models for downstream tasks. Our study demonstrates three key benefits: *Accuracy:* Experiments on four diverse EEG benchmarks demonstrate consistent performance gains across both single- and multi-dataset pretraining settings, achieving up to 11% improvement in Cohen’s Kappa over strong baselines. *Generalization:* Moreover, as a plug-and-play component, it consistently boosts the performance of diverse foundation models, including BIOT and LaBraM. *Scalability:* By operating at the single-channel level rather than relying on the strict 10–20 EEG system, our method has the potential to be device-agnostic. Experiments on ear-EEG sleep staging, which differs from the pretraining data in signal format, channel configuration, recording device, and task, show that our tokenizer outperforms baselines by 14%. A comprehensive token analysis reveals strong class-discriminative, frequency-aware, and consistent structure, enabling improved representation quality and interpretability. Code is available at https://anonymous.4open.science/r/TFM-Token-FE33.

Despite their impressive performance, generative image models trained on large-scale datasets frequently fail to produce images with seemingly simple concepts -- e.g., human hands or objects appearing in groups of four -- that are reasonably expected to appear in the training data. These failure modes have largely been documented anecdotally, leaving open the question of whether they reflect idiosyncratic anomalies or more structural limitations of these models. To address this, we introduce a systematic approach for identifying and characterizing "conceptual blindspots" -- concepts present in the training data but absent or misrepresented in a model's generations. Our method leverages sparse autoencoders (SAEs) to extract interpretable concept embeddings, enabling a quantitative comparison of concept prevalence between real and generated images. We train an archetypal SAE (RA-SAE) on DINOv2 features with 32,000 concepts -- the largest such SAE to date -- enabling fine-grained analysis of conceptual disparities. Applied to four popular generative models (Stable Diffusion 1.5/2.1, PixArt, and Kandinsky), our approach reveals specific suppressed blindspots (e.g., bird feeders, DVD discs, and whitespaces on documents) and exaggerated blindspots (e.g., wood background texture and palm trees). At the individual datapoint level, we further isolate memorization artifacts -- instances where models reproduce highly specific visual templates seen during training. Overall, we propose a theoretically grounded framework for systematically identifying conceptual blindspots in generative models by assessing their conceptual fidelity with respect to the underlying data-generating process.

Deception is a pervasive feature of human communication and an emerging concern in large language models (LLMs). While recent studies document instances of LLM deception, most evaluations remain confined to single-turn prompts and fail to capture the long-horizon interactions in which deceptive strategies typically unfold. We introduce a new simulation framework, LH-Deception, for a systematic, empirical quantification of deception in LLMs under extended sequences of interdependent tasks and dynamic contextual pressures. LH-Deception is designed as a multi-agent system: a performer agent tasked with completing tasks and a supervisor agent that evaluates progress, provides feedback, and maintains evolving states of trust. An independent deception auditor then reviews full trajectories to identify when and how deception occurs. We conduct extensive experiments across 11 frontier models, spanning both closed-source and open-source systems, and find that deception is model-dependent, increases with event pressure, and consistently erodes supervisor trust. Qualitative analyses further reveal emergent, long-horizon phenomena, such as ``chains of deception", which are invisible to static, single-turn evaluations. Our findings provide a foundation for evaluating future LLMs in real-world, trust-sensitive contexts.

The Clock and Pizza interpretations, associated with architectures differing in either uniform or learnable attention, were introduced to argue that different architectural designs can yield distinct circuits for modular addition. In this work, we show that this is not the case, and that both the uniform and trainable attention architectures implement the same algorithm via topologically and geometrically equivalent representations. Our methodology goes beyond the interpretation of individual neurons and weights. Instead, we identify all of the neurons corresponding to each learned representation and then study the collective group of neurons as one entity. This method reveals that each learned representation is a manifold that we can study utilizing tools from topology. Based on this insight, we can statistically analyze the learned representations across hundreds of circuits to demonstrate the similarity between learned modular addition circuits that arise naturally from common deep learning paradigms.

Autoregressive (AR) models, despite their remarkable successes, encounter limitations in image generation due to sequential prediction of tokens, e.g. local image patches, in a predetermined row-major raster-scan order. Prior works improve AR with various designs of prediction units and orders, however, rely on human inductive biases. This work proposes Basis Autoregressive (BAR), a novel paradigm that conceptualizes tokens as basis vectors within the image space and employs an end-to-end learnable approach to transform basis. By viewing tokens $x_k$ as the projection of image $\mathbf{x}$ onto basis vectors $e_k$, BAR's unified framework refactors fixed token sequences through the linear transform $\mathbf{y}=\mathbf{Ax}$, and encompasses previous methods as specific instances of matrix $\mathbf{A}$. Furthermore, BAR adaptively optimizes the transform matrix with an end-to-end AR objective, thereby discovering effective strategies beyond hand-crafted assumptions. Comprehensive experiments, notably achieving a state-of-the-art FID of 1.15 on the ImageNet-256 benchmark, demonstrate the ability of BAR to overcome human biases and significantly advance image generation, including text-to-image synthesis.

Chess has played a uniquely important historical role as a testbed domain for artificial intelligence. Applying new architectures to improve absolute chess performance, and more recently to predict human moves at specified skill levels, has therefore garnered attention in the machine learning literature. Current approaches to these problems employ transformer models with widely varying architectural designs, and use unintuitive tokenization schemes that are not amenable to interpretability techniques, which hinders their applicability for teaching and human-AI interaction. We introduce Chessformer, a novel chess transformer model design that consists of an encoder-only model which processes chessboard squares as input tokens, instead of moves or the entire position, a dynamic positional encoding scheme that allows the model to flexibly adapt to the unique geometries present in chess, and an attention-based policy output design. We show that Chessformer advances the state of the art in all three major chess modeling goals: it significantly improves the chess-playing performance of a state-of-the-art chess engine, it surpasses the previous best human move-matching prediction performance with a much smaller model, and it enables substantial interpretability benefits. Our unified approach constitutes a broad advance across several important tasks in chess AI, and also demonstrates the benefits of carefully adapting transformers' tokenization systems, output systems, and positional encodings to reflect the structure of a domain of interest.

Learning in high-dimensional MDPs with complex state dynamics became possible with the progress achieved in reinforcement learning research. At the same time, deep neural policies have been observed to be highly unstable with respect to the minor variations in their state space, causing volatile and unpredictable behaviour. To alleviate these volatilities, a line of work suggested techniques to cope with this problem via explicitly regularizing the temporal difference loss to ensure local $\epsilon$-invariance in the state space. In this paper, we provide theoretical foundations on the impact of $\epsilon$-local invariance training on the deep neural policy manifolds. Our comprehensive theoretical and experimental analysis reveals that standard reinforcement learning inherently learns counterfactual values while recent training techniques that focus on explicitly enforcing $\epsilon$-local invariance cause policies to lose counterfactuality, and further result in learning misaligned and inconsistent values. In connection to this analysis, we further highlight that this line of training methods break the core intuition and the true biological inspiration of reinforcement learning, and introduce an intrinsic gap between how natural intelligence understands and interacts with an environment in contrast to AI agents trained via $\epsilon$-local invariance methods. The misalignment, inaccuracy and the loss of counterfactuality revealed in our paper further demonstrate the need to rethink the approach in establishing truly reliable and generalizable reinforcement learning policies.

While large reasoning models have shown remarkable ability to generate long chains-of-thought (CoTs) in English, we still lack understanding of how these long-form reasoning abilities transfer to the vast majority of the world’s languages. In this work, we systematically investigate four key stages of model development–scaling, pretraining, post-training, and inference–to understand how long CoT capabilities extend beyond English. We compare two reasoning settings across nine non-English target languages: En-CoT, where models process target-language inputs, but reason in English; and Target-CoT, where models both process inputs and generate long CoTs in the target language. We find that scaling reasoning model size improves multilingual task performance in En-CoT, but Target-CoT performance lags behind. This gap widens for tasks requiring long, multi-step CoTs such as mathematical reasoning. Shifting to pretraining, we find that adding a specialized reasoning stage enhances En-CoT performance but degrades Target-CoT, whereas broad multilingual pretraining improves both modes simultaneously. Given the scarcity of high-quality reasoning traces in languages other than English, we explore synthetic data curation approaches for post-training. We demonstrate that fine-tuning on reasoning traces automatically translated from gold English traces outperforms fine-tuning on target-language traces distilled from large reasoning models. Finally, we report disparities in inference efficiency between languages and uncover language-specific failure modes in CoTs. We release models, datasets, and code to foster further research.