ICML 2026 — 티어별 논문 리스트 (abstract 전문)

Video subtitle removal is essential for content localization and media re-editing, yet existing mask-guided diffusion methods face critical limitations: training inefficiency requiring extensive annotations, inference complexity demanding explicit mask sequences, and static prior utilization unable to adapt to quality variations. We present CLEAR, a lightweight adapter-based framework addressing these challenges through three technical innovations. First, self-supervised prior learning (Stage I) extracts occlusion guidance from video pairs using pixel differences as weak supervision, eliminating annotation dependency while learning generalizable subtitle features across languages. Second, LoRA-based adaptive refinement (Stage II) enables parameter-efficient training that preserves pre-trained visual priors while achieving true mask-free end-to-end inference. Third, adaptive focal weighting dynamically adjusts prior influence based on local quality assessment. CLEAR achieves superior multilingual subtitle removal with only 0.77% trainable parameters.

With the recent fast development of generative models, instruction-based image editing has shown great potential in generating high-quality images. However, the quality of editing highly depends on carefully designed instructions, placing the burden of task decomposition and sequencing entirely on the user. PhotoAgent formulates autonomous image editing as a long-horizon decision-making problem. It reasons over user aesthetic intent, plans multi-step editing actions via tree search, and iteratively refines results through closed-loop execution with memory and visual feedback — without requiring step-by-step user prompts. To support reliable evaluation, we introduce UGC-Edit, an aesthetic evaluation benchmark of 7,000 photos and a learned aesthetic reward model. PhotoAgent significantly outperforms existing methods in both instruction faithfulness and visual quality across diverse editing scenarios.

OCE reformulates editing-based erasure as multiplicative parameter updates from a geometric perspective, applying layer-wise orthogonal transformations derived from a closed-form solution to enable precise concept erasure while preserving neuron magnitude and angular geometry. To address conflicting constraints in multi-concept erasure, OCE introduces a subspace-level objective with structured subspace manipulation, yielding a more effective and scalable erasure. The core limitation of editing-based methods is reliance on additive parameter updates: concept semantics primarily depend on neuron direction rather than magnitude while overall generative capacity relies on angular geometry, and additive updates inherently entangle these factors introducing unintended interference. OCE outperforms existing methods in concept erasure and non-target preservation, erasing up to 100 concepts in 4.3 seconds while preserving generation quality.

Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. We present Motive (MOTIon attribution for Video gEneration), a motion-centric, gradient-based data attribution framework that scales to modern, large, high-quality video datasets and models. We use this to study which fine-tuning clips improve or degrade temporal dynamics. Motive isolates temporal dynamics from static appearance via motion-weighted loss masks, yielding efficient and scalable motion-specific influence computation. On text-to-video models, Motive identifies clips that strongly affect motion and guides data curation that improves temporal consistency and physical plausibility. With Motive-selected data, we improve both motion smoothness and dynamic degree on VBench, achieving a 74.1% human preference win rate. This is the first framework to attribute motion rather than visual appearance in video generative models.

To achieve real-time interactive video generation, current methods distill pretrained bidirectional video diffusion models into few-step autoregressive models. This distillation faces an architectural gap when full attention is replaced by causal attention. We propose Causal Forcing, which uses an autoregressive teacher for ODE initialization to bridge the architectural gap, and then applies the same DMD (Distribution Matching Distillation) procedure as in Self Forcing. This yields high-quality real-time video generation with only 1–2 sampling steps per frame, enabling streaming and controllable rollout. Follow-up: Causal Forcing++ (arXiv:2605.15141) extends the pipeline with causal consistency distillation for scalability.

Vision-Language Models frequently produce self-reflective commentary during reasoning, such as claims about examining figures more carefully. This research investigates whether such statements represent genuine visual re-examination or merely learned textual patterns. We developed VisualSwap, a probing framework that swaps images after a model generates reasoning to test detection capabilities. We created VS-Bench containing 800 image pairs, and tested Qwen3-VL, Kimi-VL, ERNIE-VL. The findings reveal significant failures: models struggle to detect image changes, with accuracy declining substantially. Advanced thinking models show greater vulnerability than instruction-following variants, and increased model scale doesn't resolve the issue. However, multi-turn user interactions restore visual grounding, whereas self-generated reflections during continuous generation do not. Attention analysis shows self-reflection doesn't increase attention to visual tokens, while user instructions substantially elevate such attention — suggesting current VLMs prioritize language patterns over genuine visual processing.

The paper addresses a fundamental challenge in Vision-Language Models: understanding whether performance failures stem from perception or reasoning deficits. We propose that the root cause of this trade-off is an ambiguity in modality credit assignment. Our solution introduces a reinforcement learning framework featuring Perception Verification, which employs "blindfolded reasoning" to evaluate perceptual accuracy independently. The approach incorporates Structured Verbal Verification to replace variable LLM evaluation with deterministic algorithmic assessment. These components integrate into a Modality-Aware Credit Assignment mechanism that routes rewards to the specific source of error — either bad seeing or bad thinking — enabling a single VLM to achieve simultaneous performance gains across a wide task spectrum.

The frontier of Large Language Model capabilities has shifted from single-turn code generation to agentic software engineering. While post-training methods dominate, agentic mid-training remains under-explored despite offering scalability advantages. Our solution introduces agent-native data comprising two trajectory types: contextually-native trajectories providing broad coverage, and environmentally-native trajectories from executable repositories offering authentic interaction data. The approach tackles distribution mismatch between static training data and dynamic development environments. Our 32B and 72B models achieve 56.1% and 58.5% resolution rates on SWE-Bench Verified. Authors plan to open-source a significant portion of datasets, recipes, and model checkpoints.

Multimodal agents offer a compelling path to automating complex document-intensive workflows, yet a critical question remains: do these architectures demonstrate genuine strategic reasoning, or simply conduct stochastic trial-and-error search? To address this, we introduce Agentic Document VQA (MADQA), a benchmark of 2,250 human-authored questions grounded in 800 heterogeneous PDF documents. Guided by Classical Test Theory, we design it to maximize discriminative power. We introduce a novel evaluation protocol for measuring the accuracy-effort trade-off. Using this framework, we find that humans show strong metacognitive calibration, adapting or abandoning failed strategies, whereas frontier agents often persist in unproductive loops with diminishing returns. Best agents match human accuracy but succeed on largely different questions and rely on brute-force search to compensate for weak strategic planning, failing to close nearly 20% gap to oracle.

Existing online benchmarks for mobile GUI agents remain largely app-centric and task-homogeneous, failing to reflect the diversity and instability of real-world mobile usage. VenusBench-Mobile is a challenging online benchmark for evaluating general-purpose mobile GUI agents under realistic, user-centric conditions. It builds two core evaluation pillars: what to evaluate via user-intent-driven task design that reflects real mobile usage, and how to evaluate through a capability-oriented annotation scheme for fine-grained agent behavior analysis. Evaluation reveals large performance gaps relative to prior benchmarks. Diagnostic analysis further shows that failures are dominated by deficiencies in perception and memory, which are largely obscured by coarse-grained evaluations. Moreover, even the strongest agents exhibit near-zero success under environment variations.

In closed-loop multi-turn agent reinforcement learning, LLM agents exhibit reasoning collapse, where reasoning shifts toward generic templates, weakly coupled to inputs. Such collapse is easy to miss with entropy or surface diversity metrics since reasoning text still varies but becomes input-agnostic. We propose an information-theoretic decomposition of reasoning variable Z's variation into conditional entropy H(Z|X) and mutual information I(X; Z). Template collapse occurs when H(Z|X) stays high while I(X; Z) drops, yielding diverse-looking but generic reasoning. To make I(X; Z) a reproducible diagnostic, we introduce an MI-style retrieval protocol treating each reasoning trace as a query to retrieve its source; accuracy degrades toward chance under collapse. We propose reward-variance-aware filtering to prioritize high-signal updates. Improves input dependence, stability, and performance across multi-turn environments including VLMs.

RL, particularly RL from verifiable reward (RLVR), has become a crucial phase of training LLMs and a key focus of current scaling efforts. However, optimization practices in RL largely follow those of next-token-prediction stages, despite fundamental differences between RL and these stages. One such practice is the use of AdamW, widely adopted despite its high memory overhead. Our analysis shows that both momentum and adaptive learning rate of AdamW are less influential in RL than in SFT. Confirming our hypothesis, the substantially more memory-efficient SGD — known to perform poorly in supervised learning — matches or even outperforms AdamW in RL for LLMs. Remarkably, full fine-tuning with SGD updates fewer than 0.02% of model parameters without any sparsity-promoting regularization, more than 1,000× fewer than AdamW.

Safe deployment of increasingly capable AI agents may require visibility into how they make decisions. We introduce three evaluation frameworks and a novel metric for assessing monitorability — the ability to detect problematic behavior through chain-of-thought analysis. Key findings: reasoning transparency outperforms action-only monitoring in practical settings; current models demonstrate generally reliable but imperfect monitorability. Longer reasoning chains typically improve detection capabilities. A smaller model at higher reasoning effort can yield higher monitorability, at greater inference compute cost. Monitorability can be improved by asking models follow-up questions and giving their follow-up CoT to the monitor. OpenAI released datasets and reference code from this work in April 2026.

Training against white-box deception detectors has been proposed as a way to make AI systems honest, however such training risks models learning to obfuscate their deception to evade the detector. We constructed a realistic coding environment where reward hacking via hardcoding test cases naturally occurs, and showed that obfuscation emerges. Two obfuscation strategies emerge: (1) Obfuscated activations — model outputs deceptive text while modifying internal representations to no longer trigger the detector; (2) Obfuscated policy — model outputs deceptive text that evades the detector, typically by including a justification for the reward hack. Sufficiently high KL regularization and detector penalty can yield honest policies, establishing white-box deception detectors as viable training signals for tasks prone to reward hacking.

Aligning LLMs with the diverse spectrum of human values remains a central challenge: preference-based methods often fail to capture deeper motivational principles. Our framework addresses three key gaps in value-based alignment: extraction that ignores hierarchy, evaluation that lacks calibrated intensity measurement, and insufficient steerability control. The proposed system comprises three components: (1) HIVES, a hierarchical value embedding space that captures intra- and cross-theory value structure; (2) VIDB, the Value Intensity DataBase, a large-scale resource of value-labeled texts with intensity estimates derived from ranking-based aggregation; and (3) an anchor-based evaluator that produces consistent intensity scores by ranking outputs against VIDB panels. Through testing across ten models and four value theories, we discovered asymmetries in steerability and composition laws for multi-value control.