#01Nov 4, 2026
cs.CV
Turns Matter: Advancing Streaming Video RL via Turn-Level Credit
Shuai Dong, Kanghui Tian, Linghan Chen and 20 more
Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.
#02Nov 4, 2026
cs.CV
LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
Haoxiang Cao, Yuxuan Zhang, Penghui Du and 16 more
We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.
#03Nov 4, 2026
cs.CV
Trick-or-Treat: Adversarial Evaluation of Multimodal Medical Reasoning with Synthetic Images
Oishi Banerjee, Eujin Lee, Sung Eun Kim and 4 more
We used medical image synthesis to create a challenging multimodal evaluation framework that exposes hallucination behaviors in frontier models. Evaluating multimodal reasoning has historically been difficult because existing medical datasets rarely pair different image modalities, such as radiographs and external clinical photographs. To address this gap, we introduce Trick-or-Treat, an adversarial multimodal medical benchmark that pairs real chest radiographs with synthetic clinical photographs to challenge models' cross-modal reasoning. For each of a wide range of conditions, we created two pairs of images, where the pairs share a single chest radiograph but contain different external photographs. In matched pairs, the chest radiograph and photograph both contain findings suggesting the same disease (e.g., varicella pneumonia suggested by lung opacities and crusted skin lesions seen in chickenpox). In mismatched pairs, the photograph depicts an alternative clinical finding that would not lead to the same unifying diagnosis (e.g. replacing crusted skin lesions with ordinary acne). Synthetic photograph generation allows clinical findings to be varied while the person's features and photographic setting stay consistent, reducing confounding. In an investigation of three frontier models, we found that models hallucinated clinically significant photograph findings in 45-57.5% of mismatched pairs, up from 25-37.5% for the same photographs shown alone. When faced with mismatched pairs, models tended to ignore mismatches and instead hallucinated findings that better matched the radiograph (e.g. misreading acne as crusted skin lesions to justify diagnosing chickenpox). Trick-or-Treat both demonstrates concerning error pattern in frontier models' medical reasoning and provides a proof of concept for the use of synthetic images in adversarial multimodal evaluation.
#04Nov 4, 2026
cs.CV
From Pixels to Actions - Towards Native Vision-Language-Action Policies
Weichen Fan, Haiwen Diao, Penghao Wu and 14 more
Vision-language-action models are typically built on a simple premise: perception is first compressed by visual encoders, and action reasoning begins only afterwards. Yet manipulation often hinges on precisely the visual signals most vulnerable to this abstraction, e.g., local geometry, boundaries, depth, and subtle state changes. This raises a fundamental question: should action policies inherit a fixed visual endpoint, or directly access evolving visual computation? We introduce NEO-VLA, a structurally encoder-free vision-language-action framework. Images are tokenized by lightweight patch convolutional layers and evolve jointly with language throughout the backbone, allowing a flow-matching action expert to draw from instruction-conditioned visual representations at different depths. This simple change reveals a consistent pattern. Native architectures outperform encoder-based counterparts under controlled comparisons; full fine-tuning reduces but does not erase the gap, and additional scale alone cannot recover it. More strikingly, layer-wise probing shows that the action-critical signals often peak well before the final visual representation. Across diverse manipulation settings and distribution shifts, NEO-VLA consequently achieves stronger and more robust control. Together, these results suggest that embodied control benefits from access to visual representations across the computational hierarchy, rather than a fixed encoder output.
#05Nov 4, 2026
cs.CV
TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing
Ruikang Li, Zixuan Chen, Jiarui Wu and 6 more
Strong single-turn editing performance does not guarantee reliable multi-turn interaction: as edited outputs are repeatedly reused, small unintended changes can accumulate into severe visual drift. This long-term drift is often mitigated through training on multi-turn data or cross-turn rollouts, potentially altering pretrained priors, or through specialized inference-time mechanisms that are difficult to scale to foundation models. We introduce TailSteer, a final-stage training framework that improves multi-turn robustness while retaining the generative and editing capabilities acquired through earlier training. TailSteer follows a simple principle: keep early denoising dynamics close to those of the base model and concentrate preservation corrections near the trajectory tail, where fine-grained deviations can be suppressed. These corrections are learned from the editor's own on-policy denoising rollouts, using unchanged source content as a preservation reference without requiring curated ground-truth editing pairs. Experiments across the FLUX, SenseNova, and Qwen families show that TailSteer substantially reduces cumulative drift over 15-20 editing turns while largely preserving base-model performance on extensive text-to-image and single-image editing benchmarks. These results demonstrate that long-horizon editing robustness can be unlocked in foundation models after TailSteer, rather than by relearning their existing capabilities.