RoSeViT introduces a role-separated Vision Transformer that decouples high-level instruction inference from visual workspace computation using controller and workspace tokens. This design allows nonlocal workspace interactions to be routed through the controller, reducing complexity compared to dense ViTs. Empirical results show RoSeViT improves ARC-1 accuracy from 54.5 to 56.6 with a backbone replacement and up to 62.6 when integrated into the full system.
arXiv:2606. 30185v1 Announce Type: new Abstract: Improving vision-language models (VLMs) on visual reasoning typically requires retraining or hand-designed prompts and tools.
Improving vision-language models (VLMs) on visual reasoning typically requires retraining or hand-designed prompts and tools. We present Dynamo, a training-free framework that adapts a frozen VLM without any weight updates.
arXiv:2610.14403v2 Announce Type: replace Abstract: Learning reusable skills can improve visual reasoning agents, yet applying a skill still requires the agent to implement or invoke a procedure and...
arXiv:2607. 12815v1 Announce Type: new Abstract: Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces.
T3R is a training‑free, two‑stage token refinement framework designed to improve the efficiency and robustness of Vision‑Language‑Action (VLA) models. The first stage uses segmentation‑guided pruning to remove irrelevant visual patches, while the second stage applies saliency‑guided attention biasing to focus on action‑critical regions and language cues. Experiments on LIBERO‑Spatial, LIBERO, SIMPLER, and RoboTwin show that T3R can eliminate 85% of visual tokens with minimal loss in success rate and even improve performance under distribution shifts, achieving up to a 3.3× speedup in large‑scale settings.
arXiv:2609.24362v1 Announce Type: new Abstract: Sandboxed computer environments support multi-step reasoning with tools, executable programs, and persistent files, yet their extension from language m...
arXiv:2607. 29586v1 Announce Type: cross Abstract: The Abstraction and Reasoning Corpus (ARC) tests whether a model can infer an unseen transformation from a few input-output examples and apply it to a new grid.
Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces. We ask whether CoT requires continued access to image tokens, or whether it mainly operates over visual information already made available earlier in the forward pass.
The paper introduces NEO-VLA, an encoder‑free vision‑language‑action framework that tokenizes images with lightweight patch convolutions and lets visual and language features evolve together throughout the backbone. By allowing a flow‑matching action expert to access instruction‑conditioned visual representations at multiple depths, NEO‑VLA consistently outperforms encoder‑based counterparts across diverse manipulation tasks and distribution shifts. Layer‑wise probing reveals that action‑critical signals peak early in the visual hierarchy, underscoring the benefit of accessing representations across the computational pipeline rather than relying on a fixed encoder output.
The paper introduces IG‑VLA, a vision‑language‑action framework that learns to imagine task‑relevant future scene evolution in a latent spatiotemporal space, guiding action prediction without generating full pixel‑level videos. It further compresses this future reasoning into a compact Scene Gist Token via Scene Gist Memory, enabling efficient inference. Experiments on LIBERO, LIBERO‑Plus, and VLABench show that IG‑VLA improves success rates by nearly 6% and speeds up inference up to 6.38× compared to strong baselines.
arXiv:2610.08133v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models achieve strong robotic manipulation performance but incur high computational costs from processing long token seq...
arXiv:2606. 01503v1 Announce Type: cross Abstract: Unified vision-language models (VLMs) integrate visual understanding and visual generation within a single autoregressive backbone, but their joint training is computationally expensive and largely overlooked from an efficiency perspective.