C2-Routing cuts VLM visual tokens by routing crop or compress
With a 1,024-token cap already applied, the method reports 36.53% fewer FLOPs and a 0.77% mean relative score drop.
Stories per day · 30 days
8 RSS
Relatedrobotics9vla6benchmark5huggingface5embodied-ai4qwen3agent2reinforcement-learning2robot-agents2scaling-laws2spatial-reasoning2vision-language23d13d-composition1adversarial1agentic-ai1agentic-reasoning1agents1
91 stories, newest first · one card per story
Stories tagged #vlm
PLatour Enables Prefill-to-Autoregressive Cross-Model Transfer
PLatour framework shows promise in balancing capabilities and costs but gains may not require true sender-receiver communication.
SleuthVLM outperforms Gemini on omni video benchmarks
SleuthVLM reaches 75.0% accuracy on omni benchmarks via transcription and agentic reasoning.
Visual perturbations alter VLA actions without changing text rationale
TNAS-guided attacks induce collisions in 52.6% of driving scenes while keeping rationales word-for-word identical.
H2RB bridges human-to-robot VLA transfer
H2RB co-trains VLAs on unpaired human and robot data for 51.8 object transfer success.
Visual perturbations decouple actions from identical rationales in VLAs
Alpamayo and InstructVLA show TNAS steering shifts actions 4.64x-47.8x more than text perturbations while preserving rationales.
D&A Enables VLMs as Zero-Shot Robotic Reward Models
Decouples progress estimation into objective description and rubric-based assessment to reduce confirmation bias in zero-shot VLM evaluations for robotics.
HeLM Improves Non-Markovian Manipulation in VLA Models
HeLM achieves 12.8 percentage point intra-episode success gain on 10 non-Markovian tasks with 56 variants.
DR Pixel enables page-native deep research with vision-language models
DR Pixel is an open-source agent that reasons directly over document page images, outperforming OCR-based methods on scientific papers.
DPDM Improves Agent Decision Making via Shared Predictor
DPDM raises mean accuracy from 78.80% to 84.92% across 50 benchmarks using one frozen predictor.
CaPtain turns small VLMs into robot manipulation agents
CaPtain trains small open VLMs into code-as-policy agents achieving 73.8% success on LIBERO benchmarks and running on real hardware without frontier models.
Mage-VL-Cowork-4B Achieves State-of-the-Art on OSWorld-Verified
The 4B hybrid CUA is trained through a five-stage curriculum using skill-conditioned generation workflows that synthesize 293806 tasks.
PhysEvolve trains VLMs on executable worlds for physical reasoning
From Qwen3-VL-8B-Instruct, the paper reports PhysBench Throwing up 8.8 and ContPhy Goal-Pit up 24.4 points.
OSMA-Bench finds robot policies fail obstacle-conditioned safe motion
VLM and VLA policies can rank obstacles by safety but do not change clearance, speed, or smoothness to match.
LOLPlay Benchmark Evaluates Real-Time Visual Agents in League of Legends
GPT-6 Astra achieves 25.0% task success rate at low and medium reasoning effort across 324 evaluations.
IGG benchmark asks GUI agents to reveal hidden targets before clicking
Humans score 94.6%, GPT-5.5 79.4%, and the best open-source VLM only 31.6% on 601 tasks.
Sensorimotor Alignment proposed as offline proxy for robot policy success
The training-free S^2 metric correlates with closed-loop robot task success and can rank VLM backbones before VLA fine-tuning.
Frontier VLM Agents Show Limited Success in Robot Generalist Tasks
Closed-weight models lead in all five domains while open-weight VLMs show lower planning and manipulation success.
New offline metric measures sensorimotor alignment in AI models
Sensorimotor Alignment S2 correlates with robot task success and identifies strong vision-language-action models before expensive rollouts.
S-Agent 2.0 adds spatial tools to general coding agents
With Qwen3.8-27B it scores 66.5 on MMSI-Bench and 82.3 on ReVSI, above SpatialClaw.
Paper studies physical recursive self-improvement in VLM policies
PhysRSI-eval scores VLMs on 18 manipulation tasks under a fixed budget and tests whether public scores predict later physical gains.
Token readout choice confounds POPE VLM hallucination scores
A two-token yes/no logit lookup never matches LLaVA-1.5-7B outputs, shifting adversarial F1 by 6.13 points.
InfiniDream generates editable photorealistic 3D worlds with VLMs
Evaluations on 100 object and 20 scene tasks report gains on four perceptual metrics.
Paper studies physical recursive self-improvement in VLM agents
PhysRSI-eval scores 18 manipulation tasks and finds close public averages can hide late or inefficient gains.