I don't remember where I've got that specific picture from, but here's a more detailed study: vlmsarebiased.github.io
Vision Language Models are Biased
Vision Language Models are Biased: state-of-the-art VLMs score 100% on familiar subjects but only 17.05% when counting in subtly modified, counterfactual images.
vlmsarebiased.github.io
November 8, 2026 at 4:30 PM
I don't remember where I've got that specific picture from, but here's a more detailed study: vlmsarebiased.github.io
Extremely cool to see how some of the frontier model work (ex. System 1 / Decision Models) are starting to make their way into robotics, *usefully*.
VFlash: Enabling System-1 VLMs for Physical AI · NVIDIA Research
VFlash enables System-1 VLMs for Physical AI through fast, accurate parallel visual planning, accelerating robot and autonomous driving inference.
research.nvidia.com
November 6, 2026 at 12:47 AM
Extremely cool to see how some of the frontier model work (ex. System 1 / Decision Models) are starting to make their way into robotics, *usefully*.
You got an angle where you could set up a second camera and take license plate photos? Barring that I bet the VLMs could do a good job of calling out car models at least
November 9, 2026 at 11:09 AM
You got an angle where you could set up a second camera and take license plate photos? Barring that I bet the VLMs could do a good job of calling out car models at least
Takeaways:
- Training can build strange visual associations that models don't verbalize, a problem for trace-based oversight of agents in multimodal envs
- Visual "neuralese" looks unlikely for now, but as models align meaning w/ each other, messages get harder for humans to read.
- Training can build strange visual associations that models don't verbalize, a problem for trace-based oversight of agents in multimodal envs
- Visual "neuralese" looks unlikely for now, but as models align meaning w/ each other, messages get harder for humans to read.
A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMs
Do vision-language models (VLMs) share our ability to communicate visually? We share preliminary findings, starting from a surprising observation: simple visual symbols like a red heart can strongly i...
arxiv.org
November 5, 2026 at 2:54 AM
Takeaways:
- Training can build strange visual associations that models don't verbalize, a problem for trace-based oversight of agents in multimodal envs
- Visual "neuralese" looks unlikely for now, but as models align meaning w/ each other, messages get harder for humans to read.
- Training can build strange visual associations that models don't verbalize, a problem for trace-based oversight of agents in multimodal envs
- Visual "neuralese" looks unlikely for now, but as models align meaning w/ each other, messages get harder for humans to read.
Amazon Marketplace evaluates A+ Content quality with SageMaker AI and VLMs. Amazon pairs a vision-language model with OCR signals and a CatBoost classifier trained on expert annotations.
Amazon Marketplace evaluates A+ Content quality with SageMaker AI and VLMs - GenAInews
Amazon pairs a vision-language model with OCR signals and a CatBoost classifier trained on expert annotations.
genainews.tech
November 6, 2026 at 2:51 AM
Amazon Marketplace evaluates A+ Content quality with SageMaker AI and VLMs. Amazon pairs a vision-language model with OCR signals and a CatBoost classifier trained on expert annotations.
Can VLMs play Pac-Man? Or other Atari games? This paper describes AtariBench, a benchmark that evaluates how well VLMs track and understand changes in games.
If you know the story of @demishassabis, you might have guessed correctly that this paper is co-written by researchers at Google DeepMind. 👀👇
If you know the story of @demishassabis, you might have guessed correctly that this paper is co-written by researchers at Google DeepMind. 👀👇
November 9, 2026 at 7:30 PM
Can VLMs play Pac-Man? Or other Atari games? This paper describes AtariBench, a benchmark that evaluates how well VLMs track and understand changes in games.
If you know the story of @demishassabis, you might have guessed correctly that this paper is co-written by researchers at Google DeepMind. 👀👇
If you know the story of @demishassabis, you might have guessed correctly that this paper is co-written by researchers at Google DeepMind. 👀👇
@hikikomorphism.bsky.social: have u seen arxiv.org/abs/2610.36449? seems potentially relevant to your metacog work
A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMs
Do vision-language models (VLMs) share our ability to communicate visually? We share preliminary findings, starting from a surprising observation: simple visual symbols like a red heart can strongly i...
arxiv.org
November 4, 2026 at 8:57 PM
@hikikomorphism.bsky.social: have u seen arxiv.org/abs/2610.36449? seems potentially relevant to your metacog work
Paper: shm.to/9C7Bsz1
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
PANORAMA: Panoptic Grounded Captioning via Mask Proposal Selection
Intelligent systems that act in the world require image understanding that is both comprehensive and spatially grounded. Current vision-language models (VLMs) can generate fluent and detailed image ca...
shm.to
November 5, 2026 at 8:39 AM
Paper: shm.to/9C7Bsz1
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
Hongbin Lin et al.
#arXiv #cs.AI
Hongbin Lin et al.
#arXiv #cs.AI
Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
Reasoning traces improve vision-language models on complex multimodal tasks, but forcing every input to follow a long reasoning process leads to overthinking and unnecessary inference cost. We study adaptive fast/slow thinking for VLMs, where one model answers directly for simple inputs and reasons…
arxiv.org
November 9, 2026 at 9:26 PM
Chongyu Qu, Ritchie Zhao, Yufan He, Dong Yang, Zhengyi Lu, Junchao Zhu, Tianyuan Yao, Juming Xiong, Junlin Guo, Yanfan Zhu, Yuechen Yang, Daguang Xu, ...
TACT-KV: Tri-Axis Cosine Transform for Compressing Volumetric KV Caches in Medical VLMs
https://arxiv.org/abs/2611.07593
TACT-KV: Tri-Axis Cosine Transform for Compressing Volumetric KV Caches in Medical VLMs
https://arxiv.org/abs/2611.07593
November 8, 2026 at 10:38 PM
Chongyu Qu, Ritchie Zhao, Yufan He, Dong Yang, Zhengyi Lu, Junchao Zhu, Tianyuan Yao, Juming Xiong, Junlin Guo, Yanfan Zhu, Yuechen Yang, Daguang Xu, ...
TACT-KV: Tri-Axis Cosine Transform for Compressing Volumetric KV Caches in Medical VLMs
https://arxiv.org/abs/2611.07593
TACT-KV: Tri-Axis Cosine Transform for Compressing Volumetric KV Caches in Medical VLMs
https://arxiv.org/abs/2611.07593
Bidirectional highlight sync between Markdown and PDF is brilliant.
The trickiest trap is text drift: OCR/VLMs normalize ligatures and hyphen breaks, breaking string matching. Linking Markdown token spans to glyph bounding boxes keeps annotations pinned accurately to the PDF canvas.
The trickiest trap is text drift: OCR/VLMs normalize ligatures and hyphen breaks, breaking string matching. Linking Markdown token spans to glyph bounding boxes keeps annotations pinned accurately to the PDF canvas.
November 7, 2026 at 10:42 PM
Bidirectional highlight sync between Markdown and PDF is brilliant.
The trickiest trap is text drift: OCR/VLMs normalize ligatures and hyphen breaks, breaking string matching. Linking Markdown token spans to glyph bounding boxes keeps annotations pinned accurately to the PDF canvas.
The trickiest trap is text drift: OCR/VLMs normalize ligatures and hyphen breaks, breaking string matching. Linking Markdown token spans to glyph bounding boxes keeps annotations pinned accurately to the PDF canvas.
🚨🚨📢📢 I'm beyond excited to announce PANORAMA, our new work on panoptic grounded captioning! Current VLMs write fluent captions, but grounding them in pixels remains challenging. PANORAMA describes both foreground objects and background regions, grounding each referring phrase with pixel-level masks.
November 2, 2026 at 1:46 PM
🚨🚨📢📢 I'm beyond excited to announce PANORAMA, our new work on panoptic grounded captioning! Current VLMs write fluent captions, but grounding them in pixels remains challenging. PANORAMA describes both foreground objects and background regions, grounding each referring phrase with pixel-level masks.
D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models
Vision-Language Models (VLMs) hold promise for zero-shot video task-progress estimation in robot learning, yet existing frameworks frequently suffer from…
Marius Memmel et al.
Vision-Language Models (VLMs) hold promise for zero-shot video task-progress estimation in robot learning, yet existing frameworks frequently suffer from…
Marius Memmel et al.
D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models
Vision-Language Models (VLMs) hold promise for zero-shot video task-progress estimation in robot learning, yet existing frameworks frequently suffer from confirmation bias, hallucinating progress in f
arxiv.org
November 3, 2026 at 11:17 PM
D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models
Vision-Language Models (VLMs) hold promise for zero-shot video task-progress estimation in robot learning, yet existing frameworks frequently suffer from…
Marius Memmel et al.
Vision-Language Models (VLMs) hold promise for zero-shot video task-progress estimation in robot learning, yet existing frameworks frequently suffer from…
Marius Memmel et al.
Qu, Zhao, He, Yang, Lu, Zhu, Yao, Xiong, Guo, Zhu, Yang, Xu, Landman, Tang, Huo: TACT-KV: Tri-Axis Cosine Transform for Compressing Volumetric KV Caches in Medical VLMs https://arxiv.org/abs/2611.07593 https://arxiv.org/pdf/2611.07593 https://arxiv.org/html/2611.07593
November 6, 2026 at 6:40 AM
Qu, Zhao, He, Yang, Lu, Zhu, Yao, Xiong, Guo, Zhu, Yang, Xu, Landman, Tang, Huo: TACT-KV: Tri-Axis Cosine Transform for Compressing Volumetric KV Caches in Medical VLMs https://arxiv.org/abs/2611.07593 https://arxiv.org/pdf/2611.07593 https://arxiv.org/html/2611.07593
How Do Artificial Agents Make Decisions in Light of Visual Data?
Keyty Loza, Shayan Monadjemi
#arXiv #cs.AI #cs.HC
Keyty Loza, Shayan Monadjemi
#arXiv #cs.AI #cs.HC
How Do Artificial Agents Make Decisions in Light of Visual Data?
Vision-language models (VLMs) have demonstrated significant advances in reasoning with visual data, which is particularly important given that much of human knowledge is encoded visually in scientific articles and other sources. However, a gap remains in understanding how artificial agents, powered…
arxiv.org
November 5, 2026 at 9:26 PM
How Do Artificial Agents Make Decisions in Light of Visual Data?
Keyty Loza, Shayan Monadjemi
#arXiv #cs.AI #cs.HC
Keyty Loza, Shayan Monadjemi
#arXiv #cs.AI #cs.HC
How Do Artificial Agents Make Decisions in Light of Visual Data?
Vision-language models (VLMs) have demonstrated significant advances in reasoning with visual data, which is particularly important given that much of human knowledge is encoded visually in scientific articles and other sources. However, a gap remains in understanding how artificial agents, powered…
arxiv.org
November 5, 2026 at 10:40 PM
Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
Hongbin Lin et al.
#arXiv #cs.AI
Hongbin Lin et al.
#arXiv #cs.AI
Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
Reasoning traces improve vision-language models on complex multimodal tasks, but forcing every input to follow a long reasoning process leads to overthinking and unnecessary inference cost. We study adaptive fast/slow thinking for VLMs, where one model answers directly for simple inputs and reasons…
arxiv.org
November 9, 2026 at 10:59 PM
Training AIs to Solve CAPTCHAs - Schneier on Security www.schneier.com/blog/archive...
Training AIs to Solve CAPTCHAs - Schneier on Security
Interesting research: Abstract: Vision-language models (VLMs) can solve visual CAPTCHAs without task-specific training, but the agents built on them approach every challenge from scratch. For such an…
www.schneier.com
November 6, 2026 at 8:30 AM
Training AIs to Solve CAPTCHAs - Schneier on Security www.schneier.com/blog/archive...
PANORAMA is fully open-source!
Paper: shm.to/9C7Bsz1
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
We found that it generalises to grounding robotic arms/grippers and the manipulated objects, so you can use it to improve your VLA/WM!
Paper: shm.to/9C7Bsz1
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
We found that it generalises to grounding robotic arms/grippers and the manipulated objects, so you can use it to improve your VLA/WM!
November 4, 2026 at 8:50 AM
PANORAMA is fully open-source!
Paper: shm.to/9C7Bsz1
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
We found that it generalises to grounding robotic arms/grippers and the manipulated objects, so you can use it to improve your VLA/WM!
Paper: shm.to/9C7Bsz1
Project page: shm.to/Pmyj02o
Code: shm.to/lRcAvgY
HF: shm.to/uPa4MCC
Dataset: shm.to/B2mWgRN
We found that it generalises to grounding robotic arms/grippers and the manipulated objects, so you can use it to improve your VLA/WM!
Depth plus VLMs sounds like a solid way to stop my robots from hugging walls, but I still have to nag them to pick up the wrench.
November 3, 2026 at 11:01 PM
Depth plus VLMs sounds like a solid way to stop my robots from hugging walls, but I still have to nag them to pick up the wrench.
An MIT-licensed 3B VLM running fully on Axera NPUs, no GPU or cloud, is what smart-camera builders need. But the model card is bare: no benchmarks, no quantisation figures. Hard to see offline OCR beating cloud APIs until Axera publishes numbers.
Bringing Vision-Language AI to the Edge: Qwen2.5-VL 3B on Axera NPUs
Running vision-language models (VLMs) on the edge has traditionally been a trade-off between capability and hardware cost. For developers building smart cameras or IoT devices, relying on cloud APIs…
aichina.news
November 8, 2026 at 12:36 PM
An MIT-licensed 3B VLM running fully on Axera NPUs, no GPU or cloud, is what smart-camera builders need. But the model card is bare: no benchmarks, no quantisation figures. Hard to see offline OCR beating cloud APIs until Axera publishes numbers.
Every run is scored on held-out items (macro-F1, mAP, WER, Dice, whatever fits), and you can push a model to the Hugging Face Hub with a model card.
Already have a model running somewhere? Bridges for Label Studio ML, Roboflow, HF Inference, CVAT and OpenAI-compatible VLMs score it the same way.
Already have a model running somewhere? Bridges for Label Studio ML, Roboflow, HF Inference, CVAT and OpenAI-compatible VLMs score it the same way.
November 5, 2026 at 4:37 PM
Every run is scored on held-out items (macro-F1, mAP, WER, Dice, whatever fits), and you can push a model to the Hugging Face Hub with a model card.
Already have a model running somewhere? Bridges for Label Studio ML, Roboflow, HF Inference, CVAT and OpenAI-compatible VLMs score it the same way.
Already have a model running somewhere? Bridges for Label Studio ML, Roboflow, HF Inference, CVAT and OpenAI-compatible VLMs score it the same way.
Empty model card, no Int4 evals, and an MIT claim clashing with upstream Apache-2.0. Still, a 2B Qwen3-VL in 1–1.5 GB on an Axera NPU fills a real gap. Until someone publishes accuracy numbers, I'd call this a demo, not a deployment — anyone with bench results, prove me wrong.
Running Qwen3-VL Locally: A New Edge-Friendly VLM for Axera Hardware
For developers building the next generation of smart cameras, autonomous robots, and IoT devices, the ability to run vision-language models (VLMs) locally is a significant leap forward. We are seeing…
aichina.news
November 7, 2026 at 10:49 AM
Empty model card, no Int4 evals, and an MIT claim clashing with upstream Apache-2.0. Still, a 2B Qwen3-VL in 1–1.5 GB on an Axera NPU fills a real gap. Until someone publishes accuracy numbers, I'd call this a demo, not a deployment — anyone with bench results, prove me wrong.
Robots need to know where things are, how far away, and what to do next. OpenCV Live, Thu Nov 5, 9am PT: RealSense depth + Jetson + ROS 2 + VLMs, with live AgenticROS demos.
opencv.org/from-pixels-...
opencv.org/from-pixels-...
From Pixels to Physical AI: Building the Visual Cortex for Intelligent Robots - OpenCV Live! 232
Live Stream Date/Time: Thursday, November 5th, 2026 @ 9am Pacific. Computer vision has long been about helping machines recognize what’s in front of them. Physical AI raises the bar: robots also need ...
opencv.org
November 3, 2026 at 7:46 PM
Robots need to know where things are, how far away, and what to do next. OpenCV Live, Thu Nov 5, 9am PT: RealSense depth + Jetson + ROS 2 + VLMs, with live AgenticROS demos.
opencv.org/from-pixels-...
opencv.org/from-pixels-...