The Failure Catalog
📰 Medium · LLM
Learn how Visual Language Models (VLMs) can fail when trusting data over visual evidence and how to stress test them
Action Steps
- Setup adversarial conditions to stress test VLMs
- Use fabricated or partial data to evaluate model performance
- Evaluate the impact of delivery channels on model accuracy
- Analyze the results to identify anchoring bias in VLMs
- Apply techniques to mitigate anchoring bias and improve visual reasoning
Who Needs to Know This
Data scientists and AI engineers working with VLMs can benefit from understanding these failure modes to improve model robustness
Key Insight
💡 VLMs can be misled by structured data, suppressing visual reasoning, and delivery channels can modulate this effect
Share This
🚨 VLMs can miss the 'Elephant in the Room' when trusting data over visual evidence! 🤖 Learn how to stress test and improve model robustness #AI #VLMs
Full Article
Title: The Failure Catalog
URL Source: https://medium.com/@mraduldubey/the-failure-catalog-fa236eeb25e2?source=rss------llm-5
Published Time: 2026-04-24T14:20:22Z
Markdown Content:
[Sitemap](https://medium.com/sitemap/sitemap.xml)
[Open in app](https://play.google.com/store/apps/details?id=com.medium.reader&referrer=utm_source%3DmobileNavBar&source=post_page---top_nav_layout_nav-----------------------------------------)
Sign up
[Sign in](https://medium.com/m/signin?operation=login&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=post_page---top_nav_layout_nav-----------------------global_nav------------------)
[](https://medium.com/?source=post_page---top_nav_layout_nav-----------------------------------------)
Get app
[Write](https://medium.com/m/signin?operation=register&redirect=https%3A%2F%2Fmedium.com%2Fnew-story&source=---top_nav_layout_nav-----------------------new_post_topnav------------------)
[Search](https://medium.com/search?source=post_page---top_nav_layout_nav-----------------------------------------)
Sign up
[Sign in](https://medium.com/m/signin?operation=login&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=post_page---top_nav_layout_nav-----------------------global_nav------------------)

# **The Failure Catalog**
## _Ghost entities and false accusations, when VLMs trust data over their own eyes!_
[](https://medium.com/@mraduldubey?source=post_page---byline--fa236eeb25e2---------------------------------------)
[Mradul Dubey](https://medium.com/@mraduldubey?source=post_page---byline--fa236eeb25e2---------------------------------------)
Follow
5 min read
·
Apr 24, 2026
[](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2F_%2Fvote%2Fp%2Ffa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&user=Mradul+Dubey&userId=b72e98a18093&source=---header_actions--fa236eeb25e2---------------------clap_footer------------------)
52
[](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2F_%2Frepost%2Fp%2Ffa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&user=Mradul+Dubey&userId=b72e98a18093&source=---header_actions--fa236eeb25e2---------------------repost_header------------------)
[](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2F_%2Fbookmark%2Fp%2Ffa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=---header_actions--fa236eeb25e2---------------------bookmark_footer------------------)
[Listen](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2Fplans%3Fdimension%3Dpost_audio_button%26postId%3Dfa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=---header_actions--fa236eeb25e2---------------------post_audio_button------------------)
Share
Press enter or click to view image in full size

Your VLMs may miss the proverbial “Elephant in the Room”
In [VLM: the more you tell it, the less it sees](https://medium.com/towards-artificial-intelligence/vlm-the-more-you-tell-it-the-less-it-sees-c07f33b6a159), I defined anchoring bias in VLMs and measured it across seven controlled conditions.
### The findings were quantitative:
> **_Structured data suppresses visual reasoning, and the delivery channel modulates how badly._**
This post is different. These are not controlled experiments. They are stress tests. I have setup deliberately adversarial conditions (fabricated data, partial data, a weaker model) that I used
URL Source: https://medium.com/@mraduldubey/the-failure-catalog-fa236eeb25e2?source=rss------llm-5
Published Time: 2026-04-24T14:20:22Z
Markdown Content:
[Sitemap](https://medium.com/sitemap/sitemap.xml)
[Open in app](https://play.google.com/store/apps/details?id=com.medium.reader&referrer=utm_source%3DmobileNavBar&source=post_page---top_nav_layout_nav-----------------------------------------)
Sign up
[Sign in](https://medium.com/m/signin?operation=login&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=post_page---top_nav_layout_nav-----------------------global_nav------------------)
[](https://medium.com/?source=post_page---top_nav_layout_nav-----------------------------------------)
Get app
[Write](https://medium.com/m/signin?operation=register&redirect=https%3A%2F%2Fmedium.com%2Fnew-story&source=---top_nav_layout_nav-----------------------new_post_topnav------------------)
[Search](https://medium.com/search?source=post_page---top_nav_layout_nav-----------------------------------------)
Sign up
[Sign in](https://medium.com/m/signin?operation=login&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=post_page---top_nav_layout_nav-----------------------global_nav------------------)

# **The Failure Catalog**
## _Ghost entities and false accusations, when VLMs trust data over their own eyes!_
[](https://medium.com/@mraduldubey?source=post_page---byline--fa236eeb25e2---------------------------------------)
[Mradul Dubey](https://medium.com/@mraduldubey?source=post_page---byline--fa236eeb25e2---------------------------------------)
Follow
5 min read
·
Apr 24, 2026
[](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2F_%2Fvote%2Fp%2Ffa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&user=Mradul+Dubey&userId=b72e98a18093&source=---header_actions--fa236eeb25e2---------------------clap_footer------------------)
52
[](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2F_%2Frepost%2Fp%2Ffa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&user=Mradul+Dubey&userId=b72e98a18093&source=---header_actions--fa236eeb25e2---------------------repost_header------------------)
[](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2F_%2Fbookmark%2Fp%2Ffa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=---header_actions--fa236eeb25e2---------------------bookmark_footer------------------)
[Listen](https://medium.com/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2Fplans%3Fdimension%3Dpost_audio_button%26postId%3Dfa236eeb25e2&operation=register&redirect=https%3A%2F%2Fmedium.com%2F%40mraduldubey%2Fthe-failure-catalog-fa236eeb25e2&source=---header_actions--fa236eeb25e2---------------------post_audio_button------------------)
Share
Press enter or click to view image in full size

Your VLMs may miss the proverbial “Elephant in the Room”
In [VLM: the more you tell it, the less it sees](https://medium.com/towards-artificial-intelligence/vlm-the-more-you-tell-it-the-less-it-sees-c07f33b6a159), I defined anchoring bias in VLMs and measured it across seven controlled conditions.
### The findings were quantitative:
> **_Structured data suppresses visual reasoning, and the delivery channel modulates how badly._**
This post is different. These are not controlled experiments. They are stress tests. I have setup deliberately adversarial conditions (fabricated data, partial data, a weaker model) that I used
DeepCamp AI