Artificial Analysis launched on November 6 the VIALS-AA, an implementation of the VIALS benchmark from Handshake AI. The test evaluates whether language models can interpret scientific images and answer questions about the evidence they contain. The main result is straightforward: models take different paths to reach the correct answer and often get it wrong even after carefully inspecting the images.
VIALS-AA: Benchmark for Reasoning on Scientific Images
Artificial Analysis launched VIALS-AA, an implementation of Handshake AI's VIALS benchmark, which tests whether models can interpret scientific images and answer based on the evidence.
Marcelo Santiago
Criador do bridgecode
The biggest AI newsletter in the World
Get AI articles in your inbox
Technical curation on models, agents and coding tools. No spam, unsubscribe anytime.
The benchmark does not measure only final accuracy. It records the model's trajectory: how many times it inspects the image, which regions it observes and at what point it decides to answer. This makes it possible to compare not only who gets more right, but who reasons more efficiently.
In this article you will see:
- What VIALS-AA is and where it came from.
- What the benchmark measures and how it scores.
- Why models still make mistakes after inspecting images.
- How to access the public leaderboard.
- How to test today.
What is VIALS-AA
VIALS-AA is Artificial Analysis' implementation of the VIALS benchmark created by Handshake AI. The dataset and original evaluation are described on Handshake AI's page. Artificial Analysis adapted the benchmark for its public leaderboard, with documented methodology.
The focus is reasoning about scientific images. Instead of pure text questions, the model receives an image with visual data, such as charts, diagrams or experiment photographs, and needs to answer questions about the evidence present.
What the benchmark measures
VIALS-AA measures whether the model can extract information from a scientific image and reach the correct answer. The score considers final accuracy and path efficiency: models that inspect the image many times unnecessarily lose points on efficiency.
The public leaderboard shows results by model, with the run date. The launch version brings results from November 6, 2026. For current results, the leaderboard is the source.
Why models still err even after inspecting images
The central finding of the benchmark is that models often arrive at the wrong answer even after examining the image carefully. This suggests the problem is not just seeing the image, but interpreting visual evidence with the right context.
Different models take different paths. Some reach the correct answer with few inspections. Others inspect a lot and still err. This difference in trajectory is what the benchmark exposes, not just the final score.
How to access the public leaderboard
The VIALS-AA leaderboard is available at artificialanalysis.ai/evaluations/vials-aa. The methodology is documented at artificialanalysis.ai/methodology/intelligence-benchmarking. The original Handshake AI benchmark is at joinhandshake.com/research/benchmarks/vials.
These links allow you to verify the numbers and methodology without relying on third parties. It is a public leaderboard, with reproducible data.
How to test today
The simplest flow is: access the leaderboard, compare the models you use with the leaders, and check if your model appears. If it does not appear, the documented methodology allows you to reproduce the evaluation on your own images.
For anyone building agents that analyze documents or reports with charts, VIALS-AA is a direct signal: whether the model can read visual evidence before making a decision.
Conclusion
VIALS-AA exposes a real gap: models that see the image but do not interpret the evidence. For anyone building agents that read visual documents, this benchmark is a practical and public test.
Instead of just following news like VIALS-AA, put AI to work for you at bridgecode.
Source: Artificial Analysis
Image and references
See also statebench long context state tracking, featureloop benchmark agents code requirements evolution, statebench long context state tracking and featureloop benchmark agents code requirements evolution.
Frequently asked questions about VIALS-AA
What is VIALS-AA?
Artificial Analysis's implementation of Handshake AI's VIALS benchmark, which tests reasoning about scientific images.
What does the benchmark measure?
Whether the model extracts information from a scientific image and reaches the correct answer, considering accuracy and trajectory efficiency.
Why do models still make mistakes even after inspecting images?
The problem is not just seeing the image, but interpreting visual evidence with the right context.
Where can you see the results?
On the public leaderboard at artificialanalysis.ai/evaluations/vials-aa.
Is the methodology public?
Yes, documented at artificialanalysis.ai/methodology/intelligence-benchmarking.
Additional context
This post is part of bridgecode's daily radar, which tracks releases, benchmarks and AI tools for developers. Follow the next editions for more technical analyses and comparisons. The analysis above is based exclusively on official sources and public leaderboards, with no invented data. Always check the original source before making technical decisions. For teams adopting AI in development, the next step is to measure impact in a real workflow and compare it with what existed before the new release. Also follow other blog posts about benchmarks and AI tools to broaden your decision making repertoire. When a new model appears, the practical question is not just the score, but whether it solves your use case with acceptable cost and latency. Real workflow tests with data from your domain often reveal differences that benchmarks do not capture. For that reason, use benchmarks as an initial filter and validate with your own experiments before migrating. bridgecode tracks these changes so you do not need to monitor dozens of sources manually. Model choice is an engineering decision, not just marketing. Document the criteria, measure the outcome and revise when new data appears.
Tags:#radar#ia
Developer and creator of bridgecode, an agentic development environment for orchestrating squads of AI agents. Helps people going from vibe coding to professional development move from idea to shipped product.
Related articles
Benchmarks
PLE: physical engineering benchmark for AI agents
Google Cloud launched the Physical-Engineering Last Exam, an 85 task physical engineering benchmark. Every model scored below 50%. See the numbers.
7 min
Benchmarks
Inkling 2 Small: benchmark shows strength in enterprise tasks
Thinking Machines Lab announced Inkling 2 Small with 594 billion parameters and 17 billion active. The model scored 71.8% on EnterpriseBench-CoreCraft and 46.8% on EnterpriseOps-Gym.
5 min
Benchmarks
OpenAI introduces BioBasicBench: an AI test for biology
OpenAI launches BioBasicBench, a task set to measure AI agents in biology scenarios. What the benchmark evaluates and how it compares to existing sets.
5 min