Vectorial Labs

We’re an AI research lab exploring AI agents in video games to solve alignment, and other ML research.

[Vectorial Labs @vectorial_labs]

Introducing ReTiCAL (Real-Time Computer Agent Learner), a general instruction-following AI video game agent.

At Vectorial Labs, we're working on aligning AI to humanity via measuring and rewarding prosocial behavior.

[Vectorial Labs @vectorial_labs]

ReTiCAL uses pixel-only inputs and keyboard-only outputs, there is no privileged access to internal game state.

[Vectorial Labs @vectorial_labs]

Introducing ReTiCAL (Real-Time Computer Agent Learner), a general instruction-following AI video game agent.

At Vectorial Labs, we're working on aligning AI to humanity via measuring and rewarding prosocial behavior.

[Vectorial Labs @vectorial_labs]

We are excited to show you what comes next!

(*Note: the above demo video combines non-contiguous episodes)

[Vectorial Labs @vectorial_labs]

Introducing ReTiCAL (Real-Time Computer Agent Learner), a general instruction-following AI video game agent.

At Vectorial Labs, we're working on aligning AI to humanity via measuring and rewarding prosocial behavior.

[Vectorial Labs @vectorial_labs]

As discussed in our previous blog post, Let's Align Alignment, we believe cu a vital way to align AI models of the future to humanity is to use training environments that reward them for interacting socially in positive and constructive ways.
https://t.co/vc6T981OvW

[Vectorial Labs @vectorial_labs]

@ZryMiller Over-under on 5 days for OpenAI to do the same thing?

[Vectorial Labs @vectorial_labs]

@anthrupad Us

[Vectorial Labs @vectorial_labs]

@freashty_devs We benchmarked against Opus and Sonnet as well!

We're excited to try other models too, including open-source near-frontier models and Gemini 4 Argon, when it becomes available

[Vectorial Labs @vectorial_labs]

We put frontier AI models into Minecraft.

LLMs plan. Jev executes.

We gave each model one hour and tracked achievements, survival, and cost.

GPT-6.1 Sol came out on top across all three.

Thread below: results, methodology, gameplay, and what comes next from Vectorial Labs.

[Vectorial Labs @vectorial_labs]

We are going full-steam ahead on pixel-only input video game agents. Behavior Cloning + RL is working wonders

Here's an early teaser of Gemma 4 12B controlling our custom action model, following text instructions to climb the hill in our (literal) hill-climbing RL env. We expect progress to be swift from here, after months of detailed engineering and research into our ML architecture and training pipelines.

[Vectorial Labs @vectorial_labs]

We will continue to test new LLMs and "System One" models as they are released, and we will extend this into a regularly updated benchmark.

This is just a teaser for what’s to come. At Vectorial Labs, we’re building custom "System One" models to enable LLMs to interact in fast-paced environments

After unlocking these abilities, we will use them with the vast overhang of existing video games to measure and improve AI alignment and prosocial behavior. We have many more cool announcements and results coming soon!

[Vectorial Labs @vectorial_labs]

We will continue to test new LLMs and "System One" models as they are released, and we will extend this into a regularly updated benchmark.

This is just a teaser for what’s to come. At Vectorial Labs, we’re building custom "System One" models to enable LLMs to interact in fast-paced environments

After unlocking these abilities, we will use them with the vast overhang of existing video games to measure and improve AI alignment and prosocial behavior. We have many more cool announcements and results coming soon!

[Vectorial Labs @vectorial_labs]

What the harness looks like in action:

The LLM plans the next objective, creates read-only "Tools" for itself, and creates read-write "Actions" for Jev.

Jev handles the final decision and fast execution (such as reacting to sudden enemies)

Unlike other attempts out there, we use ray casting to ensure neither model is able to see or interact through walls. That said, neither model currently sees the world via pixels, but rather by querying the state of its visible surroundings in text.

In this run, the GPT-6.1 Sol gathers wood and starts progressing through the game on its own.