Confluence Labs

Roadmap & Position in Foundation Models

Builds foundation models that learn from minimal data, achieving SOTA on ARC-AGI-2 at $11.77/task.

Company Overview

Builds foundation models optimized for learning efficiency, enabling AI systems to rapidly adapt to new tasks with minimal data, particularly in data-sparse scientific fields. Achieved state-of-the-art on ARC-AGI-2 benchmark with an open-source solver (97.9% SOTA at $11.77/task).

What They're Building

The company's public product roadmap & what they're committed to building.

Open-sourced ARC-AGI-2 solver (97.9% SOTA at $11.77/task) on GitHub. Approach uses LLMs to write code describing transformations, structured to optimally resemble training data, enabling long-horizon work. Focus on data-sparse domains: hardware engineering, drug design, physics research. Approaching from two angles: hypothesis generation and experiment design automation.

Latest Intelligence

Zeitgeist tracks private signals to determine where the company is heading strategically.

Inference-Time Learning Targets Sparse Science

May 13, 2026

Confidence:

High

New Intel: Confluence Labs is applying learning-efficient foundation models to hardware, drug design, and physics after strong ARC-AGI-2 results. The bet is that sparse scientific domains need models that adapt from minimal data.

Founder and Key Execs

Brent Burdick, Co-Founder

Co-Founder

Niranjan Baskaran, Co-Founder

Co-Founder

Founder Force Multiplier

Brent Burdick is a self-taught engineer who left college in 2022 to focus on learning efficiency. Open-source, benchmark-topping ARC-AGI-2 solver demonstrates technical depth most stealth-stage startups cannot match.

Funding History

2025 | Brent Burdick and Niranjan Baskaran co-found Confluence Labs.
2026 | Accepted into Y Combinator W26 batch.
2026 | Open-sources ARC-AGI-2 solver with 97.9% SOTA at $11.77/task.
2026 | Recruiting interdisciplinary collaborators (hardware engineers, biologists, materials scientists).

Competitors

Foundation Models

Mistral, Cohere, AI21 Labs.

Efficient Fine-Tuning

Together AI, Predibase, Lamini.

Scientific AI

Recursion, Isomorphic Labs.

Reasoning/AGI

Numenta, Liquid AI, Ndea (YC W26).

Confluence Labs

's Moat:

Open-source ARC-AGI-2 solver at 97.9% demonstrates the learning efficiency approach at public benchmark level. The moat is methodological: few-shot adaptation to new tasks with minimal data is a fundamentally different capability than scaling pre-training, and Confluence's architecture encodes that advantage.

How They're Leveraging AI

Continual learning adaptation

Rapidly adapts foundation models to new scientific domains using minimal experimental data, accelerating discovery in biology and materials science.

Test-time task adaptation

Applies test-time training and refinement loops so models can solve novel abstract reasoning tasks on the fly without retraining from scratch.

Multi-agent orchestration

Orchestrates specialized AI sub-agents that collaborate to solve complex scientific problems where no single model has sufficient training data.

AI Use Overview:

Using continual learning adaptation that retains knowledge across domains, test-time task adaptation for zero-shot transfer, and multi-agent orchestration for complex reasoning.

More Similar Companies

Entire

Git-native AI code explainability and session context capture

The ex-GitHub CEO is building the compliance layer for AI-generated code, with personal relationships to every enterprise buyer who will need it.

Pinecone

Managed vector database and knowledge infrastructure for production AI apps.

A category winner pitch rests on Pinecone turning vector search into the default memory layer for RAG, agents, and enterprise knowledge apps.

Approxima

Lets product teams go from idea to deployed software in under an hour with AI agents.

Most AI coding tools target greenfield features. Approxima goes after the unglamorous maintenance work (bug fixes, incremental updates) that eats 60%+ of engineering time, with sandbox validation that lets agents merge to production without human review.

21st Labs

Helps developers ship AI apps 10x faster with purpose-built components and agent tools.

AI coding tools need a trusted component layer to ship production-ready UI, and their 1.4M developer distribution gives them a head start before Vercel or GitHub bundle one in.

Back To All Companies >