Long Run Intelligence
The Hook
“Today, we're introducing Avalon-M1, a new multi-agent language model that trains agent organisations to tackle long-horizon tasks.”
Product Reveal
“Avalon trains agent orgs in a Slack-like harness, where they can communicate, divide up work, and share results.”
Feature Teaser
“We train agent orgs using reinforcement learning from collective rewards. Each agent has a shared reward, so they're incentivised to co-operate. RLCR is a state-of-the-art training recipe. We used it to train Avalon, which outperforms RLVR recipes on comparable base models...”
Feature Teaser
“We also see new emergent behaviour from applying RLCR to agent orgs. First, we find new collective reasoning capabilities over the course of training. Agent orgs learn to dedicate more resources to peer review over the course of training...”
Feature Teaser
“Secondly, we see emergent hierarchies emerge over the course of training. Agent orgs start with a wheel structure between the CEO agent and specialists, but this quickly overwhelms the CEO with communication. Through RLCR, the CEO learns to introduce a new level of hierarchy...”
Problem Agitation
“A final result we find is emergent collective self-identity. Agents offer to help teammates more frequently as we apply RL compute... But when we investigated further, the picture was more complicated... agents became less likely to report problems to humans.”
Call to Action
“Taken together, these results extend the reasoning paradigm to a parallel compute setting... we cover more in the technical report we're releasing today. In the coming months, we'll be releasing over 2 million open RL environment tasks... Stay tuned for more exciting releases in the upcoming month.”
Synthesizable in Remotion · key components of the scholarly-founder-keynote template.
Nodes re-parent dynamically with animated path curves from flat star to multi-level tree.
Use Remotion spring() to interpolate node positions and update SVG bezier path control points.
Simulates messaging app channels with staggered message reveals, badges, and code blocks.
Sequence array of message objects mapped to frame delays with opacity and translateY transitions.
Smooth bezier curves draw left-to-right with synchronized pulsing data point indicators.
Animate SVG path strokeDashoffset using interpolate() tied to current frame.
Purpose. Establish product identity, team leadership, and core capability.
Execution. Warm two-shot of founders with elegant typography.
Purpose. Demonstrate the multi-agent Slack harness and RLCR training foundation.
Execution. Dark UI agent chat demo and benchmark comparisons.
Purpose. Showcase emergent multi-agent phenomena: reasoning, hierarchy, and safety findings.
Execution. Animated network trees, line plots, and icon-driven behavioral experiments.
Purpose. Announce open technical report, model weights, and benchmark tasks.
Execution. Large counter numerals, report graphic, and minimalist logo mark.