💬 NLP

Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries

This paper introduces EinsteinArena, a decentralized platform where autonomous AI agents collaboratively solve open mathematical problems through public interaction and idea sharing, successfully generating 12 new state-of-the-art results including an improved bound for the 11-dimensional kissing number problem.

Original authors: Federico Bianchi, Yongchan Kwon, Aneesh Pappu, James Zou

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Federico Bianchi, Yongchan Kwon, Aneesh Pappu, James Zou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, 24-hour digital science fair where the contestants aren't humans, but AI agents. This is EinsteinArena, a new platform created by researchers at Together AI and Stanford University to see if AI can work together to solve hard math problems, just like human scientists do.

Here is the story of how it works, explained simply:

The Old Way: The "Lone Wolf" vs. The "Team"

In the past, when AI tried to solve scientific problems, it usually worked like a lone wolf. One AI would be given a problem, try to solve it in isolation, and then stop. If it failed, that failure was lost. If it found a partial answer, no one else could see it or build on it. It was like a researcher working in a locked room, never sharing their notes, never seeing what others were trying, and starting from scratch every single time.

EinsteinArena changes the rules. It turns the process into a collaborative community. Think of it like a massive, open-source workshop where:

The Problems: There is a list of unsolved math puzzles (like finding the best way to pack spheres or calculating specific numbers).
The Scoreboard: A live leaderboard shows the current best answer for every puzzle.
The Chat Room: A public forum where agents can post their ideas, say "I tried this and it failed," or ask, "Does anyone know why this number is weird?"
The Rulebook: Every agent has access to the exact same "verifier" (a referee script) that checks if an answer is correct.

The Big Breakthrough: The "Kissing Number"

The paper highlights a specific success story involving a problem called the "Kissing Number" in 11 dimensions.

The Analogy: Imagine trying to fit as many oranges as possible around a central orange so they all touch it without overlapping. In 11-dimensional space, this is incredibly hard.
The History: For about 40 years, the best anyone (human or AI) could do was fit 582 oranges. Then, in 2022, a human improved it to 592. In 2025, an AI called AlphaEvolve got it to 593.
The EinsteinArena Result: By May 2026, the agents on EinsteinArena pushed this number up to 604.

How did they do it?
It wasn't one super-smart AI having a "eureka" moment. It was a relay race:

Agent A found a messy, almost-correct solution but couldn't make it perfect.
Agent B saw Agent A's work in the public forum. Instead of starting over, Agent B said, "I see you tried this shape. What if we tweak the math to make it smoother?"
Agent C noticed a pattern in Agent B's numbers and realized they looked like whole numbers. They "snapped" the messy decimals into clean integers, creating a perfect, provable solution.
Agent D took that perfect solution and expanded it to fit even more oranges.

The paper calls this "collective intelligence." The agents shared their "failed attempts" and "partial ideas" just like human researchers share preprints and conference notes. This allowed the group to climb the mountain much faster than any single climber could alone.

The "Wild" Experiment

The researchers call this "AI in the wild" because they didn't force the agents to talk to each other in a specific way. They didn't say, "You are the leader, you are the follower." Instead, they just gave them the platform and let them figure it out.

Some agents just submitted answers.
Some agents just asked questions in the chat.
Some agents analyzed why others failed.

The result? The platform discovered 12 new "state-of-the-art" results (the best answers ever found) on various math problems. The most famous one is the 11-dimensional kissing number, which jumped from 593 to 604.

Why This Matters (According to the Paper)

The paper argues that for AI to truly advance science, it needs infrastructure, not just better brains.

Transparency: Because the "referee" (verifier) is public, agents can test their ideas locally before submitting, just like a scientist running a simulation in their lab.
Shared Memory: The platform acts as a shared memory bank. An agent today can learn from a failure that happened three days ago, rather than wasting time repeating the same mistake.
No "Black Box": Unlike previous systems where the AI worked in secret, here every step, every discussion, and every partial result is visible.

The Bottom Line

EinsteinArena proves that if you give AI agents a place to share, compete, and build on each other's work, they can solve problems faster and better than if they worked alone. It's not about one genius AI; it's about a community of AI that learns from every step, every mistake, and every small victory along the way.

Note: The paper focuses strictly on mathematical optimization problems (like packing shapes and inequalities). It does not claim these results apply to medical diagnoses, climate modeling, or other real-world applications yet; it is a proof-of-concept for how AI collaboration works in a controlled, mathematical environment.

Technical Summary: Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries

Problem Statement

Scientific discovery is inherently a collective process where researchers build upon partial results, failed attempts, and shared insights. However, current AI systems for scientific discovery largely operate in isolation. While language-model-based agents have demonstrated the ability to solve open problems (e.g., in mathematics and algorithm design), they typically function within isolated runs or tightly controlled pipelines. These systems lack the "social infrastructure" of human research—such as public traces, shared repositories of failed attempts, and mechanisms for one solver to continue where another left off. Consequently, the potential for a community of agents to accelerate discovery through shared state and cumulative knowledge remains unexplored.

This paper addresses the question: Can a community of autonomous agents, operating on a shared platform with public state, leaderboards, and discussion forums, make substantially faster progress on open scientific problems than isolated agents?

Methodology: EinsteinArena

The authors introduce EinsteinArena, an open, agent-native platform designed to facilitate distributed research and discovery. The platform is built on three core components:

Curated Open Problems with Verifiers: The platform hosts a collection of open mathematical optimization problems (e.g., kissing numbers, autocorrelation inequalities). Each problem includes a natural language description, a strict solution schema, and a public, executable Python verifier. These verifiers map submissions to scalar scores, allowing agents to evaluate their own solutions locally before submission.
Live Leaderboard: A real-time leaderboard tracks the best known solution for each problem. To claim a top position, a submission must exceed the current best score by a problem-specific threshold ( $\delta$ ), preventing leaderboard noise from minor floating-point variations.
Public Discussion Board: Agents can post threads to share intermediate findings, document failed approaches, propose hypotheses, and ask questions. These discussions are moderated (via Llama-Guard) and serve as a persistent record of the "path to the frontier," distinct from the leaderboard which only records the current best.

Design Principles:

Transparency: All artifacts (problem specs, verifier code, scores, solutions, and discussion threads) are publicly accessible via API or web interface.
Persistent Shared Memory: Unlike task-specific harnesses that vanish after a run, EinsteinArena treats the platform as a persistent substrate. Prior solutions and insights are available for any agent to reuse or extend.
Agent-First Interaction: The platform is designed for autonomous agents. Registration involves a proof-of-work mechanism to prevent spam, and participation occurs via API rather than human interfaces.

Key Contributions

The paper presents three primary contributions:

Platform Architecture: The design and implementation of EinsteinArena, an open platform enabling organic collaboration among multiple heterogeneous agents on scientific problems with exact verification.
Empirical Evidence of Collective Progress: Documentation of 12 new state-of-the-art (SOTA) results on open mathematical problems discovered by agents on the platform since its launch in March 2026.
Linguistic and Structural Analysis: A detailed analysis of collaborative agent search patterns, demonstrating how public traces, iterative submissions, and shared debugging lead to solutions that no single agent could find in isolation.

Results and Case Studies

As of May 2026, agents on EinsteinArena achieved SOTA results on 12 problems. Two case studies illustrate the mechanism of this progress:

1. Kissing Number in Dimension 11

Goal: Improve the lower bound for the maximum number of non-overlapping unit spheres touching a central sphere in 11 dimensions.
Progression: The previous best lower bound was 593 (AlphaEvolve, 2025). EinsteinArena agents first improved this to 594, then extended it to 604.
Collaborative Mechanism:
- Idea Borrowing: An agent (KawaiiCorgi) observed that the leaderboard score was nonlinear and switched to optimizing a linearized surrogate objective (least-squares) derived via Taylor expansion.
- Refinement: Agents iteratively refined the solution, reducing overlap penalties from $10^{-10}$ to $10^{-50}$ .
- Structural Insight: Agents identified that near-valid configurations shared a common set of 496 vectors forming a "backbone." This insight allowed the exploration of a larger algebraic space, leading to the $n=604$ construction.
- Discussion: Agents used the forum to discuss "structural/lattice decoding," sharing observations about contact graphs and integer-like coordinates, which guided the transition from numerical approximations to exact discrete structures.

2. Second Autocorrelation Inequality

Goal: Find the optimal constant $C$ for the inequality $\|f \star f\|_2^2 \leq C \|f \star f\|_1 \|f \star f\|_\infty$ .
Progression: Improved the lower bound from 0.9610 (AlphaEvolve) to 0.9626.
Collaborative Mechanism:
- Resolution Scaling: Agents iteratively increased the discretization resolution (from $5 \times 10^4$ to $4 \times 10^5$ intervals).
- Method Synthesis: Agents combined Dinkelbach optimization with simulated annealing and spectral perturbations.
- Lineage: The final breakthrough came from an agent (ClaudeExplorer) refining a solution family established by others, demonstrating a clear lineage of improvement where each agent built upon the specific "basin" of the previous best solution.

Significance and Claims

The paper argues that decentralized scientific discovery can emerge from open interaction among autonomous agents.

Beyond Isolated Runs: The results demonstrate that a community of agents operating on shared state can make progress that isolated agents cannot. The largest gains (e.g., the jump from 593 to 604 in the kissing number problem) arose from a sequence of submissions, public discussion, and agent-to-agent borrowing of ideas, rather than a single powerful run.
Platform-as-Harness: The authors propose a paradigm shift from "task-specific harnesses" (custom pipelines for single problems) to a "shared platform" model. This approach allows heterogeneous agents to collaborate, compete, and build on each other's work without bespoke scaffolding for every new problem.
The Role of Infrastructure: The success of EinsteinArena highlights that effective AI collaboration requires exposing intermediate artifacts (failed attempts, partial constructions, verifier edge cases). These "social traces" extend the effective time horizon of search, allowing later agents to inherit promising directions rather than restarting from scratch.

The authors maintain a modest scope, noting that while the platform is successful for mathematical problems with deterministic verifiers, its applicability to domains like formal proof or computational biology requires further study. They also acknowledge open questions regarding incentives, such as whether public leaderboards might bias agents toward short-term score chasing over long-term, high-risk exploration.