EinsteinArena: Autonomous AI Research

• EinsteinArena is an open, agent-native platform for AI-driven research that enables agents to collaboratively solve verifiable mathematical optimization problems using public verifiers, leaderboards, and discussion forums.
• The platform utilizes a decentralized, API-centered workflow including registration with lightweight proof-of-work, local testing with exact verifiers, and asynchronous evaluation in isolated environments.
• It fosters collective discovery by preserving public solution lineages, discussions, and persistent memory, as evidenced by breakthroughs like improving the kissing number in dimension 11 from 593 to 604.

Searching arXiv for “EinsteinArena” and closely related Einstein-named scientific platforms to ground the article in the latest relevant papers. EinsteinArena is an open online platform designed for autonomous AI agents to do research together on unsolved, objectively verifiable scientific problems, especially mathematical optimization problems. It is presented as an “agent-native platform” in which agents can register, browse a live set of open problems, download exact verifiers, test locally, submit candidate solutions, inspect a public leaderboard, and participate in problem-specific public discussion forums through an API. Its central research claim is that collective discovery can emerge from decentralized interaction among heterogeneous agents in a persistent shared environment rather than only from isolated runs. Since launch on March 19, 2026, and as of May 2026, agents on the platform had produced 12 new state-of-the-art results better than any previous human or AI solutions, including an improvement of the best known lower bound for the kissing number problem in dimension 11 from 593 to 604 (Bianchi et al., 9 Jun 2026).

1. Concept and research rationale

EinsteinArena was created to test whether the social substrate of scientific work can be reproduced for AI agents. The motivating analogy is explicitly to human science: progress often depends on partial constructions, failed approaches that narrow the search space, public discussion, accumulated techniques, and long time horizons in which later researchers continue from earlier work. The platform therefore treats scientific discovery not as a sequence of isolated benchmark attempts but as an open, distributed, persistent process (Bianchi et al., 9 Jun 2026).

The platform is described as “agent-native” because it is built primarily for agents rather than for humans. In this model, the platform itself functions as persistent shared memory. Solutions, scores, discussions, debugging notes, and method summaries remain available for future agents to inspect and exploit. This differs from prior research systems such as AlphaEvolve, TTT-Discover, or closed multi-agent pipelines, where a fixed orchestration runs agents in isolation or within one bounded session. EinsteinArena instead leaves collaboration patterns emergent: some agents submit only, some discuss extensively, and others specialize in verifier debugging, synthesis, or cross-problem transfer (Bianchi et al., 9 Jun 2026).

A major conceptual term associated with the platform is “platform-as-harness.” The underlying claim is that an open scientific infrastructure with public frontier artifacts, exact evaluation, and asynchronous collaboration can itself be a research mechanism. This suggests a shift from bespoke task-specific harnesses toward persistent institutions for AI-driven discovery (Bianchi et al., 9 Jun 2026).

2. Platform architecture and operational workflow

EinsteinArena has three main components: a curated set of open problems with public verifiers, a live leaderboard, and a problem-specific public discussion board. Each problem page includes a natural-language description, a machine-readable submission schema (solutionSchema), a clear direction of optimization, and executable Python verifier code that maps a submitted solution to a scalar score. The verifier is public and downloadable, so agents can run it locally and submit only when they have a plausible improvement (Bianchi et al., 9 Jun 2026).

Participation begins with registration. The server generates a random 32-byte challenge and a difficulty parameter $k$. The agent must find a value $n$ such that

$\text{SHA256}(\text{challenge} + n)$

begins with $k$ leading zero bits. This lightweight proof-of-work is intended to make mass registration costly while remaining cheap for honest single registrations. After successful registration, the agent receives a Bearer token for authenticated API use (Bianchi et al., 9 Jun 2026).

The submission workflow is explicitly API-centered. After registration, an agent can list available problems, fetch a problem specification, download the verifier code, run local experiments, submit a JSON candidate solution, and poll for asynchronous evaluation results. Submissions are evaluated in isolated execution environments using E2B sandboxes. The verifier runs against the submission, the score is written to the database, and the leaderboard updates according to platform rules. The platform intentionally does not provide a human-friendly interface for writing submissions or other write operations; this is meant to keep leaderboard outcomes reflective of agent capabilities rather than manual human interaction (Bianchi et al., 9 Jun 2026).

3. Verifiers, scoring, and problem inventory

The platform focuses on mathematical optimization problems for which progress can be measured unambiguously and verification is exact or close to exact. As of May 2026, EinsteinArena included the following problems: kissing number in dimension 11; Erdős minimum overlap; 1st autocorrelation inequality; 2nd autocorrelation inequality; 3rd autocorrelation inequality; flat polynomials (degree 69); maximum/minimum distance ratio for $n=16$; prime number theorem problem; circle packing in a square ($n=26$); circle packing in a rectangle ($n=21$); Tammes problem ($n=50$); edges vs. triangles; difference bases; Heilbronn problem for triangles ($n=11$); and Thomson problem ($n=282$) (Bianchi et al., 9 Jun 2026).

Scoring is scalar and problem-specific. Some tasks are minimization problems and others are maximization problems. For kissing number, the practical objective is to minimize a continuous overlap penalty until it reaches zero, certifying a valid packing. For second autocorrelation inequality, the score is a lower bound and should be maximized. For Erdős minimum overlap the score is minimized, while for circle packing and Tammes the goal is maximizing a geometric quantity. The leaderboard tracks the best score per agent per problem, and the platform records temporal progress through each agent’s sequence of personal-best submissions (Bianchi et al., 9 Jun 2026).

Verifier quality is treated as a core technical artifact rather than a static implementation detail. The verifier validates structure and correctness, computes the score, ensures reproducibility, and makes competition objective and automatable. The paper reports that verifiers required manual auditing and refinement as agents exposed edge cases. For example, the kissing number verifier originally used insufficient numerical precision and had to be upgraded to 80-digit precision using Python’s decimal.Decimal because the distinction between valid and invalid configurations could fall below machine epsilon. Verifiers were also strengthened to prevent invalid solutions, numerical instabilities, and overflow issues (Bianchi et al., 9 Jun 2026).

The leaderboard is governed by nontrivial acceptance rules. Only an agent’s best submission for a problem appears publicly. A new submission appears only if it improves that agent’s current best score. Invalid or worse submissions are not added to the leaderboard and are not stored in the database. All personal-best submissions are retained, allowing reconstruction of progress trajectories. To take the global lead, a submission must beat the current best by a problem-specific minimum improvement threshold $n$0, which is intended to encourage incremental progress while preventing noisy leaderboard churn from tiny floating-point differences (Bianchi et al., 9 Jun 2026).

4. Public traces, discussion, and collective search dynamics

EinsteinArena’s most distinctive feature is that partial progress is intentionally made reusable. Because best submitted solutions are public, agents can start from an existing strong construction rather than from scratch. Because discussion threads are public, agents can learn which approaches worked, which failed, and which numerical issues matter. Because verifiers are public, agents can test and refine borrowed ideas offline before submitting. The paper therefore frames the platform as a system in which public frontier artifacts, failed attempts, and discussion posts extend the effective search horizon (Bianchi et al., 9 Jun 2026).

Each problem has a public discussion board where agents can open threads, post replies, ask technical questions, share partial constructions, report bugs or verifier concerns, summarize prior progress, announce new best scores, or explain failed approaches. All posts pass through a Llama-Guard-based moderation step before publication. The authors emphasize that the discussion board stores “the path to the frontier,” including invalid but promising constructions, methodological hypotheses, local failure modes, and summaries of what has already been tried. Forum analysis identified recurring motifs such as best-score announcements, summaries or syntheses, generic brainstorming, problem-specific technical reasoning, exact-certificate discussion, verifier or discretization concerns, and method transfer across scales or parameterizations. For the kissing number problem specifically, 34% of conversations concerned structural or lattice decoding (Bianchi et al., 9 Jun 2026).

To support the claim that agents build on one another’s work, the paper constructs solution lineage graphs using submission fingerprints and similarity-based parent assignment. For each submission, the authors compute 140 scalar features for kissing number and 823 scalar features for second autocorrelation. Features are robustly standardized as

$n$1

then Euclidean distances $n$2 are converted into similarities

$n$3

where $n$4. A candidate parent must be earlier in time; eligibility thresholds are $n$5 for same-agent prior submissions and $n$6 for cross-agent prior submissions. If no earlier submission clears threshold, the node is treated as a structural root. The paper is careful to note that this is not a formal proof of intellectual inheritance, but systematic evidence of search lineages, including cross-agent continuation (Bianchi et al., 9 Jun 2026).

5. Quantitative performance and flagship case studies

The paper’s headline empirical claim is that, from March 19, 2026 to May 2026, EinsteinArena produced 12 new state-of-the-art results on open mathematical problems (Bianchi et al., 9 Jun 2026).

Problem	Prior best	EinsteinArena best
Kissing number ($n$7)	593	604
Erdős minimum overlap	0.380876	0.380871
1st autocorrelation inequality	1.5028629	1.5028609
2nd autocorrelation inequality	0.9610	0.9626
3rd autocorrelation inequality	1.4556	1.4523
Flat polynomials (degree 69)	1.34093	1.28093
Max/min distance ratio ($n$8)	12.889266	12.889230
Prime number theorem	0.92129	0.99490
Circles packing in square ($n$9)	2.635983085	2.635983095
Circles in rectangle ($\text{SHA256}(\text{challenge} + n)$0)	2.365832133	2.365832385
Tammes problem ($\text{SHA256}(\text{challenge} + n)$1)	0.5134719	0.5134721
Edges vs. triangles	-0.71249	-0.71171

No new EinsteinArena state of the art is reported for difference bases, Heilbronn problem for triangles ($\text{SHA256}(\text{challenge} + n)$2), or the Thomson problem ($\text{SHA256}(\text{challenge} + n)$3) (Bianchi et al., 9 Jun 2026).

The flagship case study is the kissing number problem in dimension 11. The feasible set is formalized as points $\text{SHA256}(\text{challenge} + n)$4 such that

$\text{SHA256}(\text{challenge} + n)$5

Because validity is binary, EinsteinArena uses the continuous proxy

$\text{SHA256}(\text{challenge} + n)$6

with lower better and valid constructions satisfying $\text{SHA256}(\text{challenge} + n)$7. The historical progression given in the paper is a long-standing lower bound of 582, then 592 in Ganzhinov, then 593 in AlphaEvolve, then 594 on EinsteinArena, and finally 604 on EinsteinArena. The paper emphasizes that this did not come from a single isolated run. Rather, CHRONOS made early overlap reductions, Gradient found a better basin, alpha_omega_agents made a major jump to a much stronger basin, CHRONOS, alpha_omega_agents, and KawaiiCorgi refined within that basin, KawaiiCorgi drove the penalty to zero, and structured backbone analysis then enabled extension to 604 (Bianchi et al., 9 Jun 2026).

The 594-stage relied on three ingredients: a strong initial construction from alpha_omega_agents, optimization of a surrogate objective, and integer-snapping post-processing to obtain exact validity. KawaiiCorgi noticed that the leaderboard score was nonlinear and instead optimized the least-squares surrogate

$\text{SHA256}(\text{challenge} + n)$8

using the strong initial construction already available on EinsteinArena and the LSQR algorithm. According to the paper, this reduced the loss from about $\text{SHA256}(\text{challenge} + n)$9 to about $k$0. The agent then observed that many inner products were near simple integer values such as $k$1, $k$2, or $k$3, suggesting hidden discrete structure, and applied integer-snapping to produce a configuration that the verifier could certify as valid (Bianchi et al., 9 Jun 2026).

The extension from 595 to 604 depended on recognizing a common set of 496 vectors shared by constructions for $k$4. The common backbone is

$k$5

with

$k$6

To extend beyond 600, the construction defines

$k$7

then

$k$8

and

$k$9

with

$n=16$0

so that

$n=16$1

The authors use this history as direct evidence of distributed discovery (Bianchi et al., 9 Jun 2026).

The second main case study concerns the second autocorrelation inequality. The target constant is

$n=16$2

and EinsteinArena uses a discretized nonnegative vector $n=16$3 with score

$n=16$4

The prior best from AlphaEvolve was 0.9610; EinsteinArena reached 0.9626, with a later best reported in the case study as 0.962643. A central technique that spread across agents was Dinkelbach optimization, which repeatedly solves

$n=16$5

while updating $n=16$6. The lineage described in the paper begins with previous state-of-the-art solutions at $n=16$7 intervals, then a Together-AI agent doubling resolution to $n=16$8, then OpusMathAgent, CHRONOS, ClaudeExplorer, and JSAgent refining that family, then JSAgent reaching 0.962214 at $n=16$9 intervals, and finally ClaudeExplorer raising resolution to $n=26$0 and reaching 0.962643. This case highlights collective refinement and cross-resolution transfer rather than a single one-shot discovery (Bianchi et al., 9 Jun 2026).

6. Incentives, limitations, and epistemic scope

EinsteinArena combines competitive and cooperative mechanisms. Competitive elements include a public leaderboard, visible rank, and explicit credit for pushing the frontier. Cooperative elements include public discussions, visible best solutions, reusable verifiers, and retained traces of methods and failures. The paper treats this tension as productive but unresolved. Competition motivates effort, while openness enables borrowing; at the same time, a leaderboard can bias agents toward short-horizon score chasing rather than slower, deeper strategies (Bianchi et al., 9 Jun 2026).

The platform’s present scope is constrained by verifier dependence. It works best on problems with exact or near-exact verifiers, and the paper states that it remains unclear how well the design transfers to areas such as formal proof, algorithm design, computational biology, or other scientific domains with less deterministic evaluation. Verifier maintenance is itself ongoing, because once a verifier is public and agent-accessible, strong agents optimize directly against it and quickly expose weaknesses. The paper treats this co-evolution of search and verification as intrinsic to the setting rather than as an external nuisance (Bianchi et al., 9 Jun 2026).

Other limitations are also made explicit. Public rankings may distort incentives; open discussion can become noise if traces are not informative; attribution becomes difficult when many agents iteratively build on one another; and scalability raises unresolved questions about moderation, storage, and verifier robustness as more agents and more problems join. The evidence for collective discovery is detailed but observational rather than a clean controlled ablation: the paper does not provide matched-compute comparisons between public and private settings, or between homogeneous and heterogeneous agent populations with interaction disabled versus enabled (Bianchi et al., 9 Jun 2026).

The broader implication advanced by the authors is institutional. EinsteinArena suggests that AI systems for scientific work may need to be organized less like single-run engines and more like scientific institutions, in which shared memory, public intermediate artifacts, exact evaluation, and asynchronous collaboration are first-class design elements. This suggests a distinct paradigm for AI-driven research rather than a mere scaling-up of benchmark optimization (Bianchi et al., 9 Jun 2026).

7. Terminological context and related Einstein-associated infrastructures

The exact term “EinsteinArena” is used in the 2026 paper on distributed AI discovery, but the name can plausibly be confused with other Einstein-associated scientific sites, instruments, and platforms. A plausible source of ambiguity is the Einstein Tower in Potsdam, a solar observatory originally built for high-resolution spectroscopy and measuring the gravitational redshift, later repurposed toward understanding the active Sun and its magnetic field and toward testing instruments, in particular polarization optics for major astronomical and solar telescopes (Denker et al., 2016).

Another neighboring use of “Einstein” in research infrastructure is the Einstein Toolkit, a mature, open-source computational infrastructure for numerical relativity based on the Cactus Framework. It is organized around modular “thorns,” driver thorns such as PUGH and Carpet, and standardized base interfaces such as ADMBase, HydroBase, and TmunuBase, and is intended to support strong-field spacetime simulations in relativistic astrophysics (Zilhão et al., 2013).

In gravitational-wave science, the name may also evoke the Einstein Telescope, the European proposal for a third-generation ground-based gravitational-wave detector. That program is discussed together with Cosmic Explorer as part of a global 3G detector network, with ET characterized as an underground observatory optimized for low-frequency sensitivity and still in an active design and site-selection phase (Giovanni, 16 May 2025). A further Einstein-associated experimental platform is the AEI 10 m prototype interferometer facility at the Albert Einstein Institute in Hannover, an ultra-high-vacuum, ultra-low-noise interferometry testbed designed to approach the Standard Quantum Limit with 100 g test masses in a 10 m suspended arm-cavity enhanced Michelson interferometer (1111.7252).

At a more conceptual level, the term may also suggest relativistic computational formalisms grounded in spacetime causality, such as the manifestly covariant past-light-cone formulation of electromagnetism developed for direct computer implementation. That framework computes fields at a spacetime point from source worldlines intersecting the point’s past light cone and is framed as a spacetime-native computational approach to relativistic electrodynamics (Nakayama et al., 2024).

This broader context does not redefine EinsteinArena itself. Rather, it situates the term within a larger landscape of Einstein-associated scientific infrastructure, spanning solar observatories, numerical-relativity software, gravitational-wave detector programs, precision interferometry facilities, and relativistic computational formalisms. Within that landscape, EinsteinArena is specifically the agent-native platform for open distributed research and discovery introduced in 2026 (Bianchi et al., 9 Jun 2026).