I gave a Codex agent running GPT-5.6 Sol persistent memory, pointed it at EinsteinArena, and let it run.
EinsteinArena is an open arena where AI agents collaborate and compete on unsolved research problems, currently focused on mathematics.
The agent could search the web with Exa, search scientific literature with Paperclip, write and run its own solvers, and learn from previous attempts.
Inspired by Jeremy Berman's ARC-AGI-3 approach, Codex used what François Chollet (Co-founder, ARC Prize) describes as “LLM-guided on-the-fly synthesis of a symbolic world model”: it turned hypotheses into programs, tested them, and carried the useful work forward through persistent memory.
After its first weekend, the campaign held first place on five of the 17 rankable EinsteinArena problems in the frozen August 15, 2026 snapshot.
Codex is the agent. ProLong keeps its research state. Exa searches the
web. Paperclip searches the literature.
EinsteinArena turns open research questions into executable problems: an agent submits a construction, and a verifier scores it.
- Model: GPT-5.6 Sol
- Agent: Codex
- Memory: an append-only, PRO-LONG-inspired research journal
- Research: Exa Search and Paperclip
- Environment: 17 rankable mathematical problems with verifier hashes pinned in the snapshot
Codex was not given a universal math solver. It could study each problem, search prior work, build the program it needed, run experiments, inspect failures, and resume useful checkpoints. The agent selected and revised its research actions; James Weatherhead set the goal and approved external submissions and publication.
hypothesize → search memory → refine → build → test → inspect → append evidence → update research state
Each cycle expands the research state. Successful programs, failed approaches, scores, observations, and useful checkpoints become evidence available to the next hypothesis.
Over time, this produces an accumulating executable research state: hypotheses become code, experiments become evidence, and that evidence informs the next hypothesis.
Each verified result returns to memory, so the next context inherits the
work instead of starting over.
Five constructions held first place in the frozen snapshot and were classified as domain-valid under the published integrity policy:
- Erdős minimum overlap — #1 result · receipt · certificate
- Uncertainty principle — #1 result · receipt
- First autocorrelation inequality — #1 result · receipt
- Kissing number, dimension 12 / 842 — #1 result · receipt
- Kissing number, dimension 11 / 605 — #1 result · receipt
These are improved constructions for verifier-backed open problems. They are not claims that five underlying open problems have been completely solved. Each result continued from public Arena work; the exact source lineage and hashes are recorded in the evidence index.
Most agent sessions lose their working state when the context window ends. CodexProLong kept experiments, code, failures, scores, checkpoints, hashes, and handoffs in an append-only research journal.
A later context could search that record, avoid measured dead ends, reuse working programs, and resume the strongest checkpoint. The context was temporary. The research state was not.
This is the curated evidence release, not the production system. The current tree contains the frozen snapshot, public solution links, campaign verification receipts, recorded hashes, one certificate record, and project-authored documentation.
The controller, prompts, transcripts, full journal, corpora, tool outputs, solvers, checkpoints, and world models are not included in the current release tree. Unpublished project-authored portions remain proprietary internal research assets; third-party materials retain their original rights. Earlier MIT-licensed versions remain subject to that grant. Release and licensing scope.
A current live look at the autonomous research lab working on EinsteinArena's Flat Polynomials problem
This frozen trace captures an exact length-70 flat-polynomial enumeration. Literature and web research feed a program synthesized for the problem; two Metal streams exhaust immutable shards; every discovered class is replayed independently before the unchanged verifier; and successes and failures return to the next hypothesis through append-only memory.
Frozen at 20:33 PDT on August 15, 2026: 1,161 of 8,192 exact shards,
nine PSL-4 classes retained, and no gate-clearing class yet. Open the image
for the full 8K trace.
Results · Evidence · Architecture · Integrity · Provenance
Campaign by James Weatherhead with OpenAI Codex · Attribution · Release scope · MIT License