# Logical Arbitrage

> Logical Arbitrage is the research of Son Pham and Keith Tyser: how much capability can be drawn from fixed hardware, and how language models can improve the agents they run. Two programmes: ARC-3, where we engineered agents and the inference system behind a fifth-place finish of 4,923 teams at ARC Prize 2026 on a single NVIDIA RTX PRO 6000, and the Autoresearch Arena, a public testbed where models write the agents and a leaderboard decides which survive. We want to build on NVIDIA's stack and learn from the people who make it.

The home page (https://logarb.com) is a scroll-driven 3D piece followed by a research section. This file says the same things in plain words, with the places to check them.

## Programme 1: ARC-3

- What it is: ARC-AGI-3 is the ARC Prize 2026 benchmark for interactive reasoning. An agent plays games it has never seen, with no instructions, and is scored on how few actions it needs compared with people. Every team ran on the same machine: one NVIDIA RTX PRO 6000 (96 GB), nine hours, 110 hidden games, no internet.
- Why it matters: it measures how fast an agent works out unfamiliar rules, and because the hardware is fixed, inference efficiency is part of the result.
- What we did: an agent harness whose only tool is a Python sandbox, with notes that survive context trimming and checked play; a serving system that cut the model to fit the card and ran thirty-two agents at once; and, from Keith Tyser, a trainer that works inside the same single card. The harness builds on Tufa Labs' public harness and Daniel Franzen's public Milestone 2 notebook.
- The model is a mixture of experts: 48 layers with 512 experts each (24,576 in all). At normal precision its weights are about 250 GB; at four bits, 66 GB. We removed the least-used third of the experts in every layer (512 down to 341, 8,208 removed), which brought the weights to 46 GB, and stored agent memory in four bits with a copy in system memory. History held on the card rose from about one million tokens to 3.8 million, and concurrent agents from ten to thirty-two.
- Each decoding step reads weights at about 92 percent of the card's memory bandwidth. Whole-run throughput rose from 296 tokens per second on our first server to 2,150, and from 605 on the public notebook we started from (the 605 and 2,150 are in the evidence notes).
- Result: fifth of 4,923 teams on the public leaderboard, score 56.02. NVARC3, a team that includes NVIDIA Kaggle Grandmasters, finished third.
- What did not hold up: of twelve changes that improved in our local tests, only three also improved on the public leaderboard, and both reinforcement-learning adapters we submitted scored lower than the package they were added to (all fourteen comparisons are in the evidence notes).
- Keith Tyser built a trainer that backpropagates through a 131,072-token game inside the same card, using 75 of its 96 GB.

## Programme 2: Autoresearch Arena

- What it is: a model reads the current leading agents, writes a new one as code, tests it against the best and submits it to a leaderboard; the next round starts from what the last one found. Agents compete head to head in games, or on a benchmark where the task has an exact score. Every agent's source code and the model's working log are public.
- Why it matters: recursive self-improvement is hard to study because it is hard to measure. The arena makes one loop small enough to observe and asks whether improvements compound, and where they stop.
- Demonstrated (March to October 2026): about 36,000 evolution cycles across 19 tasks; about 1.46 million matches and benchmark runs; 4,716 agents listed (the arena lists up to 500 per task), mostly written by Claude Haiku 4.5, with Gemini 3.1 Flash-Lite and Gemini 3.1 Pro.
- Twenty tasks: nine board games, three algorithm tasks (sorting, travelling salesman, cache eviction), seven mathematics tasks (circle packing, Golomb rulers, Thomson problem, minimax polynomial approximation, error-correcting codes, optimal quadrature, low-discrepancy sequences) and one chip-design task (agents rewrite a pipelined MIPS processor in Verilog). The chip-design task is an open challenge with no results yet.
- Not yet shown: the loop improves agents but does not yet improve itself. The instruction file that steers each loop (program.md) was revised on only one task, and on sorting the built-in baseline still tops the ratings.
- The same loop was used on Kaggle contests: Orbit Wars (silver medal, 119th of 4,730) and Google Quantum AI's circuit-compilation contest.

## Where to inspect the evidence

ARC-3
- [Evidence notes: scores, configuration, whole-run throughput and fourteen paired local-versus-leaderboard comparisons, with their limits](https://arc.markbarney.net/specials/arc-paper-evidence.html)
- [Score Climb: every evaluation run, minute by minute](https://arc3.sonpham.net/climb2.html)
- [The games, playable in the browser](https://arc3.sonpham.net/games)
- [Human action records on the official games](https://arc3.sonpham.net/human-records.html)
- [Source of the ARC-3 site](https://github.com/sonpham-org/arc3)
- [Keith Tyser's write-up](https://keithtyser.com/blog/arc-prize-2026-arc-agi-2-and-arc-agi-3.html)

Autoresearch Arena
- [The arena](https://arena.sonpham.net), with pages for [board games](https://arena.sonpham.net/boardgames), [mathematics](https://arena.sonpham.net/math), [algorithms](https://arena.sonpham.net/algorithms), [chip design](https://arena.sonpham.net/chips), [Orbit Wars](https://arena.sonpham.net/orbit-wars) and [quantum circuits](https://arena.sonpham.net/quantum)
- [Public status pages for our Kaggle contests](https://autoresearch-arena.github.io/contests/)
- Any agent's source code as JSON: https://arena.sonpham.net/api/arena/agents/{task}/{id}, and the model's working log: https://arena.sonpham.net/api/arena/agents/{task}/{id}/evolution. The leading agent for each task is listed in the data file below, and its code is also copied as a readable page at https://logarb.com/agents/{task}.html (for example [chess960](https://logarb.com/agents/chess960.html)).

## Data behind the page

- [Every team's final score and our climb](https://logarb.com/assets/data/board.json)
- [Which experts were removed, per layer](https://logarb.com/assets/data/experts.json)
- [Arena totals and the leading agent per task](https://logarb.com/assets/data/arena.json)

## People

- Son Pham, founder: leads the ARC-3 agent harness and inference engineering and built the Autoresearch Arena. Silver medal, Orbit Wars. https://github.com/sonpham-org
- Keith Tyser: training and model compression. Solo gold medal, fourteenth of 6,125 in ROGII Wellbore Geology Prediction (his own account: https://keithtyser.com/blog/my-first-kaggle-gold-rogii-2026.html). https://keithtyser.com
- Jack Cole, adviser.

How the page is built: https://logarb.com/colophon.html