---
id: "2026-10-15-llm-discovered-quantum-circuits-harvard-anthropic-mythos"
url: "https://postcutoff.com/e/2026-10-15-llm-discovered-quantum-circuits-harvard-anthropic-mythos/"
as_of: "2026-10-16T12:39:00+02:00"
date: "2026-10-15"
date_precision: day
category: science
importance: 3
confidence: medium
status: [Partly confirmed]
sources: 1
editor: human editor
human_review: null
version: "2026-10-16"
---

As of: 2026-10-16 12:39 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-15-llm-discovered-quantum-circuits-harvard-anthropic-mythos/

# Harvard–Anthropic team (with Mikhail Lukin) uses Claude Mythos 5 in an evolutionary search to beat state-of-the-art quantum circuits on seven of eight tasks

arXiv 2610.22367 (15 Oct 2026) by Andi Gu, Alek Kemeny (Anthropic), Susanne Yelin and Mikhail Lukin pairs Claude Mythos 5, run inside the open-source OpenEvolve framework, with an independent evaluator that verifies and scores reversible circuits. On eight compilation tasks (decoders for decoded quantum interferometry, HHL rotations, Coulomb and Gaussian functions, chemistry and lattice-gauge simulation, modular inversion), the discovered circuits reduce both Toffoli and qubit counts versus the state of the art in seven, with Toffoli×qubit gains of 4.1× to 39×; in the eighth it found a new reversible architecture making divsteps practical for prime-field modular inversion.

## Key facts
- Paper: 'Discovering state-of-the-art quantum circuits with large language models', Andi Gu (Harvard; Anthropic STEM Fellows Program), Alek H. Kemeny (Anthropic), Susanne F. Yelin (Harvard), Mikhail D. Lukin (Harvard; QuEra), arXiv 2610.22367 [quant-ph], 15 Oct 2026
- Setup (quoted): 'We search over these circuit-generating programs with OpenEvolve, using Claude Mythos 5 to propose architectures and program rewrites … Starting from a blank program, it chooses the algorithm, representation, and circuit structure and produces the discovered circuits without substantive human intervention'; MAP-Elites keeps Toffoli/qubit trade-offs
- Table I, Toffoli×qubit gain over baseline: Reed–Solomon DQI decoder 12× (5.1M → 440k Toffolis); HHL rotation angle 32×; Coulomb potential 9.9×; fixed-point Gaussian 39×; FeS alias–Select 4.1×; non-abelian lattice-gauge simulation 7.7×; Hermitian DQI decoder: first reversible compilation (9.3·10⁵ Toffolis vs 10⁸ naive)
- Modular inversion: a new reversible architecture that makes divsteps, a classical algorithm 'previously impractical for reversible integer arithmetic', viable; separately a specialised Luo construction with 3.4× fewer Toffolis at five extra qubits
- Authors' framing: 'some apparent costs of quantum arithmetic reflect limitations of existing reversible constructions rather than of the algorithms themselves'

## What happened

Quantum algorithms often spend most of their cost on classical subroutines (arithmetic, decoding, function evaluation) that must be
written as reversible circuits. Optimizers usually only apply equivalence-preserving rewrites to a fixed design.

On 15 Oct 2026 a Harvard group led by Mikhail Lukin, with an Anthropic co-author and an Anthropic STEM Fellow as first author, posted
[arXiv 2610.22367](https://arxiv.org/abs/2610.22367). Claude Mythos 5, inside OpenEvolve, proposes whole circuit-generating programs
from a task specification and a blank start; a separate evaluator checks correctness on basis states and scores Toffoli count
times peak qubits. Across eight tasks from optimisation, chemistry, simulation and number theory, the search beat published
constructions in seven, often by an order of magnitude, for example a 12× lower cost for the Reed–Solomon decoder in decoded
quantum interferometry. In the eighth it found a reversible design for the divstep modular-inversion algorithm.

No Hacker News thread was found on 16 Oct (Algolia); X was not searched.

## Why it matters

Resource estimates for useful quantum algorithms depend on such subroutines; large cuts in their cost move estimates for
applications like DQI and lattice-gauge simulation. It is also an AlphaEvolve-style result with a frontier commercial model and an
open-source framework, by a leading experimental quantum group.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 168 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 107 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 198 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 137 days after its cutoff

## Sources

1. [arXiv 2610.22367: Discovering state-of-the-art quantum circuits with large language models](https://arxiv.org/abs/2610.22367) (arxiv.org, paper)

## Changes

- 2026-10-16 (filed): Created from the arXiv PDF

## Related

- 2026-10-14: [Mid-October 2026: arXiv results with disclosed AI help](https://postcutoff.com/e/2026-10-14-ai-assisted-results-mid-october-2026/index.md)