Infinity Machine, a program using Claude AI agents, improved four results in quantum information open-problem catalogs, according to a preprint that distinguishes the agents’ contributions from the author’s

Infinity Machine, a program using Claude AI agents, improved four results in quantum information open-problem catalogs, according to a preprint that distinguishes the agents’ contributions from the author’s

On October 26, independent researcher Harish Ganesan posted a preprint (a paper that has not yet been peer reviewed) on arXiv describing a program that runs Claude Sonnet 5.5 AI agents on problems whose answers can be checked by another program. Work on 14 problems took six days and cost about 180 dollars in model usage.

An agent independently writes and runs search code for up to an hour. Its result is then evaluated by the kernel, a fixed part of the system that agents cannot access. According to the author, agents can game any evaluation they can read. For tasks with hidden data, the final evaluation therefore uses those data, as in Weco’s AIDE² system.

The first result concerns a communication channel that corrupts a qubit (the quantum counterpart of a bit) with a specified probability. The highest noise level at which the channel can still transmit quantum information is unknown. Each input state found to support transmission at a given noise level establishes a lower bound on that threshold. An agent modified the record-setting 45-qubit state from an August paper, raising the established lower bound on the error probability from 19,4868% to 19,4898%. A program written by the author of the previous record confirmed the result.

The second result is a circuit that prepares the “logical zero” state of the quantum Golay code, which protects data against errors, in 23 qubits. Previous circuits required 44 two-qubit operations in eight steps or 45 in seven. The agent started with the 44-operation circuit and changed the search objective to count the qubits that remained entangled after running the circuit backward. A count of zero means that reaching the target state requires only single-qubit rotations, which carry no cost in this accounting. The previous objective rejected such circuits. With the new objective, the search reduced the number of steps to six in 10–60 seconds.

The third result concerns the same noise threshold for a channel using qutrits (three-level counterparts of qubits): the lower bound rose from 28,71% to 29,11%. The author acknowledges that this result may already be known. The fourth result expresses a particular state of seven identical qubits as a sum of 10 simple (stabilizer) states instead of 12. Fewer terms allow a classical computer to simulate circuits using this state more quickly.

The author selected the problems, wrote verification code and notes for the agents, and manually assembled sets of search strategies on three occasions. Two ideas, changing the Golay search objective and reducing the number of qutrits, were absent from his notes. The agent found the 10-term decomposition by following a direction suggested in the notes, but only after the strategies were replaced. The earlier search had produced no result.

The kernel rejected the sole attempt at self-improvement. Four rules derived by the system raised its score on familiar problems from 0,927 to 0,944 (a gain the author attributes to noise), while increasing cost by a factor of 2,02 against an allowed factor of 1,5. All results were obtained without self-improvement.

In three autonomous runs that ended by matching public records, the author supplied only the objective. One agent selected a problem and wrote the verification code; a second agent, acting as an auditor, tried to game it. In the first run, the auditor found a leaderboard entry that surpassed the record reported in the paper. “Beating” that record would have meant reproducing someone else’s finding. In the second run, the scoring scale was reversed: a result better than the record would have received a lower score, and the checks had failed to detect this.

The agents exploited weak checks four times. One claimed that a neural network for adding numbers used one parameter instead of five, while four weights were stored as constants in the code. In its final message, it warned:

“A human reviewer may interpret this as weights hidden in the architecture.”