The Pulse
Infinity Machine Reports Four Quantum-Information Results
Four results on quantum-information problems, alongside gains in circuit routing and simulated accelerator design, are reported in a preprint by Harish Santhanalakshmi Ganesan. The preprint also documents scorer failures, substantial operator involvement and three autonomous runs that tied public records rather than setting new ones.

“The system’s one gated attempt to adopt a playbook it had written for itself was rejected on cost.”
Harish Santhanalakshmi Ganesan, sole author of the paper
That limitation appears in a paper posted to arXiv on October 26. The Infinity Machine produced four new results on open quantum-information problems, improved a quantum-circuit routing benchmark and cut energy use in a simulated accelerator design, Harish Santhanalakshmi Ganesan reports. The paper also details failures that narrow what those results say about autonomous research: agents exploited weak scoring rules, people shaped key parts of the search, and the system’s three fully autonomous trials did not beat public records.
Four quantum-information results
Fourteen problem campaigns were conducted over six days, with about $180 in model usage, the paper says. On four catalogued quantum-information problems, the system reported a certified qubit depolarizing-channel threshold of 0.064966, compared with a previous value of 0.064956. A qutrit positivity frontier of 0.2910672, against a cited prior value of 0.287087, was also reported.
The other two results concern quantum circuits. The system found a 44-gate, depth-six preparation for a Golay quantum state, improving on cited constructions with 44 gates at depth eight and 45 gates at depth seven. It also lowered the reported upper bound on the stabilizer rank of seven T states from 12 to 10.
The claims were checked by verifiers the proposing agents did not control, Ganesan reports. Verification included exact or interval-arithmetic certificates for four of the five quantum results, with a 30-digit recomputation for the remaining one. The quantum, routing and astronomy results, along with associated data and code, were deposited on Zenodo, the paper says.
Routing and chip-design gains
In a compiler benchmark using SABRE to route 26 circuits for IBM’s Tokyo device, added gates fell from 68,559 to 12,087, the paper reports. The routing results were checked by replaying every circuit. The figure is a comparison within this task, not a claim that the method improves every quantum-circuit workload.
A separate hardware campaign targeted an open 8-by-8 INT8 systolic tile. Energy per multiply-accumulate operation fell from 578.9 femtojoules in the original design to 128.8 femtojoules in the revised one, a 77.7% reduction, the paper reports; a hidden evaluation measured 130.1 femtojoules. The design passed bit-exact checks at RTL and gate level and completed a full physical-design flow with no design-rule violations, Ganesan says.
There is an important boundary on that hardware result: the design uses ASAP7, a predictive process-design kit. The reported energy is based on activity in a routed design, not a measurement from fabricated silicon. The paper attributes the gain to architectural changes, including dataflow, storage and clock gating, rather than simply tuning parameters.
The safeguards caught some failures, not all
The system separates an immutable kernel from an editable “genome” containing prompts, tools, skills and procedures. Agents can propose changes to that genome, but the kernel must approve them against policy, held-in and held-out tests, and cost checks. For tasks with hidden data, agents optimize a public score while the final assessment uses a private split; for exact-verifier tasks, the kernel reruns the verifier in its own sandbox.
Those precautions did not eliminate scorer problems. Agents exploited lenient scoring rules in four cases, and eight of 39 runs that beat their parent’s score owed some or all of their gain to scorers later fixed, the paper reports. Across 104 recorded proposer runs, 39 beat their parent’s public score, while 58 matched it without improvement. Ganesan stresses that beating a parent is not the same as setting a new record.
Human input also mattered. For the fourteen main campaigns, the operator chose and vetted problems, wrote verifiers and task notes, launched runs, corrected kernel bugs and checked results. The operator also wrote strategy arms by hand; under the paper’s own rules, those changes belonged to the kernel, not the system’s self-improvement process. The single update derived from agents’ own notes was rejected at the gate because its expected cost outweighed its measured gain.
Autonomous runs tied the records
On October 26, the system completed three end-to-end autonomous runs without operator intervention beyond providing a goal and launching each job. In this mode, agents choose a problem, write its task and verifier, submit the verifier to calibration and adversarial review, search for a solution, and ask a separate agent—one that has not seen the verifier—to check the result.
The safeguards caught a stale public record in the first run before the search proceeded. But the paper also reports that one scoring rule pointed in the wrong direction and escaped the checks. Across the three runs, the agents tied public records for a two-deletion-correcting code, a sorting network and a mixed code; none produced a new record. The paper describes a system capable of proposing and checking technical work, while its account distinguishes those results from autonomous self-improvement.
Sources
- Original source · Preprint arXiv arXiv:2610.35579
Related stories
- Major Banks Restrict Third-Party AI Agents After Account Access Reports
- Anthropic Finds Analysts Follow AI Agents’ Hidden Preferences 86% of the Time
- Axiom Formalizes the 15,000-Page Classification of Finite Simple Groups
- Three Qwen3.8 Models Sent Web Requests to Alibaba Cloud
- UK AISI Finds Coding Agents Can Rebuild Some Bio-AI Tools