HK//HACKERSHACKATHONS · KAGGLE · FOR BEGINNERS中文
HACKERS / Kaggle

Google Quantum AI Verifiable Quantum Circuit Compilation (Kaggle): beginner guide with the public score taken apart

Nothing here is trained. You hand Kaggle three quantum circuits, written as gate lists in Google's open-source kickmix format, and a verifier runs each one on many bit-string inputs to check it does the right job. Then it scores how much smaller and cheaper your circuits are than Google's reference versions. No quantum physics background is needed. The prize pool is $150,000, and the competition runs from Oct 22, 2026 to Jan 20, 2027, with a $10,000 leaderboard milestone on Nov 20. The best public notebooks on day one all land near 3.90, and their write-ups don't break that number down. We did, using the costs the notebooks print: the decoder task is 9.5× cheaper than the reference, the 1/√x task 5.1×, and the chemistry task only 1.22×. Because the score is a geometric mean of the three, the weakest task is where a beginner's effort counts most. About a day after launch, the public leaderboard leader was at 10.02, about 17 times further along than the public portfolio when you multiply the three ratios.

DEADLINE
PRIZE150,000 UsdGoogle Quantum AI
Official page ↗

Pipeline mind map

Built for this competition from the overview, data page, rules and official metric notebook, the Google Quantum AI blog post and companion paper, and the costs printed by the public notebooks we read on Oct 23, 2026. It follows a verify-and-shrink loop; there is no model to train.

  1. 01Set up the loop
    1. 1.1Copy the host starter (needs internet for Qualtran)
    2. 1.2Score the reference circuits: 1.0
    3. 1.3Run the verifier locally on every change
  2. 02Know the three costs
    1. 2.1invsqrt: qubits × (Toffolis + T/4)
    2. 2.2dqi: R × qubits² × Toffolis
    3. 2.3dfthcsos: λ_eff × Toffolis × qubits
  3. 03Spend effort where it pays
    1. 3.1Geometric mean: fix the weakest task first
    2. 3.2Chemistry is only 1.22× in public code
    3. 3.3Qubits count squared on dqi
  4. 04Ship safely
    1. 4.15 submissions a day, 45 minutes each
    2. 4.2Leaderboard tests use a secret salt
    3. 4.3Public notebook for the Nov 20 prize

Key facts

Host
Google Quantum AI on Kaggle. Featured competition, with Kaggle points and medals ↗
Prizes
$150,000. Final leaderboard: $50,000, $25,000, $10,000, then $5,000 each for 4th to 6th. Also $10,000 for the Nov 20 milestone leader, $10,000 for the Longest Leader Prize, and $10,000 for the best submission on each of the three tasks ↗
Deadlines
Milestone Nov 20, 2026; Longest Leader period ends Jan 4, 2027; entry and team-merger deadline Jan 13, 2027; final submission Jan 20, 2027. All at 11:59 PM UTC ↗
What you submit
A zip containing invsqrt.kmx (or .kmb), dqi.kmx (or .kmb), and a dfthcsos/ folder with angle_k_0 … angle_k_<n-1> plus submission.npz (the factorization and Givens tree). Scoring takes about 45 minutes ↗
Submissions
5 a day, with 2 final picks. Teams of up to 5; private code sharing outside a team is not allowed ↗
Licences
Data, verifiers and reference circuits are under Apache 2.0, and winners license their submission and code under Apache 2.0 ↗
Milestone and Longest Leader catch
Both prizes require the notebooks behind the score, and any datasets they use, to be public and to stay public until the end ↗
Participation (Oct 23, 2026, about 05:50 UTC)
1,094 entrants · 132 participants · 130 teams · 289 submissions ↗

Problem breakdown

What the three circuits do

invsqrt (fusion energy) takes a 22-bit number x, anywhere from 1 to 3,133,452, and must turn it into a phase proportional to 1/√x, accurate to 2⁻²⁴/x turns, leaving every helper qubit clean. dqi (combinatorial optimization) is a reversible Reed–Solomon decoder for RS(4095, 4025) over GF(2¹²). It reads 70 syndrome registers, recovers a ±1 sign and cleans up. dfthcsos (quantum chemistry) loads the Givens-rotation angles for a compressed FeMoCo Hamiltonian, in batches of J angles per circuit, no circuit wider than 128 qubits. You also choose the factorization, and it must stay within 20 mHa of the reference energy.

Why it can be checked at all

Kickmix circuits use only reversible classical gates, phase gates and measure-and-fix-the-sign uncomputation. Each one maps a bit string to a bit string plus a phase, so a normal CPU can run it one input at a time. Google says the simulator handles about 60 billion gates per second per core. The verifier tests every input for invsqrt (64 shots each) and 102,400 random shots for dqi. Inputs are seeded from a hash of your submission plus a secret salt, so any change brings fresh tests. Your local run can't reproduce the leaderboard's exact tests.

The score, in one line

Each task scores reference cost ÷ your cost, and the leaderboard is the geometric mean of the three. Submitting the reference circuits scores 1.0. Halving every cost scores 2.0. Halving just one task's cost multiplies the total by 2^(1/3) ≈ 1.26, whichever task it is. A gain on a task that is still near 1× is worth exactly as much as the same factor on a task that is already 9×. That's why the weakest task is the cheapest place to win points.

The public 3.90, taken apart

We divided the reference costs by the per-task costs printed in the top public portfolio notebook (public score 3.90011). invsqrt: 2,076,774 → 404,858, a 5.13× gain (115 qubits, 3,292 Toffolis, against 260 and 7,735). dqi: 2.444 × 10¹³ → 2.577 × 10¹², a 9.48× gain (1,785 qubits; Toffolis down from 5.73 million to about 0.68 million). dfthcsos: 3.356 × 10⁹ → 2.752 × 10⁹, only 1.22× (76,741 Toffolis against 93,237, with qubits about the same at 734 against 730). 5.13 × 9.48 × 1.22 = 59.3, and the cube root of that is 3.900, which matches the leaderboard. On Oct 23, 16 of the top 50 public teams sat between 3.900 and 3.905, the numbers published in public notebooks.

How far ahead the leaders are

At about 05:50 UTC on Oct 23 the public leader scored 10.017. 10.017³ ≈ 1,005, so the leader's three ratios multiply to about 17 times the public portfolio's 59.3, roughly 2.6× more per task. Eight teams were above 7. Nobody has published how. Google's companion paper says the hosts committed SHA-256 hashes of their own best circuits and will release them after the competition, and that their agent runs found significant gains over the published baselines.

The DQI floor nobody mentions

dqi's cost includes a repetition factor R(ε⁺) that grows with the decoder's failure rate, where ε⁺ is an upper confidence bound rather than the failures actually seen. In the public notebooks we read, R was about 1.19 for the reference decoder and also for the best public decoder, even with 0 failures in 102,400 shots. One notebook reports a single observed failure raising a candidate's dqi cost from 2.5772 × 10¹² to 2.5874 × 10¹² (+0.4%). Qubits enter squared, so on this task trimming 1% of qubits saves about as much as 2% of Toffolis.

Where the public tricks are

invsqrt: replace the reference Newton-step arithmetic with lookup tables (QROM) and fewer rotations. Public versions run from 1.03× to 5.1×. dqi: cheaper multiplication in GF(2¹²) using a tower-field split. One notebook counts a 21% Toffoli cut from this alone. dfthcsos: the host's starter shows that 12 bits is the lowest angle precision the reference factorization tolerates (11 bits breaks the 20 mHa limit), with constraints Λ > 49.0 and E_gap > 3.914898. The best public chemistry circuits arrive as files in private datasets, with no generator code, so this task has the least public know-how.

What you get

Instead of a dataset you get the verifiers, the reference circuits and the paper. The verifiers in verifiers/ are the authoritative specification. Copy the host's starter notebook to build and score a submission end to end; it installs Qualtran, so it needs internet, and the dfthcsos part takes about 10 minutes on CPU.

Instruction set
X, CX, CCX, Z, CZ, CCZ, Z_POW, NEG, SWAP, CSWAP, R, HMR (measurement-based uncompute), BIT_INVERT, BIT_STORE0/1, PUSH/POP_CONDITION, plus INFO_TAG and DEBUG_PRINT
Reference costs
invsqrt 2,076,774 (260 qubits, 7,735 Toffolis) · dqi 2.444 × 10¹³ (1,893 qubits, 5.73 million Toffolis) · dfthcsos 3.356 × 10⁹ (93,237 Toffolis, 730 qubits, λ_eff 49.30), as printed by public notebooks running the official verifiers
Hard limits
Any ANGLE circuit wider than 128 qubits is rejected; encoded Hamiltonian within 20 mHa; Λ > 49.0 and E_gap > 3.914898; submission.npz at most 2²⁶ bytes
File formats
.kmx is text and .kmb binary; both are documented in the kickmix section of Qualtran's docs

submission.zip

submission.zip
├── invsqrt.kmx          # or .kmb
├── dqi.kmx              # or .kmb
└── dfthcsos/
    ├── angle_k_0.kmx    # numbered from 0, no gaps
    ├── angle_k_1.kmx
    ├── ...
    └── submission.npz   # U, W, H2_I, H1_I, Hsym_lower, tree_edges

From the competition data / overview page ↗

Metric: Geometric mean over the 3 tasks of reference cost ÷ your cost

Higher is better. invsqrt cost = qubits × (Toffolis + T gates/4). dqi cost = R(ε⁺) × qubits² × Toffolis. dfthcsos cost = λ_eff × Toffolis × qubits for the whole block encoding, where λ_eff depends on the factorization's Λ and energy gap. Reference circuits score exactly 1.0, and a circuit that fails the verifier gives no score.

About half the test data decides the public leaderboard and the other half the final standings. Because the tests are fuzzed, the halves differ in which random inputs and salts they use, not in labels. A circuit that is truly correct should score the same on both, but one that only nearly passes can fail differently. ↗

Score of the best public notebook portfolio on Oct 23, 2026 (5.13 × 9.48 × 1.22, cube root): 3.9001 ↗

Public baselines

Public notebooks we read on Oct 23, 2026 through Kaggle's public API, with vote counts from that morning. Scores are the authors' own reports or the public leaderboard; we didn't submit anything.

Notebook▲LBWhy open it
Verifiable Circuit Compilation Starter (host)221.0 (reference)Start here. The host's own end-to-end pipeline: builds the reference circuits, writes the zip, runs the verifiers. Its chemistry section explains the precision and Λ / E_gap limits better than anything else public.
Quantum Circuits | Memory & Algebra123.90427 (projected by author)Prints the reference and candidate cost of every task, which is what we used for the decomposition. Adds a phase calibration that cuts invsqrt rotations from 37 to 32.
Quantum Circuits | QROM + Tower42103.90307 (public leaderboard)Shows the two main public ideas, lookup-table invsqrt and tower-field DQI multiplication, and reports how a single dqi failure changes the cost.
DQI | Exact 42-AND Arithmetic + Full Verification5dqi only: 5.73M → 4.50M Toffoli-equivalentsThe clearest single-idea notebook: one GF(2¹²) multiplication trick, fully verified, a 21% Toffoli cut. A good model for how to test one change at a time.
Verified circuit quick wins4invsqrt 1.815×, about 1.22 overallLists invsqrt gains version by version (1.07× → 1.53× → 1.65× → 1.71× → 1.82×), which shows what each step buys.

Most-voted notebooks right now

Refreshed daily from the Kaggle API (last 2026-11-16). Not reviewed by an editor; the table above is the reviewed list.

Timeline

  1. Start date
  2. $10,000 Mid-Competition Milestone: public leaderboard leader at 11:59 PM UTC, notebook must be public
  3. End of the Longest Leader Prize period; qualifying notebooks and datasets public by then
  4. Entry and team-merger deadline
  5. Final submission deadline, 11:59 PM UTC

Step-by-step plan

  1. Day 1: score 1.0 on purpose

    Copy the host starter, run it unchanged and submit. A 1.0 shows your zip layout, file names and the 45-minute scoring loop all work. Note the three reference costs it prints.

  2. Day 1–2: write the cost model down

    Make a small table that turns qubits, Toffolis, T gates and R into each task's ratio and the geometric mean. Check it gives 3.90 for the public portfolio numbers above. From then on, judge every idea by how much it moves that one number.

  3. Week 1: pick one task

    For a beginner, invsqrt is the easiest to read (a fixed-point 1/√x), and its verifier checks every input, so there is no sampling luck. Chemistry has the most room left in public code. Leave dqi until you are comfortable with GF(2¹²) arithmetic.

  4. Week 2–4: one change at a time

    Run each idea (a lookup table, a rotation fewer, a qubit saved by reusing a register) through the local verifier, then submit. Local tests use different random draws from the leaderboard, so a circuit that only just passes locally can still fail there. With 5 submissions a day, batch changes you've already verified.

  5. Before Nov 20

    If you go for the $10,000 milestone, publish the notebook that makes your submission, and any dataset it uses, and keep them public to the end. The same applies to the Longest Leader Prize.

  6. Through Jan 20

    Choose your 2 final submissions from circuits that passed with room to spare. Credit any public notebook you build on, and remember winners release code under Apache 2.0 and must be able to reproduce it.