Hybrid Quantum Selection of Learning Rates in Language-Model Training: Termination Criteria and an Interaction-Guided Schedule Oracle
Organizations: Independent Researcher
Abstract
Learning-rate choices strongly affect language-model fine-tuning, but the best joint profile across layer groups is generally unknown. This paper studies quantum minimum finding for layer-group learning-rate selection within a hybrid classical-quantum training framework. It combines Durr-Hoyer minimum finding with BBHT search and compares fixed-budget and confidence-based stopping policies using a lookup oracle built from measured profile scores at fixed training checkpoints. The paper then introduces an Interaction-Guided Schedule Oracle that scores and marks joint schedules in superposition without a precomputed score table. Classical one-step loss measurements supply the main, pairwise and three-way interaction effects between layer groups. The circuit adds the effects for each schedule, compares the resulting score with the current baseline score, and marks improving schedules for amplitude amplification. Measured candidates are checked classically before acceptance. Uniform-rate training supplies the initial baseline; using schedules from classical adaptive methods as baselines for joint refinement is left for future work. In closed-loop fine-tuning of a 1.5-billion-parameter language model, each selected schedule drives the next 200 classical training steps. Benchmark accuracies lie within the range observed across five uniform-rate seeds. The results give a working example of how quantum superposition can support adaptive learning-rate selection during training; the remaining quantum limitations are discussed in the paper.
Figures & tables
| criterion | rounds | P(min. found) | regret | ||
|---|---|---|---|---|---|
| A global time-out | 77.3 | 8.2 0.1 | 222.2 0.2 | 1.000 | 0 |
| B per-visit budget | 29.3 | 8.1 0.1 | 29.4 0.0 | 0.999 | |
| C confidence count | 52.2 | 8.2 0.1 | 121.3 0.2 | 1.000 | 0 |
| D state-dependent stop | 28.3 | 8.1 0.1 | 25.6 0.1 | 0.999 |
| held-out loss at step | accuracy (%) | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| run | SciQ | PubMedQA | ARC-Easy | MMLU | |||||
| oracle, single batch | |||||||||
| oracle, five-batch average | |||||||||
| oracle, sampled on IBM hardware | |||||||||
| uniform rate, same seed | |||||||||
| uniform rate, five seeds (range) | – | – | – | – | |||||
| checkpoint | device | returned schedule | largest | ||||
|---|---|---|---|---|---|---|---|
| ibm_fez , ibm_marrakesh | (1.4, 0.7, 1.4, 1.4, 0.5, 1.4) | ||||||
| ibm_marrakesh | (0.7, 0.5, 1.4, 0.7, 1.4, 0.7) | ||||||
| ibm_kingston | (1.4, 1.4, 0.5, 0.5, 0.7, 0.5) | ||||||
| ibm_kingston | (0.5, 1.4, 1.0, 1.4, 0.7, 1.0) | ||||||
| ibm_marrakesh | (1.0, 0.5, 0.7, 0.7, 1.4, 1.4) |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| clock | value (as printed) | how obtained | source | literature comparison |
|---|---|---|---|---|
| , one Grover iteration at the design size | s (53–57) ns | Eq. ( 24 ): ns, median of 34 compiled schedules ( to 9; the executed job gives 31.8 and 34.5 ns) (706–747) (896–969) layers (1,602–1,716), bands over 3 seeds 2 calibration snapshots; declared | compiled schedules (this work); App. part A | IonQ Forte MS two-qubit gate median 672 s vs Heron cz 68 ns: about slower per two-qubit gate ( Chen et al., 2024 ) ; Willow surface-code cycle 1.1 s, a logical layer cycles ( Google Quantum AI and Collaborators, 2025 ) |
| , one state preparation + measurement | s | pooled iteration content s ; s , one total over the seven circuits ( , both families); circuit share –3.1 s, the rest is the backend’s inter-shot repetition delay (configured 250 s) | job metrics; backend configuration | The cited AES cost model expresses runtime through circuit depth and cycle time, with no setup term ( Sarah D., 2024 ) |
| , one classical check | s ( s later population) | median gap between consecutive replay files (restore 25 steps evaluation save): 389 gaps of one replay session, p10–p90 354.4–357.0 s, the slower population printed; later replays: 336.6 s; a GPU clock, never averaged | training-run replay timing (this work) | — (own workload; Eq. ( 21 ) , Eq. ( 20 ) ) |
| criterion | rounds | queries, improve | queries, stop | P(min. found) | regret | speedup, unit-cost | speedup, hw-informed |
|---|---|---|---|---|---|---|---|
| A DH96 global cutoff | 77.3 | 8.2 0.1 | 222.2 0.2 | 1.000 | 0 | 0.21 | 1.05 |
| B BBHT cascade budget | 29.3 | 8.1 0.1 | 29.4 0.0 | 0.999 | 1.9e-06 | 0.84 | 2.77 |
| C BBHT-confidence stop, | 52.2 | 8.2 0.1 | 121.3 0.2 | 1.000 | 0 | 0.35 | 1.55 |
| D dynamic stop (N,u,f, ), | 28.3 | 8.1 0.1 | 25.6 0.1 | 0.999 | 1.4e-06 | 0.90 | 2.86 |
| size | trials | rounds | queries | P(min. found) | speedup (unit-cost) | speedup (hw-informed) |
|---|---|---|---|---|---|---|
| 81 | 5 1,000 | 28.3 0.1 | 33.8 0.1 | 0.999 | 0.90 | 2.86 |
| 128 | 5 1,000 | 33.3 0.1 | 46.6 0.1 | 1.000 | 1.13 | 3.85 |
| 256 | 1 5,000 | 41.5 0.2 | 71.8 0.2 | 1.000 | 1.65 | 6.17 |
| single-batch measurements | five-batch measurements | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| leg | schedule | qubits | schedule | qubits | sign flips | ||||||
| (0.5, 1.4, 1.4, 0.5, 0.7, 0.5) | 129 | 87 | 264 | 24 | (0.5, 1.0, 1.4, 0.7, 1.4, 0.7) | 12 | 62 | 271 | 22 | 106 | |
| (0.5, 1.4, 0.7, 0.7, 1.4, 1.4) | 66 | 102 | 296 | 24 | (1.4, 0.7, 0.7, 0.7, 0.5, 1.4) | 25 | 75 | 320 | 22 | 76 | |
| (1.4, 0.5, 1.4, 0.7, 1.4, 1.4) | 78 | 109 | 494 | 24 | (0.7, 0.5, 0.5, 0.5, 1.4, 0.5) | 33 | 77 | 289 | 22 | 79 | |
| (1.4, 1.4, 0.5, 0.7, 1.4, 1.4) | 162 | 61 | 292 | 25 | (1.4, 0.7, 1.4, 0.5, 1.4, 0.5) | 11 | 52 | 269 | 23 | 73 | |
| (0.7, 1.4, 0.5, 1.4, 1.0, 1.4) | 28 | 76 | 304 | 24 | (1.4, 1.4, 0.5, 1.4, 0.5, 0.7) | 32 | 76 | 292 | 22 | 88 | |
| compilation | status | CZ per body, mean (min–max) | ||
|---|---|---|---|---|
| validated synthesis, chained layouts | measured, bodies, CZ counted only | ( – ) | ; , | |
| shared-AND synthesis, one iteration compiled whole, simulated backend | prototype, three routing seeds | – | – | not defined (one iteration only) |
| shared-AND synthesis, chained layouts, live backend (calibration of 11 November 2026) | measured, bodies | ( – ) | ; , |
| after the last update | circuit with the largest | |||||
|---|---|---|---|---|---|---|
| checkpoint | threshold path (round) | rounds | iterations | (round) | CZ gates | depth |
| (4) (5) (7) (11) (14) (19) | ( ) | |||||
| (1) (4) (10) | ( ) | |||||
| (3) (4) (23) | ( ) | |||||
| (4) (14) (29) | ( ) | |||||
| (14) (22) | ( ) | |||||
| checkpoint | device | jobs | IBM turnaround (s) | queue (s) | device (s) | QPU charged (s) | |
|---|---|---|---|---|---|---|---|
| sum | median / longest job | ||||||
| ibm_fez , ibm_marrakesh | / | ||||||
| ibm_marrakesh | / | ||||||
| ibm_kingston | / | ||||||
| ibm_kingston | / | ||||||
| ibm_marrakesh | / | ||||||
Learning Rate Scheduling
Learning rate scheduling studies how to vary optimizer step sizes during training to improve stability, convergence, efficiency, and generalization.
See all 65 papers