Adiabatic quantum search with atoms in a cavity driven by lasers
D Daems; S Guérin
6 sequential KUs at 500 words · six model/chunk views
Read complete Knowledge Units →Models, measurements and five complete examples
Scientific facts, extracted as Knowledge Units
Read the complete factual extraction of five scientific papers, from the first source chunk to the last. Compare 500- and 1,000-word chunks, Gemma LoRA students and the Qwen-27B teacher. This demo adds guarded back-translation to reduce copied wording and measures whether question-answering quality survives.
Full paper→Sequential 500 / 1,000-word KUs→Detect copied phrases→English → German → English→Fidelity & copying guards→Frozen QA evaluation
Full original paper91.0%182 / 200 correct
Qwen teacher · 500-word KUs + BT90.0%180 / 200 correct
Gemma12 LoRA · 500-word KUs + BT88.5%177 / 200 correct
No paper context43.0%86 / 200 correct
20 disjoint papers, 200 frozen four-choice MCQs per condition, the same Qwen2.5-7B answerer. All 28 contexts were freshly judged in this repair study. No thinking, no retraining, and no QA-based repair selection. Failed generations remain wrong in the denominator.
Weights and chunk-size profiles
Gemma 4 12B IT · KU LoRA rank 128
500-word inference profile · 1,000-word inference profile
One adapter per base model, trained on 1,000-word teacher chunks for one epoch. These links apply the same weights at two inference chunk sizes.
Gemma 4 E4B IT · KU LoRA rank 128
500-word inference profile · 1,000-word inference profile
One adapter per base model, trained on 1,000-word teacher chunks for one epoch. These links apply the same weights at two inference chunk sizes.
Teacher: Qwen/Qwen3.8-27B-FP8 · Translation: TranslateGemma 4B.
QA accuracy: before and after guarded back-translation
| KU generator | Source chunk | Raw KUs | Guarded BT | Change | Paired 95% interval | Invalid after BT |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B FP8 · teacher | 500 words | 88.5% (177/200) | 90.0% (180/200) | +1.5 pp | +0.0 to +3.0 pp | 0 |
| Gemma 4 E4B IT · no adapter | 500 words | 75.5% (151/200) | 74.5% (149/200) | -1.0 pp | -3.5 to +1.5 pp | 10 |
| Gemma 4 E4B IT · KU LoRA rank 128 | 500 words | 82.5% (165/200) | 82.5% (165/200) | +0.0 pp | -1.5 to +1.5 pp | 10 |
| Gemma 4 12B IT · no adapter | 500 words | 75.0% (150/200) | 75.0% (150/200) | +0.0 pp | -2.0 to +2.0 pp | 10 |
| Gemma 4 12B IT · KU LoRA rank 128 | 500 words | 89.0% (178/200) | 88.5% (177/200) | -0.5 pp | -1.5 to +0.0 pp | 0 |
| Qwen 3.8 27B FP8 · teacher | 1,000 words | 88.0% (176/200) | 87.5% (175/200) | -0.5 pp | -2.5 to +1.0 pp | 0 |
| Gemma 4 E4B IT · no adapter | 1,000 words | 66.5% (133/200) | 66.5% (133/200) | +0.0 pp | -3.0 to +3.5 pp | 0 |
| Gemma 4 E4B IT · KU LoRA rank 128 | 1,000 words | 79.5% (159/200) | 78.0% (156/200) | -1.5 pp | -3.5 to +0.5 pp | 0 |
| Gemma 4 12B IT · no adapter | 1,000 words | 64.5% (129/200) | 64.5% (129/200) | +0.0 pp | -3.0 to +3.0 pp | 10 |
| Gemma 4 12B IT · KU LoRA rank 128 | 1,000 words | 84.5% (169/200) | 84.0% (168/200) | -0.5 pp | -2.0 to +1.0 pp | 0 |
Paired intervals use 10,000 paper-bootstrap draws. An interval containing zero does not establish a reliable accuracy change. Repair and batching parameters were fixed before this QA run. Questions were teacher-authored and automatically source-validated; they are not independently human-curated.
Full-paper, no-context and summary reference conditions
| Context | QA accuracy | 95% interval | Invalid slots | Weights |
|---|---|---|---|---|
| No paper context | 43.0% (86/200) | 38.0–48.0% | 3 | Model / adapter |
| Complete original paper / ground-truth context | 91.0% (182/200) | 87.5–94.5% | 0 | Model / adapter |
| Qwen 27B summary | 81.0% (162/200) | 75.5–86.5% | 0 | Model / adapter |
| Gemma E4B summary · no adapter | 64.0% (128/200) | 51.5–75.0% | 20 | Model / adapter |
| Gemma E4B KU LoRA used for summary generation | 68.0% (136/200) | 54.5–80.0% | 20 | Model / adapter |
| Gemma12 summary · no adapter | 73.5% (147/200) | 67.0–79.5% | 0 | Model / adapter |
| Gemma12 KU LoRA used for summary generation | 80.5% (161/200) | 74.0–86.5% | 0 | Model / adapter |
| Gemma12 earlier summary LoRA · rank128 | 89.0% (178/200) | 83.5–93.5% | 0 | Model / adapter |
Original paper means the actual source context, not an assumed perfect answer score. These summary rows were not back-translated in this KU study.
Copying and fidelity checks
| KU generator | Words | Raw coverage ≥6 | BT coverage ≥6 | BT longest run | BT ≤5 pass | BT ≥7 flagged |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B FP8 · teacher | 500 | 10.97% | 7.85% | 31 words | 0 / 20 | 20 / 20 |
| Gemma 4 E4B IT · no adapter | 500 | 22.51% | 19.07% | 37 words | 0 / 20 | 20 / 20 |
| Gemma 4 E4B IT · KU LoRA rank 128 | 500 | 14.80% | 11.16% | 37 words | 0 / 20 | 20 / 20 |
| Gemma 4 12B IT · no adapter | 500 | 8.16% | 6.54% | 20 words | 0 / 20 | 19 / 20 |
| Gemma 4 12B IT · KU LoRA rank 128 | 500 | 13.47% | 10.18% | 35 words | 0 / 20 | 20 / 20 |
| Qwen 3.8 27B FP8 · teacher | 1,000 | 10.32% | 7.56% | 22 words | 0 / 20 | 20 / 20 |
| Gemma 4 E4B IT · no adapter | 1,000 | 21.27% | 18.25% | 41 words | 0 / 20 | 20 / 20 |
| Gemma 4 E4B IT · KU LoRA rank 128 | 1,000 | 14.19% | 9.96% | 26 words | 0 / 20 | 20 / 20 |
| Gemma 4 12B IT · no adapter | 1,000 | 8.35% | 6.30% | 15 words | 1 / 20 | 16 / 20 |
| Gemma 4 12B IT · KU LoRA rank 128 | 1,000 | 11.18% | 8.55% | 37 words | 0 / 20 | 20 / 20 |
Conservative repair acceptance
1,123 windows inserted across 190 of 200 KU document versions. Raw candidates had 315 conservative number / unit / formula flags. 0 retained document-level critical-value flags after guard/rollback.
Acceptance requires both NLI directions ≥0.9, unchanged critical-value signatures and no copied run over five words inside the repaired window. Rejected changes retain the original field. These automated checks do not certify every scientific claim.
Measured compute
Translation: 17.5 active seconds / 0.0049 active GPU-hours for 200 KU document versions. Complete experiment: 0.4500 reserved GPU-hours on four GH200 GPUs.
Active translation excludes cold load, semantic checks, QA and idle roles. Reserved allocation includes them. These are measured experimental costs, not a large-corpus throughput guarantee.
Primary copying counts retain every factual field, including protected technical names, relation targets and bibliographic information occurring inside factual attributes, exactly as in the preceding KU benchmark. Titles at document level are metadata. Preserving names can leave literal matches; no output is removed to improve the score.
Context, attribution and style: what is actually present?
The Alexandria paper describes contextual summaries, entities, attributes, relationships and sentence MinHash. This experiment uses the repository’s sequential schema: at most ten preceding KU summaries/entity names support consistent naming. It does not supply future source chunks. The recorded context_summary is a local contextual description under the current prompt; a separate summary of the entire preceding narrative is not guaranteed.
Each reader exposes source-registry title/authors alongside the actual KU metadata. Authors may also appear inside entity attributes. The original document-author envelope is empty; registry attribution is added for display only. There is no dedicated style annotation field. Navigation shows the preceding and following KU summaries as a reader aid, rather than generated lookahead annotations. Short factual descriptors are visible where the extractor recorded them.
| Generator / chunk | Exact title /20 | Author envelope /20 | Author attributes /20 | Style attributes /20 | Context / generated KUs | MinHash / generated KUs |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B FP8 · teacher · 500 | 20 | 0 | 2 | 0 | 163 / 163 | 163 / 163 |
| Gemma 4 E4B IT · no adapter · 500 | 20 | 0 | 0 | 0 | 162 / 162 | 162 / 162 |
| Gemma 4 E4B IT · KU LoRA rank 128 · 500 | 20 | 0 | 3 | 0 | 162 / 162 | 162 / 162 |
| Gemma 4 12B IT · no adapter · 500 | 20 | 0 | 0 | 0 | 162 / 162 | 162 / 162 |
| Gemma 4 12B IT · KU LoRA rank 128 · 500 | 20 | 0 | 6 | 1 | 163 / 163 | 163 / 163 |
| Qwen 3.8 27B FP8 · teacher · 1,000 | 20 | 0 | 2 | 0 | 85 / 85 | 85 / 85 |
| Gemma 4 E4B IT · no adapter · 1,000 | 20 | 0 | 0 | 0 | 85 / 85 | 85 / 85 |
| Gemma 4 E4B IT · KU LoRA rank 128 · 1,000 | 20 | 0 | 3 | 0 | 85 / 85 | 85 / 85 |
| Gemma 4 12B IT · no adapter · 1,000 | 20 | 0 | 0 | 0 | 84 / 84 | 84 / 84 |
| Gemma 4 12B IT · KU LoRA rank 128 · 1,000 | 20 | 0 | 7 | 0 | 85 / 85 | 85 / 85 |
Five complete papers
Selected before this repair evaluation, without QA-based selection. Each paper provides six complete saved views: Gemma12 LoRA, Gemma E4B LoRA and Qwen27, each at 500 and 1,000 words. Partial failures, if any, remain explicit.
Evidence for contact delocalization in atomic scale friction
D Abel; S Yu Krylov; J W M Frenken
7 sequential KUs at 500 words · six model/chunk views
Read complete Knowledge Units →THE 2-ADIC VALUATION OF STIRLING NUMBERS
Tewodros Amdeberhan; Dante V Manna; Victor H Moll
9 sequential KUs at 500 words · six model/chunk views
Read complete Knowledge Units →A six-month periodic fasting reduces microalbuminuria and improves metabolic control in patients with type 2 diabetes and diabetic nephropathy: a randomized controlled study
Alba Sulaj; Stefan Kopf; Ekaterina Von Rauchhaupt; Elisabeth Kliemank; Maik Brune; Zoltan Kender; Hannelore Bartl; Fabiola Garcia Cortizo; Katarina Klepac; Zhe Han; Varun Kumar; Valter Longo; Aurelio Teleman; Jürgen G Okun; Jakob Morgenstern; Thomas Fleming; Julia Szendroedi; Stephan Herzig; Peter P Nawroth
10 sequential KUs at 500 words · six model/chunk views
Read complete Knowledge Units →Intact finger representation within primary sensorimotor cortex of Musician's Dystonia
Anna Sadnicka; T Wiestler; K Butler; E Altenmüller; M J Edwards; N Ejaz; J Diedrichsen
12 sequential KUs at 500 words · six model/chunk views
Read complete Knowledge Units →Downloads and reproducibility
Evaluation PDF · Complete Gemma12 examples PDF · All QA scores and protocol · Copy audit · Example inventory and checksums
Project Alexandria paper · Code and evaluation repository. Demo KUs preserve saved outputs without editorial rewriting. Original paper bodies, source-containing prompts, model reasoning and MCQs are not included in this reader.
Embedded author/title/citation attributes and exact document-title strings are also measured as metadata in a separate bibliographic copying diagnostic. The historical primary all-factual-string audit above remains unchanged; this additional classification does not change repairs or QA scores.