Question Answering over Long Documents / V116
On six sealed technical dossiers of about 51,000 tokens each, EMMA’s long-document method answered 32 of 36 questions right (88.9%) with a small frozen reader, against 38.9% for the best baseline. On the 12 questions also given to a published multi-agent reading method, it answered 11 against 1 while using 1.1% of that method’s tokens.
SCOPE + LIMITATIONS
- Synthetic dossiers from one generator, six question kinds and 36 questions. The multi-agent method was designed for large readers, so with a small reader this comparison favours EMMA’s method. The reader was interrupted once and resumed from saved answers.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Generate six sealed dossiers and 36 questions in six kinds. Every method uses the same small frozen baseline reader with greedy decoding. Baselines are the full document cut to fit, retrieval of the 40 most relevant lines, and Chain of Agents, which was limited to 12 questions before scoring because each took about 24 minutes. Two ablations switch off internal components.
All five gates passed: +50.0 points over the best baseline against +10, 1.1% of the multi-agent method’s tokens against a 50% limit, and both ablations lower. First occurrences in logs were the weak kind at 2 of 6, fixed later (V118.1).
- EMMA method
- 32 of 36 (88.9%); 31% of lines read
- Full document, cut to fit
- 14 of 36 (38.9%)
- Retrieval, top 40 lines
- 12 of 36 (33.3%)
- Chain of Agents · 12 questions
- 1 of 12; EMMA 11 of 12 on the same questions
- Ablation · first internal component off
- 30 of 36 (83.3%)
- Ablation · second internal component off
- 15 of 36 (41.7%)
Question and method
Generate six sealed dossiers and 36 questions in six kinds. Every method uses the same small frozen baseline reader with greedy decoding. Baselines are the full document cut to fit, retrieval of the 40 most relevant lines, and Chain of Agents, which was limited to 12 questions before scoring because each took about 24 minutes. Two ablations switch off internal components.
Recorded results
| Condition | Recorded result |
|---|---|
| EMMA method | 32 of 36 (88.9%); 31% of lines read |
| Full document, cut to fit | 14 of 36 (38.9%) |
| Retrieval, top 40 lines | 12 of 36 (33.3%) |
| Chain of Agents · 12 questions | 1 of 12; EMMA 11 of 12 on the same questions |
| Ablation · first internal component off | 30 of 36 (83.3%) |
| Ablation · second internal component off | 15 of 36 (41.7%) |
Comparative standing and locked gates
All five gates passed: +50.0 points over the best baseline against +10, 1.1% of the multi-agent method’s tokens against a 50% limit, and both ablations lower. First occurrences in logs were the weak kind at 2 of 6, fixed later (V118.1).
Interpretation and limitations
Synthetic dossiers from one generator, six question kinds and 36 questions. The multi-agent method was designed for large readers, so with a small reader this comparison favours EMMA’s method. The reader was interrupted once and resumed from saved answers.
SOURCE PROVENANCE
Question Answering over Long Documents / V116
LABORATORY REPORT / 2026-10-26SOURCE CHECKSUM / SHA-256
0cc5a4ef87e52ac4145b46654f4f578608bd1da5f7b302322008174ab0a00b0bPublic journal edition reviewed 2026-10-28. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.