Wikipedia Closed-Book Jump
16% → 66.25%
Episode-reported move from raw-document fine-tuning to Active Reading on the linked SimpleQA-style Wikipedia setup.
Study Engine
Active Reading reframes the document as something to study, not merely reread. Click through the stages to see how the paper turns source pages into tailored training artifacts.
interactive step trace
Click a stage to spotlight its role in the curriculum
Overview Diagram
all diagrams are inline SVG
Transform Inventory
Paraphrase breaks dependence on original sentence order.
Active recall forces the answer to be produced from memory.
Timeline / association reshapes entity facts into linked structures.
Analogy / comparison increases retrieval routes inside the weights.
What Is New Here
This is mainly a curriculum paper. The model studies synthetic materials designed to teach the document, instead of only seeing the document again or getting a narrow stream of QA pairs.
Retention Grid
This heatmap is illustrative mock data shaped to match the episode’s ranking of methods: raw rereading helps least, synthetic QA helps more, and the broader Active Reading mix holds up better across rephrasings.
hover for cell detail
Mode switches the training diet behind the same question surface
Heatmap
blue → orange → red = stronger retention
Rows
The row axis stresses the same fact under different question styles: exact recall, paraphrase, temporal phrasing, association, distractor pressure, and number-heavy asks.
Columns
The column axis spreads fact types across entity, date, causal, alias, numeric, and relation-heavy pockets to mimic a bounded document archive.
Results Curve
This view mixes episode-reported numbers with a stylized scaling curve. Toggle between direct accuracy, relative lift over raw fine-tuning, and the synthetic-token growth story.
reported metrics + illustrative trend line
Switch between benchmark view and token-budget view
Performance Chart
gold-context marker included where discussed
Method Colors
Raw FT sees the original documents again.
Paraphrase adds rewordings but stays narrow.
Synthetic QA teaches explicit question-answer mappings.
Active Reading mixes study formats and keeps improving longer.
Reading Caution
The strongest chart is also the messiest attribution story. Equal-compute controls are missing, so data mix, steps, and total budget all move at once.
Deployment Lens
The paper matters most where documents are stable and low-latency closed-book answers matter. The scenario toggle shifts the recommended zone and shows why retrieval still dominates freshness-critical settings.
scenario switch + evidence gap view
Scenario changes the fit zone for methods on the left
Fit Map
hover points and risk bars
Big Missing Controls
The episode repeatedly flags equal-compute baselines, held-out documents, and broader paraphrase transfer as the experiments most needed next.
Why Teams Still Use Retrieval
Retrieval keeps provenance, freshness, revocation, and citations in the loop. Parametric memory wins on latency and packaging, but it is harder to edit cleanly.
Reference Strip
Primary paper plus the comparison literature and related episodes surfaced in the discussion.
compact source graph