What RAG-SCHOLAR does

Less context.
More answer quality.

RAG-SCHOLAR is a context-efficient retrieval system for answering research questions across large scholarly corpora. It navigates papers, sections, entities, and figures before assembling only the compact, complementary evidence needed by the answer model.

Result highlight — b = 320: 47.82 Paper F1 and 69.42 Fact F1 with 250.97 total tokens.

Explore the evidence See the method

146 scholarly papers 116 text-suite queries

Text suite · b = 320Higher is better

47.82

39.86

Paper F1 ↑

69.42

53.31

Fact F1 ↑

RAG-SCHOLAR · b = 320CompAct

Total tokens / query250.97

97.76% fewer total tokens and higher Paper F1 and Fact F1 versus CompAct.

Why RAG-SCHOLAR

A large literature is not
one long context.

Research answers often depend on details scattered across papers, sections, figures, and repeated terminology. Sending the entire collection to an answer model is expensive and lets redundant context compete with the evidence that actually resolves the question.

01 · THE PROBLEM

Evidence is dispersed.

The right source may be easy to miss, while relevant passages are buried in long PDFs or repeated across background material.

02 · OUR APPROACH

Retrieve structure, then evidence.

RAG-SCHOLAR identifies papers and section roles first, uses scholarly entities and relations to guide retrieval, then keeps a compact, complementary text-and-visual bundle for the final answer.

Two accuracy metrics

Find the right papers.
Answer with the right facts.

Paper F1 measures source identification in the final answer; Fact F1 measures factual answering. Both are macro-averaged percentages on a 0–100 scale, evaluated on their respective query sets.

QUALITY 01

Paper F1 ↑

47.82

Final-answer paper identification across 116 text queries.

QUALITY 02

Fact F1 ↑

69.42

Factual answering on the 24-query factual subset of the text suite.

How it works

Do not flatten
the literature.
Organize its evidence.

RAG-SCHOLAR keeps papers, sections, passages, figures, and scholarly entities in a navigable hierarchy. The controller selects a compact, complementary evidence bundle rather than treating a longer context as automatically better.

Parse scholarly structure

Recover papers, sections, passages, figures, and normalized scholarly entities.

Retrieve papers first

Locate coherent candidate documents before selecting their local evidence.

Build a controlled bundle

Retain relevant and complementary evidence under the chosen controller setting.

Ground the answer

Bring paper, textual, and visual evidence together for the final response.

Structured PDFs

Papers, sections, passages, figures, and local links.

offline parse

Scholarly Graph

Entities and cross-paper relations expose evidence topology.

structured retrieval

Evidence Controller

Coarse-to-fine selection limits answer-model input.

query time

Grounded Answer

Evidence supports selected papers, facts, and claims.

answer time

Text-suite evidence

Shorter evidence.
Stronger quality profile.

At b = 320, RAG-SCHOLAR improves both Paper F1 and Fact F1 over all ten Table 1 baselines, using fewer total tokens than nine of them. Choose a baseline to compare both scores and the measured token cost.

Answer quality

Paper F1 and Fact F1 · macro averages (%)

0–100 native scale

Compact evidencePage-level VLM retrievalLong-context compression

47.82

39.86

Paper F1 ↑

69.42

53.31

Fact F1 ↑

RAG-SCHOLAR · b = 320CompAct

Against CompAct: +7.96 Paper F1 and +16.11 Fact F1 percentage points.

Parallel total-token bars

Total = Pre. + Input, per query.

lower is better

TOKEN COMPARISON

RAG-SCHOLAR · b = 320

250.97

CompAct

11,181.32

0Fixed linear scale · 45,416.27 tokens

97.76% fewer total tokens than CompAct, with higher Paper F1 and Fact F1.

RAG-SCHOLAR: Pre. 0.00 + Input 250.97.
CompAct: Pre. 11,026.36 + Input 154.96. Total savings include online compression; CompAct retains a smaller input context.

b = 320 · compact quality point47.82 Paper F1 · 69.42 Fact F1 · 250.97 total tokens.

b = 640 · Paper F1 peak49.49 Paper F1 at 387.05 total tokens: highest paper identification point in the sweep.

Uncapped · Fact F1 peak71.67 Fact F1 at 565.21 total tokens: b = 320 uses 55.60% fewer input tokens, with a 2.25-point Fact F1 difference.

Multimodal evidence

The same pattern
across text and visuals.

Table 2 adds visual input. RAG-SCHOLAR uses 71.4–72.0% fewer total answer-model input tokens than all three page-level multimodal systems while increasing both Paper F1 and Fact F1.

RAG-SCHOLAR

Multimodal suite · 30 paper queries / 15 reasoning queries

Table 2

Paper F1100.00

Fact F160.00

Total input tokens1,308.50

Text 1,090.70 + visual 217.80. Smaller text input is not offset by a larger visual payload.

ColPali

ICLR 2025 · total input 4,649.97

+43.67Paper F1

+13.71Fact F1

−71.86%tokens

M3DocRAG

ICCVW 2025 · total input 4,668.37

+37.56Paper F1

+3.08Fact F1

−71.97%tokens

VisRAG

ICLR 2025 · total input 4,581.77

+55.33Paper F1

+34.67Fact F1

−71.44%tokens

Table 1 record

The full text-suite frontier,
including every baseline family.

The cyan row is the primary compact-quality operating point used throughout this page. All token values are per-query means; the reporting convention keeps online preprocessing, answer-model input, and total cost distinct.

Method / setting	Paper F1 ↑	Fact F1 ↑	Pre. tok. ↓	Input tok. ↓	Total tok. ↓	Prep. (s) ↓
RAG-SCHOLAR retention settings
b = 100	44.83	65.56	0	174.42	174.42	11.39
b = 240	45.64	66.00	0	218.20	218.20	11.38
b = 320 · compact quality point	47.82	69.42	0	250.97	250.97	11.37
b = 480	49.01	64.44	0	321.36	321.36	11.35
b = 640 · Paper F1 peak	49.49	58.47	0	387.05	387.05	11.36
Uncapped · Fact F1 peak	46.74	71.67	0	565.21	565.21	11.80
Compact evidence
CompAct (EMNLP 2024)	39.86	53.31	11,026.36	154.96	11,181.32	46.77
ReComp (ICLR 2024)	40.39	27.00	0	244.28	244.28	25.06
Page-level VLM retrieval
ColPali (ICLR 2025)	31.84	45.07	0	1,854.68	1,854.68	—
M3DocRAG (ICCVW 2025)	31.21	42.30	0	1,854.42	1,854.42	—
VisRAG (ICLR 2025)	29.97	28.05	0	1,854.46	1,854.46	—
MDocAgent (arXiv 2025)	36.67	33.19	0	1,838.84	1,838.84	—
Long-context compression
LLMLingua-2 (ACL Findings 2024)	36.78	13.19	42,729.27	2,687.00	45,416.27	205.91
LongLLMLingua (ACL 2024)	30.75	28.60	24,421.79	2,304.91	26,726.70	448.14
Selective Context (EMNLP 2023)	27.15	17.78	14,106.06	2,500.00	16,606.06	—
LLMLingua (EMNLP 2023)	19.78	11.25	23,596.27	2,234.35	25,830.62	418.99

Source: updated manuscript, Table 1. Costs are per-query means; Total = Pre. + Input. Pre. counts tokens processed by online LLM compression; Input counts the final serialized context. Generated output is excluded; dashes indicate unreported preparation times.

Metric notes

Two measures of accuracy.
A clear accounting of input cost.

Paper F1 and Fact F1 describe different parts of a scholarly answer. The controller setting b selects an operating point; the measured context length is reported separately as Input.

Paper F1

Macro F1 of the paper set named in the final answer: 116 text queries and 30 multimodal queries.

Fact F1

Macro F1 on 24 factual text queries and 15 visual-reasoning queries, respectively.

Token cost

Text suite: Total = Pre. + Input. Multimodal suite: Total = Text + Visual. Generated output is excluded.

Less context.More answer quality.

A large literature is notone long context.

Evidence is dispersed.

Retrieve structure, then evidence.

Find the right papers.Answer with the right facts.

Paper F1 ↑

Fact F1 ↑

Do not flattenthe literature.Organize its evidence.

Structured PDFs

Scholarly Graph

Evidence Controller

Grounded Answer

Shorter evidence.Stronger quality profile.

Answer quality

Parallel total-token bars

The same patternacross text and visuals.

RAG-SCHOLAR

ColPali

M3DocRAG

VisRAG

The full text-suite frontier,including every baseline family.

Two measures of accuracy.A clear accounting of input cost.

Less context.
More answer quality.

A large literature is not
one long context.

Find the right papers.
Answer with the right facts.

Do not flatten
the literature.
Organize its evidence.

Shorter evidence.
Stronger quality profile.

The same pattern
across text and visuals.

The full text-suite frontier,
including every baseline family.

Two measures of accuracy.
A clear accounting of input cost.