1 of 42

DocPrune: Document Token Pruning for Efficient Document Question Answering

2026.03.11

2 of 42

Index

  1. Motivation
  2. Method
  3. Pipeline
  4. Experiment
  5. Summary

3 of 42

1. Motivation

Task A. Long Document Question Answering

문서 : 3366개

페이지 : 41114개

Question

For which film did Ben Piazza play the role of Mr. Simms?

페이지 : 4개

페이지 : 6개

.....

4 of 42

1. Motivation

Task A. Long Document Question Answering

문서 : 3366개

페이지 : 41114개

Question

For which film did Ben Piazza play the role of Mr. Simms?

페이지 : 4개

페이지 : 6개

.....

Stage 1. Evidence Page Retreival

Stage 2. Question Answering

5 of 42

1. Motivation

Task A. Long Document Question Answering

Stage 1. Evidence Page Retrieval

Stage 2. Question Answering

Top-K Retrieved Pages

The title is W. C. Handy.

Head

Vision Encoder

LLM

...

...

...

Document token

Prompt token

Question token

Top-K Retrieved Pages

3rd

2nd

1st

Answer

Describe

the

image

What

is

the

title

.

Prompt

Question

Document Encoder

...

...

Document

Question �Encoder

What

is

the

title

?

Question

?

N pages

...

Question embedding

Retrieval

...

...

...

Document embeddings

N pages

Select

Output�token

RAG-QA

6 of 42

1. Motivation

Task B. Token Pruning

Output : 10 Token

38% ↓

Input : 16 Token

7 of 42

1. Motivation

Task B. Token Pruning

Output : 10 Token

38% ↓

Input : 16 Token

Point 1. How to define unnecessary tokens?

Point 2. Which layer to drop these tokens?

8 of 42

1. Motivation

Task B. Token Pruning

Stage 1. Evidence Page Retrieval

Stage 2. Question Answering

Top-K Retrieved Pages

The title is W. C. Handy.

Vision Encoder

LLM

...

...

...

Document token

Prompt token

Question token

Answer

Describe

the

image

What

is

the

title

.

Prompt

Question

Document Encoder

...

...

Document

Question �Encoder

What

is

the

title

?

Question

?

N pages

...

Question embedding

Retrieval

...

...

...

Document embeddings

N pages

Select

Output�token

QA

Head

3rd

2nd

1st

K Pages

9 of 42

1. Motivation

Task B. Token Pruning

Stage 1. Evidence Page Retrieval

Stage 2. Question Answering

Top-K Retrieved Pages

The title is W. C. Handy.

Vision Encoder

LLM

...

...

...

Document token

Prompt token

Question token

3rd

2nd

1st

Answer

Describe

the

image

What

is

the

title

.

Prompt

Question

Document Encoder

...

...

Document

Question �Encoder

What

is

the

title

?

Question

?

N pages

...

Question embedding

Retrieval

...

...

...

Document embeddings

N pages

Select

Output�token

QA

Head

Output : 10 Token

Top-K Retrieved Pages

Point 1. How to define the objective of pruning?

→ Question-Dependent Objective

10 of 42

1. Motivation

Task B. Token Pruning

Stage 1. Evidence Page Retrieval

Stage 2. Question Answering

Top-K Retrieved Pages

The title is W. C. Handy.

Vision Encoder

LLM

...

...

...

Document token

Prompt token

Question token

3rd

2nd

1st

Answer

Describe

the

image

What

is

the

title

.

Prompt

Question

Document Encoder

...

...

Document

Question �Encoder

What

is

the

title

?

Question

?

N pages

...

Question embedding

Retrieval

...

...

...

Document embeddings

N pages

Select

Output�token

QA

Head

Output : 10 Token

38% ↓

Top-K Retrieved Pages

Layer 2, 3, 5...

Point 1. How to define the objective of pruning?

Point 2. Which layer to drop these tokens?

24ECCV : FastV

→ Question-Dependent Objective

→ Heuristic selection ( Layer 2, 3, ... )

11 of 42

1. Motivation

Task B. Token Pruning

Stage 1. Evidence Page Retrieval

Stage 2. Question Answering

Top-K Retrieved Pages

The title is W. C. Handy.

Vision Encoder

LLM

...

...

...

Document token

Prompt token

Question token

3rd

2nd

1st

Answer

Describe

the

image

What

is

the

title

.

Prompt

Question

Document Encoder

...

...

Document

Question �Encoder

What

is

the

title

?

Question

?

N pages

...

Question embedding

Retrieval

...

...

...

Document embeddings

N pages

Select

Output�token

QA

Head

Output : 10 Token

38% ↓

Top-K Retrieved Pages

Layer 2, 3, 5...

Point 1. How to define the objective of pruning?

Point 2. Which layer to drop these tokens?

Point 3. Image-targeted pruning design

→ Question-Dependent Objective

→ Heuristic selection ( Layer 2, 3, ... )

12 of 42

2. Method

Observation

Method

13 of 42

2. Method

O1. Dominance of background regions

14 of 42

2. Method

Q1. In which film did Ben Piazza play Mr. Simms?

Q2. What animals race in the Kentucky Derby?

A. Horses

A. Mask

O1. Dominance of background regions

1. Foreground

2. Background

  1. Divide into 32 x 32 patches
  2. Labeled background if all pixels match

→ background : 36%

15 of 42

2. Method

M1. Background Token Pruning

 

16 of 42

2. Method

O2. Sparse localization of supporting evidence

Q. In which film did Ben Piazza play Mr. Simms?

Q. What animals race in the Kentucky Derby?

A. Horses

A. Mask

  1. Answers rely on small, specific regions

17 of 42

2. Method

O2. Sparse localization of supporting evidence

Q. In which film did Ben Piazza play Mr. Simms?

Q. What animals race in the Kentucky Derby?

A. Horses

A. Mask

  1. Answers rely on small, specific regions.

  • Top 10% tokens hold 50–80% attention.

18 of 42

2. Method

O2. Sparse localization of supporting evidence

Q. In which film did Ben Piazza play Mr. Simms?

Q. What animals race in the Kentucky Derby?

A. Horses

A. Mask

  1. Answers rely on small, specific regions.

  • Top 10% tokens hold 50–80% attention.

  • Inputting top 10% of tokens based on the �20th layer's attention

-1.1

-0.8

19 of 42

2. Method

M2. Question-Aware Token Pruning

1. compute the cosine similarities and aggregated by text tokens

2. smooth the similarity map using Gaussian filtering

3. remain only relevance tokens exceed the threshold

20 of 42

2. Method

O3. Layer-dependent pruning effect

L14

L0 – L14 : random > attn based

L15 – L27 : random < attn based

21 of 42

2. Method

O3. Layer-dependent pruning effect

L0 – L14 : random > attn based

L15 – L27 : random < attn based

L14

1. Weak identification of salient visual tokens before layer 14

2. Underdeveloped inter-token interactions before layer 14

22 of 42

2. Method

O3. Layer-dependent pruning effect

L0 – L14 : random > attn based

L15 – L27 : random < attn based

L14

1. Weak identification of salient visual tokens before layer 14

2. Underdeveloped inter-token interactions before layer 14

How can we know layer 14?

23 of 42

2. Method

L2 Norm ↑ → Comprehension ↑

O3. Layer-dependent pruning effect

→ Model Comprehension

24 of 42

2. Method

O3. Layer-dependent pruning effect

→ Model Comprehension

L2 Norm ↑ → Comprehension ↑

25 of 42

2. Method

O3. Layer-dependent pruning effect

→ Model Comprehension

L2 Norm ↑ → Comprehension ↑

F1 ↑

F1 ↑

26 of 42

2. Method

O3. Layer-dependent pruning effect

→ Model Comprehension

L2 Norm ↑ → Comprehension ↑

27 of 42

2. Method

M3. Comprehension-Aware Token Pruning

1. determine the optimal layer with comprehension threshold

2. tokens with attention values below are dropped

28 of 42

2. Method

1

2

3

Dominance of background regions

Sparse localization of supporting evidence

Layer-dependent pruning effect

1

2

3

Background Token Pruning

Question-Aware Token Pruning

Comprehension-Aware Token Pruning

Observation

Method

Summary. Observation vs Method

29 of 42

3. Pipeline

Overview of Pipeline

30 of 42

3. Pipeline

Overview of Pipeline

1. Maximize Information Preservation

1

1

31 of 42

3. Pipeline

Overview of Pipeline

1. Maximize Information Preservation

2. Leverage Retrieval-Guided Feature

2

32 of 42

3. Pipeline

Overview of Pipeline

1. Maximize Information Preservation

2. Leverage Retrieval-Guided Feature

3. Query-Centric Pruning

3

33 of 42

3. Pipeline

Query-Dependency Only

Heuristic Selection

Image-Centric Design

Previous

OURS

Pruning Objective

Layer Selection

Target Domain

Previous vs OURS

34 of 42

3. Pipeline

Previous vs OURS

Query-Dependency Only

Heuristic Selection

Image-Centric Design

Pipeline-aware Adaptive Pruning

Previous

OURS

Pruning Objective

Layer Selection

Target Domain

35 of 42

3. Pipeline

Previous vs OURS

Query-Dependency Only

Heuristic Selection

Image-Centric Design

Pipeline-aware Adaptive Pruning

Comprehension-Based Optimal Selection

Previous

OURS

Pruning Objective

Layer Selection

Target Domain

36 of 42

3. Pipeline

Previous vs OURS

Query-Dependency Only

Heuristic Selection

Image-Centric Design

Pipeline-aware Adaptive Pruning

Comprehension-Based Optimal Selection

Document-Centric Design

Previous

OURS

Pruning Objective

Layer Selection

Target Domain

37 of 42

4. Experiment

E1. Qualitative Analysis-A

38 of 42

4. Experiment

E1. Qualitative Analysis-B

39 of 42

4. Experiment

E2. Ablation Study

+1.2

x2.2

+0.6

x2.3

+1.0

x2.2

+1.2

x2.2

+1.0

x3.3

+1.5

x3.2

40 of 42

4. Experiment

E3. Main Table

Performance on M3DocRAG

41 of 42

4. Experiment

E3. Main Table

Performance on MMLongBench

Performance on VDocRAG

42 of 42

5. Conclusion

  • We introduced DocPrune, a training-free progressive token pruning framework

  • DocPrune exploits the structured layout of documents and employs a comprehension-aware criterion to select pruning layers

  • Through three complementary stages(BTP, QTP, CTP), DocPrune progressively eliminates the redundant tokens in a way of reflects the objective of each stage.

  • The experiments show that DocPrune consistently outperforms prior pruning methods even with lower computation