Reading papers for Ph.D. preparation is different from reading papers for a course.

For a course, the goal is often to understand the method and reproduce the main derivation. For research, the goal is more demanding:

• identify the real problem being solved;
• understand why the method works;
• see what assumptions the paper relies on;
• find what remains unsolved;
• translate the paper into possible experiments or new research questions.

This note summarizes the reading workflow I currently use while preparing for Ph.D. applications in 3D perception, collaborative perception, and occupancy world models.

1. Read for Questions, Not Only Answers

The most useful part of a paper is not always the final architecture.

A good paper usually contains a sharper question:

What bottleneck prevents current systems from working better, and what representation or learning signal makes that bottleneck easier?

For my direction, I try to map every paper into one of several questions:

1. Representation: How is 3D space represented? BEV feature, voxel grid, occupancy volume, token set, implicit field, or object-centric memory?
2. Observation: What does the agent actually observe? Single camera, multi-view images, LiDAR, multi-agent messages, history, or future supervision?
3. Communication: If multiple agents are involved, what is transmitted, when, and under what budget?
4. Temporal reasoning: Does the model only reconstruct the current frame, or does it maintain memory and predict the future?
5. Deployment constraint: What are the limits in bandwidth, latency, memory, compute, or robustness?

This framing helps me avoid reading papers as isolated tricks. Each paper becomes evidence about how the field is trying to solve a larger problem.

2. The Three-Pass Reading Workflow

I use a three-pass process.

2.1 Fast Triage

The first pass usually takes 20 to 40 minutes.

I read:

• title;
• abstract;
• introduction;
• main figure;
• method overview;
• main quantitative table;
• ablation table;
• limitation section if available.

The output is a quick decision:

• deep read now if the paper is central to my current project;
• save for later if it is relevant but not urgent;
• skim only if it is useful background;
• skip if it does not connect to my research questions.

This prevents the reading list from becoming an infinite queue.

2.2 Deep Reading

For important papers, I read the method and experiments slowly.

I focus on six items:

1. Problem formulation: What are the inputs, outputs, assumptions, and constraints?
2. Representation: What is the core data structure that makes the method possible?
3. Learning objective: What losses or supervision signals shape the model?
4. System design: Which modules matter, and how are they connected?
5. Evidence: Which experiments actually support the main claim?
6. Failure cases: Where does the method become weak?

For example, in collaborative occupancy prediction, I ask:

• Is the transmitted message a dense feature map, a sparse feature, a voxel representation, or a token set?
• Is communication sender-driven, receiver-driven, or jointly optimized?
• Does the model account for pose error, delay, occlusion, or limited bandwidth?
• Does the evaluation report communication cost clearly?
• Is the method robust when fewer agents are available?

2.3 Research Extraction

The last pass is the most important one.

After understanding the paper, I force myself to write down:

• one sentence describing the paper’s core insight;
• one reason the method works;
• one assumption that may break;
• one experiment I would add;
• one idea that could connect to my own work.

If I cannot extract at least one research question from a paper, I probably understood the surface but not the research value.

3. A Paper Reading Template

I usually convert each important paper into a small structured note.

Paper:
Problem:
Core idea:
Representation:
Method:
Loss / objective:
Datasets:
Main results:
Ablations:
Assumptions:
Limitations:
Possible extensions:
Connection to my work:
Implementation difficulty:

This template is simple, but it is useful because it separates what the paper says from what I think about the paper.

The “connection to my work” field is especially important. A paper should not just enter my memory; it should enter my research system.

4. A Utility Score for Paper Priority

To decide which papers deserve deep reading, I sometimes use an informal score:

where:

• (R(p)) is relevance to my current project;
• (N(p)) is novelty of the idea;
• (I(p)) is implementation value;
• (S(p)) is strategic value for long-term research direction.

This is not a real metric, but it reminds me that not every famous paper is equally useful at a particular moment.

For example:

• a paper may be extremely influential but far from my current experiments;
• a workshop paper may contain a practical ablation that directly helps my project;
• a survey may be valuable for positioning, even if it has no implementation details.

The goal is not to read everything. The goal is to read the right things deeply enough.

5. Reading for Implementation

A paper becomes much more concrete when I ask how I would implement it.

I usually write an implementation checklist:

1. What data preprocessing is required?
2. What tensors are passed between modules?
3. What is the shape of each major representation?
4. What losses are needed?
5. Which baseline must be reproduced first?
6. Which ablations are essential?
7. What logging is needed to debug the method?

For 3D perception, tensor shapes are especially important.

An occupancy model may move through:

If I cannot write down these transformations clearly, I probably do not understand the method well enough to build on it.

6. Reading for Ablations

Ablation studies reveal what the authors believe is important.

When reading ablations, I ask:

• Which module gives the largest gain?
• Does the gain still hold under stronger baselines?
• Is the improvement caused by the proposed idea or by extra compute?
• Are communication, latency, and memory measured fairly?
• Are there missing ablations that would challenge the claim?

For my work on token communication, important ablations include:

• fixed Top-K tokens vs. adaptive token selection;
• dense feature sharing vs. token communication;
• with or without temporal memory;
• sender-driven vs. receiver-driven communication;
• different bandwidth budgets;
• performance under occlusion and sparse-agent settings.

Good ablations are not just supporting evidence. They are also a map for future work.

7. Reading for Research Taste

Research taste is the ability to judge which problems are worth solving.

Paper reading helps build this ability, but only if I read actively.

I try to notice:

• whether the problem is fundamental or only benchmark-specific;
• whether the method simplifies the problem in an elegant way;
• whether the contribution would still matter if the dataset changed;
• whether the idea connects to a broader direction;
• whether the limitations suggest a natural next step.

In my current direction, I am increasingly drawn to papers that connect perception with constraints:

• limited bandwidth;
• partial observability;
• temporal uncertainty;
• deployment cost;
• downstream planning value.

These constraints make research more difficult, but also more meaningful.

8. Closing Thoughts

The best paper reading habit I have found is to always produce an artifact.

After reading a paper, I want at least one of the following:

• a structured note;
• an implementation checklist;
• an ablation idea;
• a diagram;
• a research question;
• a paragraph that can later enter my research statement.

This changes paper reading from passive consumption into active research preparation.

For Ph.D. applications, this matters because the goal is not to show that I have read many papers. The goal is to show that I can identify problems, reason about methods, design experiments, and grow a coherent research direction.