Week 08
Paper Draft Complete

• [ ] Complete full paper draft
• [ ] Create all figures and tables
• [ ] Prepare code repository structure
• [ ] Begin code documentation

1. Paper Draft Complete

All sections written:

Introduction (1.5 pages):

• Query-document asymmetry problem framed
• HyDE critique (hallucination in specialized domains)
• QUID as "semantic anchoring" positioned
• Contributions clearly stated

Related Work (1.5 pages):

• Query expansion: Rocchio → RM3 → doc2query → HyDE
• Dense retrieval: DPR → Contriever → BGE
• Text diffusion: D3PM → MDLM → LLaDA
• Gap identified: no diffusion for retrieval

Method (1.5 pages):

• Problem formulation with equations
• QUID algorithm pseudocode
• Retrieval pipeline description

Experiments (3 pages):

• Datasets and baselines (Table 1)
• Main results (Table 2)
• Ablations (Table 3)
• Analysis with figures

Conclusion (0.5 pages):

• Contributions summarized
• Limitations acknowledged
• Future work suggested

2. Figures Created

• Figure 1: QUID architecture diagram
• Figure 2: Bar chart - nDCG@10 across methods/datasets
• Figure 3: Scatter plot - semantic drift vs. improvement
• Figure 4: Quality vs. latency tradeoff curve

3. Tables Finalized

• Table 1: Dataset statistics
• Table 2: Main results (7 methods × 4 datasets)
• Table 3: Ablation summary

4. Code Repository Structure

QUID/
├── README.md # Installation, usage, citation
├── requirements.txt # Dependencies
├── LICENSE # MIT License
├── src/
│ ├── quid.py # Main QUID class
│ ├── diffusion.py # LLaDA generation
│ ├── retrieval.py # BGE-M3 retrieval
│ └── evaluation.py # BEIR evaluation
├── experiments/
│ ├── run_beir.py # Reproduce main results
│ ├── ablations.py # Reproduce ablations
│ └── analysis.py # Mechanism analysis
└── configs/
 └── default.yaml # Hyperparameters

1. Paper Draft Complete

All sections written:

Introduction (1.5 pages):

• Query-document asymmetry problem framed
• HyDE critique (hallucination in specialized domains)
• QUID as "semantic anchoring" positioned
• Contributions clearly stated

Related Work (1.5 pages):

• Query expansion: Rocchio → RM3 → doc2query → HyDE
• Dense retrieval: DPR → Contriever → BGE
• Text diffusion: D3PM → MDLM → LLaDA
• Gap identified: no diffusion for retrieval

Method (1.5 pages):

• Problem formulation with equations
• QUID algorithm pseudocode
• Retrieval pipeline description

Experiments (3 pages):

• Datasets and baselines (Table 1)
• Main results (Table 2)
• Ablations (Table 3)
• Analysis with figures

Conclusion (0.5 pages):

• Contributions summarized
• Limitations acknowledged
• Future work suggested

2. Figures Created

• Figure 1: QUID architecture diagram
• Figure 2: Bar chart - nDCG@10 across methods/datasets
• Figure 3: Scatter plot - semantic drift vs. improvement
• Figure 4: Quality vs. latency tradeoff curve

3. Tables Finalized

• Table 1: Dataset statistics
• Table 2: Main results (7 methods × 4 datasets)
• Table 3: Ablation summary

4. Code Repository Structure

QUID/
├── README.md # Installation, usage, citation
├── requirements.txt # Dependencies
├── LICENSE # MIT License
├── src/
│ ├── quid.py # Main QUID class
│ ├── diffusion.py # LLaDA generation
│ ├── retrieval.py # BGE-M3 retrieval
│ └── evaluation.py # BEIR evaluation
├── experiments/
│ ├── run_beir.py # Reproduce main results
│ ├── ablations.py # Reproduce ablations
│ └── analysis.py # Mechanism analysis
└── configs/
 └── default.yaml # Hyperparameters

1. Page limit pressure: Had to move some analysis to appendix.

1. Figure sizing: Balanced readability vs. space with 2-column figures.

• Complete draft is crucial milestone for iteration
• Algorithm pseudocode makes method reproducible
• Code release preparation takes significant effort

• Revise paper based on self-review
• Complete code documentation
• Verify reproducibility