Coreference resolution remains challenging for pronouns with distant or 
ambiguous antecedents. Prior neural approaches struggle with these 
difficult cases due to limited context modeling. We introduce 
LongContext-Coref, a retrieval-augmented coreference model that 
dynamically retrieves relevant context from document history. On the 
OntoNotes 5.0 benchmark, LongContext-Coref achieves 83.4 F1, improving 
over the previous state-of-the-art by 1.2 points. On the challenging 
WinoBias dataset, we reduce gender bias by 34% while maintaining 
accuracy. Qualitative analysis reveals that our model successfully 
resolves pronouns requiring world knowledge, a known weakness of 
prior approaches.

├── Introduction
│   ├── Task motivation
│   ├── Prior work limitations
│   ├── Your contribution
│   └── Contribution bullets
├── Related Work
├── Method
│   ├── Problem formulation
│   ├── Model architecture
│   └── Training procedure
├── Experiments
│   ├── Datasets (with statistics)
│   ├── Baselines
│   ├── Main results
│   ├── Analysis
│   │   ├── Error analysis
│   │   ├── Ablation study
│   │   └── Qualitative examples
│   └── Human evaluation (if applicable)
├── Discussion / Limitations
└── Conclusion

Table 3: Human Evaluation Results (100 samples, 3 annotators)
─────────────────────────────────────────────────────────────
Method        | Fluency | Coherence | Factuality | Overall
─────────────────────────────────────────────────────────────
Baseline      |   3.8   |    3.2    |    3.5     |   3.5
GPT-3.5       |   4.2   |    4.0    |    3.7     |   4.0
Our Method    |   4.4   |    4.3    |    4.1     |   4.3
─────────────────────────────────────────────────────────────
Inter-annotator κ = 0.72. Scale: 1-5 (higher is better).

Video calling has become essential for remote collaboration, yet 
current interfaces poorly support the peripheral awareness that makes 
in-person work effective. Through formative interviews with 24 remote 
workers, we identified three key challenges: difficulty gauging 
colleague availability, lack of ambient presence cues, and interruption 
anxiety. We designed AmbientOffice, a peripheral display system that 
conveys teammate presence through subtle ambient visualizations. In a 
two-week deployment study with 18 participants across three distributed 
teams, AmbientOffice increased spontaneous collaboration by 40% and 
reduced perceived isolation (p<0.01). Participants valued the system's 
non-intrusive nature and reported feeling more connected to remote 
colleagues. We discuss implications for designing ambient awareness 
systems and the tension between visibility and privacy in remote work.

├── Introduction
│   ├── Problem in human terms
│   ├── Why technology can help
│   └── Contribution summary
├── Related Work
│   ├── Domain background
│   ├── Prior systems
│   └── Theoretical frameworks
├── Formative Work (often)
│   ├── Interviews / observations
│   └── Design requirements
├── System Design
│   ├── Design rationale
│   ├── Implementation
│   └── Interface walkthrough
├── Evaluation
│   ├── Study design
│   ├── Participants
│   ├── Procedure
│   ├── Findings (quant + qual)
│   └── Limitations
├── Discussion
│   ├── Design implications
│   ├── Generalizability
│   └── Future work
└── Conclusion

├── Introduction
├── Related Work
├── Methods
│   ├── Participants
│   ├── Procedure
│   ├── Data collection
│   └── Analysis method (thematic, grounded theory, etc.)
├── Findings
│   ├── Theme 1 (with quotes)
│   ├── Theme 2 (with quotes)
│   └── Theme 3 (with quotes)
├── Discussion
│   ├── Implications for design
│   ├── Implications for research
│   └── Limitations
└── Conclusion

Participants valued the ambient nature of the display. As P7 described: 
"It's like having a window to my teammate's office. I don't need to 
actively check it, but I know they're there." This passive awareness 
reduced the barrier to initiating contact.

**Implication 1: Support peripheral awareness without demanding attention.**
Ambient displays should be visible in peripheral vision but not require 
active monitoring. Designers should consider calm technology principles.

**Implication 2: Balance visibility with privacy.**
Users want to share presence but fear surveillance. Systems should 
provide granular controls and make visibility mutual.

Fraud detection in e-commerce requires processing millions of 
transactions in real-time while adapting to evolving attack patterns. 
We present FraudShield, a graph neural network framework for real-time 
fraud detection that scales to billion-edge transaction graphs. Unlike 
prior methods that require full graph access, FraudShield uses 
incremental updates with O(1) inference cost per transaction. On a 
proprietary dataset of 2.3 billion transactions from a major e-commerce 
platform, FraudShield achieves 94.2% precision at 80% recall, 
outperforming production baselines by 12%. The system has been deployed 
at [Company], processing 50K transactions per second and preventing 
an estimated $400M in annual fraud losses. We release an anonymized 
benchmark dataset and code.

├── Introduction
│   ├── Problem and impact
│   ├── Technical challenges
│   ├── Your approach
│   └── Contributions
├── Related Work
├── Preliminaries
│   ├── Problem definition
│   └── Notation
├── Method
│   ├── Overview
│   ├── Technical components
│   └── Complexity analysis
├── Experiments
│   ├── Datasets (with scale statistics)
│   ├── Baselines
│   ├── Main results
│   ├── Scalability experiments
│   ├── Ablation study
│   └── Case study / deployment
└── Conclusion

Table 4: Scalability Comparison (runtime in seconds)
──────────────────────────────────────────────────────
Dataset     | Nodes  | Edges  | GCN   | GraphSAGE | Ours
──────────────────────────────────────────────────────
Cora        |  2.7K  |  5.4K  |  0.3  |    0.2    |  0.1
Citeseer    |  3.3K  |  4.7K  |  0.4  |    0.3    |  0.1
PubMed      | 19.7K  | 44.3K  |  1.2  |    0.8    |  0.3
ogbn-arxiv  | 169K   | 1.17M  |  8.4  |    4.2    |  1.6
ogbn-papers | 111M   | 1.6B   |  OOM  |   OOM     | 42.3
──────────────────────────────────────────────────────

Our contributions are:
• We identify [problem/insight]
• We propose [method name] that [key innovation]
• We demonstrate [results] on [benchmarks]
• We release [code/data] at [URL]