literature-review

---

name: literature-review description: Generate structured literature reviews on research topics by synthesizing findings from multiple papers. Use when conducting research surveys, identifying themes and gaps, or preparing academic writing on any scientific topic. allowed-tools: Bash, Read, Grep, Glob, TodoWrite

<quick_start> To generate a literature review:

Copy/asta:literature-review "federated learning for healthcare applications"

Optional depth levels:

• --depth quick (5-10 papers, fast overview)
• --depth standard (15-25 papers, balanced)
• --depth comprehensive (30+ papers, thorough) </quick_start>

<success_criteria>

• Identifies and synthesizes key papers on the topic
• Groups findings into coherent themes
• Notes methodological approaches and trends
• Identifies gaps and future research directions
• Provides proper citations for all claims </success_criteria>

Depth Guidelines:

1. Use mcp__asta__snippet_search with the main topic to find relevant passages
2. Identify key papers from the results
3. Use mcp__asta__get_paper on top papers to get abstracts and TLDRs

Phase 2: Citation Expansion

1. For seminal papers (high citation count), use mcp__asta__get_paper_citations to find follow-on work
2. Use mcp__asta__get_paper_references on key papers to identify foundational work
3. Track the citation network to understand field structure

Phase 3: Synthesis

1. Group papers into thematic clusters based on:
   • Methodological approach
   • Application domain
   • Theoretical contribution
2. Identify common findings and consensus
3. Note contradictions or ongoing debates
4. Identify gaps where research is sparse

Phase 4: Reporting

1. Structure findings into clear sections
2. Provide citations for all claims
3. Highlight most influential papers
4. Suggest future research directions

<output_format> Structure the review as:

Copy# Literature Review: {Topic}

## Overview
{Brief introduction to the field and scope of this review - 1-2 paragraphs}

## Key Themes

### {Theme 1 Name}
{Summary of papers and findings related to this theme}
- {Paper1} ({Year}): {Key finding}
- {Paper2} ({Year}): {Key finding}

### {Theme 2 Name}
{Summary}

## Methodological Approaches
{Common methods, datasets, evaluation metrics across the field}

## Key Findings
{Synthesized findings noting consensus and disagreements}

## Research Gaps
{Identified gaps and opportunities for future work}

## Seminal Papers
{List of most influential papers to read first, with brief justification}

## References
{Formatted citations for all papers mentioned}

</output_format>

Overview

Explainability in deep learning has emerged as a critical research area, driven by the need to understand and trust AI systems in high-stakes applications. This review covers key approaches including attention visualization, feature attribution, and concept-based explanations.

Key Themes

Gradient-Based Methods

Post-hoc explanation methods that compute input feature importance:

• Simonyan et al. (2014): Introduced saliency maps via gradient computation
• Sundararajan et al. (2017): Integrated Gradients with axiomatic foundations
• Selvaraju et al. (2017): Grad-CAM for visual explanations in CNNs

Attention-Based Interpretability

Using attention weights as explanations:

• Bahdanau et al. (2015): Attention mechanism for sequence-to-sequence models
• Jain & Wallace (2019): Challenged attention as faithful explanation

Research Gaps

• Limited work on explanations for generative models
• Lack of standardized evaluation metrics for explanation quality
• Few studies on user comprehension of explanations

Seminal Papers

1. "Attention Is All You Need" - Foundation of attention mechanisms
2. "LIME" - Model-agnostic local explanations
3. "Integrated Gradients" - Axiomatic attribution method

<anti_patterns>

• Don't fabricate papers - Only cite papers found through Asta tools
• Don't oversimplify - Capture nuance and disagreements in the field
• Don't ignore recency - Balance seminal work with recent developments
• Don't skip citations - Every claim should be backed by a specific paper </anti_patterns>