Compare Papers (mosaic compare)

mosaic compare generates a structured comparison table across a set of cached papers. When an LLM is configured, it extracts researcher-defined dimensions (method, dataset, metric, result by default) from each paper's title and abstract. Without an LLM, only metadata fields (year, source, journal, DOI) are populated.

Quickstart

# Compare top-cited cached papers on diffusion models (LLM required for method/dataset/metric/result)
mosaic compare --query "diffusion models" --sort citations -n 15

# Save to Markdown
mosaic compare --query "transformer attention" --output comparison.md

# Compare papers from a BibTeX file along custom axes
mosaic compare --from refs.bib --dimensions "method,dataset,BLEU,limitations"

# Export as CSV or JSON
mosaic compare --query "protein folding" --output comparison.csv
mosaic compare --query "protein folding" --output comparison.json

Options

Option	Short	Default	Description
--query TEXT	-q		Filter papers from the cache by title/abstract
--from FILE			Load papers from a .bib or .csv file
--max INT	-n	20	Maximum number of papers to compare
--dimensions TEXT		method,dataset,metric,result	Comma-separated comparison axes
--output FILE	-o		Write table to file (.md, .csv, .json)
--sort TEXT			Pre-sort papers: citations (most cited first) or year (newest first)

How it works

Paper selection

Papers are taken from the local cache in three ways:

1. --query — full-text search on title and abstract.
2. --from FILE — DOIs extracted from a .bib or .csv file, then looked up in the cache.
3. No filter — all cached papers, up to --max.

After filtering, papers are optionally sorted by citation count or year and truncated to --max.

LLM extraction

When [llm] api_key and [llm] provider are set in the config, MOSAIC sends batches of 20 papers to the LLM with a structured prompt:

For each paper below extract the following dimensions: "method", "dataset", "metric", "result".
Return a JSON array of exactly N objects. Each object must have exactly these keys: …
Use "-" when a dimension cannot be inferred from the title or abstract.

The LLM fills in each field from the title and abstract (up to 300 characters). Unavailable fields are marked –.

Configure LLM

mosaic config --llm-provider openai --llm-api-key sk-...
# or Anthropic:
mosaic config --llm-provider anthropic --llm-api-key sk-ant-...

Metadata-only fallback

Without an LLM, or when the LLM call fails, MOSAIC falls back to extracting fields that are directly available in the cache:

Dimension name	Extracted from
year	paper.year
source	paper.source
journal	paper.journal
doi	paper.doi
authors	paper.short_authors
citations / citation_count / cited	paper.citation_count
anything else	–

Output formats

Extension	Format
terminal (default)	Rich table
.md / .markdown	Markdown table
.csv	CSV with header row
.json	JSON array, one object per paper

The terminal table is always printed; --output saves a copy to disk in addition.

Examples

Terminal output

mosaic compare --query "diffusion models" --sort citations -n 5

 #  Title                          Year  Authors          Method              Dataset      Metric  Result
 1  DDPM                           2020  Ho et al.        Denoising diffusion  CIFAR-10    FID     3.17
 2  Stable Diffusion               2022  Rombach et al.   Latent diffusion     LAION-5B    FID     12.6
 3  DALL-E 2                       2022  Ramesh et al.    Hierarchical GLIDE   COCO        FID     10.39
…

Markdown table

| # | Title | Year | Authors | Method | Dataset | Metric | Result |
|---|-------|------|---------|--------|---------|--------|--------|
| 1 | DDPM  | 2020 | Ho et al. | Denoising diffusion | CIFAR-10 | FID | 3.17 |

Custom dimensions

mosaic compare --from refs.bib --dimensions "task,architecture,training_data,main_limitation"

Any dimension name works; the LLM will attempt to extract it from the abstract.

Limitations

• Accuracy depends on what is stated in the abstract. Full-text extraction is not yet supported.
• Batching is limited to 20 papers per LLM call; larger sets take proportionally longer.
• Very long tables (>50 rows) can be hard to read in the terminal — use --output to save as Markdown or CSV.

Workflow example

# Step 1: collect papers
mosaic search "graph neural networks" -n 50

# Step 2: compare top-cited methods
mosaic compare --query "graph neural networks" --sort citations -n 20 \
               --dimensions "task,gnn_type,dataset,accuracy" \
               --output gnn-comparison.md

# Step 3: visualise the citation network of the same corpus
mosaic network --query "graph neural networks" --cluster --output gnn-network.md