Benchmarks Are Engines, Not Cameras

This project explores how benchmarks shape ML research, drawing on Donald MacKenzie's framework of performativity from the sociology of finance. The core idea: ML benchmarks don't just measure progress — they actively influence the research they evaluate. When a benchmark becomes the target of optimization, it gradually shifts from measurement tool to optimization target, with consequences for statistical validity, proof standards, and the direction of the field.

The project covers the multiple comparisons problem in ML (and how corrections from medicine and genomics could be adopted), proof culture fragmentation in mechanistic interpretability (where SAE and causal abstraction communities share a foundation but have developed largely independently), and the dynamics of benchmark adoption and abandonment over time.

Slides

Slides (LaTeX source) — 46 slides covering performativity in ML benchmarks, multiple comparisons and FWER/FDR, proof culture fragmentation in mechanistic interpretability, and counterperformativity in peer review.

To compile:

cd slides
pdflatex benchmarks_engines_slides_v1.tex

Paper

Draft paper in progress (TBD). Current working version in paper/.

Data

File	Description	Source
data/iclr2024_demand_compliance.csv	Experiment demand and compliance signals for 7,404 ICLR 2024 submissions	OpenReview API
data/iclr2023_demand_compliance.csv	Same for 3,792 ICLR 2023 submissions	OpenReview API
data/openalex_citation_analysis.json	Benchmark lifecycle citation counts (ImageNet, GLUE, etc.)	OpenAlex API
data/mi_citation_network_final.json	Citation network among 27 canonical MI papers	Semantic Scholar API
data/mi_citation_corpus/	Expanded 688-paper MI citation corpus (209 SAE, 40 Causal, 128 Circuits, 311 general)	Perplexity deep research
data/perplexity_citation_data/	Independent citation validation dataset (27 papers)	Perplexity deep research
data/fwer_audit_landmark_papers.json	FWER audit of landmark ML papers	Manual audit
data/mi_audit_papers.json	MI paper evidence standard audit	Manual audit
data/self_audit_comparisons.json	Self-audit of implicit comparisons	Manual audit

Scripts

Script	What it does
scripts/citation_network_analysis.py	Analyze citation links among MI papers via Semantic Scholar
scripts/citation_network_openalex.py	Fetch benchmark lifecycle data from OpenAlex
scripts/generate_lifecycle_figures.py	Generate benchmark adoption/abandonment curves
scripts/gen_lifecycle_single.py	Generate single-panel lifecycle figure for slides

Reproducing

# Benchmark lifecycle figures
pip install matplotlib
python scripts/gen_lifecycle_single.py

# Citation network analysis (requires Semantic Scholar API key)
pip install requests
python scripts/citation_network_analysis.py

Some data was sourced via Perplexity deep research and is included directly in data/ rather than being reproducible via script.

License

MIT

Citation

Tower, E. (2026). Benchmarks Are Engines, Not Cameras. Zenodo. https://doi.org/10.5281/zenodo.20815419