somaxsoma/iclr2026-ai-safety

This is a small Python project that surveys the ICLR 2026 conference to find papers about AI safety. ICLR is one of the main academic conferences for machine learning research, and OpenReview is the public site where ICLR papers are submitted and reviewed. The repo's idea is to pull every submission, ask a language model whether each one is safety research, sort the matches by topic, and then produce a chart. The README describes the workflow as four scripts that run in order. The first one, fetch_papers.py, talks to the OpenReview API and writes every ICLR 2026 submission into a CSV file called iclr2026_papers.csv. The second one, classify.py, goes through that CSV row by row, asks an LLM to decide whether each paper is AI safety research and which sub-area it falls into, then writes the labelled rows to safety_results.csv. The third script, analyze.py, takes the labelled CSV and aggregates the counts so you can see how many ICLR 2026 papers fall into each safety sub-area. The output is a single PNG chart, ai_safety_iclr2026.png, which the README embeds. The fourth script, download_pdfs.py, downloads the actual paper PDFs for the entries that were classified as safety, organising them on disk by sub-area so a reader can browse the literature directly. The README is short and does not list dependencies, an LLM provider, an OpenReview authentication method, the exact list of sub-areas, or instructions for running the pipeline. There is no install section, no example output beyond the chart image, and no explanation of how the classifier prompt is written. Anyone wanting to reproduce the analysis would need to read the four Python files to fill in those gaps. In practical terms, the repo is best understood as the author's own one-off analysis pipeline that happens to be public, rather than a packaged tool. It is useful as a starting template if you want to do similar conference-wide topic surveys on OpenReview data, or if you want to see how the author wired up the LLM classification step. The result image in the README is the headline deliverable: a snapshot of how much of ICLR 2026 is safety work and how it is distributed across subdomains.