Drbellamy/icml-2026

ICML 2026 — the whole conference, in one queryable dataset

10,815 papers (5,872 main-conference + 4,943 workshop) with titles, abstracts, authors, session times, rooms, and PDF links — plus every tutorial, invited talk, and workshop. Scraped from the official ICML 2026 virtual site and OpenReview, organized so an AI agent can answer real questions about the conference in one shot.

📱 Navigating ICML from your phone? Open Claude Code (or any agent), paste this dataset id, and ask "build me a Tuesday schedule for someone into RL and diffusion." That's it.

One line to load it

import pandas as pd
df = pd.read_parquet("hf://datasets/Drbellamy/icml-2026/data/papers.parquet")

or with 🤗 datasets:

from datasets import load_dataset
papers = load_dataset("Drbellamy/icml-2026", "papers", split="train")

Using it from Claude Code (even on your phone)

One-time setup (30 seconds, doable from your phone). Claude Code cloud sandboxes don't allow HuggingFace's file CDN by default, so enable network access once:

1. Open claude.ai/code in your phone's browser
2. Tap the cloud icon → the settings gear
3. Set Network access → Full → Save changes

That's it — now start sessions from the Claude mobile app and they can reach this dataset. (Prefer an allowlist over "Full"? Choose Custom and add huggingface.co and *.hf.co.) Cloud sessions require a Pro, Max, or Team plan.

Then, in a Claude Code session, paste one line and ask a question:

Use the HuggingFace dataset Drbellamy/icml-2026 (config "papers", it's parquet).
Build me a personalized ICML Wednesday schedule — I care about RL infra,
agents, and efficient inference. Give me titles, times, and rooms.

The agent loads the parquet, filters, and hands you a plan.

Schema — papers config (one row per paper)

column	type	description
track	string	main_conference or workshop
paper_type	string	oral, spotlight, poster, or workshop
title	string	paper title
authors	list[string]	author names in order
abstract	string	full abstract
day	string	tuesday … saturday (Seoul time)
date	string	ISO date, e.g. 2026-07-07
session	string	session / workshop name
session_start_kst	string	ISO 8601, Asia/Seoul
session_end_kst	string	ISO 8601, Asia/Seoul
location	string	room / poster board
workshop	string	workshop name (null for main track)
url	string	ICML page
pdf	string	PDF link (OpenReview) where available
openreview	string	OpenReview forum link where available

Other configs: main_conference_full (all 6,341 accepted main-track papers with PDFs, author IDs, keywords, and spotlight/regular tier), events (180 tutorials / invited talks / expo / socials, with abstracts and speaker bios), workshops (44 workshops with organizers and summaries). Per-day JSON is also in by_day/ for people who want structure over tables.

Example questions to ask an agent

• "Find every oral paper mentioning diffusion and list its time and room."
• "Build my Tuesday: I care about RL and robotics. No overlaps."
• "Which authors have the most papers at ICML 2026?"
• "List all workshop papers on interpretability with their PDF links."
• "Cluster Wednesday's poster abstracts into themes and summarize each."
• "What are the most common keywords among spotlight papers?"

Recipes

Schedule builder — "Given I care about X and Y, plan each day with no time conflicts; output a table of title / time / room / type." Topic discovery — "Embed the abstracts and cluster into 15 topics; name each and count papers." Reading list — "Rank all posters by relevance to 'long-context efficiency' and give me the top 25 with PDF links." Social graph — "Who are the most prolific authors, and which pairs co-author most often?"

Reproduce it / scrape your own conference

The full pipeline is open-source: github.com/DavidBellamy/icml2026-scraper. Any conference's virtual site + OpenReview + an agent = a queryable conference in your pocket.

Coverage & provenance

• Source: the official ICML 2026 virtual site (icml.cc/virtual/2026) and OpenReview, July 2026.
• Papers with PDF/OpenReview link: 9,801 / 10,815. The rest are papers whose titles didn't cleanly match across sources, plus workshop invited talks (no paper/PDF exists).
• Times are in Asia/Seoul (KST), the conference timezone. The main_conference_full config also carries UTC where relevant.
• 1 workshop ("Game Theory in Nature") had not published its schedule at scrape time, so its papers are absent.
• This is a community scrape, refreshed during the conference. Counts may shift slightly as organizers finalize schedules.

License / rights

Titles, abstracts, and author information are the intellectual property of their respective authors and ICML. This dataset is shared for research and conference-navigation purposes. No paper PDFs are redistributed here — only links. If you are an author or organizer and want something changed, open a discussion on this dataset.

Citation

@misc{icml2026dataset,
  title  = {ICML 2026 Conference Dataset (papers, workshops, schedule)},
  author = {Bellamy, David},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Drbellamy/icml-2026}}
}

Built with Claude Code. Any conference scrape + HuggingFace + an agent = a queryable conference in your pocket.