Conference Watch

ConferenceWatch

Produce an accurate, well-sourced snapshot of the AI/ML conferences a user cares about — with every submission deadline stage, dates, location, special sessions / CFP, official links, and a 5-year acceptance-rate trend.

Two artifacts are always produced:

1. A JSON dataset (one object per conference) — the structured source of truth.
2. A human-readable Markdown report with a short summary.

Guiding principles

• Never invent dates. A deadline goes into the JSON only if you found it on an official/authoritative source. If you infer from historical patterns, mark it approximate and say so explicitly in the report.
• Cite everything. Every edition and every acceptance-rate number carries a source_links entry. Prefer official conference sites and CFP pages; use aggregators (see references) as leads, then confirm on the official page.
• Timezones matter. Record the raw deadline text (e.g. "May 22, 2026 AoE"). Most AI venues use AoE (Anywhere on Earth, UTC-12). Keep the AoE flag.
• Today is the anchor. Use the current date to decide which edition is "upcoming" vs "past". Do not assume the model's training-cutoff year.

Workflow

Step 0 — Establish the anchor date

Determine today's date (from the environment/context). All "future vs past" decisions are relative to it. State the anchor date in the report metadata.

Step 1 — Analyze input & narrow the scope

Parse the user's request for any of these filters:

• Area / topic — e.g. NLP, CV, ML theory, RL, robotics, HCI, systems, data mining, speech, multimodal, AI4Science.
• Reputation / tier — CORE rank (A*/A/B) or "top-tier only".
• Difficulty — proxied by acceptance rate (e.g. "only <25% venues").
• Named conferences — if the user already lists venues, skip to Step 2.
• Region / timing — e.g. "conferences with deadlines in the next 6 months", or "held in Europe".

If the request is under-specified, ask 1–3 concise follow-up questions to narrow it (offer sensible defaults so the user can decline easily). Example:

To focus the search, which of these should I use?

1. Area — all AI, or a subfield (NLP / CV / RL / …)?
2. Tier — top-tier (CORE A*) only, or include A/B venues?
3. Horizon — only deadlines in the next N months, or the full year? (If you'd rather not specify, I'll default to the top AI venues across areas.)

Default when the user declines or doesn't answer: the top general AI venues plus the leading venue(s) for any area they mentioned. See references/conference-catalog.md for the curated tier list to draw from.

Step 2 — Build the initial conference list (initial JSON)

From the filters, assemble the candidate list and write the initial JSON using the schema in references/json-schema.md and the template assets/conference-data.template.json. At this stage only metadata + conference names/areas/tier need to be filled; edition/deadline/trend fields are placeholders to fill in Step 3–4.

Confirm the list with the user briefly ("I'll investigate these N conferences: …") before the (potentially long) search phase, unless they asked you to just go.

Step 3 — Research each conference: the UPCOMING edition first

For each conference, search for the next (future) edition relative to the anchor date. Query patterns that work well:

• "<CONF> <year> call for papers", "<CONF> <year> important dates", "<CONF> <year> submission deadline", "<CONF> <year> paper deadline".
• Then open the official site / CFP page to confirm; aggregators can be stale or wrong.

When exact info is found, record in the conference's upcoming_edition:

• year, location (city, country; note if virtual/hybrid), venue if known.
• dates: conference start/end.
• deadlines: every stage — abstract registration, full paper, supplementary, rebuttal, author response, notification, camera-ready, workshop/tutorial proposals, etc. Keep the raw text + AoE flag + confirmed: true.
• special_sessions: special tracks, new-this-year themes, datasets & benchmarks track, position papers, findings, industry track, journal-to-conference, etc.
• call_for_papers_url and website.
• data_confidence: "confirmed" and source_links.

If no future edition is announced yet, leave upcoming_edition.data_confidence as "approximate" and fill it via inference in Step 4.

Step 4 — Historical anchor: last 5 editions → trend + inference

For each conference, gather the last 5 editions (most recent past years):

• Populate acceptance_rate_trend with {year, submissions, accepted, acceptance_rate, source} per year — as many of the 5 as are available.
• If the upcoming edition's deadlines were not found in Step 3, infer the approximate deadline window from the historical pattern (e.g. "abstract deadline has fallen in the third week of May for the last 4 years → est. ~mid-May "). Write these into upcoming_edition.deadlines with confirmed: false and set data_confidence: "approximate". Always flag inferred dates clearly in both JSON and report.

Step 5 — Finalize JSON

Write the complete JSON to conferences.<anchor-date>.json (or a path the user specifies). Validate it against the schema: every conference has a name, tier, website; every deadline has a stage + confirmed flag; every trend/edition row that carries data also carries a source link.

Step 6 — Produce outputs

1. Answer directly in chat: a compact table or list of the most time-sensitive items (nearest deadlines first), plus the short summary.
2. Write the Markdown report using assets/report.template.md:
   • A 3–5 sentence summary synthesizing what you found (nearest deadlines, notable new tracks, acceptance-rate direction).
   • A per-conference breakdown (deadlines table, location, special sessions, links, acceptance-rate trend).
   • A clearly-labeled "Approximate / inferred" section for anything not confirmed.
   • A Sources list.

Output file naming

• Data: conferences.<YYYY-MM-DD>.json
• Report: conference-report.<YYYY-MM-DD>.md (Use the anchor date so successive runs are comparable / diffable.)

Quality checklist before finishing

• Anchor date stated; upcoming vs past decided against it.
• Every confirmed deadline has a source link; timezone/AoE captured.
• Inferred/approximate items are flagged in both JSON and report.
• Acceptance-rate trend covers up to 5 recent years with sources.
• JSON validates against the schema; report summary written.
• Missing data is shown as null with a note — never fabricated.