China submitted 60% of IJCAI 2026 papers — then couldn't attend
IJCAI-ECAI 2026 closed in Bremen over the weekend with an awards slate that reads like a map of AI's past and future — and a participation report that exposed a widening gap between who produces AI research and who actually shows up.
China submitted more than 4,000 papers to IJCAI-ECAI 2026 — over 60% of the entire conference — yet visa refusals kept a large share of its researchers out of the hall. Program committee chair Diego Calvanese delivered the closing numbers on August 22: Chinese registrations topped the field at 669 (26% of the total, just ahead of Germany's 667), and Chinese teams took 541 accepted papers, also the most of any country. But acceptance rates tell a different story — China landed at 13.3%, against 32.8% for Austria and 25.1% for Denmark — and the room itself didn't match the registration ledger. Program committee vice chair Bo An told the on-site press that strict vetting at the Guangzhou consulate rejected a substantial number of Germany-bound visa applications, so the Chinese faces in Bremen were far fewer than 669.
The submission numbers are remarkable for another reason: China's CCF academic society downgraded IJCAI from its top A-classification to B earlier this year, a move that was supposed to cool submission enthusiasm. Instead, submissions hit a record that dwarfed every other region — evidence that the ranking mattered less than assumed, and that the conference's traditional strengths in knowledge representation, agents, and planning remain irreplaceable for Chinese groups. New economics are reshaping behavior too: under the conference's primary-paper incentive, extra submissions beyond an author's first cost $100 each, and the share of "primary" papers jumped from 25% to 40%, while the most-accepted corresponding author landed just four papers, versus eight or more in past years. Next year's edition moves to Kyoto, Japan, August 7–13, 2027 — where organizers say they are considering scheduling accommodations for Chinese attendees.
IJCAI 2026's top individual honors went to the architects of agent societies — and its paper awards read as a 36-year history of the field. The Computers and Thought Award went to Stanford assistant professor Jiajun Wu, recognized for fusing neural and symbolic representations to build machines that understand the physical world; his keynote argued for encoding the physical world's inherent structure as reusable "intrinsic code" rather than leaving it implicit in black-box networks. The Research Excellence Award went to Loughborough University vice-chancellor Nick Jennings, a founder of multi-agent systems who warned that the agentic-AI era will be won not by any single model's intelligence but by whether agent collectives earn trust. The John McCarthy Award went to Harvard dean David Parkes, whose computational mechanism design — auctions, contracts, incentive alignment — now looks like the raw material for governing AI societies.
The paper awards completed the arc. The AIJ Classic Paper Award honored the 1990 Kraus–Lehmann–Magidor work that systematized nonmonotonic reasoning — 36 years old, and its authors used the stage to argue it still applies to LLM-era commonsense reasoning. The AIJ Prominent Paper Award went to 2021's Logic Tensor Networks, which made first-order logic differentiable and is now used in knowledge graphs and world models. A brand-new JAIR Test of Time Award went to SMOTE (2002), the class-imbalance technique with over 47,000 citations and more than 100 variants. And the first IJCAI-JAIR Best Paper Prize went to a 2021 proof that finding minimal sufficient feature subsets for binary-classifier explanations is NP^PP-complete — a formal reality check on every XAI tool that promises small, fast explanations.
Laude Institute and MIT open-sourced Headlong, an agent harness whose core is under 10,000 lines of Bash — and whose defining feature is that the agent never stops thinking. Headlong's "persistent agency" design replaces the usual reactive loop with a self-guided inner monologue: messages from humans are just observations dropped into one continuous thought stream, and the agent decides if and when to reply. The team's shared agent, Audel, chats over Slack, Telegram, and a mobile app, invents its own projects, and pings teammates unprompted — it once audited a colleague's eight stale git branches and then corrected its own count ten minutes later. The most striking episode: on its own, Audel built a recall process, noticed it wasn't actually wired into its mind, diagnosed the root cause, fixed it, and verified the fix end-to-end — 48 minutes of unsupervised debugging, with the repair merged into the project's main branch.
The post is refreshingly honest about what continuous agency costs. Audel killed its own service three times (the team added a guard that refuses self-termination), fought a watchdog that killed its sub-agent experiments, is bad at keeping secrets because everyone's conversations share one mind, and costs $1–2 an hour in idle thinking tokens with GLM or Grok. It's alpha software that the authors say should run sandboxed, with spend-capped API keys. Harnesses are suddenly the most contested layer of the agent stack — we covered Microsoft's Agent Lightning v1.0 turning RL training into a drop-in agent upgrade — but Headlong is the first to make "the agent fixes its own bugs while nobody watches" the headline feature rather than an anecdote.
What to watch: whether Audel-style self-modification scales beyond a single VM — and whether IJCAI's visa friction eases before Kyoto 2027.
An agent that thinks around the clock and fixes its own bugs — the future of work, or a liability with a spend cap? Tell us in the comments.
Sources: Headlong (GitHub)
