By a pre-event tally, 20.8% of IJCAI-ECAI 2026's accepted main-track papers are the work the LLM era calls 'reasoning'

In 2019, Nanjing University researcher Yu Yang said IJCAI was the best home for reasoning, planning, and knowledge work in AI, and that those directions were what the field most needed. Seven years later, the 2026 program of that same conference, run jointly with ECAI in Bremen, Germany from August 15-21, accepted 990 papers, and 148 of the accepted main-track count (20.8%) sit in the cluster Yu Yang was talking about (leiphone tally).

The cluster is the part of AI that predates the chatbot era. It covers rule-based logical reasoning, classical planning, knowledge representation (storing facts and rules so machines can reason over them), and constraint solving, the formal methods tradition that built the field before neural networks took over. The LLM industry has spent the last two years re-marketing that work as its frontier, calling it "reasoning" or "chain-of-thought" without inheriting the institutional home that actually publishes it.

IJCAI, the International Joint Conference on Artificial Intelligence, has run since 1969 and is one of the field's two oldest general-purpose venues. ECAI, the European Conference on Artificial Intelligence, is its European counterpart. Every other year the two merge; 2026 is one of those years, and the merged program is what the 148-paper count comes from.

The 148 number needs a caveat. The 82-paper figure that anchored the original analysis counts only two sub-areas: Knowledge Representation and Reasoning (67 papers) and Planning and Scheduling (32 papers), which together make up 13.3% of accepted main-track papers. Add Constraint Satisfaction and Optimization plus Search, and the cluster reaches 148 papers at 20.8%. Both figures are pre-event tallies from the Chinese-language AI trade outlet AI科技评论 (leiphone), not the official IJCAI-ECAI proceedings. The conference has not yet published its final accepted-paper list, and the run happens Aug 15-21.

The work in the cluster is not nostalgic. NDProp, a paper by Eiter, Inoue, and Moriyama in this year's program, makes Answer Set Programming (a long-standing formal language for rule-based reasoning) differentiable, so it can be trained with gradient descent. The authors report their solver runs about 42 times faster than NeurASP and 7 times faster than SLASH on MNIST arithmetic benchmarks, while matching the semantics of an earlier differentiable solver, DProp. The code is public. It is a result that does not get a product launch but does change what symbolic methods can plug into.

The institutional depth is not just headcount. Recent IJCAI distinguished paper awards have repeatedly gone to reasoning, planning, and knowledge work, suggesting the 148 papers are not a coincidence of submission mix. The SOTA in adjacent solver tracks tells a similar story. AE-Kissat-MAB won the SAT Competition 2025 main track. SAT, or Boolean satisfiability, is the problem of finding inputs that make a logical formula true, the bedrock of automated reasoning. SATLUTION, a 2025 preprint with code, reports the lowest average solving time against the 2025 winners.

The counter-argument is real. LLM-native reasoning research, the kind OpenAI, Anthropic, and DeepMind publish under the "reasoning model" label, is now the most cited work in the field, and IJCAI's 148 papers do not include any of it. A 2026 reader who only reads the marketing will reasonably conclude that the LLM lab is where reasoning research lives. The IJCAI-ECAI 2026 program is evidence that the institutional center has not moved, even as the marketing has. ECAI 2024's own proceedings scope the venue to planning, search, constraints, satisfiability, and knowledge representation, the same five areas the 2026 cluster covers.

The official proceedings drop when the conference opens on August 15. The final 148 number, and the exact share that lands in reasoning, planning, and knowledge, will be settled then.