82 篇论文撑起的判断：IJCAI 凭什么是推理、规划、知识的「第一主场」

TL;DR - A Chinese tech-media analysis of IJCAI-ECAI 2026's accepted-paper lineup, arguing the conference remains the top venue for symbolic reasoning, planning, and knowledge representation — and that these areas are being revitalized by LLM hybridization. It matters as a snapshot of where neurosymbolic AI research is concentrating as the field debates whether LLMs can actually reason.

• Scale: ~5,400+ submissions, 990 accepted (~18%). KRR (67 papers) + Planning & Scheduling (32) = 13.3% of main track; adding Constraint Satisfaction/Optimization and Search reaches 148 papers (~20.8%). The joint IJCAI/ECAI (Bremen, Aug 15–21) pairing reinforces the European symbolic-AI tradition.
• Reasoning: NDProp makes Answer Set Programming differentiable via a "Decision-Propagation" pipeline (RNNs replace false-value decisions, fuzzy-logic operators replace propagation), enabling end-to-end GPU training without supervision — reportedly ~42× faster than NeurASP and ~7× than SLASH on MNIST arithmetic. AESAT uses a mixture-of-experts LLM loop (GPT-4.5 ideation, Claude 3.7 coding, DeepSeek-R1 analysis) to evolve SAT branching heuristics; derived solver AE-Kissat-MAB won the 2025 SAT Competition main track.
• Planning: DUPLEX (Northeastern Univ. + Midea) restricts a lightweight LLM to structured information extraction into PDDL, delegating synthesis to a classical solver, with a slow-system LLM repair loop on failure. Amazon's SIPP-PP-LNS combines safe-interval priority planning with XGBoost-guided large neighborhood search, claiming 2–3 orders of magnitude speedup over diffusion planners at tens-of-milliseconds latency.
• Knowledge: CRIL (Artois) landed 7 KRR papers spanning ASP abstraction, argumentation-extension diversity, prime-implicant XAI for tree models, distributed model counting, and inconsistency measures; other work covers explainable ASP surveys and GEV's statically type-checked knowledge-graph updates.