arXiv cs.AI
·
18 hours ago
Researchers introduced a method using Inductive Logic Programming to extract symbolic representations of reinforcement learning policies and defined four explainability metrics: activation rate, feature coverage, syntactic distance, and semantic distance. The proposed metrics were tested across multiple RL domains and demonstrated capability to reveal action-specific learning dynamics, fine-grained feature importance, and coordination patterns in multi-agent RL. This work advances explainable reinforcement learning by providing objective, quantifiable measures of policy interpretability beyond user studies and subjective assessments.
arXiv cs.AI
·
18 hours ago
Researchers developed a Structured Multi-Agent RCA pipeline to identify root causes in microservice failures using telemetry data, showing that existing classical and LLM-based methods fail on the OpenRCA benchmark. The system outperformed baselines by analyzing whether failures stem from reasoning gaps or data ambiguity, revealing that evidence is present in the vast majority of cases but models struggle to reason over it correctly. The findings indicate that improving root cause analysis requires better model reasoning capabilities rather than just better data pipelines or prompt engineering.
arXiv cs.AI
·
18 hours ago
Researchers argue that autoformalization efforts should scale from individual statements to complete theories with all dependencies and definitions included in structured libraries. Current autoformalization work typically handles isolated statements, but real formalization requires formalizing entire webs of axioms, definitions, and lemmas as interdependent systems. This shift requires developing new approaches to handle theory-level complexity and maintaining consistency across formal knowledge bases.
arXiv cs.AI
·
18 hours ago
Researchers extended neuro-symbolic AI based on Belnap's intensional first-order logic by incorporating probabilistic reasoning for unknown sentences using Nilsson's probability structure. The approach introduces global and local symmetry transformations to preserve knowledge while computing probability density functions through neural networks using maximum entropy principles. This enables AGI systems to combine neural learning with symbolic reasoning while handling uncertainty in logical reasoning tasks.
