SoftReason: A Fully Differentiable Neuro-Soft-Symbolic Deductive Reasoning Architecture over High-Dimensional Perceptual Data
SoftReason keeps perception-to-deduction reasoning differentiable instead of handing off to discrete symbols.
The paper describes a trainable architecture where perceptual inputs produce probabilistic facts, while knowledge-graph triples act as high-confidence soft evidence. Its main mechanism is a learned differentiable version of the immediate-consequence operator, updating a soft interpretation tensor through probabilistic closure. The author instantiates it for knowledge-aware visual question answering, tying visual grounding, KG evidence, and deductive reasoning into one model. ArXiv · AI/CL/LG's note
The paper describes a trainable architecture where perceptual inputs produce probabilistic facts, while knowledge-graph triples act as high-confidence soft evidence. Its main mechanism is a learned differentiable version of the immediate-consequence operator, updating a soft interpretation tensor through probabilistic closure. The author instantiates it for knowledge-aware visual question answering, tying visual grounding, KG evidence, and deductive reasoning into one model. ArXiv · AI/CL/LG's note
score 4