Why this direction matters

Neural models are excellent at representation learning, while symbolic methods are strong in compositional reasoning and constraint satisfaction. Most systems still bind these two worlds at software level, which introduces significant latency and energy overhead.

Core idea

Design a heterogeneous accelerator where:

  • Tensor array clusters handle dense embedding computation.
  • Rule-matching engines execute sparse symbolic predicates.
  • Shared semantic memory supports low-latency exchange of latent vectors and logical tokens.

The key hypothesis: by moving symbolic primitives into hardware-adjacent units, we can reduce neural-to-symbolic orchestration overhead by an order of magnitude.

Candidate research questions

  1. What ISA abstraction can express both tensor ops and symbolic constraints compactly?
  2. How should memory coherence work between continuous embeddings and discrete symbol tables?
  3. Can adaptive routing decide at runtime whether a query should stay neural, symbolic, or hybrid?

Evaluation sketch

  • Workloads: theorem-guided planning, neuro-symbolic QA, and robotic task composition.
  • Metrics: joules/query, reasoning latency, and faithfulness to explicit constraints.