Exploratory architectures for reasoning systems that sit between pure neural methods and pure symbolic methods — built around the hypothesis that the right intermediate representation for LLM reasoning is a rule-mined causal graph extracted from attention, not free-form chain-of-thought.
Module 1 — ASI (Recursive CA + Attention + Symbolic)
Phased architecture:
- 2D cellular automaton as a controllable dynamical substrate.
- Transformer trained on the CA trajectories with attention hooks exposed.
- Attention tensors → binary causal matrix → temporal causal graph → FP-growth rule mining over that graph.
The bet: attention isn’t the explanation, but the causal graph built from attention is a substrate you can do symbolic discovery on.
Module 2 — ESDE (Enhanced Symbolic Discovery Engine)
Targets AIME / SAT / GSM8K via a neural-symbolic hybrid: Phi-3 / Mistral via Ollama → SketchNegotiator → RuleEngine backed by SymPy / Z3 → Contrastive Evolver → CodeBERT embeddings + FAISS MemoryGraph + FP-growth → FeedbackLoop.
Status
Prototypes, not products. Architectural boundaries are clean and the core ideas survive contact with real data, but benchmark results, ablations, and end-to-end reliability aren’t there yet. The strongest publishable slice is ASI Phases 1–3B with controlled experiments against a transformer-only baseline — fits ICLR ME-FoMo, NeurIPS SRI, or NeSy workshop.
- Main repo: rohit-ravi2/recursive-research
- HSDE (nested, separate repo): rohit-ravi2/hsde