Comfrey RRM Model: Geometric Recursive Reasoning with Self-Learning and agentic binning steering model for evals

We introduce Comfrey GRRM, a framework that makes large language model
(LLM) inference geometry-aware, depth-adaptive, and self-correcting at inference
time and during geometry-aware training. Current transformer architectures are
unable to fully utilize the embedding space because their reliance on Euclidean
geometry is ill-suited to the inherently hyperbolic structure of language representations.
As a result, large regions of the representational landscape remain unexplored
during both training and inference. Our geometry-aware approach provides access
to these previously unreached spaces.
The methodological centrepiece is Wavelet-Scale PCA ():, where layers are units,
geometric feature signals are distributional variables, and wavelet coefficient distributions
over ordered frequency scales are the bin vectors. The classical Tchebychev
interval step is replaced by an Entropy-Modulated Wavelet Interval whose
half-width t σij(1+Pij Hij η) is modulated by Shannon entropy of the scale-energy
distribution, with a Bayesian logistic probability penalised by λ = −t log N+1
N .
The wavelet substrate is the Brahimian Wavelet Family: divergence-free curl
wavelets in d-dimensional embedding space with exact Z/3Z cyclic symmetry. Five
geometric signals (Kirchhoff temporal tension, IRQ Ricci curvature, GPS geodesic
consistency, Boltzmann free energy, IRQ algebraic coherence) assemble into a 9-
dimensional ManifoldState m ∈ [0, 1]9.
A Recursive Thinking Module (, [14]) injects m into the first token embedding
between depth steps of an Adaptive Computation Time loop. A Recursive Language
1
© 2026 CC BY-NC-ND 4.0
Model Decomposer (, [11]) breaks hard queries into sub-questions solved independently
by . A Parallel Engine runs Ollama and HuggingFace backends simultaneously
and selects the geometrically superior answer. A Verifier (self-consistency,
constraint checking, entailment hardening) validates every answer before return. A
Geometric Hypernetwork maps m to per-layer LoRA deltas; Geometric uses the
manifold signals as label-free reward components. An Episodic Geometric Memory
accumulates episodes and retrieves the nearest LoRA adapter at inference time.