[ OPEN RESEARCH / PROJECTS ]
The current MLBricks project catalog — from active sequence and feed-forward bricks to research systems and experimental efficiency components.
Filter by maturity to see what is active today, what is still under research, and what remains experimental.
[ RESEARCH CATALOG ]
Recurrent state-based sequence processing for efficient full-sequence training and compact-state generation.
A production-facing sequence brick with full-sequence training, recurrent generation, and automatic execution planning.
An optimized causal-attention brick with compact latent processing, cached decode support, and automatic selection between supported execution paths.
Pushes attention efficiency through architecture and kernel design while retaining a familiar attention-facing interface.
State Oriented Unified Processing: an architecture that gives model blocks evolving state, observer memory and learned fusion, exploring whether useful capability can be achieved with less dependence on deep serial stacks.
Asks whether state, memory and fusion can carry part of the work normally pushed into more physical layers — a potential path toward lower serial depth for edge and large-scale AI.
Adaptive 4–32-bit matrix storage that measures how much precision a matrix needs, packs it at the smallest accepted width, and promotes it into a supported compute bucket at runtime.
Targets storage efficiency without blindly forcing every matrix into the same low-bit format; sensitive matrices can retain more precision when the calibration error requires it.
A mixer-conditioned recurrent feature network that uses the hidden stream, current and previous mixer outputs, persistent FFN state, and physical-depth information.
Makes feed-forward computation responsive to evolving model state and representation change instead of treating each layer as an isolated dense transform.
An adaptive RMS-based residual controller that bounds branch-update energy and candidate residual-stream growth relative to the current stream.
Prevents oversized residual growth from becoming one more reason later blocks lose relative leverage in deep models.
Vision models built around ESA, using state-based spatial processing while retaining the shared MLBricks vision runtime.
Extends state-based processing into vision workloads while keeping a common MLBricks component model.
A StateAwareFFN extension that reuses a shared state-refiner core while pass embeddings, pass gates and the evolving state make each virtual refinement distinct.
Trades some repeated compute for a much smaller parameter increase than duplicating full physical state-aware FFN blocks.
A compact pass-specific gated Micro-FFN that creates narrow virtual refinement steps instead of adding another full-width feed-forward block.
Adds effective computation depth with small pass-specific weights — useful when parameter storage is more constrained than available compute.
Bolt-based vision models that pair the shared MLBricks vision runtime with Bolt as the underlying sequence mixer.
Tests how Bolt-style optimized mixing transfers from language-model sequence workloads into visual processing.
Neural Orchestrated Retrieval Augmentation — an MLBricks research project exploring how models can receive retrieved knowledge as readable text, compact neural memory, or a combination of both.
NORA explores a flexible knowledge layer that can support conventional language models as well as state- and memory-oriented MLBricks systems without forcing every use case into one form of context.
A research routing component for coordinating specialized processing paths inside composable MLBricks systems.
Harmony explores how a model can organize work across reusable components while keeping the system flexible, inspectable, and easier to compose.
[ RESEARCH DIRECTION ]
Explore the public ESA project or contact MLBricks about research collaboration and future experimental releases.
