Research lineage
Evidence selected
what remained.
ASM-CM emerged from a sequence of hypotheses, failures, ablations and confirmations. This page preserves the scientific lineage without turning an experimental result into a universal claim.
Open questions
The program exists to measure, not presume.
Scaling
Do gains observed in smaller models persist as parameters, data and diversity increase?
Memory
Which mechanisms preserve useful context without storing the entire history?
Geometry
When does relational geometry improve transition, and when does it merely add cost?
Efficiency
Where are the crossover points between quality per token and quality per GPU hour?
Timeline
DRM → ASM
The theoretical origin remains documented while the architecture follows experimental evidence.
Directional Relational Manifolds
The geometric hypothesis introduces variable effective dimension, directional transitions and path-dependent memory.
DRM Language Emitter
The theory becomes an experimental causal emitter with latent state, directional field, flow and metric.
From hypothesis to the ASM family
Mechanisms are removed or isolated. Direct controls and selective memory show that explicit geometry is not always necessary.
ASM-R promotion
Three seeds through 100M tokens confirm ASM-R as the main token-quality architecture; ASM-S remains the efficiency variant.
Paired Transformer
At 100M tokens and seed 1, the paired baseline leads in CE, throughput and training time. Multiseed confirmation remains open.
ASM-C through 32K
Compact inference bounds cache, VRAM and incremental cost; the short MQAR gate fails and blocks long-memory claims.
ASM-CM promotion
Fast weights, selective consolidation and mixed training recover associations through 32K; three seeds and post-FP32 revalidation confirm bounded state with preserved language.
ASM 1B
The campaign proposes fixing inference, measuring real cost and scaling in gated stages toward roughly one billion parameters.
Experimental family
ASM-CM at the center. Other variants as controls and lineage.
ASM-CM
Promoted architecture: durable associative memory, preserved language quality and bounded streaming state.
View model card →ASM-R
Direct contextual transition conditioned by relational geometry; promoted among ASM variants for token-level quality.
View model card →ASM-C
Compact streaming form of ASM-R that reuses its weights and limits incremental state to a local block and completed causal state.
View model card →ASM-S
Architecture without explicit geometry, focused on selective memory and preserved as an efficiency control.
View model card →ASM-F
Geometric variant with a metric-normalized directional frame. Its first generation failed multiseed validation.
View model card →ASM-X
Explicit DRM architecture preserved as a historical and experimental reference for the theory.
View model card →In progress
From ASM-CM to external validation
- 01Test semantic retrieval beyond synthetic MQAR.
- 02Confirm ASM-CM × Transformer on external corpora and tasks.
- 03Measure prefill, decode, memory, time and energy under external protocols.
- 04Investigate autonomous write selection, consolidation and episodic retrieval.
- 05Train larger pilots only after quality, memory and efficiency gates.
Scientific roadmap
Replicate → scale → compare → publish.
The main milestone is not processing the largest possible number of tokens. It is producing a conclusion that survives review of the protocol, artifacts and limitations.
Another initiative
ASM 1B is community-funded foundational research.
It is not required for ASM-CM pilots, licensing or integration. The applied ASM-CM track remains separate from the proposal to train a larger model.
Explore the ASM 1B initiative →