Latent transitions
Test whether learned internal transition steps can replace some explicit reasoning without degrading the result a user receives.
Geometric reasoning systems / controlled model adaptation
Active research · controlled experiments completeCan latent reasoning become more efficient without becoming less reliable?
I direct a local model-research program on an NVIDIA DGX Spark powered by the GB10 Grace Blackwell Superchip. The work translates geometric and latent reasoning papers into controlled training experiments, matched baselines, fresh evaluation sets, output review, and promotion gates that can reject a shorter or faster model when its answer quality does not hold.
R1 / research problem
Shorter reasoning is useful only if the model remains correct.
The central question is not whether a model can emit fewer tokens. It is whether latent or compressed reasoning can reduce work while preserving answer quality, explanation integrity, stopping behavior, and general usefulness.
Match learning to inference
Verify that model inputs, targets, prompt prefixes, answer supervision, and stopping behavior teach the same task the model must perform at generation time.
Quality before speed
Evaluate correctness, explanation quality, natural stopping, token use, latency, integrity, and resource cleanup together before authorizing a longer run.
Research ownership
Human direction, AI-assisted implementation, evidence-based acceptance.
Question and scopeI define the research question, useful outcome, hardware boundary, frozen comparison, and stop conditions.
Experiment and evaluation directionI require matched controls, fresh questions, answer and explanation review, integrity checks, and explicit promotion criteria.
Failure analysisI challenge plausible explanations, separate stopping from correctness, and redirect the work when the training contract or task mix does not match the intended use.
Final acceptanceAI agents assist with implementation, test construction, analysis, and documentation. I retain the decision to promote, hold, redesign, or stop.
Experiment architecture
Turn a paper claim into a falsifiable local test.
Define the hypothesis
Reduce a paper-derived idea to a bounded question with a result that can be accepted or rejected.
Build matched controls
Pin the source model, data order, update budget, prompt contract, scoring rules, and comparison conditions.
Train and evaluate locally
Run controlled adaptations and fresh evaluations on the DGX Spark while preserving model and artifact identity.
Promote or hold
Compare correctness, explanation quality, stopping, work, latency, and cleanup before deciding whether to scale.
Completed evidence
Multiple controlled runs now define what did—and did not—work.
Pilot evaluationTested the initial paper-derived training approach against a pristine baseline.
Matched comparisonCompared trained and baseline behavior under controlled generation conditions.
Prompt diagnosticsSeparated stopping behavior from actual reasoning and answer quality.
Direct-answer controlTested whether cleaner formatting represented better task performance.
Task-mix auditExamined what the training data actually taught the model to do.
Each run used a predeclared evidence boundary. Training stability, checkpoint identity, evaluation coverage, output scoring, substantive review, and GPU cleanup were kept separate so that one passing layer could not stand in for overall model quality.
What the experiments established
Efficiency signals appeared, but they did not preserve quality.
Latent steps changed work—not trustThe tested latent condition reduced generated work but caused the trained checkpoint to lose cases it had previously answered correctly.
Formatting improved while answers worsenedThe direct-answer control produced cleaner, more consistent formatting while substantive correctness dropped sharply.
Prompt closure was not a reasoning repairClosing the thinking prefix improved stopping, but an added final-answer cue caused a repeated-answer collapse instead of restoring reasoning.
The dataset taught the wrong task mixThe audit found a math-only training mix with no coding, SQL, chat, or general-instruction tasks. Compact targets also discarded their source rationales.
These are negative promotion results with positive engineering value: they prevented longer training on a recipe that was shorter or cleaner without being reliably better.
Current direction
Redesign the training signal before spending more compute.
The next bounded proposal starts from the pristine model and tests a small task-aligned adapter using verified short-rationale targets, retention replay, and fresh sealed evaluation gates. The goal is to learn whether concise reasoning can retain capability across the tasks the model is actually expected to perform—not merely produce shorter text.
Pegasus Cortex remains the operational retrieval baseline for related work: bounded context retrieval, exact typed reads, provenance, and non-authorizing stored knowledge. Its role is to provide a real operating problem against which future retrieval and reasoning improvements can be evaluated.