Investigating recursive reasoning,
self-learning loops, and dense small models.
NeuLabs is an AI laboratory focused on algorithmic density over brute-force scale. We research test-time deliberation, autonomous self-improvement loops, and compact model architectures (1B–8B) capable of robust logical deduction.
Core Scientific Inquiries
We depart from traditional cluster-scale scaling laws by investing in structural reasoning efficiency and autonomous verification.
Recursive Reasoning Loops
Inference-time search where models critique, branch, and recursively audit their own logical deductions prior to final token emission.
Compact Small Language Models (SLMs)
Distilling frontier supervisory models (e.g. Claude 3.5 Sonnet) into specialized 1B–8B parameter models optimized for latency-critical and edge-native deployments.
Self-Learning & Continual Adaptation
Dynamic test-time learning where agents leverage verifiable environmental feedback to refine localized memory graphs without catastrophic forgetting.
Synthetic Reasoning Trajectories
Procedural generation of verified synthetic curricula for rigorous domains—such as computational taxation, symbolic math, and code verification.
The Recursive Deliberation Engine
Our four-phase execution pipeline decoupling computational capacity from parameter count.
Candidate Generation
Compact generative priors explore candidate trajectories across logical constraints.
Recursive Auditor
Deterministic checks and critic models evaluate deductive consistency at each reasoning branch.
Localized Error Feedback
Detected fallacies feed back into deliberation loops without parameter degradation.
Distilled Synthesis
Verified solutions output with high confidence and logit calibration.
Active Benchmarks & Initiatives
Current empirical studies and laboratory implementations undergoing evaluation.
| Code | Initiative | Status |
|---|---|---|
| NL-REC-01 |
Recursive Deliberation via Iterative Error Back-propagation
Benchmarking test-time compute scaling laws on compact models across constrained deductive logic sets.
|
In Progress |
| NL-SLM-02 |
High-Density Distillation from Frontier Supervisory Models
Transferring multi-step logical reasoning trajectories from Claude 3.5 Sonnet to edge-ready 1B–3B model weights.
|
Benchmarking |
| NL-SYN-03 |
Domain-Verified Synthetic Trajectory Curricula
Procedural pipeline generating non-leaking mathematical and statutory compliance dataset artifacts.
|
Phase 1 Complete |
“The future of artificial intelligence does not belong to ever-larger power grids, but to recursive architectures that think before they speak.”— NeuLabs Research Thesis, 2026