Pioneering
AI research.
We regularly publish our architectural discoveries, alignment methodologies, and agent benchmarks to help foster an open and verifiably safe intelligence ecosystem.
Vilcus-1: Multi-Step Inference Chains via Curated Rationales
We introduce Vilcus-1, a proprietary reasoning model trained to build dense logical rationales prior to emitting terminal tokens. By formulating logical search as a discrete optimization process, we show significant reasoning consistency gains over standard autoregressive decoding.
Zero-Hallucination Alignment via Verified Formal Proofs
This paper presents a scalable methodology to enforce structural alignment in large language models by feeding formally verified proof solvers directly into the reward modeling objective. We demonstrate a reduction in critical factual errors without degradation in open-ended generative speed.
Cordenex: Autonomous Code Synthesizers with Closed-Loop Execution
We detail Cordenex, an autonomous coding agent that utilizes a dedicated compiler feedback environment during inference to self-correct compilation and logical errors before finalized code generation. We evaluate its capabilities on highly complex long-horizon engineering tasks.