Lexsi Labs Research

Our research is dedicated to addressing fundamental challenges in making AI acceptable, safe, and aligned. We strive to ensure AI operates with certainty, consistently serving the core objectives of its users.

Featured

All Research Papers

Research Paper

Distilling Tabular Foundation Models for Structured Health Data

June 5, 2026

Research Paper

Shaping the Prior: How Synthetic Task Distributions Determine Tabular Foundation Model Quality

June 5, 2026

Research Paper

Ensembling Tabular Foundation Models - A Diversity Ceiling And A Calibration Trap

June 4, 2026

Research Paper

Pocket Foundation Models: Distilling TFMs into CPU-Ready Gradient-Boosted Trees

June 3, 2026

Research Paper

Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution

June 2, 2026

Research Paper

AlignTune: Modular Toolkit for Post-Training Alignment of Large Language Models

February 23, 2026

Research Paper

Beyond Uniform Credit: Causal Credit Assignment for Policy Optimization

February 13, 2026

Research Paper

Beyond KL Divergence: Policy Optimization with Flexible Bregman Divergences for LLM Reasoning

February 4, 2026

Research Paper

C -ΔΘ: Circuit-Restricted Weight Arithmetic for Selective Refusal

February 4, 2026

No results found.

Lexsi Labs Research

Featured Papers

Ensembling Tabular Foundation Models - A Diversity Ceiling And A Calibration Trap

Beyond KL Divergence: Policy Optimization with Flexible Bregman Divergences for LLM Reasoning

C -ΔΘ: Circuit-Restricted Weight Arithmetic for Selective Refusal

Orion-MSP: Multi-Scale Sparse Attention for Tabular In-Context Learning

TabTune: A Unified Library for Inference and Fine-Tuning Tabular Foundation Models

Interpretability as Alignment: Making Internal Understanding a Design Principle

All Research Papers

Distilling Tabular Foundation Models for Structured Health Data

Shaping the Prior: How Synthetic Task Distributions Determine Tabular Foundation Model Quality

Ensembling Tabular Foundation Models - A Diversity Ceiling And A Calibration Trap

Pocket Foundation Models: Distilling TFMs into CPU-Ready Gradient-Boosted Trees

Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution

AlignTune: Modular Toolkit for Post-Training Alignment of Large Language Models

Beyond Uniform Credit: Causal Credit Assignment for Policy Optimization

Beyond KL Divergence: Policy Optimization with Flexible Bregman Divergences for LLM Reasoning

C -ΔΘ: Circuit-Restricted Weight Arithmetic for Selective Refusal