alignment-tax

When Better Means Less: Quantifying What Benchmarks Miss Between Model Generations. 2,310 controlled comparisons show GPT-5 series lost 6.7x creativity and gained 4.4x false refusals vs chatgpt-4o-latest — invisible to standard benchmarks.

Exploring Aligment TAX and impact in LLM compute cost

Base Qwen/qwen:2.5-coder:1.5b model, fine-tuned for MAX performance on Agentic Tasks and Ecosystems. Surpassed Base Model on Python/Typescript Coding Workfoads.

From TextBlob to LLM Agents — sentiment model selection for B2B technical support with CSAT ground truth (ACL 2026 Industry Track). Paper, figures, aggregate results, poster & slides.

Code and data for "The GRPO Tax is Smaller Than You Think" (TMLR 2026): a dense-checkpoint longitudinal study of capability preservation during GRPO reasoning training.

Mechanistic Alignment & Non-Linear Representation Hardening Kernel for Foundation Models (ImmunoBench-1000 & EU AI Act Art 51/53)