ArticlesFindingsMediaAbout
GitHubX.comRSS Feed
Vansh Vazirani

Research

ThoughtWeaver: Structured-CoT in Markdown for Enhanced Reasoning

Purpose-built for swiftly exploring candidate solutions while consuming fewer tokens than conventional reasoning models.

Jan 02, 2026
Reasoning models
20 min read

Reproducing Introspective behaviour on Injected Concepts in Open-Source Large Language Models

This is my attempt at reproducing concept injection locally and investigating whether small open-source models can genuinely introspect and detect concepts artificially injected into their internal states.

Dec 02, 2025
Interpretability
10 min read

Paper Breakdowns

Engram: DeepSeek's Brilliant Architecture for Information Recall

A new conditional memory module proposed by DeepSeek that sits inside the Transformer as a third fundamental building block, alongside Attention and the Feed-Forward Network (FFN).

Apr 12, 2026
LLM architecture
11 min read

Categories

InterpretabilityLLM architectureReasoning models