Post 2 of 2: word edits, synonym swaps and LLM paraphrasing against a from-scratch watermark, and the trade-off between surviving edits and protecting the key.
read onhashnode (opens in a new tab)medium (opens in a new tab)
Post 1 of 2: implement the green-red list watermark, detect it with a proper statistical test, and measure what it costs.
read onhashnode (opens in a new tab)medium (opens in a new tab)
Add a search tool and parallel tool calls to a small-LLM ReAct agent, going from 5 to 12 correct answers, then rebuild it with LangGraph and compare.
Build a ReAct agent from scratch with Phi-4-mini and measure every change on a fixed test set: prompt fixes, reflection and constrained JSON decoding.
also onhashnode (opens in a new tab)medium (opens in a new tab)