AI writing has recognisable and irritating patterns, and is often referred to as slop. Can AI agents and specialised LLMs de-slop or humanise text? That is, can they rewrite LLM-generated text to improve its readability and make it sound more natural?
We believe that AI will eventually write as the best humans, but today’s LLMs are not there yet. We launched this arena to benchmark how well different systems perform, and to bring closer the future where models write well.
Eric Fithian
Researcher at UChicago
Kirill Skobelev
Incoming CS PhD student at Northwestern
Giulio Frey
Researcher at UChicago
Every time you use the Deslop Arena, the website picks two entries and shows you their outputs side-by-side. An entry is either an LLM using a skill or a third-party humanizer. The submission can be up to 1,000 words and 9,000 characters. The names of the entries remain hidden until you vote.
The main leaderboard ranks every entry on an Elo-like scale, and separate tabs break that out into model, skill, and humanizer boards. Each score carries a 95% bootstrap interval. In the leaderboard, you can also see a card for every skill and model.
To score submissions, we are using a score that is similar to Elo. We fit a Bradley–Terry model over all of the voting data. Bradley–Terry assumes a “strength” si for every entrant i, and the probability that entrant A beats B is given by
P(A beats B) = sA / (sA + sB)
The model further assumes that the strength of a model+skill combination decomposes as
smodel+skill = smodel × sskill
So every vote a pairing collects teaches us about both its model and its skill at once. We fit all strengths jointly by maximum likelihood, with the Baseline skill fixed as the reference (sskill = 1). A model’s score therefore means “this model running the baseline skill”, a skill’s score means “what this skill adds over the baseline, on any model”, and the combined leaderboard multiplies the two. Ratings are reported as
rating = 1000 + 400 × log10(s)
So, for example, if entrant A scores 1400 and entrant B scores 1000, the model says A should win about ten times for every time B wins. If they are 1200 and 1000, A is about a 3-to-1 favorite, and so on.
We periodically scan GitHub and skills.sh for skills that rewrite LLM-written text in English with the aim of improving its quality and have more than 1,000 GitHub stars or 1,000 skills.sh installs. If you believe we missed a cool skill, or if you operate a humanizer platform and would like to get added, please email contact@desloparena.com with the subject line “New submission”.