Renmin's DeepEvolve tops FutureX leaderboard for three weeks

A forecasting system from Renmin University of China, Nanyang Technological University and IQuest Research held the top spot on the FutureX leaderboard for three consecutive weekly evaluations, with a rating improvement the authors put at up to 195%. The paper on Hugging Face, titled *Evolving with Open Worlds*, introduces a system called DeepEvolve, assembled from three modules the team names Prometheus, Branchus and Athena.

"Most contemporary AI systems are released with frozen parameters and improve through external updates," the abstract states, pitching recursive self-improvement as the alternative: a loop in which an agent's parameters, skills and training environment all evolve from a single pool of its own trajectories.

Prometheus is the data environment. It ingests daily news from 125 websites and synthesizes verifiable forecasting questions tied to future events. Branchus is the forecaster, split into a Search team (Searcher, Scheduler, Critic) that gathers evidence and a Predict team that reasons over cached evidence to produce probabilistic forecasts. Athena is the evolution engine, combining reinforcement learning with verifiable rewards (RLVR), a skill-distillation pipeline that mines lessons from past trajectories, and adversarial task synthesis that targets the agent's weakest categories.

The numbers the paper reports: +1.3 points of choice accuracy from the RLVR loop, and a trajectory library expanding across three nested pools of 88, 186 and 262 live trajectories, with domain coverage growing from 168 to 479. For structure-synthesis problems, the distilled skills cut tool calls per question by 44%, from 32 to 18. On the Prometheus-synthesized evaluation, Qwen3.8-Max was the strongest base model, hitting 72.3% choice accuracy and a Bradley-Terry score of 62.2, with Branchus occupying the top three overall positions across all four tested LLMs. The FutureX ranking the authors report spans three weekly evaluations; the paper does not state what happens as that sliding window extends.

Code is at github.com/RUCAIBox/DeepEvolve. It lands in a crowded moment for agent research, with 515 agent stories across our tracker in the last 90 days.