On November 9, Baidu's Wenxin disclosed its recursive self-improvement (RSI) progress, unveiling ERNIE Evolve, an automated research system that splits work among agents handling direction, coding, testing, evaluation and review to propose paths, run experiments and keep what works.
On Kunlunxin P800 low-latency setups, about 168 hours and four rounds of directed exploration raised generation speed to 623.6% of baseline, lifted deterministic-inference throughput to 133.7% and cut first-token latency by 12%. Baidu says ERNIE Evolve has been applied to the upcoming Wenxin 5.5 Flash; the Base version was open-sourced on November 5.
The significance is that a lab now states publicly that a machine picked the experiments, a step beyond automation toward self-directed model improvement.
Source: Leiphone / Baidu Wenxin
