An OpenAI engineer warned the models still occasionally overfit
OpenAI evaluated its new models on PostTrainBench-Lite, a shortened version of our original benchmark that gives agents 5 hours instead of 10 to improve an open-source base model.
GPT-5.6 Sol and Terra outperform GPT-5.5, but still rely on narrow strategies and sometimes overfit to the eval (common behavior). As we’ve reported before, the real frontier is research judgment and it remains one of the most exciting challenges for responsible RSI to solve.
No Digg Deeper questions have been answered for this story yet.
Great to see that the GPT-5.6 system card reports results on PostTrainBench (Lite)! GPT-5.6 performs substantially better than GPT-5.5.
Source: https://deploymentsafety.openai.com/gpt-5-6-preview/posttrainbench-lite
Details on PostTrainBench Lite:
Great to see that the GPT-5.6 system card reports results on PostTrainBench (Lite)! GPT-5.6 performs substantially better than GPT-5.5.
@karinanguyen Improved scores on the shortened version reflect better handling of model improvement tasks. The noted limitations around narrow approaches point to important areas for future refinement in agent systems.
@karinanguyen 这个榜单估计又要被吵几天了
