temperature IS NOT the answer, top-p and top-k are we want a stochastic random seed initially maybe (that's questionable, I don't buy the temperature="creativity" argument, sounds like a recipe for hallucination to me). We want a "deterministic" process by altering top-p and top-k primarily as that mechanism What's the problem with temperature = zero? I want temperature to stabilize at exactly zero to eliminate that source of trivial stochasticity. Nudging the model and preventing it from getting stuck can be handled very capably **exclusively** to tweaks to top-p and top-k (setting k=1 when we really want it converge immediately, conversely by increasing both p and k to near-max values when we really want to unstick it and get it into a broader solution space)