4 ms·
Very interesting. Also I find these training parameters quite elegant: - Diversity: This term encourages the model to generate a diverse set of samples, preven
by mentalgear 11mo ago
Very interesting. Also I find these training parameters quite elegant:
- Diversity: This term encourages the model to generate a diverse set of samples, preventing mode collapse.
- Fidelity: This term rewards the model for making predictions that are close to the ground-truth
I'm wondering if a continuos next-vector generative approach also increase innate "reasoning" capabilities of the model, since it could potentially capture more of the semantics of the data vs just tokens.
- barrenko 11mo agoAnd may be even more adapted to sorts of RL finetuning?
- mike_hearn 11mo agoThey say this technique isn't compatible yet with RL because you can't adjust the logits. So no GRPO I guess, which is going to be the biggest issue. An LLM with no RL applied isn't going to be that useful.