Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.
I am honored to have received runner up best paper award at ICML's https://forecasting-workshop.github.io/ for my Bayesian forecasting paper. In my talk I showed some unpublished results on a more rigorous Bayesian approach. I have now added these results to the arxiv paper....
Unfortunately this method does much worse than my heuristic BLF ("direct") method on forecast bench. However, on problems where there is no crowd prior, it can sometimes make the agent "less wrong" (see RHS below). https://x.com/sirbayes/status/2076887676549259368/photo/1
The basic idea is to do recursive/online updating of the log-odds using the log-likelihood ratio, lambda_t, which is estimated by an LLM. (We also add tempering parameter alpha to reduce over-confidence.) https://x.com/sirbayes/status/2076887674951176547/photo/1
Details are in appendix J of https://arxiv.org/abs/2604.18576.
I am honored to have received runner up best paper award at ICML's https://forecasting-workshop.github.io/ for my Bayesian forecasting paper. In my talk I showed some unpublished results on a more rigorous Bayesian approach. I have now added these results to the arxiv paper....
Unfortunately this method does much worse than my heuristic BLF ("direct") method on forecast bench. However, on problems where there is no crowd prior, it can sometimes make the agent "less wrong" (see RHS below). https://x.com/sirbayes/status/2076887676549259368/photo/1
The basic idea is to do recursive/online updating of the log-odds using the log-likelihood ratio, lambda_t, which is estimated by an LLM. (We also add tempering parameter alpha to reduce over-confidence.) https://x.com/sirbayes/status/2076887674951176547/photo/1
Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.