I think this is what we would have observed in Germain's and Constance's paper respectively if decay were true

Thanks for the pointer Vinny!

Exhibit #3: back to "stable" tasks, @constancedestais.bsky.social conditioned learning rates on confidence over time, and show that the asymmetry is still there. Indeed, it increases over time. Note that the model structure would have perfectly allowed for a "symmetric decaying" pattern 4/n

Exhibit #2: learning rate bias has been reported (by us and other groups) in volatile tasks or conditions where, normatively, learning rates should not decay, and, perhaps more importantly, empirically, they indeed do not decay - if not, accuracy would not be above chance 3/n
doi.org/10.1016/j.ti...

Exhibit #1: I was aware of this possibility since our first paper on the topic, and this is why we fitted separate learning rates in the first half and the second half of the learning phase. We found no evidence of decay and robust bias in both phases 2/n

www.nature.com/articles/s41...

2/2 this is the second paper by Vidal and Moran.

osf.io/preprints/ps...

Something similar has been going on concerning learning bias versus perseveration. Initially fuelled by Kentaro Katahira and co, and more recently by Juan Vidal and Rani Moran. (see the discussion, about what counts as a plausible computation)
See this 1/2 and...
psycnet.apa.org/buy/2023-182...

I think that is where I disagree with the author. Asymmetry is apparent, only if we assume people are Bayesian (which I do not believe is the case). But then I like that a "rational" analysis (Bays) of the task leads to the emergence of asymmetry, which may explains why the it is used by the brain

I guess the missing link here is "However, we find that even if the agent updates its belief via, arguably objective, Bayesian inference, fitting the above model demonstrates both the biases". I working under the assumption that the Bayes solution is understood as normative given the task here

I am very humbled that during the past years so many smart people took seriously our research questions and results to push forward our understanding.
On the specific subject matter (bias or optimal) I am still persuaded that it is a bias, that just happens to be generally optimal 😉

Also quite coherent with what has been found by @akjagadish.bsky.social and co that asymmetric update emerges as an optimal solution in a neural-network meta-reinforcement learning agent arxiv.org/abs/2402.03969

It resonates perfectly well with what we (with @isabellehoxha.bsky.social and Léo Sperber) have shown, that the bias emerges as an evolutionary stable solution in multi-agent simulations

www.pnas.org/doi/abs/10.1...