This was the Friday poster version of Mattar's talk, same underlying model, but a much more granular, back-and-forth conversation, mostly aimed at people who already knew the paradigm and wanted to poke at the mechanics.
For each remembered episode, the model tracks three things: a Gaussian mean belief about the episode's reward, the precision of that belief (inverse variance), and a probability that the episode is relevant to the current decision. Before deliberation even starts, the agent is handed one free sample of every item, meant to capture the idea that people briefly hold candidates in working memory before actively starting to deliberate.
Sampling has a baseline cost plus a distance cost, since memories are treated as spatially organized. Moving attention to a new location costs more than continuing to sample the one you're already on, which is free on repeat. The total expected payoff is the reward from acting on the currently relevant memories, accumulated over sampling, minus the accumulated cost of everything sampled along the way.
The stopping rule follows from a value-of-computation calculation. For each candidate episode, compute the expected gain from sampling it once more and continuing optimally from there, versus the value of just stopping right now and making a choice. Sample whichever episode has the highest positive value. If everything comes back negative, stop and decide. That's the optimal policy in principle, but solving it exactly is an intractable, infinite-horizon problem, which is why the lab trained a recurrent network via reinforcement learning on roughly 250 million simulated tasks to approximate it instead. The network's own memory representations were kept deliberately dumb, relevance is just a zero-or-one flag, nothing richer, so that any behavioral richness that shows up has to come from the deliberation policy itself, not from smuggled-in memory content.
The model reproduces two clean human patterns: people take longer to decide the more candidate memories they have, and they tend to look next at nearby locations rather than far ones. Where it falls short is a forward temporal bias in human data. People are more likely to next sample something encoded shortly after the current item, which the model doesn't capture, because it has no representation of encoding context, temporal or spatial, at all. That gap gets attributed to unmodeled encoding effects rather than to a flaw in the sampling and deliberation account.
One idea floated in discussion was making the retrieval cost of a specific memory a function of the stimulus itself rather than a fixed free parameter, for instance testing whether people with mood or anxiety-related biases show a sampling bias toward negative memories because those memories are, in some sense, cheaper to retrieve, rather than because they're rated as more relevant. Right now all memories are treated as equally costly to retrieve, which is clearly a simplification. The basic prediction, longer decisions with more memories, is close to a given for any sampling-based account. The more interesting evidence is in the finer-grained, trial-by-trial dynamics that the network also reproduces.
Poster-session conversations are hard to note down cleanly since several people are talking at once, so treat the Q&A part especially as a rough paraphrase.