nucleation

LIT-1184: Factorized embedding of goal and uncertainty in the lateral prefrontal cortex guides stably flexible learning

2 min read

Properties4
StatusActive — read 2026-10-16 (NOTE-953) in full from the open-access article, with the supplementary tables for the shattering comparisons and the brain–behaviour correlations. Worth reading as a clear human-fMRI application of Bernardi et al.'s geometry tests (shattering dimensionality, cross-condition generalization) to goals and transition noise. Read it knowing that it is a reanalysis of 20 subjects with effects a few points above chance; that "uncertainty" is block-wise transition noise, coded only when a specific goal is cued; that it tested only whether the goal code is invariant to uncertainty, never the converse; that the task is action–outcome learning with experimenter-cued goals; and that goal-code invariance does not correlate with stability once corrected for multiple comparisons, although the main text says it does. Registered at the owner's request (U9 of the exchange of 2026-10-16 on active inference).
Version2
Tagsneuroscience, cognition, learning-and-conditioning
Cited as “paper” byNOTE-953 — Sung, Rigotti & Lee — factorized goal and uncertainty codes in lateral PFC

Graph View

Sung, Y., Rigotti, M. & Lee, S. W. (2025). Nature Communications 16:11643. doi:10.1038/s41467-025-66677-w.

Key takeaways

From the reading, NOTE-953.

  • Task. Action–outcome learning: two choices per trial through a tree of states to a coloured coin. A cue at trial start says which coin pays (specific goal) or that any does (non-specific). Transitions alternate between (0.9, 0.1) and (0.5, 0.5) in short uncued blocks.
  • Behaviour. Noise lowered optimality only without a specific goal (0.850 against 0.619). With one, there was no change (p = 0.384).
  • Agents. Model-free agents trade flexibility against stability; model-based agents, which score their transition model by a goal-independent state prediction error and replan from the cued goal, do not. The agents are not fitted to the humans.
  • “Factorized”, operationally. Shattering dimensionality over six goal × uncertainty classes, with nonlinear SD below goal and linear SD (all above chance), and goal CCGP across noise levels not differing from within-level goal SD. In vlPFC and dlPFC the uncertainty SD is not separable from the nonlinear SD.
  • Brain and behaviour. Goal SD in vlPFC and dlPFC correlates with flexibility, stability and performance. Goal CCGP correlates with flexibility and performance, but its correlation with stability does not survive FDR correction (q = 0.054), although the main text (p. 7) reports it. Uncertainty SD predicts nothing.

Why this is here

Registered at the owner’s request (U9), for the will-organization line and CLAIM-342. The registration offered factorized goal and uncertainty codes as a candidate for how a mind keeps its scoring of evidence apart from its current goal (QUESTION-121). The reading narrows that.

  • The paper tests whether uncertainty bends the goal code, not whether the goal bends the evidence-side code.
  • Uncertainty is coded only when a specific goal is cued.
  • “Uncertainty” is a block-level noise estimate, not a score of evidence.

The nearer analogue to independent scoring is in its simulation: the model-based agent scores its transition model without reference to the goal and avoids the trade-off, while the model-free agent scores through the goal-defined reward and suffers it. That evidence still arrives through the agent’s own choices. The task is an action–outcome loop, so the paper says nothing about CLAIM-341’s loop without action. It carries no instruction for machine-learning practice.

Backlinks