Review Papers Claim by Claim

 Quick Explanation Copied

 Long Explanation

Core Claim and Evidence Structure

This Trends in Cognitive Sciences review challenges the textbook view that short-term memory (STM) is maintained exclusively by stimulus-selective persistent spiking in frontoparietal cortex. The authors synthesize electrophysiology, human imaging, and recurrent neural network (RNN) modeling to argue that (i) persistent activity is highly variable, weak, or absent in many STM tasks; (ii) information can be stored 'silently' via short-term synaptic plasticity (STSP), building on the Mongillo synaptic theory of working memory; and (iii) passive maintenance and active manipulation of information recruit different substrates — manipulation reliably demands persistent activity Masse et al. 2020 TiCS review.

The strongest cited human evidence is the 'pinging' paradigm: task-irrelevant impulses or TMS partially rescue decoding accuracy for memoranda that had decayed from BOLD/EEG readout, implying latent hidden states rather than erasure. The authors' own RNN work (biologically constrained, Dale's law, STSP-governed synapses) shows delayed match-to-sample is solvable activity-silently, whereas delayed match-to-rotated-sample always requires persistent activity, with manipulation load predicting delay-decoding strength across nine tasks.

Notably, a 2025 intracranial study (30 epilepsy patients, 866 neurons) provides causal support for the classical view the review questions: hippocampal stimulation during maintenance abolished persistent category-selective firing, increased population distance-to-attractor (t=4.1, df=42.01, p<0.001), and impaired behavior (accuracy 95.3% overall; slower RTs, p<0.001), demonstrating persistent attractor dynamics causally support at least some human WM maintenance Hippocampal stimulation causal WM study (2025). These are complementary, not contradictory: the review itself predicts task-dependence, and a load-1/3 Sternberg task with actively probed items sits on the manipulation-heavy end.

Strengths and Blind Spots

• Strengths: coherent framework linking task demands to substrate; advocates RNNs as hypothesis generators; explicitly acknowledges that fMRI/MEG aggregate signals may lack resolution to prove purely silent storage, and that a recent study found weak but significant decoding of unattended items from BOLD.
• Blind spots: the review's flagship 'pinging' studies are indirect — reactivation on cueing cannot fully exclude weak residual spiking. Later modeling shows STSP alone fails flexible selection and distractor resistance in early visual cortex, whereas top-down feedback succeeds Feedback vs STSP modeling study (2025) — a mechanism the review underweights relative to STSP. Species generalization (mouse ALM, macaque LIP/PFC, humans) and the confound that persistent activity may encode motor plans rather than memoranda remain unresolved.

What Would Change the Conclusion

The review's central thesis is falsifiable: demonstrating that persistent activity is necessary for all STM tasks regardless of demands, or directly measuring synaptic efficacies in vivo showing silent traces are unnecessary, would overturn it. Conversely, direct in vivo measurement of STSP during behavior would cement it. Confidence: moderate-to-high for the task-dependence claim; lower for the claim that purely silent storage dominates any natural task.

Author Review: Nicolas Y. MasseAuthor Review: Matthew C. RosenAuthor Review: David J. Freedman

 BGPT Paper Review

Study Novelty

Synthesizes an emerging shift (Stokes 2015, Mongillo 2008) rather than inventing it; its task-dependence framework and own RNN results are the most original contributions.

Scientific Quality

Well-sourced review (99 references) in a strong journal; appropriately caveats indirect human evidence. As a review it presents no new data, and its own group's studies receive prominent framing (self-citation emphasis).

Study Generality

Applies across species (mouse, macaque, human) and imaging modalities, framing STM/WM broadly, though anchored to delayed-response paradigms.

Study Usefulness

Highly useful for experimental design: warns that single-task studies confound maintenance with manipulation and promotes RNNs as cheap hypothesis generators.

Study Reproducibility

Key modeled and empirical studies cited have code/data in some cases (e.g., RNN and pinging studies), but the review itself contains no primary data.

Explanatory Depth

Provides mechanistic depth via state-space attractors, subspace attractors, and STSP toy models, though it stops short of a quantitative unifying principle.