acceptodds
Under review as a conference paper at ICLR 2027

Preference Conditioned State Persistence for Spiking Sequential Recommendation

Abstract

Sequential recommendation relies on recurrent states that retain interaction history while remaining responsive to the current preference context. Adaptive spiking neurons can modulate state persistence, but existing formulations primarily characterize retention coefficients rather than how those coefficients are assigned to recurrent state groups. We introduce Preference-Conditioned Dynamics (PCD), which predicts groupwise retention from the current preference representation and recurrent history. The resulting recurrent transition is determined not only by the retention coefficients but also by their assignment to fixed state groups. To isolate this assignment, we replay learned membrane-retention sequences and permute coefficients across groups while preserving the coefficient multiset and mean at every event. On Amazon Video Games, reassignment produces measurable changes in PCD ranking metrics and substantially smaller responses in a Dynamic Gated Neuron comparator with the same recurrent-cell parameter count. MovieLens 1M reproduces this contrast for normalized discounted cumulative gain and mean reciprocal rank at cutoff 10. Controlled training variants and complementary CIFAR-10 experiments further show that the learned retention structure depends on conditioning-feature correspondence and on gradient access to that feature. These results show that marginal retention statistics are insufficient to characterize the functional state-persistence structure learned by recurrent Spiking Neural Networks.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_T7kqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Preference-Conditioned Dynamics (PCD), a spiking recurrent unit for sequential recommendation that predicts group-wise membrane-retention coefficients based on current preference representations and recurrent history. It proposes frozen-sequence replay, an intervention that permutes coefficient-to-group assignments while preserving the coefficient multiset and mean at each event. Experiments on two recommendation datasets compare PCD with a parameter-matched DGN baseline, complemented by control experiments on CIFAR-10.

Critical strengths
  1. Frozen-sequence replay provides a clear way to disentangle two properties that co-vary during training: the realized retention coefficients and their assignment to recurrent state groups. The paired evaluation protocol is also carefully designed.
  2. The training-time controls are well designed. By crossing task-feature matching with gradient access and comparing intrinsic conditioning with representation-level modulation, the authors establish a clear causal framework for evaluating the proposed mechanism.
Critical weaknesses
  1. The paper's own analysis indicates that the response to coefficient reassignment depends on both coefficient heterogeneity and state discrepancy. Since DGN-CM derives retention coefficients from summaries of neuronal states, its group-wise coefficients may be relatively homogeneous. Its response on Video, which is nearly indistinguishable from zero, may therefore reflect coefficient homogeneity rather than an absence of learned correspondence between coefficients and states.
  2. On Video, the PCD response is no larger than the run-to-run standard deviation of the same metric reported in Table B2. On ML-1M, the interaction interval for Recall@10 includes zero, and several paired seeds favor the comparator. Effects of this magnitude are difficult to interpret as having meaningful functional consequences for the recommendation task.
  3. According to Table B3, SASRec outperforms the proposed model on both datasets. Table D1 further shows that a single-state variant with feedback disabled and a fixed-retention variant both match or outperform the default two-state dynamic configuration. Consequently, the contributions of the additional recurrent components and the spiking implementation remain unclear.
Questions for authors
  1. How do the authors establish that the larger response observed in PCD reflects a learned correspondence between retention coefficients and recurrent states, rather than the greater group-wise dispersion of PCD's retention coefficients?
  2. How do the authors demonstrate that a perturbation no larger than the run-to-run variability of the same metric has meaningful consequences for individual users or items?
Reviewer_aHSCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Some analysis of the temporal scales involved in capturing contextual dependencies in sequential recommender systems. Spiking neuron architectures involved in some of the alternatives. Numerical results on several datasets but insufficiently explained to be able to assess rationale or quality, and the metrics of comparison are poorly defined. Unclear what, if anything, is learned in the process.

Critical strengths

None

Critical weaknesses

The writing is close to incomprehensible.

Relevance of the cited papers to the actual statements made is minimal, e.g. unclear what the Quax reference (which argues that continuous time RNNs where the membrane time constants are more precisely parametrize describe brain data better) have anything to do with "distributing memory in recurrent units", Helfer paper describes context dependent modulation of neural dynamics,but cited for a contextual embedding LIF(?).

Questions for authors

In what way is the premise here different from older work like SS4Rec (Xiao 2025)?

What is the primary reference for Dynamic Gated Neuron (DGN) and in what way are these used in your approach?

How is the group retention coefficient defined?

Can you please define PCD in more operational terms, for a specific model architecture and task?

how can the LIF encoder be simultaneously stateless for the l window of representing a signal and simultaneously have memory of a persistent state?

more directly, what is the goal of embedding h (i think?) into s? What do z and u represent in equation (1)? if I is a current why does it simultaneously depend on both the approximate rate z and the instantaneous spikes s?

What are all of the unexplained variables in Fig1?

How does coefficient reassignment as a metric relate to the practical application of sequential recommendations? Can you explain the consequences of this metric with regards to the practical application?

Reviewer_Lm7XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how retention coefficients are assigned to different groups of recurrent states in spiking neural networks. The authors propose Preference-Conditioned Dynamics (PCD), which adjusts these coefficients based on the current preference representation and past neuronal activity. The main idea is that the assignment of retention coefficients matters, not just their values. To test this, the authors keep the learned coefficients fixed and shuffle their assignments across state groups. They evaluate the resulting changes on two recommendation datasets and conduct additional experiments on CIFAR-10 to study the effects of feature alignment and gradient flow. The results suggest that the learned relationship between retention coefficients and recurrent states can affect model predictions.

Critical strengths
  1. The frozen-sequence replay experiment separates coefficient values from their assignment to state groups, which is more informative than simply looking at average retention values or distributions.
  2. The authors include paired experiments, confidence intervals, feature reassignment, and gradient detachment, which strengthen the reliability of their observations.
Critical weaknesses
  1. Insufficient evidence for meaningful learned assignment.Equations (3)–(5) already show that coefficient reassignment can change recurrent states. The observed performance drop therefore does not necessarily demonstrate a meaningful learned assignment. Moreover, PCD and DGN-CM use different conditioning inputs, making the comparison less conclusive. Stronger matched baselines and direct analysis of the learned assignments are needed.
  2. Limited practical gains and insufficient experimental analysis. PCD offers only small improvements over simpler models, and some simplified variants perform similarly or better (Tables B2, B3, and D1). The experiments mainly report aggregate performance changes, without sufficiently explaining what useful temporal structure the model learns. These results make the practical value of the additional mechanism unclear.
Questions for authors
  1. Could the authors compare PCD with a baseline using the same conditioning inputs as PCD?
  2. How much of the reassignment effect can be explained by coefficient variability and state discrepancy, as suggested by Eq. (5)?
  3. Why do some simpler recurrent configurations achieve similar or better performance in Table D1?
  4. Could the authors visualize how retention assignments evolve across recurrent groups and interaction sequences?
  5. What is the actual computational overhead of PCD compared with simpler alternatives?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.