{"id":"52339c73-9058-47e0-b6db-a874d86b8898","arxiv_id":"2508.14081","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding sleep-like replay consolidation after each task reduces catastrophic forgetting in Equilibrium-Propagation-trained recurrent networks, matching or exceeding backpropagation-trained baselines on five image benchmarks.","lead":"A new training recipe gives recurrent neural networks a 'sleep phase' after each new task, replaying old lessons so the networks forget less. On five standard image benchmarks, the sleep-equipped networks matched or beat conventional training, a step toward lifelong learning machines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Resource-matching and ImageNet scope are unverified; headline superiority may be an artifact.","rationale":"The reader's weakest assumption—baseline matching—is correct and central. My concern sharpens it by adding the compute-matching dimension (EP's settling iterations vs. BPTT's truncated passes) and by emphasizing that the ImageNet scope is essential to interpret the superiority claim. The abstract-only status means there is no way to verify these details, so the UNVERDICTED verdict stands. I do not see an internal inconsistency or an obviously false claim, but the empirical evidence is not yet checkable. The proposed concrete test would resolve the concern and either support or undermine the headline.","tokens_in":1038,"tokens_out":2214,"duration_ms":25865,"concrete_test":"Obtain the authors' code and hyperparameters; re-run both arms with identical architecture, per-task epoch budget, batch size, optimizer, sleep-epoch count, replay buffer size, and compute-matched EP versus BPTT (e.g., equal number of forward-backward passes). Clarify ImageNet subset; if it is not ImageNet-1K, either rerun on ImageNet-1K or explicitly rescope the claim. If EP+SRC no longer matches/exceeds BPTT+SRC under these matched conditions, the headline superiority is unsupported; if it does, the claim holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that MRNN-EP with SRC matches or outperforms MRNN-BPTT with SRC across five benchmarks. Since the abstract reports no architecture, hyperparameters, epochs per task, sleep-epoch counts, replay sample numbers, compute budgets, or ImageNet dataset scope, the most load-bearing risk is that the comparison is not resource-matched. EP requires iterative equilibrium settling per example; if EP models received more optimizer steps, more rehearsal/sleep iterations, or used fewer classes in the ImageNet subset, the result is an artifact of unequal compute or data, not algorithmic advantage. The claim of 'surpassed BPTT on ImageNet' is uninterpretable without stating whether it is full ImageNet-1K or a subset. This is not an internal inconsistency but an empirical verification gap that directly underpins the generalization of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sleep-like replay consolidation (SRC) algorithm for Equilibrium Propagation (EP) trained recurrent neural networks (RNNs), intended to mitigate catastrophic forgetting in class-incremental learning. The abstract reports that EP-trained multilayer RNNs (MRNN-EP) with SRC significantly improve retention, perform on par with BPTT-trained RNNs equipped with SRC on MNIST, and surpass BPTT-based models on Fashion MNIST, Kuzushiji-MNIST, CIFAR10, and ImageNet. Combining SRC with awake rehearsal is reported to further boost long-term retention. The abstract also claims superiority over feedforward networks with established regularization techniques. This review is based solely on the abstract, as the full text was not available.","tokens_in":1173,"tokens_out":2005,"duration_ms":22971,"significance":"If the reported results hold, the paper would make a meaningful contribution to biologically plausible continual learning: it shows that a sleep-like replay mechanism can be integrated with EP-trained RNNs and could close the gap with BPTT, while also improving over standard regularization baselines. The proposed SRC mechanism is conceptually simple and potentially reusable across architectures and learning algorithms. The explicit combination of sleep-phase consolidation and awake rehearsal is a worthwhile direction. However, because the abstract reports no experimental details, the significance cannot yet be assessed beyond the plausibility of the idea. The manuscript would benefit from full disclosure of the experimental protocol, resource matching, and statistical analysis.","major_comments":[{"comment":"The central claims—'SRC significantly improves resilience' and 'significantly better compared to feedforward networks'—are made without any statistical support. No effect sizes, error bars, number of runs/seeds, or significance-test details are provided. For a learning method paper, this is load-bearing: the reader cannot determine whether the observed improvements are robust or within run-to-run noise. The full manuscript must include these details, ideally with confidence intervals or paired tests across multiple seeds.","section":"Abstract"},{"comment":"The class-incremental learning protocol is not specified. The abstract does not state the number of tasks, per-task class counts, task ordering, or whether ImageNet refers to ImageNet-1K or a subset. The ImageNet result in particular is uninterpretable without knowing the subset size and class split. This directly affects the generality of the claim that SRC-equipped MRNN-EP 'surpassed' BPTT-based models. The full methodology must describe the exact continual-learning protocol for each benchmark.","section":"Abstract"},{"comment":"The comparison between MRNN-EP and MRNN-BPTT is not shown to be resource-matched. To establish that SRC removes the continual-learning disadvantage of EP, the EP and BPTT arms must be matched in architecture, parameter count, training epochs per task, sleep-phase consolidation epochs, rehearsal sample counts, and computational budget. If, for example, the EP models received more sleep epochs or more rehearsal iterations, then 'on par' or 'surpassed' could reflect unequal resource allocation rather than algorithmic advantage. The manuscript must report these budgets explicitly.","section":"Abstract"},{"comment":"The feedforward baselines with 'several well-established regularization techniques' are not named. Without identifying which techniques (e.g., elastic weight consolidation, synaptic intelligence, etc.), their hyperparameters, and their training budgets, the claimed superiority over these baselines cannot be reproduced or fairly evaluated. Provide a full baseline specification.","section":"Abstract"}],"minor_comments":[{"comment":"Naming consistency: use 'Fashion-MNIST' and 'Kuzushiji-MNIST' with hyphens, as in the original dataset names, to avoid ambiguity.","section":"Abstract"},{"comment":"The abstract uses 'awake replay' but does not define it in relation to SRC. A brief clarification of 'awake rehearsal' vs 'sleep replay' would improve readability.","section":"Abstract"},{"comment":"No mention of code or data availability. For reproducibility, the full manuscript should state whether code and trained models will be released.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"This review is based only on the arXiv abstract because the full text was not available. The central idea is plausible and relevant to the journal's scope, but the abstract alone does not meet the evidentiary bar for acceptance. The missing experimental details (protocol, baselines, resource matching, statistical analysis) are load-bearing; they are not cosmetic. I recommend requesting the full manuscript and, if the details are provided, a full technical review. Given the current lack of verifiability, I cannot recommend acceptance at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper puts a known idea—sleep-like replay and rehearsal as a cure for catastrophic forgetting—into a new setting: equilibrium-propagation-trained RNNs. That is a real gap worth filling, and the head-to-head against BPTT with the same consolidation scheme is the right question to ask. The abstract reports a clear win: EP matches BPTT on MNIST and beats it on four other image benchmarks when both use SRC. If those numbers hold under matched architectures, training budgets, and consolidation resources, this would be a meaningful result for the biologically plausible training community.\n\nThe trouble is that none of the load-bearing details are in the abstract. There are no run counts, error bars, effect sizes, or significance tests. The class-incremental protocol isn't described—how many classes per task, what order, and crucially, whether the ImageNet comparison uses a subset or the full dataset. The comparison could be fair and the effect real, but the abstract gives me no way to distinguish that from an unequal-resource artifact. For instance, EP settles to equilibrium and may require more compute per task; if the BPTT baselines did not receive equivalent sleep epochs or rehearsal samples, the superiority claim is not about algorithms.\n\nThere is also a conceptual concern that may or may not be resolved in the full text: if the sleep phase replays stored old-task examples, then the mechanism is at bottom rehearsal, and the distinction from plain \"awake\" rehearsal needs to be demonstrated, not just asserted. The abstract claims SRC is distinct, but without ablations, I can't assess it.\n\nTo be clear, these are verification gaps, not detected errors. The central claim may hold up. The paper deserves a serious referee who can request the missing methods and statistics. The idea is coherent and the work appears honest, but I would not cite it yet as evidence for anything beyond a proposal. I'd bring it to a reading group for discussion, and I'd encourage the editor to send it out—the referee will need to press hard on protocol matching and ImageNet scope.","headline":"Plausible and potentially important proposal for EP-trained RNNs, but the abstract alone cannot support the headline empirical claims; needs a rigorous full-text review.","tokens_in":1705,"tokens_out":1152,"would_cite":false,"duration_ms":13608,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sleep-like replay consolidation lets equilibrium-propagation-trained RNNs match or beat backpropagation-based models across five image benchmarks, rescuing them from catastrophic forgetting.","keywords":["equilibrium propagation","catastrophic forgetting","continual learning","sleep replay consolidation","awake rehearsal","recurrent neural networks","class-incremental learning","backpropagation through time"],"falsifier":"Re-run the class-incremental benchmarks with matched resources: identical architecture, per-task epochs, replay buffer, and SRC sleep epochs for both EP and BPTT. If BPTT with the same consolidation matches or beats EP on Fashion-MNIST, CIFAR10, or ImageNet, the claimed EP advantage over BPTT would be disproven.","tokens_in":869,"feed_emoji":"🧠","tokens_out":2058,"duration_ms":23888,"temperature":0.7,"pith_summary":"This paper tries to show that a sleep-like replay consolidation (SRC) algorithm applied after each new task can stop catastrophic forgetting in recurrent neural networks trained with Equilibrium Propagation (EP), a biologically plausible learning rule. The authors report that EP-trained multilayer RNNs equipped with SRC perform on par with BPTT-trained RNNs on MNIST and surpass them on Fashion-MNIST, Kuzushiji-MNIST, CIFAR10, and ImageNet in class-incremental learning. Adding awake rehearsal further improves long-term knowledge retention. If correct, this means biologically plausible learning rules can shed their continual-learning disadvantage and even overtake standard backpropagation when given the same consolidation mechanism.","feed_headline":"Sleep-like replay stops forgetting in EP-trained RNNs","feed_subtitle":"A sleep-phase consolidation step lets biologically plausible networks match or beat backprop across five image benchmarks.","key_machinery":"Sleep-like replay consolidation (SRC): an offline algorithm that reactivates previously learned information in an EP-trained RNN after each task, simulating memory consolidation during sleep. It is paired with awake rehearsal, a form of replay during active learning. The mechanism prevents the overwriting of old task knowledge by re-exposing the network to its own prior experience in a compressed or replayed form.","core_discovery":"The central claim is that sleep-like replay consolidation, which replays learned information during a simulated off-line phase after each task, significantly improves the resilience of EP-trained RNNs to catastrophic forgetting. In class-incremental learning, a multilayer RNN trained by EP and consolidated with SRC matches a BPTT-trained counterpart on MNIST and surpasses BPTT-based models on Fashion-MNIST, Kuzushiji-MNIST, CIFAR10, and ImageNet. Combining SRC with rehearsal during awake learning further boosts retention of old knowledge while the network continues learning new tasks.","pith_inferences":["The same sleep-like replay consolidation may transfer to other biologically plausible learning rules, such as predictive coding or direct feedback alignment, since the replay operates on stored activations rather than on the learning rule itself.","The awake-rehearsal boost hints that a continuous interplay between online replay and offline consolidation, closer to mammalian memory systems, could further close the gap to backpropagation in non-stationary settings.","A testable extension is to measure the replay buffer size and replay schedule, since the reported advantage over BPTT may depend on the amount of stored old data; varying this could reveal a trade-off between memory cost and forgetting resilience.","The ImageNet result, if replicated with a full class-incremental protocol rather than a subset, would make EP an unexpectedly strong candidate for large-scale continual learning."],"forward_implications":["Sleep-like replay consolidation is a reusable, task-agnostic mechanism that can be appended to EP training to reduce catastrophic forgetting without changing the learning rule.","EP-trained RNNs with SRC become competitive with, and on several benchmarks superior to, BPTT-trained models in continual learning.","Awake rehearsal provides an additive benefit, suggesting that combining offline consolidation with online replay yields the best long-term retention.","The results encourage using biologically plausible training algorithms in continual learning settings where backpropagation's high memory and non-local updates are undesirable.","The class-incremental results across five datasets, including ImageNet, indicate the mechanism scales to larger and more complex task sequences."],"supporting_citations":[],"fun_headline_variants":["Sleep-like replay stops catastrophic forgetting in EP-trained RNNs","EP-RNNs with sleep-like replay outlearn backprop on 5 benchmarks","Awake rehearsal + sleep replay: EP-RNNs retain knowledge better","Sleep during off-line phases helps EP-RNNs learn continually","Equilibrium Propagation RNNs benefit from sleep-like replay to forget less"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The comparison assumes that the EP and BPTT networks were given exactly the same architecture, parameter count, training epochs per task, replay buffer size, and sleep-consolidation budget, and that the continual-learning protocol (including the ImageNet scope and class ordering) was identical and representative.","fun_headline_variants_meta":{"raw":{"variants":["Sleep-like replay stops catastrophic forgetting in EP-trained RNNs","EP-RNNs with sleep-like replay outlearn backprop on 5 benchmarks","Awake rehearsal + sleep replay: EP-RNNs retain knowledge better","Sleep during off-line phases helps EP-RNNs learn continually","Equilibrium Propagation RNNs benefit from sleep-like replay to forget less"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000937,"raw_usage":{"total_tokens":3872,"prompt_tokens":801,"completion_tokens":3071,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":2977}},"tokens_in":545,"tokens_out":3071,"duration_ms":25701,"temperature":1.0,"reasoning_tokens":2977,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:07:39.191553+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the class-incremental benchmarks with matched resources: identical architecture, per-task epochs, replay buffer, and SRC sleep epochs for both EP and BPTT. If BPTT with the same consolidation matches or beats EP on Fashion-MNIST, CIFAR10, or ImageNet, the claimed EP advantage over BPTT would be disproven.","supporting_citations":[],"review_version":1}