{"id":"df0661e2-efd2-46b0-8624-79b4b08a8541","arxiv_id":"2607.26648","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The sparsity dividend of spiking networks is task-dependent: perception and attention can run at 2–5% firing without quality loss, while recurrent sequence models are pinned near 50% firing.","lead":"A study compares otherwise identical neural networks with normal and spiking neurons to see when the spike version really saves energy. It finds perception can run with very few spikes, but recurrent language models need about half their neurons firing, and attention sidesteps this only by using lots of memory.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The formal 'cannot' does not cover the experimental recurrent SNN: Prop. 1 assumes readout sees only instantaneous binary spikes, but the LIF RNN has an analog membrane state and multi-timestep dynamics, so the ~50% floor may be a training/readout artifact rather than a task-structural ceiling.","rationale":"The reader and I converge on the single-step channel assumption as the weak point. I sharpen it: the theorem's assumptions are not just unquantified; they are plausibly violated by the experimental LIF RNN, which has an analog membrane potential and a multi-timestep inference loop. Therefore the formal bound cannot carry the 'cannot' claim. The empirical floor is still interesting, and the paper's synthetic sweeps (rising floor with N, C, falling with H) are consistent with the bound's qualitative structure, which is why I keep CONDITIONAL rather than REJECT. A direct test with a temporal readout would settle whether the floor is task-structural or a readout/training artifact.","tokens_in":5379,"tokens_out":6451,"duration_ms":72315,"concrete_test":"Retrain the §4.2 recurrent LM with the readout computed from a short temporal average of hidden spikes (e.g., over the last 4–8 timesteps) or from the LIF membrane potential, keeping everything else fixed, and run the two-sided probe at ρ*=0.10. If bpc stays near the 0.50-firing level while hidden firing falls below ~50%, the ceiling is an artifact of the instantaneous spike readout rather than a property of the task.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central 'cannot' claim rests on Prop. 1 (§3), whose proof counts distinguishable instantaneous binary vectors s_t and explicitly assumes that both recurrent and readout pathways depend on the state only through s_t. The recurrent LM in §4.2 is an LIF network trained with surrogate gradients and run for T character timesteps; its state includes the continuous membrane potential, and nothing in the paper verifies that the readout is a function of s_t alone. If the readout or training exploits the analog state or spike timing across the T steps, the bound loosens by up to a factor of T (and in the analog-membrane limit, is vacuous), so it cannot rule out a temporal code at <50% firing. The paper's own §6 Limitation (3) concedes the bound is single-step. Consequently the headline contrast—recurrent language 'cannot' go below ~50% while attention 'sparsifies freely'—is not established as a task-structural law; it may be an artifact of the instantaneous-spike readout, the surrogate-gradient training objective, or the specific RNN architecture. The empirical floor is reproducible but the 'cannot' is overclaimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that the attainable sparsity of a spiking neural network is a property of the task, not of the SNN itself. Using a matched-architecture protocol (only the hidden unit changed: continuous vs. LIF), plus a two-sided target-firing-rate regularizer, it measures the minimum firing rate that preserves output quality across feed-forward perception, recurrent character-level language modeling, and a spiking Transformer. It reports that perception sparsifies to ~5%, recurrent language modeling cannot be pushed below ~50%, and a spiking Transformer reaches ~2%, leading to a claimed dichotomy: recurrent compression hits a firing floor, while attention pays a KV-cache memory wall. The paper formalizes the observation with a counting bound, rho >= H_b^{-1}(log2 M / H), and reports confirmatory sweeps of memory load, state width, and task difficulty. The contribution is framed as a task-conditional energy rule for neuromorphic hardware rather than a universal sparsity dividend.","tokens_in":5685,"tokens_out":6100,"duration_ms":69317,"significance":"If the result holds, it is a useful corrective to unqualified claims that SNNs save energy through sparsity. The matched-architecture protocol and the two-sided probe are a notably clean way to separate attainable sparsity from incidental sparsity, and the information-theoretic floor is a simple, parameter-free statement that connects task structure to a minimum firing rate. The paper also makes a credible qualitative case that recurrence and attention pay on different axes. Strengths include a released code repository, a controlled experimental design, and falsifiable monotonicity predictions. However, the quantitative ``cannot'' claim in the recurrent setting is not supported by the formal bound as stated, because the bound is single-timestep and the recurrent experiments operate over many timesteps with analog membrane states. The energy proxy table also contains an unexplained numerical inconsistency. The central qualitative message is defensible, but the strongest claims need either additional theory/evidence or careful restriction.","major_comments":[{"comment":"The formal firing-floor bound assumes that recurrent and readout pathways depend on the state only through the instantaneous binary spike vector s_t in a single timestep. The recurrent language experiment runs for T character timesteps and uses LIF neurons with continuous membrane potentials, so a temporal code across T steps could in principle carry up to T times more messages, with capacity H*T*H_b(rho). The paper itself concedes this in Limitation (3), and the ``Scope'' note after Corollary 1 only rules out reading the continuous membrane potential, not spike timing across timesteps. As a result, the headline ``cannot go below ~50%'' is not established as a task-structural law: the observed floor could be an artifact of the instantaneous-spike readout, the surrogate-gradient training objective, or the specific RNN architecture. Please either extend the bound to T-step codes, verify em","section":"§3 Prop. 1; §4.2; §6 Limitation (3)"},{"comment":"The energy and op numbers appear mutually inconsistent. With MAC = 4.6 pJ and AC = 0.9 pJ, an SNN with ops/ANN = 0.99 (the ρ*=.05 row) should consume roughly 0.99 * 0.9 / 4.6 ≈ 0.19x the ANN energy, not 5.05x. Similarly, the no-reg SNN row reports ops/ANN = 0.52 but energy/ANN = 2.7x. Either the two columns count different things (e.g., energy includes T replay, leakage, control, or hardware overhead while ops do not), or the table is mislabeled. This is load-bearing because the input-floor argument and the recurrence-vs-attention cost dichotomy are expressed in these reported multipliers. Please define precisely what each column includes and reconcile the arithmetic.","section":"§4.1 Table; §4.4; §2 energy proxy"},{"comment":"The confirmatory sweeps are qualitative monotonicity checks rather than quantitative tests of the bound. They corroborate the direction of ρ_min with N, H, and C, but no predicted versus observed firing rates are compared, and the bound is not used quantitatively for the experiments. This is not circular, but the word ``confirm'' is stronger than what the evidence supports. I recommend phrasing these results as consistency checks and, where possible, reporting the bound's numerical predictions for the specific M, H, and T used in the synthetic tasks.","section":"§4.3–§4.5 and Fig. 2/4/5"}],"minor_comments":[{"comment":"The firing column entries ``.16/.27'' and ``.050/.051'' are not defined. Do these refer to two layers, two timesteps, or mean/median? Please label the columns or add a footnote.","section":"§4.1 Table"},{"comment":"The notation H_b^{-1} is rendered incorrectly in the abstract (``H −1 b '') and the inverse branch of the binary entropy function should be explicitly restricted to ρ ≤ 1/2 to avoid ambiguity.","section":"Abstract, §3"},{"comment":"The caption mentions a threshold ``γ=80%'' but the text does not define γ or explain how it was chosen. Please define it in §4.3.","section":"Fig. 2 caption"},{"comment":"``No accuracy cost'' is relative to the unregularized SNN, not to the ANN: Table 4.1 shows SNN at ρ*=.05 attains 76.6% vs. ANN 88.1%. Please state this explicitly or rephrase to avoid misleading readers into thinking spiking matches analog accuracy.","section":"Abstract, §4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with a useful controlled protocol and a simple, teachable bound. However, the central ``cannot'' claim currently exceeds what the theory proves, because the recurrent experiments operate in a regime the bound does not cover (multi-timestep, analog membrane). The authors already acknowledge this in Limitation (3), so the fix may be mostly a matter of reframing and/or adding a small experiment that checks whether readout depends on s_t alone. The energy-table inconsistency is more troubling and should be resolved before publication, as it could signal a definitional confusion that affects the applied conclusions. I would not reject; the qualitative story is likely salvageable and valuable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper's real value is the protocol: holding architecture fixed and swapping only the neuron model, then probing attainable sparsity with a two-sided rate regularizer. That is a clean way to separate what the task permits from what a given training run happens to produce. The empirical contrast is striking: perception drives to 5% firing, the LIF recurrent LM is pinned near 50% no matter how hard the probe pushes, and the spiking Transformer drops to 2%. The layer-wise input floor is also a good point—dense frame replay eats the savings, so event-driven input is the real win condition for neuromorphic hardware. Code and logs are released, which is more than many SNN papers do.\n\nThe synthetic copy-task sweeps in §4.3 and §4.5 are consistent with the counting bound rho >= H_b^{-1}(log2 M/H), and the fact that difficulty (class count) raises the floor too is a nice corrective to a naive memory-only story. The bound itself is elementary, and the paper doesn't overstate its derivational novelty.\n\nThe soft spots are real, and the stress-test note is right. Prop. 1 assumes the recurrent and readout pathways see only the instantaneous binary spike vector s_t. The LIF in the recurrent LM has a continuous membrane potential and runs for many timesteps; the readout could in principle integrate the full spike train. The paper concedes this in Limitation (3), which is honest, but it means the bound does not quantitatively explain the ~50% floor. The 'cannot' claim therefore rests on a single architecture and a single task, and it is not yet a task-structural law. The Transformer memory wall is also asserted from the KV-cache size, not measured on-chip; the paper itself lists that as next-step work.\n\nSo I would not accept the strong version of the claim as-is. But the paper deserves a serious referee: the question—when does sparsity actually buy energy?—is important, the probe methodology is reusable, and the empirical dichotomy is a genuine contribution that will shape follow-up work. A careful reviewer should ask for either a multi-step bound or direct evidence that the readout is a function of instantaneous spikes, plus a couple more recurrent architectures and tasks before the word 'cannot' is used. I'd bring it to a reading group in the neuromorphic/SNN community, and I'd cite the probe and the task-dependence result (with the caveat noted) if I worked in that area.","headline":"Useful probe protocol and a real empirical contrast, but the formal firing-floor bound is too weak to justify the recurrent 'cannot' claim.","tokens_in":6121,"tokens_out":3568,"would_cite":true,"duration_ms":39153,"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":"This paper establishes that the sparsity a spiking network can achieve is set by the task's representational load, not by the neuron model: perception firing drops to 5%, recurrent language modeling is floored near 50%, and a Transformer sp","keywords":["spiking neural networks","sparsity","energy efficiency","firing-rate regularization","recurrent neural networks","transformers","information-theoretic bound","neuromorphic computing"],"falsifier":"Run the recurrent copy task with the readout allowed to see T consecutive spike vectors and measure whether the minimum firing needed for a fixed accuracy falls roughly by the factor predicted by capacity H·T·H_b(ρ); if the floor stays near 50%, the single-step bound is not the operative constraint.","tokens_in":5284,"feed_emoji":"⚡","tokens_out":5039,"duration_ms":50283,"temperature":0.7,"pith_summary":"The paper asks, for otherwise identical networks, how much energy spiking actually buys and what determines that number. Holding architecture fixed and swapping only the hidden unit between a continuous and a leaky-integrate-and-fire neuron, the author uses a two-sided firing-rate probe to push activity as low as possible before quality breaks. Feed-forward perception sparsifies to about 5% firing with no accuracy loss, recurrent character-level language modeling cannot go below about 50%, and a spiking Transformer sparsifies to 2% with no quality loss. The paper formalizes this with an information-theoretic floor, ρ ≥ H_b⁻¹(log₂ M / H), and confirms that the floor rises with memory load and difficulty and falls with state width. The practical consequence is that neuromorphic hardware wins on sparse event-driven perception, while sequence models pay either a firing floor or a memory wall.","feed_headline":"Recurrent spiking nets hit a 50% firing floor","feed_subtitle":"Same probe reaches 5% on perception and 2% on transformers—sparsity is task-conditional, not a hardware bonus.","key_machinery":"The firing-floor bound ρ ≥ H_b⁻¹(log₂ M / H), where H_b is the binary entropy function, H is the hidden width, and M is the number of distinct messages the spike state must distinguish per step. It works by counting: a binary vector of length H with at most k = ρH ones indexes at most 2^{H H_b(ρ)} messages, so separating M messages requires H H_b(ρ) ≥ log₂ M. The bound turns 'sparsity is task-dependent' into a quantitative rule and predicts the sweeps over memory load N, hidden width H, and class count C. It carries a load-bearing assumption that the recurrent and readout pathways depend on the state only through the instantaneous binary spike vector s_t, a single-step channel.","core_discovery":"The central discovery is a task-structural sparsity ceiling. With the same architecture and only the neuron model changed, a two-sided firing-rate probe drives a convolutional SNN from 27% to 5% hidden firing with accuracy on par with the unregularized network, while the same probe on a recurrent character-level language model moves firing only from 63% to about 50%—the recurrent state must stay active to carry information. A spiking Transformer on the same language task sparsifies freely to 2% firing with no bit-per-character change, showing the ceiling is a property of recurrent compression, not sequence modeling. The explanatory mechanism is the bound ρ ≥ H_b⁻¹(log₂ M / H): H binary spike","pith_inferences":["If the single-step bound is the real explanation, a recurrent spiking network that encodes information in spike timing across multiple timesteps should beat the ~50% floor; testing that directly would separate the bound's regime from a purely empirical limit.","The floor's dependence on class count C, not just memory, suggests that any task with many output classes or a large vocabulary will resist sparsification, even without long-range memory.","The measured AC