{"id":"63998969-6d5d-4968-b82e-f9a0804b15c1","arxiv_id":"2511.14917","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Prospective dynamics, implementable via adaptive currents, make slow leaky networks track their instantaneous target activity and thereby synchronize teaching signals in deep online learning.","lead":"A theoretical/computational paper shows that adding a 'prospective' derivative term to leaky neural dynamics lets a network track the instantaneous activity it would have if signals traveled instantly, so teaching signals arrive on time. A generalist might read it because it offers a concrete biophysical mechanism—adaptive currents—that could let hierarchical brain circuits learn from delayed feedback.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fast-adaptation assumption is load-bearing for the biological implementation; the paper itself flags it as outside the standard adaptive-neuron regime.","rationale":"I read the manuscript in good faith. The core theoretical result is the tracking theorem, and the proof is a clean identity: because τ d_t(s−fθ) = −(s−fθ), the mismatch between state and current target decays exponentially. This derivation is correct in the intended setting, where the implicit dynamics can be solved. The experiments are consistent with the theory, although no code is provided and error-bar reporting is sparse. The most load-bearing vulnerability is the step from ideal prospective dynamics to real neurons. The adaptive current in Eq. (7) approximates a perfect differentiator only as τ_a→0; in the biological regime τ_a ≥ τ, the residual in Eq. (8) is not negligible, and the paper explicitly concedes that this is a nonstandard adaptive-neuron regime. The proposed sodium-inactivation mechanism is plausible but unsupported by direct physiological measurements. This does not invalidate the mathematical/algorithmic contribution, but it makes the 'neurons with adaptive current' claim conditional. The reader identified precisely this weakness, and my assessment leaves the verdict at CONDITIONAL.","tokens_in":26788,"tokens_out":13037,"duration_ms":145285,"concrete_test":"Re-run the delayed-reaching and cartpole experiments (Sections 3.2–3.3) using the adaptive dynamics of Eq. (7) with τ_a/τ = 1 and τ_a/τ = 10, holding all hyperparameters and seeds fixed. If learning performance degrades to the leaky baseline as τ_a increases, the fast-adaptation regime is quantitatively load-bearing; if performance remains above leaky, the central claim is more robust to this physiological uncertainty.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2 itself is sound in the intended feedforward/contractive setting, but the advertised biological implementation rests on an unverified limit. Equation (7) yields prospective dynamics only when τ_a is small enough for the high-pass filter (f−a)/τ_a to approximate d_t f; Equation (8) makes the residual explicit: the adaptive tracking error is O(τ τ_a d²_t f), not zero. Brette & Gerstner (2005) adaptive-IF models place τ_a at or above the membrane time constant τ, and the paper's own Discussion admits that fast adaptation is 'outside the regime in which adaptive neurons traditionally lie' and that the hypothesis τ_a < τ is 'critical to the theory of Section 2 and its experimental counterpart.' The only biological support offered is a preprint (Brandt et al. 2024) proposing sodium inactivation, with no direct measurement that the resulting effective τ_a is below τ. Since the headline claim is that neurons 'with an adaptive current' synchronize teaching signals, the biological version of the paper is carried by this fast-τ_a assumption. If cortical adaptation cannot actually be faster than the membrane time constant, the implementation claim fails even though the abstract tracking theorem remains true.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formalizes the temporal delay of neural activities relative to the instantaneous equilibrium trajectory as a tracking problem. It proves that leaky integrator dynamics have an asymptotic tracking error bounded by γτ/μ (Theorem 1) and that adding a prospective term τ d_t f_θ(s_t,t) to the dynamics makes the residual s_t - f_θ(s_t,t) decay exponentially (Theorem 2), so that prospective neurons can effectively synchronize teaching signals. The authors then propose that adaptive currents, implemented as a high-pass filter of the input current, provide a biologically plausible approximation of prospective dynamics, and they analyze time-constant mismatches. The framework is tested on a teacher-student task with several learning rules, an inverted-pendulum control task, and a delayed-reaching task with memory-storing leaky neurons trained by an online RTRL-type rule.","tokens_in":27038,"tokens_out":5316,"duration_ms":59979,"significance":"If the results hold, the paper makes a useful contribution: it gives a clean, parameter-free characterization of the tracking error of leaky neurons and of the exponential synchronization of prospective dynamics, generalizes earlier prospective-coding theories (neuronal least action, latent equilibrium) to non-energy-based systems, and demonstrates the benefit of prospective dynamics across several learning algorithms. The derivations in the appendices are mostly clear, the tightness construction for Theorem 1 is a nice addition, and the experiments span a broader set of tasks than most papers in this area. The strength of the paper is the theoretical core: Theorem 1 and the exponential identity of Theorem 2 are stated without hidden parameters and are validated by simulation. The biological and experimental claims are, however, less robust, as detailed below.","major_comments":[{"comment":"The second part of Theorem 2, limsup ||s_t - s*_t|| = 0, does not follow from the stated assumptions. The exponential decay only shows that s_t - f_θ(s_t,t) → 0, i.e., that s_t approaches the set of fixed points of s = f_θ(s,t). Invertibility of ∂_s f_θ and Lipschitz continuity in t do not imply uniqueness of the fixed point or continuous selection of a branch; for example, f(s)=s^2 has two fixed points and the same finite-dimensional residual can decay while the state tends to either root. The intended feedforward/contractive setting is fine, but the theorem as stated is too broad. Please add the missing contraction/uniqueness assumption (e.g., the spectral condition used in Theorem 1) or prove convergence to the specific s*_t used in the paper.","section":"Theorem 2 / Appendix A.3"},{"comment":"The biological implementation claim depends critically on the adaptation time constant τ_a being much smaller than the membrane time constant τ. The high-pass filter (f-a)/τ_a approximates the derivative only in the limit τ_a→0, and Eq. (8) makes the residual O(τ τ_a d_t^2 f) explicit. The paper itself states that τ_a < τ is 'outside the regime in which adaptive neurons traditionally lie' and that the hypothesis is 'critical to the theory of Section 2 and its experimental counterpart.' The only supporting evidence offered is a preprint (Brandt et al. 2024) with no direct measurement that effective τ_a is below τ. This does not invalidate the mathematical tracking theorem, but it means the advertised 'adaptive neurons can implement prospective dynamics' headline is conditional on an unverified biological assumption. Please either provide direct evidence or reframe the contribution as a th","section":"Section 2.1 and Discussion 'Neural implementation of adaptive dynamics'"},{"comment":"The Euler-like implicit scheme in Eq. (17) is the backbone of all learning experiments, yet the authors state they are 'not aware of any theoretical guaranties' for it. Since prospective dynamics are described by a differential-algebraic equation, a naive scheme can be inaccurate or unstable in parameter regimes used in the paper, especially for the oscillatory hEP experiments. The comparison to Euler integration of the prediction-correction dynamics is reassuring but not reported quantitatively. Please add a convergence check against a high-accuracy solver for at least the teacher-student and cartpole setups, or provide an analysis of the scheme's consistency/stability.","section":"Methods, 'Simulation of prospective dynamics'"},{"comment":"The experimental evidence for the 'large variety of learning rules' claim rests on Table 1, which reports point estimates with no error bars. The learning rate is grid-searched separately for each method on only 2 seeds, so the ranking of methods may partly reflect per-method hyperparameter overfitting. For the recurrent/hEP results, the oscillation parameters |β|, ω_β, and τ_ema are also chosen by grid search. Please report mean and variance over at least 5–10 seeds, and state whether the qualitative ranking (prospective > leaky, BP ≈ prospective) is stable across the grid. This is not fatal to the theory, but it is load-bearing for the claim that prospective dynamics transfer across learning rules.","section":"Table 1 / Section 3.1"}],"minor_comments":[{"comment":"The text says 'For the recurrent BP and holomorphic EP learning experiments (Figure 1)', but the recurrent results are shown in Table 1, not Figure 1. Please correct the reference.","section":"Methods, Section 3.1"},{"comment":"The assumption that f_θ is 'Lipschitz continuous in t' should be stated more precisely, e.g., with respect to the norm on the state space and uniformly for the relevant trajectory; this will also clarify the proof of the limsup claim.","section":"Theorem 2 statement"},{"comment":"The panels E and F show time-constant mismatch; please add a legend or caption detail clarifying whether τ' is varied on the x-axis and in which units (multiples of τ). The current text refers to 'the negative peak around 0' in a way that is hard to verify from the figure.","section":"Figure 3"},{"comment":"Table 1 contains 'T rain loss' and 'T est loss'; the abstract contains 'Correspondance'. These typos should be fixed.","section":"Typography"},{"comment":"The statement 'the prospective and leaky dynamics match for τ'=0' is correct but slightly misleading, since the mismatch analysis is local around τ'=τ. It would be helpful to note explicitly that the comparison to leaky dynamics at τ'=0 is outside the range where Eq. (11) is guaranteed to hold.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for a q-bio/machine-learning interface journal. The core tracking theorem is sound in the intended feedforward setting but needs an extra uniqueness/contraction assumption. The biological implementation claim is the weakest point; as a referee I would not block publication on the absence of direct measurements, but the paper must make the conditionality clearer. The numerical integration scheme is a practical concern that can be addressed with a convergence check. The experimental section, especially Table 1, needs more seeds and error bars before the 'across learning rules' claim is fully supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper for one thing: the clean tracking theorem. Prospective dynamics of the form τṡ = −s + f(s,t) + τ d_t f(s,t) give an exponential decay of the residual ||s − f(s,t)||, so the state converges to the instantaneous target trajectory. Theorem 2 states this exactly, and the proof is straightforward once you see the rewriting. That is a real mathematical contribution, generalizing the neuronal least-action work beyond energy-based systems. Theorem 1, a tight bound for leaky integrators, is also useful.\n\nWhat the paper does well: it takes prior ideas (Mi, Haider, Senn) and puts them in a general framework with clear assumptions. The adaptive-current implementation as an approximate differentiator is nicely analyzed; the first-order mismatch analysis giving an O(τ τ_a d² f) error is a genuine extension. The numerical experiments match the theory across tracking and several learning rules, which gives confidence that the mechanism works in the intended regime.\n\nSoft spots, in proportion: the biological implementation requires adaptation faster than the membrane time constant, τ_a << τ, and the paper itself admits this is outside the standard adaptive-neuron regime. The only support is a preprint, with no direct measurements. If that assumption fails, the biological story collapses, but the abstract theorem still stands. Also, the integration scheme in Eq. 17 has no theoretical guarantees, as the authors note. Experiments use per-method learning-rate grid searches on only 2 seeds, no code is available, and Table 1 lacks error bars. One post hoc reward-shaping choice in the cartpole task is pragmatic but should not be hidden.\n\nOverall, read the paper as a mathematical/algorithmic contribution and it holds up well. Read it as a claim about cortical implementation and it needs independent support. The authors are honest about the gaps, which is a good sign. I would send this to a serious referee; the theory is solid, and the caveats are the right ones to test.","headline":"The math is the contribution: prospective dynamics provably eliminate tracking lag, but the biological implementation rests on an unverified fast-adaptation assumption.","tokens_in":27556,"tokens_out":1753,"would_cite":true,"duration_ms":18950,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92B20","37N25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Slow-integrating neurons equipped with an adaptive current can track an instantaneous target trajectory exactly, eliminating the cumulative delays that make teaching signals arrive out of sync in hierarchical networks.","keywords":["prospective neurons","teaching signal synchronization","adaptive currents","online learning","credit assignment","leaky integrator networks","equilibrium propagation","working memory"],"falsifier":"In a single leaky neuron receiving a time-varying input, measure the steady-state lag between input and voltage. Leaky dynamics predict a constant phase lag growing like gamma tau / mu; prospective dynamics predict the residual error ||s_t - f_theta(s_t,t)|| decaying exponentially to zero and the voltage leading or matching the instantaneous target. If an adaptive neuron with tau_a < tau still shows a nonzero steady-state phase lag, or if the implicit Euler scheme in Eq. 17 diverges on a simple linear target where the ODE is stable, the central practical claim fails.","tokens_in":26643,"feed_emoji":"🧠","tokens_out":4835,"duration_ms":52869,"temperature":0.7,"pith_summary":"This paper claims that the temporal lag of slowly integrating neurons—which grows with network depth and makes error signals arrive out of sync with the activity they should modify—can be removed by giving each neuron a 'prospective' input that estimates what its input will be one time constant in the future. The mathematical core is a theorem showing that if the state follows prospective dynamics, the residual error between state and instantaneous input decays exponentially and the state eventually coincides with the ideal target trajectory, so the delayed network behaves like its instantaneous counterpart. The paper then shows that a biophysically plausible adaptive current, which subtracts a low-pass filtered copy of the input, approximates this prospective term and improves tracking whenever its time constant is fast enough. On that basis, the paper demonstrates that prospective dynamics synchronize teaching signals across backpropagation, feedback alignment, and equilibrium-propagation style rules, and that they enable online learning in a motor-control task and in a delayed-reach memory task. A sympathetic reader would care because this offers a single mechanism—neural adaptation—that could solve a critical timing problem for learning in the brain.","feed_headline":"Adaptive neurons can erase the lag that breaks teaching-signal timing","feed_subtitle":"A neuron that subtracts a filtered copy of its own input tracks its target exactly, letting slow networks learn online.","key_machinery":"The key object is the prospective dynamics equation: tau s_dot_t = -s_t + f_theta(s_t,t) + tau d_t f_theta(s_t,t), where the added term is a first-order prediction of where the input is heading. Rewriting it as tau d_t[s_t - f_theta(s_t,t)] = -(s_t - f_theta(s_t,t)) exposes the mechanism: the tracking error itself obeys an exponentially stable ODE, which is why the state locks onto the instantaneous target. The biophysical implementation is the adaptive current: a low-pass filtered copy of the input acts as a high-pass filter, and when the adaptation time constant tau_a is much smaller than the membrane time constant tau it reproduces the derivative term, up to corrections of order tau tau_a","core_discovery":"The central discovery is that the delay problem for hierarchical learning is formally a tracking problem: neural activity should track the trajectory s*_t defined by s*_t = f_theta(s*_t,t), the instantaneous equilibrium of the same network. Leaky integrator dynamics tau s_dot = -s + f_theta(s,t) can approach s* only up to a tube of radius proportional to gamma tau / mu, so tracking error—and thus teaching-signal misalignment—grows with neuron time constant and input speed. Adding the prospective term tau d_t f_theta(s_t,t) changes the dynamics to tau s_dot = -s + f_theta(s,t) + tau d_t f_theta(s_t,t), and Theorem 2 shows that s_t - f_theta(s_t,t) decays as exp(-t/tau) and that limsup ||s_t -","pith_inferences":["If the fast-adaptation premise can be met in real neurons (e.g., by fast sodium inactivation), the theory predicts a measurable phase advance of voltage responses relative to slowly varying inputs, and a corresponding improvement in temporally precise learning; this is a direct experimental handle.","The tracking theorem is not limited to feedforward networks; it applies to any equilibrium-defining dynamics with invertible Jacobian, so recurrent circuits and deep equilibrium networks should show the same synchronization benefit—an extension the paper notes but does not pursue experimentally for large recurrent models.","Complementarity with eligibility traces suggests a two-mechanism picture: traces delay the presynaptic term, prospective errors advance the postsynaptic term; combining them might extend online learning to substantially deeper recurrent networks than either mechanism alone.","A practical engineering translation is that a single first-order high-pass filter inside each processing unit could let online gradient-based learning replace batched or offline training in slow physical or analog hardware, provided the integration step is numerically stable."],"forward_implications":["Any learning rule whose updates require equilibrium activities and error signals—backpropagation, feedback alignment, direct feedback alignment, holomorphic equilibrium propagation, predictive coding—can in principle be run online in slowly integrating neurons with prospective dynamics, matching instantaneous-network performance.","Teaching signals no longer need to be delayed-matched with eligibility traces or forward models; synchronization is achieved at the level of neural dynamics themselves.","Memory can be kept in slow, non-prospective neurons while only their errors are prospective; the paper shows such a hybrid network learns a delayed-reach task that pure prospective or pure leaky networks cannot solve.","Reward-based online learning in near-continuous time (actor-critic on the inverted pendulum) becomes feasible with tau = 100 ms neurons, with graceful degradation under +/-10% time-constant mismatch and failure only as the prospective term vanishes.","The decoupling of activity and error dynamics suggests architectures with prospective errors but non-prospective activity, widening the class of biologically plausible learning rules."],"fun_headline_variants":["Adaptive neurons erase teaching-signal lag in deep nets","Prospective neurons compensate for cumulative delays in learning","Neurons that anticipate inputs sync teaching signals","Slow networks can learn fast with prospective coding"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The biological implementation rests on the premise that an adaptive current can act as an accurate local differentiator, i.e., that the adaptation time constant tau_a is much smaller than the membrane time constant tau—a regime the paper itself notes lies outside where adaptive neurons traditionally operate; the numerical simulation scheme in Eq. 17 also has no theoretical convergence guarantees.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive neurons erase teaching-signal lag in deep nets","Prospective neurons compensate for cumulative delays in learning","Neurons that anticipate inputs sync teaching signals","Slow networks can learn fast with prospective coding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1298,"prompt_tokens":763,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":475}},"tokens_in":507,"tokens_out":535,"duration_ms":6206,"temperature":1.0,"reasoning_tokens":475,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:30:40.888727+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a single leaky neuron receiving a time-varying input, measure the steady-state lag between input and voltage. Leaky dynamics predict a constant phase lag growing like gamma tau / mu; prospective dynamics predict the residual error ||s_t - f_theta(s_t,t)|| decaying exponentially to zero and the voltage leading or matching the instantaneous target. If an adaptive neuron with tau_a < tau still shows a nonzero steady-state phase lag, or if the implicit Euler scheme in Eq. 17 diverges on a simple linear target where the ODE is stable, the central practical claim fails.","supporting_citations":[],"review_version":1}