{"id":"7d7aee41-4196-40b0-9253-7023f7fd4a6f","arxiv_id":"2607.17922","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PRIME recovers plasticity in shared-parameter multi-agent RL by resetting only neurons that are simultaneously forward-dormant and gradient-silent, improving IQM return by 24.9% over MAPPO under phase changes.","lead":"When disaster-response drones keep retraining while the mission changes, parts of their shared neural network quietly fall asleep and the network loses the ability to learn. PRIME resets only the neurons that are both inactive and abandoned by the training signal, reporting 24.9% higher average return than the standard MAPPO controller on a UAV emergency-communication testbed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's perturbation term conflates silent neurons with modified scalar parameters, undercounting reset cost and undermining the central theoretical claim.","rationale":"The Reader's conditional verdict is appropriately cautious, but its weakest_assumption focused on empirical scope (one simulator, two seeds) and on Assumption 2's strong convexity. I identify a more internal, load-bearing defect in the theoretical validation: Theorem 1's d*_t is defined as the number of silent neurons, while the perturbation is applied to their incoming weights, bias, and outgoing columns. The supplementary's per-neuron energy calculation ignores the outgoing-column zeroing, and the comparison to blanket perturbation d is therefore between neurons and scalar parameters. This does not invalidate the empirical comparison of PRIME vs MAPPO, and it may be repairable by redefining the tracking metric in output/function space or by counting modified parameters honestly; hence the verdict should remain CONDITIONAL rather than moving to ACCEPT or REJECT. The proposed instrumentation would settle whether the reported 'small silent-subspace dimension' actually corresponds to a small fraction of modified parameters and perturbation energy.","tokens_in":34586,"tokens_out":11728,"duration_ms":120417,"concrete_test":"Instrument the PRIME training loop (seed 42, Change mode) to record at each reset event the modified parameter indices and the squared L2 norm of the actual parameter change: Σ_{i∈S_l} (||Δw_{l,i}||² + Δb² + ||Δ[W_{l+1}]_{:,i}||²). Compute the time-averaged perturbation energy and the ratio of modified parameter count to total parameter count d. If the average energy per silent neuron scales with d_out (i.e., with the outgoing column norm) or the modified-parameter fraction is not small (e.g., >0.1), then Theorem 1's d*_t = |S_t| is a miscount and the theoretical claim must be revised; if the energy is ≈ η²γ²_r |S_t| and the parameter fraction is small, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing defect is in the perturbation accounting of Theorem 1 / Assumption 4. Assumption 4 sets d*_t = |S_t|, the number of silent neurons, and models the reset as noise projected onto a 'silent subspace of dimension d*_t', so the third term of (17) is η²γ²_r d*_t and is compared to blanket perturbation d. But PRIME's actual reset (Eqs. 12–14) modifies, per silent neuron i, the d_in incoming weights, the bias, and the entire outgoing column [W_{l+1}]_{:,i}. In parameter space the modified subspace has dimension |S_t|·(d_in + 1 + d_out), not |S_t|. The supplementary justification that 'per-neuron perturbation energy is 1' counts only the Kaiming incoming weights and explicitly sets aside outgoing zeroing as output continuity; however Theorem 1 tracks parameter error e_t = w_t − w*_t, not output error. Zeroing an outgoing column changes e_t by the current column norm. Thus the claim that reset cost scales with a small d*_t rests on comparing |S_t| (neurons) to d (scalar parameters) on inconsistent footings. For the reported architecture (1,184 actor parameters, H=32), even 10 silent neurons with d_in≈20 touch 10×(20+1+32)=530 parameters, roughly 45% of d, not a small fraction. The Noise-vs-hard-reset ablation does not repair this: matched behavioral outcome does not imply matched parameter-space perturbation energy. The theorem may survive if rephrased in function/output space, but as stated the theoretical pillar of the central claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies plasticity loss in shared-parameter cooperative MAPPO under sustained non-stationarity, in the context of UAV-assisted emergency communication networks. It reports diagnostic measurements (dormancy persistence, forward/backward decoupling, cross-agent dormancy disagreement), proposes PRIME, a periodic neuron-level reset method that resets only neurons that are both forward-dormant and backward-silent, aggregates detection statistics over the full team batch, and reinitializes with an output-preserving reset. The authors report a 24.9% IQM improvement over vanilla MAPPO on a custom phase-switching UAV-ECN simulator, extensive ablations along seven design axes, and a dynamic regret bound (Theorem 1) intended to show that the perturbation cost scales with the silent-subspace dimension rather than the full parameter count.","tokens_in":34824,"tokens_out":10200,"duration_ms":96242,"significance":"If the results hold, PRIME is a practical, low-overhead plasticity-maintenance method for a realistic and under-studied regime: shared-parameter cooperative MARL under non-stationarity. The paper's strengths include a systematic motivation study, a clearly specified algorithm, and a broad ablation matrix that isolates the training-gradient signal and team-level aggregation as the key design choices. The authors are also unusually transparent about limitations (single simulator, two seeds, deferred benchmarks). However, the theoretical pillar as stated has a load-bearing dimensional flaw, and the main experimental configuration appears to include a phase-triggered reset not present in the algorithm description. These issues can likely be repaired, but they require changes to the paper's claims and presentation.","major_comments":[{"comment":"Theorem 1's perturbation accounting conflates neurons with scalar parameters. Assumption 4 sets d*_t = |S_t| (number of silent neurons) and models the reset as noise projected onto a subspace of dimension d*_t. But the reset in Eqs. (12)-(14) modifies, for each silent neuron i, the d_in incoming weights, the bias, and the entire outgoing column [W_{l+1}]_{:,i}; in parameter space the modified subspace has dimension |S_t|(d_in+1+d_out), not |S_t|. Since the theorem tracks parameter error e_t = w_t - w*_t, zeroing an outgoing column changes e_t by the current column norm and cannot be discarded as 'output continuity'. Appendix C's justification counts only the Kaiming incoming energy and ignores outgoing weights; it also does not reconcile the Gaussian proxy's support with the actual reset support. With H=32 and Section V-C reporting Policy Layer 0 silent fractions around 20-30%, even a co","section":"Section IV-E, Assumption 4, Eq. (17), Eqs. (12)-(14)"},{"comment":"There is a mismatch between the described algorithm and the evaluated configuration. Algorithm 1 specifies a purely periodic reset (c mod F=0) with no external phase information. However, Section V-D states that 'the main experimental runs (variant A in Fig. 10(b)) additionally trigger a full reset sweep at each boundary.' The headline 24.9% IQM gain (72.626 vs 58.135) and the results in Section V-B/Table II therefore appear to be produced by variant A, not by the pure periodic procedure that the abstract and contributions claim. Variant B, which is exactly Algorithm 1, yields an even higher IQM (75.301), so the design claim is empirically salvageable, but the paper must report the clean-algorithm configuration as primary and clearly state which variant generated each figure and table. As written, readers cannot determine whether the main results validate the paper's central claim of pha","section":"Section V-D, Fig. 10(b), Algorithm 1"},{"comment":"The empirical evidence is confined to one custom simulator with U=3, N=20, H=32, and two random seeds. The paper explicitly defers larger teams/networks, standard MARL benchmarks (SMAC, MPE), and bootstrap confidence intervals to future work. Two seeds is a weak basis for the quantitative claims, especially because the AuxGrad ablation shows a 15.5-point cross-seed spread, and several figures report only min-max bands without confidence intervals. This does not invalidate the within-simulator comparison, but it sharply limits the generalizability claims in the abstract and introduction ('first method in multi-agent communication systems'). At minimum, the authors should add at least one more seed or report bootstrap CIs, and either temper the general claims or include a standard benchmark with larger networks.","section":"Section V-A, V-B"}],"minor_comments":[{"comment":"Typo: 'Substituting the update rule the update rule of the main paper' and the reference to 'Section IV-F' should be 'Section IV-E'.","section":"Supplementary Appendix B"},{"comment":"'UA V' should be 'UAV' for readability.","section":"Abstract / throughout"},{"comment":"The legend uses 'LD', 'LZG', 'LDI' and the formula 'LDI = LD LZG' without defining the intersection symbol; the notation is clear only after reading the main text.","section":"Figure S10"},{"comment":"The text says 'The sweep spans 47 IQM points' and then lists four values; this should be '4'.","section":"Section V-D, AuxGrad paragraph"},{"comment":"The notation 'τ_aux^g' appears in the text but not in Table II, and the threshold values are not fully annotated in the table header.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper's simulator is explicitly 'following the UAV-ECN setup of [22]', the authors' own companion paper, so the evaluation environment is not independent. The phase-trigger discrepancy in Section V-D is the most serious presentation issue: the headline result is generated by a configuration that uses the external phase signal, contrary to the algorithm description. The theorem's assumption-based d*_t definition also risks being seen as definitional rather than derived. These are fixable, but the revision must be substantive, not cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a worthwhile paper on plasticity loss in shared-parameter cooperative MARL, and the empirical work is mostly careful, but the theoretical claim in Theorem 1 doesn't hold up as written. The reset touches d_in+1+d_out parameters per silent neuron, not one, so the 'perturbation cost scales with small d*_t' statement is not supported by the argument as presented.\n\nWhat's genuinely new: the paper gives the first systematic look at dormant neuron accumulation in shared-parameter CTDE, shows that forward dormancy and gradient silence decouple in three of four layers, and that per-agent dormancy judgments disagree 5–18%. That's a real, useful observation. Using the live PPO gradient as the backward signal is a good idea and costs nothing extra; team-batch aggregation is well motivated and the SingleSlice ablation gives a convincing 8.1 IQM point demonstration. The ablations are extensive and well-displayed, and the authors are honest that the reset operator is exchangeable. That is good experimental practice.\n\nThe soft spots: the theory. Assumption 4 models the reset as noise on a subspace of dimension d*_t = |S_t|, but each reset changes the incoming weight vector (d_in scalars), the bias, and zeroes the entire outgoing column. In parameter space that's |S_t|(d_in+1+d_out) modified scalars, not |S_t|. The supplementary's claim that per-neuron energy is 1 counts only the Kaiming incoming weights; zeroing the outgoing column changes the error vector by the column's current norm. So the comparison to blanket perturbation on dimension d is apples to oranges. The theorem might be salvageable as an output-space or function-space tracking bound, but as stated it overclaims. The empirical story does not collapse—the method still works in the narrow testbed—but the theory section needs rework.\n\nThe other limitations are mostly acknowledged. One simulator, U=3, H=32, two seeds is thin for a 24.9% headline, and the phase-triggered sweeps in variant A sit oddly with the abstract's periodic-only claim (variant B rescues it). The AuxGrad comparison required a seed-42 threshold sweep and the winning setting stalls on seed 43; that's transparently reported but still fragile. No PE-MAMoE comparison, no code/data. All of this tempers the empirical confidence but doesn't undermine the core idea.\n\nWho this is for: researchers working on plasticity loss, continual RL, or MARL with shared parameters; also anyone building non-stationary UAV communication controllers. It deserves a serious referee—send it out, but expect the theory to be the main battleground.","headline":"Solid method paper with honest empirical work, but Theorem 1 overclaims by counting neurons where the reset modifies whole parameter slices.","tokens_in":35511,"tokens_out":4539,"would_cite":true,"duration_ms":41453,"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":"Under sustained non-stationarity, shared-parameter cooperative MARL loses learning capacity as neurons go dormant; PRIME restores it by resetting only neurons that are simultaneously forward-dormant and backward-silent, gaining 24.9% IQM re","keywords":["multi-agent reinforcement learning","plasticity loss","dormant neurons","silent neuron detection","UAV-assisted emergency communication","non-stationary environments","shared-parameter CTDE","selective reinitialization"],"falsifier":"Run PRIME against plain MAPPO on standard cooperative MARL benchmarks with larger networks (more hidden units and agents) and a wider seed population, and measure the per-layer forward-dormant versus backward-silent gap. If the 24.9% IQM advantage does not persist, or if the gap vanishes so the intersection collapses to a forward-only criterion, the claim that the bidirectional team-aggregated gate carries the gain would be contradicted. A second targeted check: on the paper's own setup, replace the live PPO gradient with an output-sensitivity proxy; the paper reports this proxy stalls on one","tokens_in":34237,"feed_emoji":"🧠","tokens_out":6620,"duration_ms":60960,"temperature":0.7,"pith_summary":"The paper tries to establish that the main hidden bottleneck in non-stationary cooperative multi-agent reinforcement learning is internal: as objectives shift, neurons in the shared policy progressively fall dormant, and a policy that still looks functional is quietly losing the capacity to learn. It argues that the standard repair, resetting dormant neurons, is unsafe in shared-parameter training because many inactive-looking neurons still carry strong training gradients, and whether a neuron looks dormant depends on which agent's observations it processes. PRIME therefore resets only neurons that are simultaneously forward-dormant and backward-silent, using the gradient the training loss has already produced, aggregated over the full team batch. If the paper is right, a lightweight periodic reset of exactly this intersection preserves useful representations while restoring plasticity, which matters for UAV emergency networks where disaster phases shift demand abruptly.","feed_headline":"Reset only silent neurons lifts drone-network returns 25%","feed_subtitle":"Disaster-phase shifts silently kill policy neurons; reset only the truly silent ones and beat MAPPO by 24.9% IQM return.","key_machinery":"The silent neuron intersection S_l = D_l ∩ G_l: D_l is the forward-dormant set (normalized mean absolute activation ≤ τ_d, aggregated over the full B×n team batch), and G_l is the backward-silent set (normalized mean absolute gradient of the live PPO training loss ≤ τ_g). Only neurons in this intersection are reinitialized, via Kaiming-uniform incoming weights, zeroed bias, zeroed outgoing weight column, and cleared Adam state. This machinery carries the argument by guaranteeing that every reset targets capacity that is genuinely expendable, which keeps the perturbation term of the regret bound proportional to the silent-subspace dimension rather than the full parameter count.","core_discovery":"The paper claims that in shared-parameter cooperative MAPPO under sustained non-stationarity, the only safe neuron-reset set is the intersection of forward-dormant and backward-silent neurons, aggregated over the full team batch. Forward dormancy alone is misleading: in three of four hidden layers, the forward-dormant set is substantially larger than the backward-silent set, so a forward-only reset would discard neurons the optimizer is still actively steering. PRIME's criterion reads the backward signal from the gradient the PPO training loss has already deposited, and reinitializes only the intersection set with an output-preserving reset. On a phase-switching UAV emergency communication s","pith_inferences":["Editorial extension: the measured decoupling between forward dormancy and training-gradient silence may be a generic signature of distributional shift in shared-parameter MARL; one could instrument other cooperative domains for the same gap and use it as an early-warning indicator before phase changes.","Editorial extension: the exchangeability of the reset operator suggests future work can focus on detection quality and trigger schedule rather than the intervention itself; making the reset period adaptive to online dormancy or feature-rank monitoring is a natural next step.","Editorial extension: a direct testable extension is to measure whether the same intersection criterion helps in standard cooperative MARL benchmarks and larger actor networks; if the forward/backward decoupling gap shrinks there, the gain attributed to the gradient gate would be expected to shrink correspondingly."],"forward_implications":["PRIME can be added to any shared-parameter MAPPO/CTDE pipeline without changing the network architecture; detection is purely periodic and needs no external signal about when the environment changes.","In the paper's phase-switching UAV emergency communication simulator, PRIME improves IQM return by 24.9% over MAPPO (72.626 vs 58.135) and keeps the dormant neuron fraction at 10-20% versus MAPPO's 40-45%.","Ablations locate the performance gain in the backward signal being the live PPO training gradient and in team-level B×n aggregation, not in the reset operator: a matched-cadence stochastic perturbation of the same silent set performs on par, while restricting detection to a single agent's slice costs 8.1 IQM points.","The dynamic regret bound shows the cost of PRIME's resets scales with the time-averaged silent-subspace dimension rather than the full parameter count, so the improvement over blanket perturbation is expected to grow with network size.","The forward-only baseline is worse than plain MAPPO in change mode (39.410 IQM vs 58.135), supporting the paper's claim that resetting dormant-but-gradient-active neurons destroys learning in progress."],"fun_headline_variants":["Rescue dormant neurons, not all: PRIME lifts UAV net returns 25%","Only reset silent neurons: PRIME beats MAPPO by 25% IQM","Dual-check neuron reset: PRIME gains 25% in UAV networks","Team-aware reset: PRIME improves drone comms by 25%","Verify both signals before reset: PRIME's 25% UAV boost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The decisive premise is that the single custom simulator configuration (three UAVs, twenty users, 32 hidden units, one cyclic phase schedule, two seeds) fairly represents shared-parameter cooperative MARL generally; the regret bound additionally assumes the clipped PPO surrogate is locally strongly convex inside its trust region.","fun_headline_variants_meta":{"raw":{"variants":["Rescue dormant neurons, not all: PRIME lifts UAV net returns 25%","Only reset silent neurons: PRIME beats MAPPO by 25% IQM","Dual-check neuron reset: PRIME gains 25% in UAV networks","Team-aware reset: PRIME improves drone comms by 25%","Verify both signals before reset: PRIME's 25% UAV boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000644,"raw_usage":{"total_tokens":2826,"prompt_tokens":803,"completion_tokens":2023,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":1931}},"tokens_in":547,"tokens_out":2023,"duration_ms":12051,"temperature":1.0,"reasoning_tokens":1931,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:38:05.772422+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PRIME against plain MAPPO on standard cooperative MARL benchmarks with larger networks (more hidden units and agents) and a wider seed population, and measure the per-layer forward-dormant versus backward-silent gap. If the 24.9% IQM advantage does not persist, or if the gap vanishes so the intersection collapses to a forward-only criterion, the claim that the bidirectional team-aggregated gate carries the gain would be contradicted. A second targeted check: on the paper's own setup, replace the live PPO gradient with an output-sensitivity proxy; the paper reports this proxy stalls on one","supporting_citations":[],"review_version":1}