{"id":"5add8099-d35f-403d-a004-37ffbf95c6b3","arxiv_id":"2512.18540","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A graph-neural-network policy parameterization guarantees closed-loop stability by construction and transfers from small to large networks without retraining.","lead":"The paper makes distributed reinforcement-learning controllers stable by construction: each agent's action is the product of a bounded direction chosen by a graph neural network and a magnitude produced by a stable filter. This lets a policy trained on a small network run on larger, differently connected networks without retraining, which matters for robot swarms and other safety-critical multi-agent systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 1's proof assumes the LRU magnitude operator is L_p-stable from Λ's eigenvalues alone, but the LRU's NN readout is unconstrained; nonzero bias or non-Lipschitz NN makes M(0)∉ℓ_p, so 'stability for any parameters' is unsupported.","rationale":"The reader's conditional verdict is reasonable, but their weakest-assumption pick (Assumption 1 / incremental gain) is not the most load-bearing issue. Assumption 1 is explicitly stated and is a scope condition. The more serious problem is internal to Corollary 1: the proof that M∈L_p does not follow from the stable eigenvalues of Λ because the LRU contains an unconstrained neural network. For any NN with nonzero bias, the magnitude operator is not L_p-stable even at zero input, so the policy does not lie in the claimed stabilizing class. This directly undermines the central claim that stability is guaranteed for arbitrary finite parameters. The gap is easily patched by adding zero-at-zero/Lipschitz assumptions on NN, but as written the theorem is false without them. Therefore the paper remains conditional: the framework is promising, but the main stability theorem needs either an additional assumption or a modified construction (e.g., subtracting NN(0) or constraining biases). The concrete test can settle whether the released implementation actually satisfies the missing condition.","tokens_in":12902,"tokens_out":11572,"duration_ms":113324,"concrete_test":"Evaluate the LRU defined in (15) on zero input w=0 with zero initial state ξ_0=0 for a network whose NN has nonzero bias (e.g., all weights zero, bias b≠0). Compute ||LRU(0)||_p; if nonzero, M is not L_p. Then simulate the closed loop (1) with a scalar L_p-stable plant, e.g., x_{t+1}=0.5x_t+u_t, w=0, using the proposed policy, and check whether x_t converges to 0. If not, Corollary 1 is false. As a quicker check, inspect the released GitHub code to see whether the NN inside the LRU has bias terms and whether the magnitude |M_t| vanishes as w→0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central stability-by-design claim rests on Corollary 1, which requires M∈L_p for the policy (13). The proof (Appendix A.1) asserts γ(LRU)<∞ solely from |λ_i(Λ)|<1 in (15). This is not sufficient: the LRU in (15) is not a linear system but a linear recurrence followed by an arbitrary neural network NN(·;φ) and a feedthrough Fz_t. No assumption is stated on NN (e.g., Lipschitz, zero-at-zero). If NN has a nonzero bias, then with w=0, z=0, and zero initial state, the output LRU(0) is a nonzero constant sequence, which is not in ℓ_p. Hence M∉L_p, and the proof of Corollary 1 breaks; the guarantee that 'any finite parameter choice' stabilizes fails. Even with zero bias, an unbounded-gain NN (e.g., non-Lipschitz activations or unbounded weights) can map an ℓ_p input outside ℓ_p. The paper later uses σ(0)=0 for GNN layers in Lemma 1, but imposes no analogous condition on NN. This is an internal gap in the main theorem, not merely a scope limitation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed stochastic control policy class that combines Graph Neural Networks (GNNs) with a Youla-like magnitude-direction parameterization. The magnitude term is a stable operator built from a GNN followed by a Linear Recurrent Unit (LRU), and the direction term is a bounded stochastic output of a GNN-RNN. The authors claim closed-loop L_p stability by construction under Assumption 1, robustness to graph-topology and weight perturbations, and compatibility with PPO. Experiments on a multi-agent navigation task show improved reward, sample efficiency, and transfer to larger/unseen topologies relative to an InforMARL baseline.","tokens_in":13258,"tokens_out":4585,"duration_ms":45462,"significance":"If the theoretical claims were fully established, the parameterization would be a valuable contribution: it would give stability guarantees that hold before, during, and after training, while retaining the expressivity and scalability of GNNs. The paper includes attempted proofs, a PPO integration recipe, and reproducible code, and the empirical comparison is relevant. However, the central stability and robustness proofs contain gaps that, as currently written, invalidate the headline claims. The underlying idea is promising and likely repairable with additional assumptions, but this version does not yet support its main conclusions.","major_comments":[{"comment":"The proof of Corollary 1 asserts γ(LRU)<∞ solely from |λ_i(Λ)|<1 in Eq. (15). This is insufficient because the LRU in Eq. (15) is not a linear system: it contains an arbitrary neural network NN(·;φ) and a feedthrough Fz_t. If NN has a nonzero bias, then for w=0 (and hence z=0, ξ=0) the output LRU(0) is a nonzero constant sequence, which is not in ℓ_p. Thus M∉L_p and the claim that any finite parameter choice stabilizes the closed loop fails. Even with zero bias, NN must be Lipschitz to map ℓ_p inputs to ℓ_p outputs. The paper must add explicit assumptions (e.g., NN is Lipschitz, NN(0)=0, no bias, and F is bounded) or restrict the parameterization accordingly.","section":"§3.1, Eq. (15); Appendix A.1"},{"comment":"Theorem 2 uses the inequality ||F(u,w)-F(û,w)|| ≤ γ(F)||u-û||. Assumption 1 only asserts that the map (u,w)↦x has finite ℓ_p gain, which does not imply an incremental gain. This inequality is therefore not justified. The proof of Theorem 2 requires the plant operator F to have a finite incremental ℓ_p gain (a Lipschitz condition). As written, Theorem 2 is not proven. This is load-bearing for the robustness claim.","section":"Appendix A.3, Eq. (A.9)"},{"comment":"The stability-by-design guarantee depends critically on Assumption 1, which requires the plant map (u,w)↦x to already be in L_p, i.e., the plant is open-loop stable or pre-stabilized. The paper does not construct or learn such a pre-stabilizing controller; it only assumes one exists. For open-loop unstable plants, no stability guarantee is provided. This is a significant scope limitation that should be stated more prominently, and the paper should discuss how the pre-stabilizer can be obtained in practice beyond the proportional controller used in the experiments.","section":"§2.1, Assumption 1"}],"minor_comments":[{"comment":"The notation |M_t(w_{t:0})|·D_t is dimensionally unclear; it should be stated whether this is an elementwise product and what the dimensions of M_t and D_t are.","section":"§3.1, Eq. (13)"},{"comment":"The diagonal normalization term Γ(Λ) and the real operator ℜ are not defined. Please clarify their roles.","section":"§3.1, Eq. (15)"},{"comment":"The input W_t to the GNN is not defined; the state matrix X_t was defined earlier, but W_t is used without specification. Also, the finite-gain condition on Φ in Corollary 1 should be tied to the GNN architecture assumptions used in Lemma 1.","section":"§3.1, Eq. (16)"},{"comment":"Experimental details are sparse: hyperparameters, architecture sizes, training lengths, and the exact baseline configuration should be reported. The claim of 'state-of-the-art' performance is stronger than a single comparison on one task supports.","section":"§4"},{"comment":"There is a sign inconsistency in the expansion of the Lipschitz bound: the term ||∆S||·||W_{L-1}||·|Ĥ_{L-1}|| appears with a plus sign on the left and a plus sign in the aggregated Δ_{L-1}; verify the derivation.","section":"Appendix A.2, Eq. (A.4)"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work (Furieri et al., 2025) for Theorem 1 and the MAD parameterization; the novelty lies in the GNN-based distributed instantiation. The two proof gaps identified above are fixable in principle, but they currently undercut the central theoretical claims. I recommend major revision rather than reject because the proposed architecture is plausible and the experiments are encouraging."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe headline: the paper promises stability by design for distributed GNN policies via a Youla-like magnitude-direction parameterization, but the main theorem as stated doesn't hold. The LRU block in the magnitude path contains an unconstrained neural network, and the proof of Corollary 1 infers L_p stability of the LRU solely from the eigenvalues of Λ. That's not enough: a nonzero bias in NN gives a nonzero constant output on zero input, which is not in ℓ_p, so u need not be in ℓ_p even when w is. The central claim that any finite choice of parameters in this class stabilizes the closed loop is therefore unsupported. This is a load-bearing gap, not a scope note.\n\nTheorem 2 has a second problem: it bounds ||x − x̂|| by γ(F)||u − û||, which requires an incremental gain for F, but the paper only assumes finite L_p gain. That doesn't follow.\n\nWhat the paper does well: the idea of separating stable magnitude from a freely-learned stochastic direction is a sensible way to keep stability guarantees while retaining GNN expressivity, and the empirical transfer results (training on 5 agents, deploying on 1–10 with unseen topologies) are clean and useful. The PPO integration is a nice practical touch. The writing is clear and the related work is honestly placed.\n\nThe soft spots beyond the proof gaps: the whole guarantee relies on Assumption 1—the plant must already be L_p stable or pre-stabilized—and the paper doesn't formalize how the base controller in the experiments fits into the theory. Also the robustness bound is very loose and, as the authors acknowledge, nonzero even without perturbations because of the stochastic direction term.\n\nNet: this is a promising direction but the theoretical core needs repair. I'd send it to review, not desk reject, because the approach is worth engaging with. The authors should add a proper stability condition on the NN (e.g., zero bias, Lipschitz with bounded gain), and either prove an incremental-gain bound for F or rework Theorem 2. With those fixes, it could be a solid paper.","headline":"The paper's stability-by-design guarantee collapses because the LRU magnitude operator is not L_p stable for arbitrary NN parameters; the empirical work is good but the core theorem needs a fix.","tokens_in":13733,"tokens_out":3801,"would_cite":false,"duration_ms":38146,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93A14","93C10","93D05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a distributed control policy parameterization that embeds GNNs into a Youla-like magnitude-direction decomposition, so that closed-loop stability is guaranteed by construction for any parameter choice, provided the pla","keywords":["distributed control","graph neural networks","stability guarantees","Youla parameterization","reinforcement learning","multi-agent systems","magnitude-direction policies","L_p stability"],"falsifier":"Deploy the policy on an open-loop unstable plant without any pre-stabilizing base controller: trajectories diverge, disproving the universality of the stability guarantee. Alternatively, construct a plant F with finite L_p gain but no incremental gain (e.g., a nonlinearity with unbounded derivative) and check whether the deviation bound (22) still holds; if it fails, Theorem 2's robustness claim is invalid.","tokens_in":12769,"feed_emoji":"🕸️","tokens_out":4250,"duration_ms":43055,"temperature":0.7,"pith_summary":"The paper tackles a persistent problem in learning-based distributed control: neural policies trained with reinforcement learning are expressive but usually come with no stability guarantee, so a small parameter change or a different network size can destabilize the system. It proposes a policy parameterization that writes each control input as the product of a magnitude and a direction, where the magnitude is produced by a stable GNN-LRU operator acting on disturbances and the direction is a stochastic GNN-RNN output of local observations. Because the magnitude is stable and the direction is bounded, closed-loop L_p stability is guaranteed by construction for any parameter choice, as long as the plant itself is stable or has been pre-stabilized. This lets training focus purely on performance while keeping the system stable before, during, and after optimization, and it allows policies trained on small systems to transfer to larger ones.","feed_headline":"Stable-by-design GNN policies scale to larger networks","feed_subtitle":"A magnitude–direction split keeps closed-loop systems stable at any parameter setting, so training only has to optimize performance.","key_machinery":"The central object is the Youla-like magnitude-direction parameterization: the control input is factored as a magnitude times a direction. The magnitude is an L_p-stable operator M, realized by a GNN-LRU (a linear recurrent unit with a GNN front-end and stable eigenvalues), which acts on the disturbance sequence w. The direction D_t is sampled from a distribution parameterized by a separate GNN-RNN that reads local observation histories. The constraint |D_t|≤1 ensures the input norm is bounded by the magnitude, so if M has finite L_p gain, the closed loop is stable by construction. This decouples stability (carried by the magnitude) from expressivity (carried by the direction).","core_discovery":"The central claim is that the polar-decomposed policy (13)-(17) — u_t = |M_t(w_{t:0})|·D_t with |D_t|≤1, M realized by a GNN-LRU with stable eigenvalues, and D_t sampled from a distribution parameterized by a GNN-RNN — is guaranteed to produce a closed-loop map w→(x,u)∈L_p whenever the plant satisfies Assumption 1 and the GNN feature extractor has finite L_p gain. Stability therefore never depends on the training procedure or on parameter values; it is a property of the parameterization. A second result (Theorem 2) shows that finite perturbations in the graph support matrix or weights produce trajectory deviations bounded by a quantity that scales with the perturbation size, implying that ch","pith_inferences":["If the guarantee holds at scale, stability-constrained RL could replace safety-filter/fallback-controller stacks for networked systems, leaving layered safety to handle only safety (e.g., collisions), not stability.","The magnitude-direction split suggests a modular design: any feature extractor with finite gain could substitute for the GNN-LRU front-end, opening the door to other graph architectures with similar guarantees.","The non-tight bound in Theorem 2, caused by stochastic direction sampling, could be sharpened with a probabilistic analysis; the paper itself flags this as future work.","In practice, the disturbance sequence w_i may not be directly measurable at each node, so the method may require disturbance observers or an output-feedback extension to be deployable when only partial observations are available."],"forward_implications":["For any network in the described class, closed-loop stability holds before, during, and after training; training only optimizes performance, never stability.","Finite changes in graph topology or model weights produce bounded deviations in closed-loop trajectories, so policies transfer to larger networks and unseen topologies without risk of divergence.","The change-of-variable log-probability computation makes the stochastic policy trainable with policy-gradient reinforcement learning methods such as PPO.","Experiments on a multi-agent navigation task show higher rewards and lower variance than a state-of-the-art MARL baseline, while preserving stability.","Because the magnitude term vanishes at equilibrium, the closed-loop system converges rather than oscillating around the target."],"fun_headline_variants":["GNN policies keep network control stable by design","Stable GNN control: scale without stability worries","Polar-decomposed GNNs guarantee closed-loop stability","Make GNN policies stable by construction, not training","Graph neural policies with provable network stability"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The plant must already be stable or pre-stabilized (Assumption 1), so the entire stability guarantee collapses for open-loop unstable systems unless an external stabilizing controller is provided; Theorem 2 further requires the plant map to have an incremental gain, not just a finite gain.","fun_headline_variants_meta":{"raw":{"variants":["GNN policies keep network control stable by design","Stable GNN control: scale without stability worries","Polar-decomposed GNNs guarantee closed-loop stability","Make GNN policies stable by construction, not training","Graph neural policies with provable network stability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1149,"prompt_tokens":647,"completion_tokens":502,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":391,"completion_tokens_details":{"reasoning_tokens":427}},"tokens_in":391,"tokens_out":502,"duration_ms":5212,"temperature":1.0,"reasoning_tokens":427,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:57:31.099746+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the policy on an open-loop unstable plant without any pre-stabilizing base controller: trajectories diverge, disproving the universality of the stability guarantee. Alternatively, construct a plant F with finite L_p gain but no incremental gain (e.g., a nonlinearity with unbounded derivative) and check whether the deviation bound (22) still holds; if it fails, Theorem 2's robustness claim is invalid.","supporting_citations":[],"review_version":1}