{"id":"f411f914-7cd2-41d8-94bc-40dd2cf56108","arxiv_id":"2607.25082","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"A pointer-network actor plus GNN critic jointly handles agent and task openness in MARL without fixed bounds, with proofs of well-definedness and strong wildfire results.","lead":"PLATO is a multi-agent RL method that lets agents pick among a changing set of tasks by pointing at them, while a graph critic tracks who is present. It targets open systems like wildfire response where firefighters and fires come and go, without padding or retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"Appendix Lemma 3 proves AO rank-consistency for an attention-based \"agent pointer\" query, but Algorithm 1 (lines 10–11) implements a statistics-based query (mean/var/min/max); the formal AO guarantee does not cover the mechanism actually implemented and evaluated.","rationale":"The reader's verdict (CONDITIONAL, high confidence) is broadly right: the paper is a competent methods contribution with a real but compositional novelty, honest scope limitations (one action per task, full observability, single domain), and a careful evaluation protocol (seed-matched paired tests, Bonferroni families, Optuna-tuned baselines). I agree with the reader's weakest-assumption choice as far as it goes, but it identifies scope limits the authors themselves state, whereas the concern I raise is an internal inconsistency the authors do not flag: the appendix's central AO lemma (Lemma 3) and the illustrative §A.4 example formalize an attention-based agent pointer that Algorithm 1 does not implement, and its proof leans on Lemma 1 in a way that is not valid as stated. This is exactly the kind of load-bearing gap a stress-test should catch, because the \"proved well-defined and rank-consistent under AO\" part of the strongest claim is doing rhetorical work in the abstract and conclusions. That said, the damage is bounded: Theorem 1's well-definedness survives (it is proved for the statistics query), the empirical results are unaffected (they used the statistics query), and the fix is either a one-lemma re-proof for the statistics query (if an analogous invariance even holds — the count-blindness/aliasing issue suggests only weaker guarantees are available) or switching the actor to the attention read and re-running. Hence the verdict should remain CONDITIONAL, with an added explicit condition: reconcile Lemma 3/§A.4 with Algorithm 1 before the formal AO guarantees can be cited as supporting the deployed architecture. I do not weight the self-benchmark aspect (MOASEI/free-range-zoo from the same group) as heavily, since baselines were tuned with equal budget and seed-matched testing, but a second-domain evaluation remains the right long-term check for the generality claim.","tokens_in":38661,"tokens_out":3117,"duration_ms":98191,"concrete_test":"Two-step check. (1) Textual: confirm whether any configuration in Appendix B/Table 5 or the code instantiates Lemma 3's attention read (MLP_n keys, ego-key query, softmax read) instead of Alg. 1 line 10's [µ,Var,min,max] statistics; if none does, Lemma 3 and §A.4 describe an unevaluated mechanism. (2) Numerical: instantiate the statistics-based query with small feature vectors (e.g., f_N=2, two tasks with fixed keys, random W_K, W_q, v, MLP_s). Find agent feature sets N and N\\{j} such that the induced queries q(N), q(N\\{j}) flip the sign of u_x − u_x′ for two tasks x,x′. If such a flip exists (it should be easy to find by random search), the rank-consistency claim the paper relies on for AO adaptation fails for the implemented actor, and Lemma 3 must be re-proved for the statistics query or the architecture changed to the attention read.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim rests on two pillars: (a) formal guarantees that PLATO is well-defined and adapts correctly under AO+TO, and (b) empirical superiority on MOASEI wildfire. Pillar (a) has an internal mismatch. Algorithm 1 and Section 4.2 build the team query as d_t = [µ, Var, min, max](o_N) followed by q = MLP_s(d_t); Theorem 1 (A.2) and Lemma 1 (A.1) prove well-definedness/permutation invariance for exactly this statistics-based query. But Lemma 3 (A.3) — the lemma that supplies the substantive AO guarantee, rank-consistency of relative agent weightings under arrival/departure — is stated and proved for a different mechanism: per-agent keys k_j = MLP_n(o_j), additive-attention scores e_j against the ego key k_i, and an attention-weighted read q_t = Σ_j α_j k_j. That attention mechanism appears nowhere in Algorithm 1 or Section 4. Worse, Lemma 3's proof of (ii) asserts that permutation invariance of the attention read \"is exactly Lemma 1,\" which is not what Lemma 1 shows (Lemma 1 covers the four statistics, not an attention read). The consequence is real, not cosmetic: the rank-consistency/invariance-to-membership property in Lemma 3(iii) and the worked example in A.4 hold only for the attention read. For the implemented statistics query, an agent departure changes µ/Var/min/max essentially arbitrarily, so no analogous guarantee exists — a departure can re-rank two remaining tasks' scores through the query in ways the appendix's narrative explicitly claims cannot happen (\"a change in the team membership by itself cannot re-rank the remaining agents\"). Additionally, the statistics query is count-blind under feature agreement (n identical agents and m ≠ n identical agents yield identical d_t), an aliasing mode the paper's \"multiple statistics reduce aliasing\" remark does not cover. This does not touch Theorem 1 (well-definedness still holds for the statistics query) or the empirical tables, but it means the paper's formal story about *how* PLATO adapts to agent open","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper introduces PLATO, a CTDE multi-agent RL method for open agent systems in which both the agent set (AO) and task set (TO) change online. The actor is a one-step pointer network: each agent builds a query from a permutation-invariant team summary (mean/var/min/max statistics over agent features, passed through an MLP) and scores the currently available tasks plus a NO-OP key via additive attention, so the policy's support always equals the current task set. A centralized bipartite agent–task GNN critic handles variable state shape. The setting is formalized as a Task-and-Agent-Open Markov Game (TaAgO-MG), with proofs of permutation invariance (Lemma 1), well-definedness over unbounded spaces (Theorem 1), support-adaptive rank-consistent pointing under TO and AO (Lemmas 2–3), and a complexity bound (Theorem 2). Experiments on the MOASEI wildfire domain (setups S0–S3, grids 2×3 to 5×5) compare against DICG, DGN, MOHITO, and four heuristics, showing PLATO generally strongest under combined openness (S2–S3) natively and more consistent in zero-shot grid transfer, with an encoder/scorer ablation.","tokens_in":39179,"tokens_out":3446,"duration_ms":126894,"significance":"If the results hold, this is a useful contribution: content-based pointing is a clean, natural mechanism for open action spaces, and joint AO+TO without padding or fixed bounds is genuinely under-addressed. The manuscript ships several things that raise confidence: explicit well-definedness and permutation-invariance arguments, a per-timestep complexity analysis, seed-matched evaluation (50 execution seeds × 3 checkpoints), Shapiro–Wilk-adaptive tests with clearly defined Bonferroni families, equal-budget Optuna tuning for all baselines, a 2×2 ablation, and a thorough reproducibility appendix including compute budget. The empirical story is honestly reported, including cells where baselines win. Impact is tempered by evaluation on a single domain from the authors' own ecosystem and by a formal gap in the AO guarantee (Major Comment 1).","major_comments":[{"comment":"Lemma 3, which supplies the substantive AO guarantee (rank-consistency of relative agent weightings under arrival/departure), is stated and proved for an attention-based 'agent pointer': per-agent keys k_j = MLP_n(o_j), additive-attention scores against the ego key k_i, and query q_t = Σ_j α_j k_j. This mechanism appears nowhere in Algorithm 1 or §4.2, which implement a statistics-based query: d_t = [µ, Var, min, max](o_N), q_t = MLP_s(d_t). For the implemented query, an agent departure changes µ/Var/min/max essentially arbitrarily, so no analogue of Lemma 3(iii) holds — a departure can re-rank two remaining tasks' scores through the query, contrary to the narrative in A.4 ('a change in the team membership by itself cannot re-rank the remaining agents'). Additionally, Lemma 3's proof of (ii) claims permutation invariance of the attention read 'is exactly Lemma 1,' but Lemma 1 covers the","section":"Appendix A.3 (Lemma 3) vs. Algorithm 1 lines 10–11 / §4.2"},{"comment":"All methods are tuned with the trial objective equal to mean validation return on the 3×3 grid under S3, and native 3×3 S3 is then a headline result (Table 1). Because tuning is symmetric across methods the comparison is not unfair, but (i) the reported 3×3 S3 numbers are optimistically biased for every method relative to untuned setups, and (ii) all other grid/setup cells use hyperparameters selected on a single, most-open configuration, which may interact with each method's sensitivity to openness. Please state this explicitly, and ideally report a held-out validation (e.g., tune on S3 but also show a tuning-free robustness check, or tune per-grid) so the reader can gauge how much of the margin is tuning-induced.","section":"§5 (Metrics) and Appendix B (Hyperparameter tuning)"},{"comment":"The claim of 'more consistent zero-shot generalization than state-of-the-art baselines' is only partially supported by Table 3. DICG leads S1 and S3 at 4×4 (trained on 2×3) and S1–S2 at 4×4 (trained on 3×3); MOHITO leads S0 at 4×4; and PLATO's zero-shot S3 return trained on 2×3 tested on 3×3 is 4.95±28.12 versus DICG's 25.31±18.21 — a large, significant deficit. PLATO's advantage is concentrated in S2–S3 at 5×5 and in consistency across metrics (putouts, NOOP), which is a real but narrower claim. The abstract and conclusion should be revised to state precisely where PLATO wins and where it does not, rather than a blanket consistency claim.","section":"Abstract / §6 (Generalizability) / §8, Table 3"}],"minor_comments":[{"comment":"The one-action-per-task assumption (selecting a task uniquely determines the non-NO-OP action) is load-bearing for Lemma 2 and the actor interface. It is acknowledged as future work, but the limitation should be flagged earlier (e.g., in the abstract or introduction), since many OASYS domains have multi-action tasks.","section":"§4.1 / §3"},{"comment":"Full observability of relevant agents and tasks is assumed; combined with the fixed train seed (300) resetting every training episode to the same initial configuration, the diversity of training conditions is narrower than the 'openness' framing suggests. A short discussion of how sensitive results are to the fixed initial configuration would help.","section":"§8 / §3"},{"comment":"The conclusions of proofs (3) and (4) as typeset ('(s,∆A_1),(s,∆A_2)∈π' and '(s,v_1),(s,v_2)∈V') appear to assert the opposite of the stated uniqueness conditions (3) and (4); presumably exclusivity is meant. Please correct the notation.","section":"Appendix A.2, proof of Theorem 1"},{"comment":"The reward bounds and the integer program (Eqs. 1–7) are presented without a clear connection to the method or experiments — they are used neither for evaluation nor for shaping. Clarify their purpose or move to a separate analysis section.","section":"Appendix A.6"},{"comment":"MOHITO is tuned with hidden dimension 32 and batch size 16, versus 128/2048 for PLATO; the differing search ranges ({32,64,128} vs {64,128,256}) may handicap MOHITO. A sentence justifying the asymmetric ranges (beyond 'anchored to published defaults') would strengthen the fairness claim.","section":"Table 5"},{"comment":"Typos/typesetting: 'reewards' (A.6); missing spaces around 'open agent systems' in §1 ('referred to asopenagent systems'); 'I' stray character at the start of Proof (2) in A.2; Figure 1 is low-resolution and adds little beyond the text; Tables 19–21 and 31–36 are very dense — consider graphical summaries.","section":"Various"},{"comment":"The 'first MARL architecture to handle both AO and TO jointly' claim should be scoped more carefully given GPL (Rahman et al., 2021) for AO in ad hoc teamwork and MOHITO's hypergraph being, by the authors' own admission, expressive enough to represent AO. 'First to jointly model, train, and evaluate AO+TO without bounds' would be more defensible.","section":"§1 / §7"}],"recommendation":"major_revision","confidential_remarks":"The domain (MOASEI wildfire), the closest baseline (MOHITO), the TaO-MG formalism being extended, and much of the motivating literature come from the authors' own research group. This is not improper, but it means the empirical validation is entirely in-house; the 'first' and 'state-of-the-art' framings would be more convincing with one third-party environment or baseline. The Lemma 3 / Algorithm 1 mismatch (Major Comment 1) is the one issue I would insist on before acceptance; it is fixable without new experiments if the authors choose to downscope the AO claims, or with moderate additional work if they implement the attention-based agent pointer."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent OASYS/MARL methods paper that actually ships a joint AO+TO architecture and beats tuned baselines on MOASEI wildfire, including zero-shot grid transfer. The formal story about agent openness, though, overclaims relative to the code path in Algorithm 1.\n\nWhat is new is the packaging, not the parts. Pointer nets, MAPPO, GNN critics, and TaO-MG/MOHITO are prior art they cite. PLATO’s contribution is set-conditioned pointing over the live task set, a team-summary query for changing teams, a bipartite agent–task critic that changes shape, the TaAgO-MG extension, and a careful wildfire eval (S0–S3, native + zero-shot, Optuna-tuned baselines, ablations, seed-matched tests). That is real work for the subfield. The one-action-per-task interface is a clean design choice they own in the conclusion.\n\nWhat holds: Theorem 1 / Lemma 1 for the statistics encoder (mean/var/min/max → MLP) are fine—permutation invariance and well-definedness over unbounded sets. Empirical tables look honest; PLATO is strong under S2–S3 and more consistent on larger/unseen grids than DICG, DGN, and MOHITO. Circularity is low; self-cites are domain/baseline, not the performance claim.\n\nSoft spot that matters: Appendix Lemma 3 proves AO rank-consistency for an attention-weighted “agent pointer” (per-agent keys, ego query, α-weighted read). Algorithm 1 and §4.2 implement none of that—only the four team statistics. Lemma 3 even leans on Lemma 1 incorrectly for the attention read. So the nice claim that “team membership alone cannot re-rank remaining agents” is true for the lemma’s mechanism, not for the deployed query: a departure can move µ/Var/min/max and reshuffle task scores arbitrarily. Count-aliasing under identical features is also unaddressed. This does not sink well-definedness or the numbers; it means the formal AO narrative and the worked example in A.4 do not cover what was trained. Secondary limits (one action per task, full observability, single domain) are stated and proportionate.\n\nWho it’s for: people building open multi-agent policies and anyone comparing against MOHITO-style openness. Worth a serious referee if the authors align Lemma 3 with the stats actor (or implement the agent pointer) and keep claims scoped. I’d engage; I would not treat the AO lemmas as load-bearing until fixed.","headline":"Solid joint AO+TO MARL system with real experiments; the appendix’s AO “agent pointer” lemma does not match the statistics query actually implemented.","tokens_in":40288,"tokens_out":643,"would_cite":true,"duration_ms":18723,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A pointer-network actor and a shape-changing GNN critic let multi-agent policies keep working when agents and tasks appear or disappear, without fixed bounds or retraining.","keywords":["open agent systems","multi-agent reinforcement learning","agent openness","task openness","pointer networks","graph neural network critic","CTDE","wildfire suppression"],"falsifier":"Train and evaluate on the same wildfire setups but give each fire a multi-action menu (e.g., suppress intensity levels or tools); if PLATO then loses its advantage over bounded or task-open-only baselines, or if zero-shot returns collapse when agent or task counts far exceed those seen in training, the central claim fails.","tokens_in":39818,"feed_emoji":"🔥","tokens_out":1039,"duration_ms":26600,"temperature":0.7,"pith_summary":"Standard multi-agent reinforcement learning assumes fixed sets of agents and tasks, but many real systems are open: teammates join or leave and jobs appear or vanish mid-operation. This paper argues that those two forms of openness can be handled together by replacing fixed-index action heads with content-based pointing and by scoring the open global state with a graph whose nodes change with the current team and task set. The resulting method, PLATO, is formalized in a Task-and-Agent-Open Markov Game and proved well-defined and permutation-invariant over unbounded agent–task spaces. On a wildfire-suppression benchmark built for open multi-agent evaluation, it matches or beats strong baselines when openness is present and transfers more consistently to unseen grid sizes without retraining. A sympathetic reader cares because openness is the default in fielded teams, and prior methods either pad to artificial limits or address only one dimension of change.","feed_headline":"Pointer policies handle open agent-task teams without bounds","feed_subtitle":"PLATO points at live tasks and scores a changing agent-task graph, beating fixed-space MARL on wildfire","key_machinery":"Content-based pointer actor: a fixed-length query built from team statistics (mean, variance, min, max) is scored against keys for every current task plus NO-OP, so the policy support always equals the live task set; paired with a centralized bipartite agent–task GNN critic whose node sets grow and shrink with composition.","core_discovery":"PLATO is presented as the first multi-agent actor–critic architecture that jointly handles agent openness and task openness without padding, masking, or fixed bounds: each agent forms a permutation-invariant team query and points, via additive attention, at the currently available tasks (plus a no-op), while a centralized bipartite GNN critic evaluates the open agent–task graph. The authors formalize the setting as a TaAgO-MG, prove the actor and critic are well-defined over the resulting unbounded spaces, and show stronger native returns and more consistent zero-shot grid transfer than DICG, DGN, and MOHITO under endogenous and exogenous task openness with temporary agent exit.","pith_inferences":["The one-action-per-task restriction is the natural next stress test: a hierarchical or two-stage pointer (task then action-within-task) would show whether the openness story survives richer action menus.","Partial observability of agents or tasks would force the team statistics and task keys to be built from beliefs rather than facts, which is a direct extension the conclusion already flags.","The same pointer-plus-variable-graph pattern may transfer to other open domains (disaster response, on-demand logistics) where both workforce and job queues fluctuate.","Ablation hints that recurrent team memory plus additive scoring helps most under joint openness out of distribution, suggesting history is a cheap substitute for explicit openness models when transfer is the goal."],"forward_implications":["Policies need not reserve fixed action slots or retrain when the live task set changes; support tracks the current tasks by construction.","Agent arrivals and departures can be absorbed through a permutation-invariant team summary without rewriting the actor architecture.","Centralized critics can evaluate open multi-agent states by letting the agent–task graph change shape rather than padding to a maximum size.","Zero-shot transfer to larger spatial layouts becomes a fair test of openness handling rather than a forced architecture change.","TaAgO-MG gives a shared formal target for methods that claim to handle agent and task openness together."],"fun_headline_variants":["PLATO points at open tasks without padding or fixed bounds","Pointer actor plus GNN critic tackles agent and task openness","Unbounded MARL: PLATO handles changing agent-task teams","PLATO beats fixed-space baselines on open wildfire teams","First actor-critic jointly open to agents and tasks, no masks"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"Each task has exactly one associated action plus a shared no-op, so selecting a task fully determines what the agent does; if tasks had multi-way action menus, the current pointer interface would not apply as stated.","fun_headline_variants_meta":{"raw":{"variants":["PLATO points at open tasks without padding or fixed bounds","Pointer actor plus GNN critic tackles agent and task openness","Unbounded MARL: PLATO handles changing agent-task teams","PLATO beats fixed-space baselines on open wildfire teams","First actor-critic jointly open to agents and tasks, no masks"]},"model":"grok-4.5","effort":"low","cost_usd":0.001953,"raw_usage":{"total_tokens":1008,"prompt_tokens":915,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":19528000,"prompt_tokens_details":{"text_tokens":915,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":20,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":915,"tokens_out":73,"duration_ms":2475,"temperature":1.0,"reasoning_tokens":20,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T01:48:00.689889+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate on the same wildfire setups but give each fire a multi-action menu (e.g., suppress intensity levels or tools); if PLATO then loses its advantage over bounded or task-open-only baselines, or if zero-shot returns collapse when agent or task counts far exceed those seen in training, the central claim fails.","supporting_citations":[],"review_version":1}