{"id":"409801c4-d9ef-4b2c-95aa-2947db3c0d03","arxiv_id":"2608.09130","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MARA combines flow-matching loss prediction with a multi-agent transformer allocator to improve task completion rates in discrete-node computational resource efficient learning.","lead":"A new scheduler called MARA uses generative prediction of learning curves and multi-agent reinforcement learning to decide which training tasks should get compute time on discrete nodes, completing 63.46% of tasks in simulation versus 54.93% for a strong baseline. The paper shows that combining better loss-curve forecasting with coordinated node assignment helps when tasks arrive dynamically with deadlines and unknown resource needs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's ablations do not establish that the flow-matching predictor drives MARA's completion-rate gain: the FM-vs-WLS gap inside the same MAT allocator is only 1.39 points and is reported without any significance test, whereas replacing the LARA allocator accounts for roughly five times that…","rationale":"I focused on the paper's internal ablations rather than the simulation-fidelity limitation already emphasized by the reader. The simulation concern is real but explicitly acknowledged in the conclusion and is an external-validity issue; it does not invalidate the claim within the paper's own defined setting. The FM attribution, by contrast, is internal, testable from the reported numbers, and central to how the paper frames its contribution. The controlled comparison MARA versus MARA-WLS holds the MAT allocator and reward shaping fixed, so it isolates the predictor; the 1.39-point gap is small and unreported for significance, while the allocator change accounts for a roughly five-times-larger gap. This means the reader's strongest claim, interpreted as the combination of FM and MAT materially improving completion, is only partially supported. The headline 8.54-point margin over LARA is not in dispute, but the causal story is. A concrete paired significance test on the existing evaluation data would settle whether the FM component earns its place in the central claim. The reader's rationale already noted this concern, though their weakest_assumption field focused on simulation fidelity, hence partial agreement.","tokens_in":14158,"tokens_out":15255,"duration_ms":151749,"concrete_test":"Recompute the MARA-versus-MARA-WLS comparison from Table 2 as a paired test using the 15 training-seed/evaluation-root combinations (or the raw per-episode logs): construct a 95% bootstrap or permutation confidence interval for the 1.39-point mean difference, pairing by evaluation root and training seed where applicable. If the CI includes zero, revise the abstract and Section 5.8 to state that FM's end-to-end benefit is not statistically detectable and attribute the improvement to the MAT allocator and dense shaping; if the CI excludes zero, the current 'flow-matching-guided' framing survives. Report the test statistic and CI in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, as the reader formulates it, is that the combination of flow-matching demand prediction and a multi-agent transformer allocator materially improves task completion. The paper's own controlled comparison in Table 2 isolates the predictor: MARA (FM+MAT+dense shaping) averages 63.46% versus 62.07% for MARA-WLS (WLS+MAT+dense shaping), a 1.39-point gap. With reported sample standard deviations of 3.61 and 2.91 over the same 15 training-seed/evaluation-root combinations, this gap is comparable to one standard error; the paper reports no significance test. Meanwhile, MARA-WLS exceeds LARA by 7.14 points and MARA exceeds LARA-FM by 2.82 points, so the dominant source of the headline 8.54-point margin is the MAT allocator and the dense shaping in Table 5, not the flow-matching predictor. Section 5.8 concludes that 'improved forecasts affect downstream allocation,' but the quantitative support is at best weak; the abstract and contribution list lead with flow matching. The completion-rate claim itself may be true, but the causal attribution in the central claim is not supported by the reported data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies allocation of discrete compute nodes among concurrently arriving learning tasks whose required training effort is unknown. It proposes MARA, which couples a conditional flow-matching predictor of future loss trajectories with a multi-agent transformer allocator, trained via PPO with a potential-based dense reward. Experiments on simulated workloads derived from LARA report that MARA completes 63.46% of tasks at the training load, outperforming the adapted LARA baseline by 8.54 percentage points and remaining ahead under heavier loads and held-out task sets, while flow matching reduces remaining-resource prediction error relative to WLS.","tokens_in":14376,"tokens_out":13080,"duration_ms":123617,"significance":"If the reported results hold, the paper offers a useful step toward discrete-node CoRE-Learning and a clean integration of generative trajectory forecasting with cooperative multi-agent allocation. The evaluation has real strengths: shared pre-generated workloads, equal weighting of task sets, frozen evaluation curves, and a controlled stress test that separates infeasible-task waste from allocation quality. The paper also gives a transparent derivation of the potential-based shaping reward. However, the empirical support for the specific claim that flow matching drives the end-to-end gain is weak, and the objective-preserving theorem depends on an unverified terminal condition. These two issues are load-bearing for the paper's central claims.","major_comments":[{"comment":"The paper's central attribution that improved forecasts drive end-to-end gains is not supported by the reported numbers. The controlled predictor comparison inside the same MAT allocator gives MARA 63.46% versus MARA-WLS 62.07%, a difference of 1.39 points, with sample standard deviations of 3.61 and 2.91 over the same 15 training-seed/evaluation-root combinations; no significance test is reported, and the gap is comparable to one standard error. By contrast, the MAT allocator and the dense shaping reward account for much larger margins (MARA-WLS exceeds LARA by 7.14 points; dense shaping exceeds sparse by 5.59 points). The authors should report paired tests, confidence intervals, or effect sizes for the FM-versus-WLS comparison, and if the evidence remains weak they should temper the causal claims in the abstract and contribution list.","section":"Section 5.3 and Section 5.8, Table 2"},{"comment":"The objective-preserving property is derived only under the assertion that the active set is empty at both episode boundaries, so Φ0=ΦT=0. The simulator description says it runs for 500 steps with Bernoulli task arrivals; tasks arriving near the end can still be active at time T and can have deadlines beyond step 500. In that case the telescoping sum in Eq. (15) leaves a residual βΦT, and the shaped return is not equal to the number of completed tasks. The authors should either restrict the workload generation so all tasks resolve before T, define the terminal state to have an empty active set, or extend the proof to account for a nonzero terminal potential.","section":"Section 4.3, Eq. (15), Appendix A.3"}],"minor_comments":[{"comment":"There are typos in the text: 'telative log targets' should be 'relative log targets' and 'separetely' should be 'separately'.","section":"Section 4.1"},{"comment":"The index range of y_j is not stated; please specify that j runs from 1 to H and explain how the boundary at j=0 is handled by the cumulative-sum initialization.","section":"Equations (4)-(5)"},{"comment":"The stress test reports zero standard deviation for MARA and Oracle; please explain how the 15 training-seed/evaluation-root combinations are realized for this constructed scenario and why no variation is observed.","section":"Section 5.4, Table 3"},{"comment":"The operational diagnostics are reported as means with sample standard deviations, but the number of episodes and the pairing structure are not given; please provide these details so the reader can judge the precision of the comparisons.","section":"Section 5.7, Table 4"},{"comment":"The phrase 'preserving multimodal future behavior' is misleading because the final remaining-demand estimate is a median of sampled trajectories; please clarify that the distribution is used only to derive a robust scalar estimate.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and has a carefully structured evaluation. The two major issues are fixable in revision: adding rigorous statistical support for the FM-versus-WLS comparison and fixing or clarifying the terminal condition in the shaping proof. I would not reject on novelty grounds, but the abstract and contribution list currently overstate the role of flow matching relative to the evidence. The fit to DAI is good."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is legitimate and the simulator result is probably real, but the abstract and contributions sell flow matching as the engine of the gain, and the paper's own controlled comparison doesn't back that. The real story is that the MAT allocator with dense shaping buys most of the improvement over LARA, with FM adding a small increment.\n\nWhat's actually new: this is the first multi-agent treatment of discrete-node CoRE-Learning with dynamic arrivals and unknown demand, and it's a natural problem. The conditional flow-matching predictor for learning-curve extrapolation is a sensible choice, and the prediction tables (Table 1) show FM clearly beats WLS on remaining-batch MAE and budget-success accuracy, especially on ID and RL OOD. The evaluation design is genuinely careful: pre-generated workloads, equal-weight aggregation across task sets, separate training of predictor and policy, and a correct potential-shaping derivation (Eqs. 13–15). The infeasible-task stress test is a nice probe that shows the method can let go of hopeless tasks.\n\nThe soft spots are in the attribution. In Table 2, swapping FM to WLS inside the same MAT allocator changes the average by only 1.39 points, with sample standard deviations of 3.61 and 2.91—the gap is within a standard error and no significance test is reported. Meanwhile, MARA-WLS beats LARA by 7.14 points, so the dominant term in the 8.54-point headline margin is the allocator plus dense shaping, not the predictor. Section 5.8's claim that 'improved forecasts affect downstream allocation' is only weakly supported by the controlled comparison. Also, the ViT OOD evaluation uses flow-sampling settings selected on the held-out family's validation data, which undercuts the zero-shot narrative for that task set. The paper should either remove the tuning leak or stop calling it zero-shot.\n\nThe other concern is reproducibility and realism: no code or data are released, and the whole evaluation runs on simulated curves with homogeneous nodes and cost-free preemption, which the authors acknowledge. That's a real limit, but not a fatal one if the artifacts are released.\n\nWho should read it: anyone working on learning-aware cluster schedulers or MARL for allocation. It deserves review, but the authors should be asked to reframe the contributions, add significance tests on the ablations, fix the ViT OOD tuning leak, and release code and workloads. I'd accept it for review with those conditions.","headline":"The simulator result is likely real, but the paper overclaims the flow-matching component's contribution; the real driver is the MAT allocator and dense shaping.","tokens_in":14952,"tokens_out":4089,"would_cite":true,"duration_ms":35402,"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":"MARA predicts future loss trajectories with conditional flow matching and assigns tasks to discrete nodes via an autoregressive multi-agent transformer, completing 63.46% of tasks at its training load—8.54 points above the LARA…","keywords":["computational resource efficient learning","multi-agent reinforcement learning","flow matching","learning-curve prediction","resource allocation","loss trajectory forecasting","discrete-node scheduling"],"falsifier":"A deployment or simulation that adds node heterogeneity and per-preemption switching costs to the paper's calibrated task streams and then measures the MARA-vs-LARA completion gap; if the gap collapses from the reported 8.5 points to near zero or reverses, the central transfer claim fails.","tokens_in":13834,"feed_emoji":"🎯","tokens_out":7890,"duration_ms":113140,"temperature":0.7,"pith_summary":"This paper addresses scheduling of concurrent learning tasks on a pool of discrete compute nodes when each task's required training effort is unknown until observed. It argues that in this setting the allocator must both predict how many batches each task still needs and coordinate node assignments under a one-task-per-node constraint. MARA combines conditional flow matching to generate a distribution over future loss trajectories and decode remaining demand, with a multi-agent transformer that assigns tasks to nodes autoregressively. The paper reports that at the training load MARA completes 63.46% of tasks, 8.54 percentage points above the LARA baseline, and that the margin persists under heavier loads and on held-out task families. The result matters because it suggests that moving from a fixed parametric extrapolator and divisible-throughput view to a generative predictor and discrete coordinated allocation improves deadline-constrained completion.","feed_headline":"New scheduler completes 63.5% of tasks, 8.5 points above LARA","feed_subtitle":"Flow-matching loss prediction plus cooperative multi-agent node assignment; margin holds under heavier loads.","key_machinery":"Two mechanisms carry the argument, joined by a shaping identity. The first is a conditional flow-matching (CFM) predictor: a GRU encoder turns a task's observed loss history into a context, and a velocity network learns to transport Gaussian noise to a 32-bin vector of interval-averaged future losses along the linear path $z_u = (1-u)z_0 + u z_1$; at inference, threshold crossings of sampled trajectories are decoded into remaining-batch estimates and the median is used as the demand signal. Because the target vector is threshold-independent, the same predictor can be decoded for any target loss and horizon. The second is a Multi-Agent Transformer (MAT) allocator: each node is an agent that selects tasks autoregressively in fixed order under masks that enforce the constraint that no task occupies more than one node, trained with PPO and GAE. The identity that makes dense training feedback safe is potential-based shaping: with $\\gamma=1$ and zero boundary potentials, $\\sum_{t=0}^{T-1} r_t = \\sum_k \\mathbb{I}[k \\text{ succeeds}]$, so the shaped reward leaves the undiscounted completion objective unchanged.","core_discovery":"The paper's central claim is that in discrete-node Computational Resource Efficient Learning—where tasks arrive over time, each must reach a target loss before a deadline, and each node serves one task per step—the combination of a conditional flow-matching predictor and an autoregressive multi-agent allocator yields the highest task-completion rate among realizable methods. At the training load, MARA completes 63.46% of tasks on average, surpassing LARA by 8.54 percentage points, and the margin holds between 8.43 and 8.84 points under heavier arrival loads. The authors argue the gain comes from both parts: flow matching reduces remaining-batch prediction error (52.1% lower capped MAE on in-distribution tasks than weighted least squares), and the multi-agent allocator makes better coordinated use of the predictions, since MARA outperforms LARA-FM by 2.83 points with the same predictor. A deliberately constructed stress test shows MARA matching the independent-feasibility upper bound (65%) by refusing to waste service on infeasible tasks, whereas FIFO and EDF complete none.","pith_inferences":["Beyond the paper: feeding the full flow-matching distribution, not just the median crossing, to the allocator as a risk signal is a testable extension that might raise completion on the ViT OOD set where median estimates are weakest.","Beyond the paper: the homogeneous-node, cost-free-preemption assumption is where real clusters differ; adding per-preemption penalties and node-speed heterogeneity to the simulator would reveal whether the 8.5-point margin survives switching costs.","Beyond the paper: real training has run-to-run stochasticity, unlike the deterministic loss curves used here; training the flow predictor on noisy prefixes would test whether the median crossing decoder preserves the completion gains.","Beyond the paper: the equal-weight average across task sets masks the small ViT OOD prediction improvement, so a dedicated study targeting slow-declining loss families could locate the method's next bottleneck."],"forward_implications":["A single trained flow-matching predictor can answer 'how many batches to target' for any loss threshold and horizon, so the same module can serve schedulers with changing service-level targets without retraining.","The allocator trained at one arrival probability retains its completion-rate margin at heavier loads, indicating zero-shot load generalization within the simulated environment.","The dense log-progress reward changes credit assignment but not the undiscounted completion objective, so the policy is genuinely maximizing the number of successful tasks rather than a surrogate.","The infeasible-task stress test shows that accurate demand estimates plus coordinated matching can prevent wasted service on hopeless tasks, a failure mode that simple FIFO and EDF policies exhibit completely."],"supporting_citations":[{"why":"Defines Computational Resource Efficient Learning and the learnability-with-time-sharing framework that motivates the problem.","marker":"[22]"},{"why":"Provides the LARA baseline, its WLS predictor, and the task bundles whose models and datasets MARA inherits and recalibrates for discrete nodes.","marker":"[17]"},{"why":"Supplies conditional flow matching, the generative framework used to model distributions over future loss trajectories.","marker":"[7]"},{"why":"Provides the Multi-Agent Transformer autoregressive sequence-modeling policy that MARA uses for node–task matching.","marker":"[18]"},{"why":"Gives the potential-based reward-shaping invariance theorem that justifies the dense progress reward without changing the completion objective.","marker":"[10]"},{"why":"Supplies generalized advantage estimation used to train the MAT allocator.","marker":"[15]"},{"why":"Supplies PPO, the policy-optimization algorithm used to train the MAT allocator.","marker":"[16]"}],"fun_headline_variants":["MARA: flow-matching scheduler completes 63.5% tasks, +8.5 pts over LARA","Flow-matching multi-agent allocator: 63.5% completion, beats LARA by 8.5 pts","Flow-matching predictor + multi-agent scheduler: 63.5% completion, +8.5 pts","New flow-matching scheduler: 63.5% of tasks, beats LARA by 8.5 points","MARA: discrete-node scheduler with flow matching hits 63.5% completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the simulated loss curves and the homogeneous, cost-free preemption and migration model faithfully representing real concurrent learning workloads; if real loss trajectories, hardware heterogeneity, or switching costs differ, the measured completion gains may not transfer to actual clusters.","fun_headline_variants_meta":{"raw":{"variants":["MARA: flow-matching scheduler completes 63.5% tasks, +8.5 pts over LARA","Flow-matching multi-agent allocator: 63.5% completion, beats LARA by 8.5 pts","Flow-matching predictor + multi-agent scheduler: 63.5% completion, +8.5 pts","New flow-matching scheduler: 63.5% of tasks, beats LARA by 8.5 points","MARA: discrete-node scheduler with flow matching hits 63.5% completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001187,"raw_usage":{"total_tokens":4901,"prompt_tokens":946,"completion_tokens":3955,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":3822}},"tokens_in":562,"tokens_out":3955,"duration_ms":29334,"temperature":1.0,"reasoning_tokens":3822,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:53:19.205864+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A deployment or simulation that adds node heterogeneity and per-preemption switching costs to the paper's calibrated task streams and then measures the MARA-vs-LARA completion gap; if the gap collapses from the reported 8.5 points to near zero or reverses, the central transfer claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Computational Resource Efficient Learning and the learnability-with-time-sharing framework that motivates the problem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the LARA baseline, its WLS predictor, and the task bundles whose models and datasets MARA inherits and recalibrates for discrete nodes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies conditional flow matching, the generative framework used to model distributions over future loss trajectories."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Multi-Agent Transformer autoregressive sequence-modeling policy that MARA uses for node–task matching."},{"cited_title":"Ng, Daishi Harada, and Stuart J","cited_arxiv_id":null,"evidence_quote":"Gives the potential-based reward-shaping invariance theorem that justifies the dense progress reward without changing the completion objective."},{"cited_title":"Jordan, and Pieter Abbeel","cited_arxiv_id":null,"evidence_quote":"Supplies generalized advantage estimation used to train the MAT allocator."}],"review_version":1}