{"id":"250c21d9-c53b-47b9-9953-84e0e0b47715","arxiv_id":"2608.12146","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RoutePack coordinates layer-wise expert placement and attention-aware data packing from rollout routing replay, reporting 8.85% and 14.89% trainer-measured throughput gains on two Ling-3.0 MoE RL workloads.","lead":"RoutePack is a scheduler for reinforcement-learning training of Mixture-of-Experts language models that chooses where experts live on GPUs and which samples go into the same training batch together, using routing information captured during rollout. On two Ling-3.0 models it reports 8.85% and 14.89% higher measured token throughput than a baseline, though the models and code are not public.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (15) optimizes max over shards of accumulated row cost, but synchronized execution rows make the step makespan the sum over rows of per-row shard maxima; these can disagree, so the 'exact' planner score may select layouts with worse throughput.","rationale":"The reader's weakest assumption is that routing replay may not match the training-time routing, a real applicability concern that the authors acknowledge in Section 6.2 and partly mitigate through R3-style replay and count refresh. My read identifies a different, more internal problem: the paper's own execution model and its scoring objective are inconsistent. The phrase 'synchronized scheduling unit' is unambiguous, and Eq. (15) optimizes a quantity that is not the makespan of synchronized rows. This is not an external artifact or a missing comparison; it is a mathematical mismatch between the stated semantics and the 'exact' score. The paper has real strengths: the hierarchical decomposition is clearly described, state-consistent materialization is thought through, the LPT placement stage has a sound aggregate invariant, and the empirical measurements are internally coherent. Those strengths do not resolve the objective mismatch, because the packing stage's contribution is exactly what the questionable score controls. I do not recommend rejecting the paper: the concern is correctable by either adopting sum_r max_g J_{r,g} as the primary score or explicitly motivating an asynchronous-shard execution model. The reader's conditional verdict already leaves room for revision, so I keep the verdict unchanged but add this technical condition to the revision list.","tokens_in":20274,"tokens_out":15036,"duration_ms":144777,"concrete_test":"Ask the authors to confirm whether a global barrier exists between execution rows. If it does, recompute the planner score on the Section 5.2 packing comparisons with the primary component replaced by sum_r max_g J_{r,g} instead of max_g sum_r J_{r,g}, using the same routing traces and fixed row count. If the selected layouts differ, or if the reported packing increments change, then Eq. (15) is misaligned with the execution model and the packing objective must be revised. A minimal two-shard, two-row instance (J_A=[[100,0],[0,100]] vs J_B=[[60,60],[60,60]]) already shows the two scores order layouts oppositely, so the check is decisive without new hardware.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 defines an execution row as 'the synchronized scheduling unit: it contains one microbatch for every DP slot,' and Figure 1 calls the layouts 'EDP-aware synchronized rows.' Under that execution model, all EDP shards must complete row r before row r+1 starts, so the optimizer-step makespan is sum_r max_g J_{r,g}, not max_g sum_r J_{r,g}. Yet Section 3.5 defines the 'exact planner score' primary component as max_g U_g = max_g sum_r J_{r,g} (Eq. 15), and Section 4.1.3 optimizes the same tuple with a scalar surrogate. The two quantities can disagree: with two shards and two rows, layout A with J=[[100,0],[0,100]] has Score primary 100 but synchronized makespan 200, while layout B with J=[[60,60],[60,60]] has Score primary 120 but makespan 120; Eq. (15) prefers A. Thus the lexicographic score is not exact for the paper's stated execution semantics, and the packing search may select layouts with worse trainer-measured throughput. This directly affects the attribution of the packing increment (4.86% on Tiny, 3.98% on Flash) and the abstract's claim that the objective 'minimizes the accumulated cost of the slowest EDP shard.' If the authors intend no per-row barrier across EDP shards, that assumption must be stated explicitly, because it contradicts the synchronized-row definition and the 'EDP-aware synchronized rows' language in Figure 1.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents RoutePack, a hierarchical planner for MoE RL training that uses rollout-time routing replay to jointly decide layer-wise expert placement and whole-sample data packing. The planner first applies LPT-based expert placement per MoE layer on aggregate routing demand, then packs samples into a fixed, capacity-feasible number of execution rows and searches their DP layout with a lexicographic objective combining an attention proxy and per-layer physical EP-rank peaks, using diverse seeding and parallel population annealing. The evaluation on two Ling-3.0-family checkpoints reports trainer-measured token throughput gains of 8.85% and 14.89% over baseline, decomposed into gains from expert rerouting and from routing-aware packing. The paper also derives a sufficient condition for CPU-side packing not to extend the training-admission critical path.","tokens_in":20633,"tokens_out":7394,"duration_ms":62846,"significance":"If the reported results hold, RoutePack is a useful contribution to MoE RL systems: it demonstrates a concrete way to coordinate expert placement with attention-aware packing under a shared routing-replay signal, and the invariant in Eq. (4) crisply justifies why packing cannot remove aggregate expert skew. The state-consistent materialization contract is a thoughtful design that preserves logical top-k routing and existing kernels. However, the headline claim depends on load-bearing technical choices that need correction or clarification, and the empirical evidence is thin; the significance is therefore currently conditional on resolving the issues below.","major_comments":[{"comment":"The 'exact planner score' is not exact for the execution model stated in Section 2.1. If an execution row is a synchronized scheduling unit with one microbatch per DP slot, all EDP shards in a row must finish before the next row starts, so the optimizer-step makespan is sum_r max_g J_{r,g}, not max_g sum_r J_{r,g} as in Eq. (15). These quantities can disagree: with two shards and two rows, layout A with J=[[100,0],[0,100]] has max_g sum_r = 100 but synchronized makespan 200, while layout B with J=[[60,60],[60,60]] has max_g sum_r = 120 but makespan 120; Eq. (15) prefers A. The packing search therefore may select layouts with worse trainer-measured throughput, and the abstract's claim that the objective 'minimizes the accumulated cost of the slowest EDP shard' is justified only if EDP shards are asynchronous pipelines without a cross-shard row barrier. The authors should either revise the objective to sum_r max_g J_{r,g} under the stated semantics, or explicitly state and defend the asynchronous-shard execution model, which currently contradicts the 'synchronized rows' language.","section":"Section 3.5, Eq. (15); Section 2.1; Figure 1"},{"comment":"The significance tests treat each optimizer-step throughput measurement as an independent sample, but all measurements within a configuration come from a single run and are serially correlated (sequence lengths and system state evolve). With 49 and 96 valid steps, the effective sample size for the Mann–Whitney tests is much smaller than the raw count, so the reported adjusted p-values (largest 0.0110) are not a reliable basis for the claim that every pairwise comparison is significant. The authors should either report run-level replication (e.g., multiple seeds per configuration) or account for autocorrelation with a block bootstrap or a time-series model; at minimum, they should present the autocorrelation of step-wise throughput and temper the significance claim.","section":"Section 5.2, Table 3"},{"comment":"The attention proxy coefficients alpha_s and beta_s are free parameters, but the paper never states their values, how they were calibrated, or a sensitivity analysis. Because the joint objective trades attention cost against expert peaks (Table 4 shows attention increasing while joint cost decreases), the reported load-balance decomposition depends on the relative scale of these coefficients. If the coefficients are arbitrary, the 'joint bottleneck' metric is not anchored to measured attention time. Please report the coefficients used for KDA and MLA, their calibration source, and a sensitivity check showing that the ranking of layouts is stable over a plausible range.","section":"Section 3.4, Eq. (12); Section 5.3, Table 4"},{"comment":"It is not stated whether the reported trainer-measured token throughput includes the planner's wall time, state materialization, and data admission. Section 6.4 admits the implementation does not yet realize the Section 4.2 schedule and that process/IPC/state-transfer costs may be exposed. If the planner runs outside the timed region, the headline 8.85%/14.89% gains overestimate the end-to-end speedup. The paper should specify what is included in the throughput measurement and, if planning is excluded, report its measured overhead on the same testbeds.","section":"Section 5.2; Section 6.4"}],"minor_comments":[{"comment":"State how many optimizer-step records were excluded due to job restarts, and confirm that the means in Table 3 are computed only over the valid steps.","section":"Section 5.1"},{"comment":"The EP balance efficiency and Attention/Joint columns are reported without variance or significance; adding per-step standard deviations or confidence intervals would make the decomposition more convincing.","section":"Section 5.3, Table 4"},{"comment":"Since the exact checkpoints are not public, releasing the routing traces and a reference implementation (or a synthetic trace generator) would materially improve reproducibility.","section":"Section 6.4"},{"comment":"The lexicographic tuple in Eq. (15) is called 'exact', but it is exact only in the sense of evaluating the stated proxy, not in the sense of matching true makespan; consider renaming it 'the planner score' to avoid overclaiming until the semantics in Eq. (15) are resolved.","section":"Section 3.5"}],"recommendation":"major_revision","confidential_remarks":"The paper's main results rely on two non-public checkpoints and include no released code or traces, which makes independent verification difficult. The referee report focuses on the exact-score mismatch under the stated synchronized-row semantics and on the statistical treatment of a single-run time series; both should be resolved before acceptance. The paper is within the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RoutePack is a genuinely useful systems paper: it identifies the coupling between expert placement and attention-aware packing in MoE RL, and it has a coherent hierarchical planner that keeps the row count fixed and preserves training semantics. The formal invariant in Eq. (4) is a nice touch—packing cannot remove aggregate expert skew—and the decomposition of measured gains into reordering vs. packing is exactly the right experiment. The related-work survey is careful, and ReLibra is positioned honestly as the closest prior art.\n\nThe main soft spot is a possible internal inconsistency in the packing objective. The paper defines an execution row as the synchronized scheduling unit (one microbatch per DP slot). Under that model, the step makespan should be sum over rows of the per-row maxima across EDP shards, i.e. sum_r max_g J_{r,g}. But Eq. (15) scores layouts by lexicographic minimization of max_g sum_r J_{r,g} (the slowest shard's total work). These two quantities disagree. A simple two-shard, two-row example shows the score can prefer a layout whose synchronized makespan is 200 over one whose makespan is 120. So the 'exact' planner score is not exact for the paper's stated execution semantics. Either the authors assume no per-row barrier across shards (which contradicts 'synchronized rows') or the packing search is optimizing the wrong quantity. This directly affects the claimed packing increment (4.86% / 3.98%) and the abstract's claim of minimizing the accumulated cost of the slowest shard. I'd want this clarified before trusting the numbers.\n\nThe other weaknesses are the usual ones: two proprietary checkpoints, one dataset, 49/96 valid steps, and step-level samples from a single run used in a Mann-Whitney test. The attention proxy coefficients are not reported. None of these are fatal in a systems paper, but they do cap my confidence at 'plausible' until the artifacts are out.\n\nBottom line: yes, send it to peer review. The idea is solid, the measurements are promising, and the objective question is resolvable. If the authors can fix Eq. (15) or clearly state the barrier assumption, this is a solid paper for a systems or MLSys venue.","headline":"Real gains and a new mechanism combination, but the packing objective's 'exact' score contradicts the paper's own synchronized-row execution model, and the evidence base is thin.","tokens_in":21161,"tokens_out":5514,"would_cite":true,"duration_ms":47353,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RoutePack uses rollout routing replay to jointly schedule expert placement and data packing in MoE reinforcement-learning training, reporting 8.85% and 14.89% trainer-measured token-throughput gains on two Ling-3.0 checkpoints.","keywords":["Mixture-of-Experts","reinforcement learning","expert placement","data packing","routing replay","load balancing","population annealing","distributed training"],"falsifier":"Two observations would settle the claim. First, inject a controlled replay-execution mismatch and execute each optimizer step with routing recorded $k$ policy updates earlier, then measure the throughput gain as $k$ grows; if the gain does not decline systematically, the exact-demand premise is not what carries the result. Second, run a synthetic workload with uniform expert popularity and equal-length responses, where the paper's own analysis predicts little room beyond length-only FFD packing; a large gain there would falsify the skew-rebalancing explanation.","tokens_in":20084,"feed_emoji":"⚙️","tokens_out":17448,"duration_ms":142459,"temperature":0.7,"pith_summary":"MoE reinforcement-learning training has two coupled load-balancing problems: the sequence lengths packed into each data-parallel microbatch set the dense attention tail, while token routing sets the sparse expert tail on expert-parallel ranks, and fixing either one alone can expose the other as the new straggler. The paper exploits a signal specific to MoE RL: rollout-time routing replay records every sample's length and per-layer expert demand before the optimizer step runs. RoutePack is a hierarchical planner that first places logical experts on physical ranks from aggregate demand, then packs whole samples into a fixed, minimum-feasible number of token-capped execution rows, scoring layouts by the accumulated joint attention-and-expert cost of the slowest expert-data-parallel (EDP) shard, the group of data-parallel slots that share one expert-parallel communicator. On two Ling-3.0 checkpoints the coordinated controls raise trainer-measured token throughput by 8.85% and 14.89% over the identity-placement, length-only baseline, with packing still contributing after reordering is applied. The case is that coordinated placement plus packing is a practical, semantics-preserving speedup for sparse RL training.","feed_headline":"15% faster MoE RL training with replay-guided packing","feed_subtitle":"Two Ling-3.0 workloads gain 8.85% and 14.89% with logical routing and row counts unchanged.","key_machinery":"The carrying mechanism is the packing-invariant lower bound of Equation (4), together with the lexicographic planner built around it. The bound states that, once a layer-wise placement $\\pi_l$ is fixed, summed row-local expert peaks cannot fall below the busiest physical rank's optimizer-window total; this is why sample reordering has a hard floor and expert placement must be decided first. Layer-wise longest-processing-time (LPT) scheduling fixes one permutation per MoE layer from the aggregate replayed counts; a fixed-bin feasibility stage certifies, or best-known bounds, the minimum row count $R^*$ under hard constraints (token capacity, nonempty cells, sample bijection, rectangular $R\\times D$ shape); and parallel population annealing then searches sample-to-cell assignments with a scalar energy that mirrors the exact lexicographic score. The joint row cost inside the score is the attention proxy $\\sum_s \\max_{d\\in D_g}(\\alpha_s \\bar{T}_{r,d}+\\beta_s \\bar{Q}_{r,d})$ over attention stages plus the per-layer busiest-rank expert load $\\sum_l \\max_p W_{r,g,l,p}$, evaluated per EDP shard so that attention tails stay cell-local while expert tails aggregate across the shard's data-parallel slots. State-consistent materialization is the piece that lets the placement take effect without breaking training semantics: expert state is copied into pre-existing slot tensors at a step boundary, and only the logical-to-physical dispatch changes.","core_discovery":"RoutePack's central claim is that routing replay turns data layout into a load-balancing control: because the planner knows the exact per-sample, per-layer expert counts $a_{i,l,e}$ before the optimizer step, it can coordinate where experts reside with which samples execute together. Placement and packing are not interchangeable. For a fixed placement, the summed row-local expert peaks are bounded below by the most-loaded physical rank's total over the optimizer window, $\\sum_r \\sum_g \\max_p W_{r,g,l,p}(x;\\pi_l) \\ge \\max_p L^{\\mathrm{agg}}_{l,p}(\\pi_l)$, so no permutation of samples can remove aggregate expert skew; conversely, a balanced aggregate placement does not stop correlated samples from overloading one rank within a single row. The planner therefore acts hierarchically: layer-wise longest-processing-time placement lowers the aggregate bound, a certified minimum row count $R^*$ is fixed, and fixed-row population annealing searches sample-to-cell assignments against the lexicographic objective (slowest EDP shard's accumulated joint cost, then total cost, then worst row-local tail), where the joint cost combines a window-normalized attention proxy with the busiest physical EP rank's load at each MoE layer. State-consistent materialization commits the placement by remapping logical experts to physical slots while preserving logical top-$k$ routing, capacity, sample coverage, equal microbatch counts, and communicator topology. The measured consequence is trainer-level token-throughput gains of 8.85% on Ling-3.0-Tiny and 14.89% on Ling-3.0-Flash, decomposing into 3.80% and 10.50% from reordering plus 4.86% and 3.98% from routing-aware packing.","pith_inferences":["The mechanism's sensitivity can be measured directly: record routing from a policy several updates older than the one that trains, run the same packing pipeline, and track how the throughput gain decays with that replay lag; the paper's own Section 6.2 admits plan quality depends on replay matching executed routing, so this decay quantifies how much of the reported gain rests on the exactness assu","Because the joint objective deliberately trades a slightly higher attention proxy for lower expert tails, sweeping the attention weights in the row cost at a fixed row count should trace a Pareto frontier between attention-optimal and expert-optimal packings, a natural extension the current implementation does not explore.","The decomposition suggests a domain-scaling prediction: workloads with stronger routing diversity (coding, STEM, mixed-domain tasks, which the paper lists as future work) should show larger reordering gains, since LPT attacks an aggregate skew bound that grows with rank-popularity concentration, while single-domain math data may understate the effect.","The same exact pre-step demand signal could drive other optimizer-window controls beyond placement and packing, such as choosing which experts share a physical rank for communication locality or ordering expert-state transfers by memory pressure; RoutePack's pattern is the general one of planning from recorded demand, not the specific two controls."],"forward_implications":["Expert reordering and routing-aware packing are complementary controls, not substitutes: each contributes independently on both models, so systems that do only length-based packing or only expert reordering leave measurable throughput on the table.","Aggregate balance and row-local balance are different targets: LPT removes more than 99% of the global EP-rank coefficient of variation, while packing cuts the worst row-local EP peak by roughly 11%, so suppressing both tails requires both controls.","The layout transformation is semantically transparent: because logical top-$k$ routing, sample coverage, capacity, equal microbatch counts, and communicator topology are preserved, any learning rule whose update is invariant to ordering and physical ownership of the sample multiset can consume the same packed layout without changing the objective.","Host-side planning need not extend training admission: when $T_{\\mathrm{LPT}}+T_{\\mathrm{pack}} \\le \\max(T_{\\mathrm{LPT}},T_{\\mathrm{aux}})+T_{\\mathrm{actor}}$, CPU packing overlaps placement-aware state materialization and adds no admission straggler.","The payoff is workload-dependent: under uniform routing, near-equal response lengths, or very small planning windows, the paper expects little room beyond standard length-only packing, and the search budget can be turned down or skipped."],"supporting_citations":[{"why":"ReLibra is the closest routing-replay-based MoE training system; it establishes replay as a load-balancing signal that RoutePack extends with capacity-feasible rows and attention modeling.","marker":"[18]"},{"why":"It supplies the R3 routing-replay mechanism that records consistent per-token expert assignments, the exact demand RoutePack consumes.","marker":"[22]"},{"why":"UltraEP justifies the physical-rank abstraction of summing a rank's local expert load and taking the maximum over ranks, which the expert cost model adopts.","marker":"[35]"},{"why":"DeepEP is the dispatch/combine backend that RoutePack retains; the planner changes only the physical expert indices presented to it.","marker":"[38]"},{"why":"AReaL's length-oriented FFD bin allocator is the packing baseline that RoutePack must beat while preserving capacity and row counts.","marker":"[5]"},{"why":"Libra is the attention-work-scheduling precedent whose sequence-derived cost model the joint objective extends to sparse expert demand.","marker":"[34]"},{"why":"Graham's LPT scheduling is the layer-wise placement algorithm that lowers the packing-invariant aggregate bound.","marker":"[15]"},{"why":"Population annealing is the search method behind the fixed-row packing solver and its parallel CPU execution.","marker":"[16]"}],"fun_headline_variants":["RoutePack: 15% faster MoE RL via routing-aware packing","Routing-aware packing speeds MoE RL by up to 15%","Expert placement and packing cut MoE RL time by ~15%","Replay-guided packing boosts MoE RL throughput by 15%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on treating the recorded routing replay, each sample's sequence length and per-layer expert counts, as exactly the demand the upcoming optimizer step will execute, so if an asynchronous or off-policy runtime reroutes with a different model version, placement and packing are optimized for the wrong demand and the measured gains are not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["RoutePack: 15% faster MoE RL via routing-aware packing","Routing-aware packing speeds MoE RL by up to 15%","Expert placement and packing cut MoE RL time by ~15%","Replay-guided packing boosts MoE RL throughput by 15%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000878,"raw_usage":{"total_tokens":3936,"prompt_tokens":1226,"completion_tokens":2710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":842,"completion_tokens_details":{"reasoning_tokens":2634}},"tokens_in":842,"tokens_out":2710,"duration_ms":16945,"temperature":1.0,"reasoning_tokens":2634,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:14:17.929683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Two observations would settle the claim. First, inject a controlled replay-execution mismatch and execute each optimizer step with routing recorded $k$ policy updates earlier, then measure the throughput gain as $k$ grows; if the gain does not decline systematically, the exact-demand premise is not what carries the result. Second, run a synthetic workload with uniform expert popularity and equal-length responses, where the paper's own analysis predicts little room beyond length-only FFD packing; a large gain there would falsify the skew-rebalancing explanation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DeepEP is the dispatch/combine backend that RoutePack retains; the planner changes only the physical expert indices presented to it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AReaL's length-oriented FFD bin allocator is the packing baseline that RoutePack must beat while preserving capacity and row counts."},{"cited_title":"Libra: Taming Attention Workload Skew in Long-Context LLM Training with Bounded Sequence Pool","cited_arxiv_id":"2607.23250","evidence_quote":"Libra is the attention-work-scheduling precedent whose sequence-derived cost model the joint objective extends to sparse expert demand."}],"review_version":1}