{"id":"d3a4ca31-bf0f-45ab-bff2-22572e56653f","arxiv_id":"2608.11152","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MISA-T is a router-level admission policy that allocates KV-cache capacity per workload class and by residency time, improving mixed RL rollout throughput by 35-53% versus a tuned vLLM Router baseline.","lead":"An RL rollout serving system called MISA-T adds admission control on top of prefix-aware routing, bounding how many sessions of each workload type may occupy KV cache. In tests on two large language models it raises rollout throughput by up to 53 percent and cuts end-to-end training iteration time by about 23 percent, without changing the trainer's workload mix.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MISA-T's session caps are computed from moving-window estimates of KV footprint and residency (Eqs. 10-13); the paper never stress-tests time-varying workload demand, so the central admission-control claim is not established beyond the quasi-stationary trace.","rationale":"The paper is carefully structured, and the ablation ordering Session Admission < MISA < MISA-T is internally consistent with the proposed mechanism; that is genuine supporting evidence. My concern is not that the equations are wrong, but that the online estimators are the one component whose failure mode is acknowledged in Limitations yet never exercised in the evaluation. The trace statistics in Table 2 are static, and the end-to-end experiment reports only aggregate outcomes, so a rapid class-demand shift could break the cap-setting premise without any measured signal. This does not refute the reported numbers on the specific trace; rather, it means the central claim should be scoped to quasi-stationary workload conditions until a workload-shift stress test and reproducible artifacts demonstrate otherwise. The reader's weakest assumption identifies the same load-bearing premise, and the conditional verdict remains appropriate: accept conditionally on artifact release and variance reporting, not as an established general result.","tokens_in":12427,"tokens_out":9479,"duration_ms":103568,"concrete_test":"Replay a mixed-rollout trace with the same class profiles; after one hour, abruptly shift RLVR output p50 length by +50% and increase agent tool-gap occupancy from 22.7% to 50%. Run MISA-T once with its normal moving-window estimators and once with an oracle controller that is given the new \\bar{k}_b and \\hat{T}_b values instantly but otherwise uses identical logic. If the windowed controller's rollout throughput or prefix-hit rate drops by more than 10% relative to the oracle, estimator lag is a first-order limitation; if the two tracks remain within a few percent, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that \\bar{k}_b and \\hat{T}_b in Eqs. (10)-(13) track current per-class KV block-time demand. Both quantities are estimated from recently completed sessions via moving windows. If a class's sequence lengths or residency times move faster than those windows (e.g., RLVR output length grows during training, or an agent benchmark lengthens tool-execution gaps), the class caps are stale: the controller allocates protected KV to yesterday's demand, over-admits the class whose footprint was underestimated, and evicts reusable prefixes of other classes. The paper's Limitations paragraph acknowledges dependence on labeled requests and timely snapshots, but the evaluation contains no time-varying workload-shift test; Table 2 is a static trace and the 50-iteration run only reports aggregate mixture, not a regime with abrupt demand changes. Since the reported gains (53.3%, 43.6%, 35.6%) are exactly the payoff of correct cap setting, the central claim is not robustly supported if estimator lag is significant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the problem of scheduling heterogeneous RL rollout workloads (RLVR, RLHF, and agentic) on a shared LLM inference service. It proposes MISA-T, a routing-layer admission policy that (i) adaptively caps the number of admitted sessions per instance, (ii) partitions protected KV-cache capacity among workload classes based on estimated spatial demand, and (iii) weights these allocations by class residency time. The authors evaluate MISA-T against a sweep-tuned cache-aware vLLM Router in rollout-only ablations on Step3.7 and Qwen3.6-35B-A3B, reporting rollout-throughput gains of 53.3% and 43.6%, and in a matched 50-iteration Step3.7 training run, reporting a 35.6% throughput gain, 22.8% lower mean iteration time, higher prefix hit rate (96.2% vs 74.5%), mixture deviation of 2.71 percentage points, and task scores within 0.5 pass@4 points.","tokens_in":12653,"tokens_out":6545,"duration_ms":55526,"significance":"If the reported gains are robust, the work fills a real gap: existing prefix-aware routers optimize placement but do not control how many heterogeneous sessions compete for KV capacity, and the paper's admission-control approach is complementary to existing serving stacks. The authors include a real production trace, an ablation that separates the three controller components, and a matched end-to-end comparison, which is more than many systems papers provide. The central idea is simple and plausible, and the results are internally consistent in direction. However, the evaluation's lack of variability reporting and the absence of a non-stationary workload test leave the magnitude of the claimed benefit not yet established to the standard required for publication.","major_comments":[{"comment":"The paper states that each reported value is the arithmetic mean of three independent runs, but it reports no variance, range, or significance test for any metric. Since the headline claims are improvements of 53.3% and 43.6% over the baseline, the lack of error bars leaves open the possibility that run-to-run variability is of the same order as the reported gains. Please report per-run values, standard deviations, or confidence intervals for the key metrics (sample rate, prefix hit rate, iteration time) in Tables 4 and 5.","section":"§5.3, Table 5"},{"comment":"The 'drain time' balancing argument in Eq. (12) concludes that allocating quota proportionally to R_{w,b} approximately balances estimated drain times across classes. This conclusion depends on the unstated assumption that each class's backlog drain time is inversely proportional to its allocated block quota. Additionally, the max(1, ·) floor in Eqs. (10) and (13) can admit a session whose footprint \\bar{k}_b exceeds the allocated quota M, which would violate the protected-capacity constraint in Eq. (3). Please either prove the balancing property under stated assumptions or weaken the claim, and justify or remove the floor.","section":"§4.3, Eqs. (12)-(13)"},{"comment":"The Qwen3.6-35B-A3B results show MISA with a prefix hit rate of 79.2%, much lower than Session Admission's 96.3%, yet the text does not explain why workload-aware allocation alone degrades prefix locality so sharply on this model. Because MISA-T differs from MISA only by the residency weighting, the unexplained MISA behavior makes it difficult to attribute the MISA-T gain to residency accounting rather than to a correction of that artifact. Please analyze this non-monotonic result or present additional ablations to isolate the cause.","section":"§5.3, Table 5 (Qwen3.6-35B-A3B)"},{"comment":"The controller's moving-window length, the shared neutral prior, and the 'conservative length priors' used for \\bar{k}_b and \\hat{T}_b are not specified anywhere in the manuscript. The claim that the same controller parameters are used across both model and hardware configurations cannot be checked or reproduced without these values. Please provide explicit parameter choices and, if the exact values are not central, a sensitivity study over reasonable ranges.","section":"§4.3, §5.1"},{"comment":"The evaluation contains no test with time-varying workload demand or sequence-length drift. All serving experiments use a fixed checkpoint and the static trace described in Tables 2 and 3, and the end-to-end run reports only aggregate mixture over 50 iterations. The caps in Eqs. (10) and (13) are computed from moving-window estimates of \\bar{k}_b and \\hat{T}_b, and the paper's own Limitations section notes reliance on timely serving-state reports. A stress test with abrupt mix changes or length growth (e.g., RLVR output length increasing during training) is needed to establish that the reported gains survive estimator lag. Please add such an experiment or explicitly scope the claims to quasi-stationary workloads.","section":"§5 (overall evaluation)"}],"minor_comments":[{"comment":"The phrase 'for a feasible offered load' is not defined; it is unclear what happens to the objective and the backlog constraint when the offered load exceeds serving capacity. Please clarify the feasibility condition.","section":"§3.3, Eq. (5)"},{"comment":"In the high-load vLLM Router row, the sample-rate delta (-55.2%) and the RPM delta (-52.5%) differ; please explain why these two throughput metrics diverge, since both count completions over time.","section":"§5.3, Table 5"},{"comment":"The caption mentions red bands marking pressure windows, but the figure does not clearly show these bands; please make the bands explicit and label the y-axis units.","section":"§5.4, Figure 4"},{"comment":"The statement 'improves mean per-replica RPM by 35.6%' lacks an explicit comparison baseline; specify whether this is relative to the no-offload MISA-T run or to the sweep-tuned vLLM Router, and add the corresponding value to the figure or text.","section":"§5.5"},{"comment":"The RLHF and RLVR session counts (599 and 809) are much smaller than the agent count (1,058); the moving-window estimates for these classes could be noisy. A brief explanation of how the shared neutral prior stabilizes early estimates would help.","section":"§2.4, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution with a plausible and well-motivated design. The main barrier to acceptance is the evaluation's lack of variability reporting and the absence of any non-stationary workload test, both of which are needed to support the magnitude of the headline gains. The unexplained MISA prefix-hit collapse on Qwen3.6-35B-A3B also deserves careful analysis before publication. These are addressable with additional experiments and analysis rather than requiring a change to the core approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"To be direct: I think the reader has it about right. The paper's core idea is genuinely new—routing-layer admission control that respects the trainer's mixture contract, with class-specific quotas and residency-time weighting. That's not a restatement of vLLM Router or the fair schedulers, which are placement-only. The formalization in Section 3 is clear, and the ablation structure (Session Admission, MISA, MISA-T) isolates the mechanism nicely; the monotone gains on two model families make the mechanism plausible. The end-to-end matched run, with mixture deviation actually improving and task scores within 0.5 points, is a good sign.\n\nThe soft spots are real, and the stress-test note lands. Equations (10)-(13) build caps from moving-window estimates of per-class KV footprint and residency. The paper's own Limitations sentence concedes the dependence on labeled requests and timely snapshots, but the evaluation never varies the workload regime: Table 2 is a static trace, the 50-iteration run reports only aggregate mixture, and there is no experiment where sequence lengths or residency shift faster than the windows. If that estimator lag matters, the 53.3% and 43.6% gains are exactly what would be fragile. That is a load-bearing gap, not a nit. I also notice Table 5 shows MISA on Qwen with 79.2% prefix hit, below the 92.4% baseline, while still delivering higher throughput; that's worth explaining. Numbers are reported without error bars or confidence intervals; the rollout-only means of three runs are not enough, and code/data are not provided. Controller thresholds are also unspecified, which makes independent reproduction harder.\n\nNone of this refutes the central claim. The mechanism is coherent, the evaluation is internally consistent, and the paper is honest about its limitations. But the numerical gains are not independently checkable as reported. I'd put this as: worth a serious referee, conditional on the authors shipping reproducible artifacts, adding variance reporting, and stress-testing the estimator under time-varying demand—at minimum a synthetic shift in sequence lengths or residency mid-run.\n\nWho gets value: systems people working on LLM serving and RL infrastructure will want to read this—it names a real gap and gives a sensible design. I wouldn't cite it in my own work until the artifacts appear. For peer review: accept with major revision.","headline":"A plausible and well-argued admission-control mechanism for mixed RL rollout serving with believable but under-supported throughput claims; deserves review with major revision.","tokens_in":13198,"tokens_out":1985,"would_cite":false,"duration_ms":18911,"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":"Admission control boosts RL rollout throughput by 53.3% over a cache-aware router","keywords":["mixed RL rollouts","admission control","KV-cache capacity","prefix reuse","workload-aware scheduling","session admission","residency time","LLM serving"],"falsifier":"Replay the Step3.7 mixed workload while delaying the session-cache snapshot reports by, say, ten times their normal interval; if the reported 35.6% end-to-end rollout-throughput gain and the 53.3% rollout-only gain collapse or reverse, the central claim fails. A lighter test is to run the same workload with a rapidly increasing output-length distribution each iteration, so the length reference lags; MISA-T should lose most of its advantage within a few iterations.","tokens_in":12239,"feed_emoji":"⚡","tokens_out":6983,"duration_ms":56575,"temperature":0.7,"pith_summary":"Mixed RL post-training streams—RLVR, RLHF, and agentic rollouts sharing one inference pool—stress serving in different ways, and prefix-cache-aware routing alone does not stop one workload from evicting another's reusable KV state. The paper proposes MISA-T, a routing-layer admission policy that decides when new sessions may enter each inference instance, splits protected KV capacity per workload class, and weights each class's demand by its observed residency time. In rollout-only ablations it reports throughput gains of 53.3% and 43.6% over a sweep-tuned cache-aware router, and in a matched 50-iteration Step3.7 RL run it raises rollout throughput by 35.6%, cuts mean iteration time by 22.8%, and keeps the consumed workload mixture within 2.71 percentage points of the trainer's target. The paper's claim is that rollout-serving efficiency can be improved by admission control alone, without changing the trainer's sampling distribution or the inference engine.","feed_headline":"Admission control boosts RL rollout throughput by 53.3%","feed_subtitle":"A routing-layer policy balances KV cache across RLVR, RLHF, and agentic workloads, cutting iteration time 22.8 percent.","key_machinery":"The load-bearing object is the per-class session cap derived from block-time demand. For each instance $w$ and class $b$, the controller forms $R_{w,b}=N_{w,b}\\bar{k}_b\\hat{T}_b$ and allocates protected KV blocks proportionally, $M^{\\mathrm{time}}_{w,b}=C_w R_{w,b}/\\sum_{b'}R_{w,b'}$, then converts blocks back into a session limit $K^{\\mathrm{time}}_{w,b}=\\max(1,\\lfloor M^{\\mathrm{time}}_{w,b}/\\bar{k}_b\\rfloor)$. This cap is what makes admission workload-aware: a class with long residency (agentic sessions) cannot exhaust the cache that decode-heavy RLVR continuations need. A secondary mechanism is overload pressure, a temporary contraction of the operational cap when prefix-hit rate degrades relative to a recent healthy reference.","core_discovery":"MISA-T's central claim is that admission is a KV-cache commitment: letting a new session start commits cache blocks that will grow, persist across tool gaps, and compete with recoverable prefixes, so the router must cap admitted sessions per workload class rather than merely picking the instance with the best prefix match. MISA-T (Mix-aware Session Admission with a Time factor) estimates each class's block-time demand as $R_{w,b}=N_{w,b}\\bar{k}_b\\hat{T}_b$ (unfinished sessions × footprint × residency), allocates the protected cache budget in proportion, and converts each share into a session cap $K^{\\mathrm{time}}_{w,b}$. An overload signal based on a confirmed drop in prefix-hit rate temporarily contracts the cap when the length reference lags. With these controls, the paper reports that MISA-T raises rollout throughput by 53.3% and 43.6% over the sweep-tuned baseline on two models, lifts prefix-hit rate from 74.5% to 96.2% in the end-to-end run, and keeps the completed workload mixture close to the trainer target.","pith_inferences":["A natural extension is to apply the same block-time cap logic to other shared commitments, such as prefill queue depth or in-flight request budgets, where heterogeneous workloads also occupy resources for unequal durations.","If residency-weighted accounting is the mechanism behind the gains, then workloads with long idle gaps (agent tool calls, human approval) should be the most sensitive to cap settings, giving a sharper ablation test than aggregate throughput.","The results suggest that serving metrics for mixed rollouts should report block-time utilization alongside prefix-hit rate, because hit rate alone cannot distinguish a cache full of useful resident state from one that is churned.","A testable extension is replacing the moving-window estimates of footprint and residency with a predictive model (for example, from prompt length and policy version), which could shrink the lag the overload signal is compensating for."],"forward_implications":["The same hardware can complete more rollout samples per minute: +53.3% and +43.6% in rollout-only runs, +35.6% end-to-end on Step3.7.","A prefix-hit rate near 96% means far less prefill work per sample, lifting prefill TPS by 33.6% to 40.4% alongside decode TPS.","The trainer-specified workload mixture is better preserved: total-variation distance drops from 4.14 to 2.71 percentage points.","Task quality is not sacrificed: pass@4 on SWE-Pro, SWE-Verified, and SWE-MTLG stays within 0.5 points of the baseline.","MISA-T composes with CPU KV offloading, keeping GPU KV utilization above 90% and adding a further 35.6% per-replica request throughput."],"supporting_citations":[{"why":"The cache-aware baseline router that MISA-T's throughput gains are measured against; the sweep-tuned static operating point is the comparison point.","marker":"[vLLM Project, 2026]"},{"why":"PagedAttention provides the block-based KV-cache model that the paper's block-time accounting builds on.","marker":"[Kwon et al., 2023]"},{"why":"RadixAttention is the prefix-reuse mechanism behind the cache hits MISA-T aims to protect.","marker":"[Zheng et al., 2024]"},{"why":"Preble is a representative placement-only routing system whose prefix-reuse/load tradeoff the paper contrasts with admission control.","marker":"[Srivatsa et al., 2025]"},{"why":"DLPM/D2LPM show prior work combining prefix reuse with fairness, motivating the class-aware allocation.","marker":"[Cao et al., 2025]"},{"why":"k-LPM adds latency constraints to prefix routing, another prior baseline the admission policy extends.","marker":"[Dexter et al., 2025]"}],"fun_headline_variants":["Admission control raises RL rollout throughput 53.3%","KV-cache admission caps deliver 53.3% RL rollout boost","Router policy ups mixed RL throughput 53.3% via admission","Admission caps on KV cache lift RL rollout throughput 53.3%","Mixed RL: admission control yields 53.3% throughput gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"MISA-T's admission caps are built from online estimates of each class's unfinished-session count, footprint, and residency; if serving-state snapshots arrive late or sequence lengths shift faster than the moving windows, protected capacity is misallocated until the estimates catch up.","fun_headline_variants_meta":{"raw":{"variants":["Admission control raises RL rollout throughput 53.3%","KV-cache admission caps deliver 53.3% RL rollout boost","Router policy ups mixed RL throughput 53.3% via admission","Admission caps on KV cache lift RL rollout throughput 53.3%","Mixed RL: admission control yields 53.3% throughput gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00117,"raw_usage":{"total_tokens":4881,"prompt_tokens":1025,"completion_tokens":3856,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":3763}},"tokens_in":641,"tokens_out":3856,"duration_ms":25629,"temperature":1.0,"reasoning_tokens":3763,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:09:01.306777+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replay the Step3.7 mixed workload while delaying the session-cache snapshot reports by, say, ten times their normal interval; if the reported 35.6% end-to-end rollout-throughput gain and the 53.3% rollout-only gain collapse or reverse, the central claim fails. A lighter test is to run the same workload with a rapidly increasing output-length distribution each iteration, so the length reference lags; MISA-T should lose most of its advantage within a few iterations.","supporting_citations":[],"review_version":1}