Machine Learning · arXiv:2611.16301 v1 · submitted 12 Nov 2026 · 74 references
LoCoJAX: Host-Local Coordination for Small-Model RL on TPUs
The paper says
“Reinforcement learning (RL) increasingly targets small language models, and executing these workloads efficiently across multiple accelerators requires effective coordination. On TPUs, Pathways-based RL uses a single controller to coordinate computation across hosts.”
more of the abstract
This execution model is effective for large models, but its coordination computational costs are harder to amortize when small models repeatedly invoke short computations. To characterize these computational costs, we sweep dense models from 0.6B to 405B parameters on TPU v7x under a decode-heavy workload and measure how host and network computational overhead vary with model size. At 0.6B, host orchestration occupies 36% of step time and network communication 2%, whereas for large models, estimated network computational costs dominate the non-compute time. Motivated by these findings, we present Local Coordination JAX (LoCoJAX), which implements local rollout and reward processing on each TPU host with synchronous global training. In controlled comparisons against a tuned and optimized Pathways baseline on Qwen3-0.6B across three workloads and 8-64 TPU v7x chips, its colocated mode (LoCoJAX-CO) is 1.3-3.6x faster per step; on GSM8K, it completes a step faster on 8 chips than Pathways does on 64. Finally, a Ray-managed disaggregated mode (LoCoJAX-DA) overlaps rollout and training on separate hosts, and placement experiments establish when this overlap compensates for dividing devices and transferring weights: only when rollout dominates the step and receives most of the hosts. On NuminaMath-CoT at 256 chips, giving rollout 75% of hosts cuts step time by about 15% compared with LoCoJAX-CO.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Honest, well-scoped systems paper: the 1.3–3.6x speedup is real for the specific baseline they built, but the headline number depends on a closed-source runtime comparison that they disclose honestly. This line is wrong
4 major objections · 9 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 74 references say.
The paper says, at §4.2, Table 3; §6.1
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.69):
“We did run others (Table E1), but tuning them as heavily as Qwen3-0.6B was infeasible. 11 References T.”
Pith says: Objection 1 of 4 · major
The central comparison (LoCoJAX vs. Pathways) relies on a baseline that the authors assembled from public components, and they acknowledge they could not tune Pathways the way they tuned LoCoJAX (§6.1). The Pathways runtime is closed source. While the authors have been careful to use equal-work verification and a pre-declared significance margin, the possibility remains that the baseline underperforms for reasons unrelated to the single-controller architecture.
more of objection 1
The paper would be strengthened by a more detailed discussion of what tuning was attempted for Pathways and why further tuning was infeasible, beyond stating that the recipe was frozen before experiments. As it stands, the headline speedup factors (1.3–3.6×) should be interpreted as conditional on the assembled baseline being representative.
The paper says, at §4.3, Figure 7; Appendix C
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
The decomposition of the step-time gap into serial request admission (~1.0 s) and per-iteration decode premium (2.15 ms vs 0.90 ms, adding ~2.25 s) is informative, but the decode premium attribution is not fully decomposed. The paper states that both stacks use the same tpu-inference sampler (Table B1 note), and that the premium arises because Pathways runs every iteration as a lockstep program across all chips while LoCoJAX hosts step their own engines.
more of objection 2
However, Appendix D shows that the Pathways proxy path carries ~15 ms per dispatch, and a decode iteration under Pathways must traverse this proxy. Equation (2) in Appendix C keeps κ and C_TP aggregate, so it is unclear how much of the 2.15 ms is device kernel time vs. host/proxy dispatch wait. If a leaner single-controller integration could reduce this premium, the headline speedup would shrink materially (the paper's own Eq. 5 shows that removing the 2.25 s premium alone drops the 32-chip GSM8K ratio from 3.56× to ~2.6×). The paper should provide a finer-grained breakdown of the decode iteration time on both stacks.
The paper says, at §4.4, Figure 6; Appendix F
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The comparison of colocated vs. disaggregated execution is less conclusive than the main comparison. Several cells are marked as provisional or post-hoc, and some splits are missing due to memory constraints or other issues. The conclusion that disaggregation is beneficial only when rollout dominates the step and receives most of the hosts is reasonable, but the evidence is somewhat thin.
more of objection 3
For example, on NuminaMath at 128 chips, the best LoCoJAX-DA result (0.87×) is provisional, and on GSM8K at 128 chips, the advantage is modest (0.90×). The paper should clarify which comparisons are statistically significant and which are not, and whether the provisional results should be considered preliminary.
The paper says, at §I (Appendix I), Tables I1–I4
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.92):
“(i) It covers one model and one dataset, Qwen3-0.6B on GSM8K, with three seeds per cell and at most 348 updates.”
Pith says: Objection 4 of 4 · major
The cost-to-learn study (Appendix I) is a valuable addition, but it has limitations that the paper acknowledges: it covers only one model and one dataset (Qwen3-0.6B on GSM8K), with three seeds per cell and at most 348 updates. Additionally, the Pathways baseline kept the template resource-manager CPU limit, which slowed its first weight transfer.
more of objection 4
The paper shows that replacing that transfer still gives 1.26–2.54×, but at 4 chips the improvement sits on the 1.25 bar and two sensitivities move a cell below it (Table I4). This appendix is informative but should be presented as preliminary, and the limitations should be more prominently stated in the main text rather than only in the appendix.
9 minor objections
The paper says, at §1, Figure 1
Pith says: minor objection 1 of 9
Figure 1 is informative but the y-axis label 'Share of step time' is cut off in the arXiv version; please ensure the figure is readable.
This objection may describe the copy Pith gave the reviewer, not the paper.
The paper says, at §3.3, §A
Pith says: minor objection 2 of 9
The measurement protocol is thorough but somewhat scattered between §3.3 and Appendix A. Consider consolidating the key elements (equal-work verification, bootstrap confidence intervals, the 1.111 margin) into a single location for clarity.
show 7 more
The paper says, at Table 3
Pith says: minor objection 3 of 9
In Table 3, the NuminaMath 16 chips row shows LoCoJAX optimized (89.4 s) being slower than LoCoJAX base (84.1 s), which is noted in the caption. This is unusual and deserves a brief explanation in the text.
The paper says, at §4.3, Figure 5
Pith says: minor objection 4 of 9
Figure 5 shows the Pathways head interval vs. step time. The caption mentions that LoCoJAX generates in two waves at 8 chips, but this is not discussed in the text. Please clarify whether this affects the comparison.
The paper says, at §6.1
Pith says: minor objection 5 of 9
The limitations section is well-written and honest. However, the statement that 'LoCoJAX is rigorously evaluated on only one model (Qwen3-0.6B)' could be softened by noting that additional results are provided in Appendix E (Tables E1, E2) for other model sizes, even if not as thoroughly tuned.
The paper says, at Appendix C, Eq. (2)
Pith says: minor objection 6 of 9
Equation (2) in Appendix C defines d_iter(t) but the notation is not fully explained. What is Q_KV? What is β_eff? These should be defined either in the equation caption or in the surrounding text.
The paper says, at Appendix D, Table D1
Pith says: minor objection 7 of 9
Table D1 lists upstream contributions. The column 'Active in' is helpful but could be clearer: does 'Both' mean both Pathways and LoCoJAX, or both stacks in the comparison? Please clarify.
The paper says, at General
Pith says: minor objection 8 of 9
The paper uses the term 'computational overhead' in a non-standard way to mean 'time outside compute' (i.e., host orchestration + network + other). This is fine once defined, but the term could be confused with algorithmic overhead. Consider using a different term like 'non-compute time' or 'coordination overhead' for clarity.
The paper says, at References
Pith says: minor objection 9 of 9
Some references are to GitHub repositories or blog posts rather than peer-reviewed publications. This is acceptable for systems work, but the paper should ensure that all claims about upstream contributions (Appendix D) are verifiable from the cited sources.
What would settle it
Run one controlled GSM8K decode iteration on the frozen Pathways baseline (same pinned commits, 64 chips) and split the measured 2.15 ms into device kernel time and host/proxy dispatch wait: take decode device time per iteration from xprof on one host (the §3.2 method, insensitive to head placement) and the iteration wall time from a timer around the tpu-inference per-iteration loop on the head host, over the same ~1,800-iteration window. If the premium over LoCoJAX's 0.90 ms is mostly host/proxy dispatch wait, the single-controller control path causes it and §4.3's attribution holds for the evaluated stack; if it is mostly device-side kernel time, the premium is an engine-integration or configuration artifact and the 1.3-3.6x magnitude overstates the coordination advantage. Then recompute the §4.3 gap and Table 3 ratios with Eq.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 13 Nov 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 9 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This is a solid systems paper that does what it says on the tin: it identifies a real bottleneck (single-controller host orchestration for small-model RL on TPUs), measures it carefully across model sizes, and builds a system that removes it. The scaling study from 0.6B to 405B is genuinely new and useful—the finding that host orchestration dominates at small sizes and network at large is clean and actionable. The measurement methodology is unusually thorough: equal-work verification, pre-declared significance margin, bootstrap CIs, separate head-node pools, and a held-out quality target for the cost-to-learn study. They also ship code and document their upstream contributions. Credit where due: this is careful engineering work with honest reporting.
The soft spot is the one the stress-test note flags: the 1.3–3.6x headline rests on a comparison against a Pathways baseline that the authors assembled and tuned themselves against a closed-source runtime. They disclose this clearly (§6.1), and they run against Google's public recipes as a secondary check (Appendix G), which helps. But the largest term in the gap—the per-iteration decode premium (2.15 vs 0.90 ms)—is measured through the authors' own Pathways-proxy integration, and Appendix D shows that proxy path carries ~15 ms per dispatch. The paper does not decompose that 2.15 ms into device kernel time versus host/proxy dispatch wait. If a leaner integration cut the premium, the headline shrinks materially: removing just that term drops the 32-chip GSM8K ratio from 3.56x to ~2.6x by their own Eq. 5. The direction of the result is robust—the serial admission term (~1.0 s) is verifiable and reproducible—but the magnitude is conditional on the quality of the Pathways integration they built.
This is a paper for systems practitioners who run small-model RL on TPUs. It deserves a serious referee: the claims are well-supported within their stated scope, the methodology is transparent, and the limitations are disclosed. The reviewer should push on the decode premium decomposition and ask for a sensitivity analysis showing how much of the headline depends on that term. But this is not a desk reject.
Recommendation: send to peer review. The paper is honest, the central mechanism is real, and the open questions are about magnitude, not existence.
The strongest objection
The central claim is that host-local coordination makes LoCoJAX 1.3-3.6x faster per step than a tuned Pathways baseline (§4.2, Table 3). Section 4.3 localizes the gap to two mechanisms: serial request admission (~1.0 s per step, in open-source code, reproducible) and a per-iteration decode premium (2.15 ms vs 0.90 ms; ~2.25 s of the ~4.7 s gap on GSM8K at 64 chips). The admission term is verifiable and solid. The decode premium is the largest term, and it is measured with both stacks running the same tpu-inference sampler (Table B1 note: both used LoCoJAX's sampler), driven either host-locally or through the closed Pathways proxy. Appendix D (family G) shows this proxy path carries ~15 ms per dispatch, with 308 dispatches per weight sync fixed from 4.6 s to 0.4 s, so ms-scale per-dispatch overhead is real on this path. A decode iteration under Pathways must traverse the same proxy, yet neither §4.3 nor Appendix C decomposes the 2.15 ms into device kernel time versus host/proxy dispatch wait (Eq. 2 keeps κ and C_TP aggregate). If a leaner single-controller integration or a better proxy cut the premium, the headline shrinks materially: removing the 2.25 s premium alone drops the 32-chip GSM8K optimized ratio from 3.56x to ~2.6x by the paper's own Eq. 5 (Appendix E.1), and in the extreme where only the verifiable ~1.0 s admission term survives, toward ~1.4x. The direction of the result is well supported by the paper's controls - same engine on both arms, LoCoJAX base versus optimized separation, the 1.111 margin, and unmodified external recipes beaten by 5.34x/2.14x (Appendix G) - so this is not a general undertuning objection. The unresolved point is quantitative: the largest mechanism is measured only through the authors' own assembly against a closed runtime, and the §6.1 limitation note does not quantify how much of the headline depends on that path.
The reader's weakest assumption was the fairness and representativeness of the assembled Pathways baseline; I share that family of concern but locate the load-bearing point more precisely. The paper is unusually careful: both stacks run the same tpu-inference sampler (Table B1 note), the base-versus-optimized split isolates O1-O4, the frozen-recipe protocol and 1.111 margin are pre-registered, and Appendix G's unmodified external recipes are beaten by even larger margins, which answers most versions of an undertuned-baseline objection. The residual, unverifiable link is the per-iteration decode premium, the largest single term of the 64-chip GSM8K gap (~2.25 s of ~4.7 s). It is the difference between the same engine driven host-locally and driven through the closed Pathways proxy; Appendix D (family G, ~15 ms per proxy dispatch, weight sync 4.6 s to 0.4 s) shows this proxy path carries ms-scale per-dispatch overhead, so a material part of the 1.25 ms per-iteration premium could be proxy implementation cost rather than an intrinsic property of single-controller coordination. Nothing in §4.3 or Appendix C decomposes the per-iteration 2.15 ms into device time versus dispatch wait, so the attribution cannot be checked from the paper. The stakes are concrete: by the paper's own Δ-model (Eq. 5), removing the premium alone lowers the GSM8K 32-chip optimized ratio from 3.56x to ~2.6x, and the extreme case where only the verifiable ~1.0 s admission term survives gives ~1.4x. The direction of the claim survives either way - LoCoJAX is faster than the documented Pathways recipe on every cell - which is why I would not move the verdict to reject or even further down; but the headline magnitude and the Discussion's architectural generalization ('host-local rollouts over a centralized Pathways controller') rest on an unmeasured quantity inside a closed runtime. This matches the reader's CONDITIONAL verdict, so I recommend no change. I partially agree with the reader because the broad 'baseline undertuned' framing is largely answered by the paper's own controls; the sharp residual is the proxy-path decode premium. The paper's acknowledged limitations (closed-source runtime, one rigorously evaluated model, excluded Gemma launches) are disclosed honestly; my concern is what remains unresolved after those disclosures - the quantitative dependence of the headline on the proxy term, which the paper does not estimate.
The referee's report
If the results hold, this paper makes a significant contribution to systems for RL training of small language models on TPUs. The characterization study (Figure 3, Table 2) provides valuable data on where time goes as model size varies, which is useful for the community. The central engineering claim—that host-local coordination can substantially outperform single-controller execution for small-model RL—is well-supported by careful measurements with equal-work verification, bootstrap confidence intervals, and a pre-declared significance margin. The paper is unusually thorough in its measurement methodology (Appendix A) and in documenting upstream contributions (Appendix D). The finding that LoCoJAX on 8 chips can outperform Pathways on 64 chips for GSM8K is striking. However, the significance is somewhat tempered by the fact that the comparison is against a single baseline (Pathways) that the authors assembled themselves, and the Pathways runtime is closed source, so the baseline may not be optimal. The paper acknowledges this limitation (§6.1). The disaggregation analysis (§4.4) is less definitive but still provides useful guidance.
Pith's reading of the paper (a machine's summary, not the authors')
The paper tries to establish that for small language models (sub-billion to few-billion parameters) undergoing reinforcement learning on TPUs, the dominant bottleneck is not compute or network communication but the orchestration work done by the single-controller runtime (Pathways). The authors first characterise this overhead by profiling GRPO across model sizes on 128 TPU v7x chips, finding that at 0.6B parameters host orchestration occupies 36% of step time while network takes only 2%. They then present LoCoJAX, a system that replaces the single controller with per-host JAX controllers and rollout engines, keeping training globally sharded. In controlled comparisons on Qwen3-0.6B across three workloads and 8–64 chips, LoCoJAX's colocated mode is 1.3–3.6× faster per step than a tuned Pathways baseline. The paper also introduces a disaggregated mode that can improve step time further when rollout dominates, but recommends colocation as the default. If the paper is right, it means that small-model RL on TPUs can be substantially accelerated by architectural changes that eliminate centralised coordination, and that the scheduler design choices inherited from large-model training are actively harmful for small models.
What carries the argument. The central mechanism is host-local coordination: replacing the single global controller (Pathways) with independent JAX controllers and rollout engines on each TPU host, while keeping training globally sharded via FSDP collectives. This changes two things: (1) request admission is parallelised across hosts instead of serialised in one process, and (2) each host's rollout engine runs independently rather than in lockstep with all other chips, reducing per-iteration decode latency. The paper also uses a Ray-managed disaggregation mode (LoCoJAX-DA) that separates rollout and training onto different hosts, introducing one-step policy staleness but allowing overlap.
Core claim. The paper's central claim is that on TPUs, the single-controller execution model (Pathways) imposes a large and scaling-independent overhead on small-model RL steps, and that moving to host-local coordination (LoCoJAX) eliminates this overhead, yielding 1.3–3.6× faster steps on Qwen3-0.6B across three workloads and 8–64 TPU v7x chips. The key evidence is a scaling study showing that at 0.6B parameters, host orchestration accounts for 36% of step time while network communication accounts for only 2%, and that this host overhead does not shrink with more chips. The paper further isolates two specific sources of the Pathways overhead: serial request admission (~1 s per step, flat across chip counts) and higher per-iteration decode latency (2.15 ms vs 0.90 ms per iteration). LoCoJAX distributes request admission across hosts and runs each host's rollout engine independently, avoiding lockstep decode across all chips. The paper also contributes a disaggregated mode (LoCoJAX-DA) that separates rollout and training hosts, and finds that it is beneficial only when rollout dominates the step and receives most of the hosts; otherwise colocation wins.
The premise it rests on. The comparison assumes that the assembled Pathways baseline is a fair and representative implementation of the single-controller execution model; the Pathways runtime is closed source, the recipe was assembled from public components, and the authors could not tune Pathways the way they tuned LoCoJAX, so the measured speedup may overstate the architectural advantage if the baseline underperforms for reasons unrelated to the single-controller design.
If the paper is right (Pith's inferences, not the authors')
Small-model RL on TPUs can be made 1.3–3.6× faster per step by replacing the single-controller runtime with host-local coordination, without changing the learning algorithm.
The host orchestration bottleneck grows as a share of step time as chips are added, meaning that scaling out with Pathways actually worsens the relative overhead for small models.
The disaggregated mode (LoCoJAX-DA) provides a further improvement of about 15% on decode-heavy workloads at 256 chips when rollout receives 75% of hosts, but colocation is the better default for balanced and prefill-heavy workloads.
The results imply that scheduler defaults built for large SPMD programs are a poor fit for small-model RL, and that architectural changes targeting host-side overhead are more impactful than optimising network or compute.
The approach extends to other small models: the paper reports speedups of 5.3× on Qwen3-1.7B and 2.1× on Gemma4-26B MoE against Google's public recipes, though with less rigorous tuning.
The findings suggest that any single-controller architecture (not just Pathways) will exhibit similar overhead for small-model RL on TPUs, because the serial admission and lockstep decode are structural consequences of centralised control, not implementation details.
The characterisation results imply that for small models, TPU RL efficiency can be improved by techniques that reduce host-side work, such as batching request admission, using lighter-weight runtimes, or offloading orchestration to devices where possible.
The disaggregation results provide a concrete decision rule for practitioners: colocate unless rollout dominates the step time and you can give rollout most of the hosts; otherwise disaggregation may hurt performance.
The paper's methodology of profiling step-time breakdowns across model sizes could be applied to other accelerator families (e.g., GPUs with single-controller frameworks) to identify similar bottlenecks for small-model RL.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Reinforcement learning (RL) increasingly targets small language models, and executing these workloads efficiently across multiple accelerators requires effective coordination. On TPUs, Pathways-based RL uses a single controller to coordinate computation across hosts. This execution model is effective for large models, but its coordination computational costs are harder to amortize when small models repeatedly invoke short computations. To characterize these computational costs, we sweep dense models from 0.6B to 405B parameters on TPU v7x under a decode-heavy workload and measure how host and network computational overhead vary with model size. At 0.6B, host orchestration occupies 36% of step time and network communication 2%, whereas for large models, estimated network computational costs dominate the non-compute time. Motivated by these findings, we present Local Coordination JAX (LoCoJAX), which implements local rollout and reward processing on each TPU host with synchronous global training. In controlled comparisons against a tuned and optimized Pathways baseline on Qwen3-0.6B across three workloads and 8-64 TPU v7x chips, its colocated mode (LoCoJAX-CO) is 1.3-3.6x faster per step; on GSM8K, it completes a step faster on 8 chips than Pathways does on 64. Finally, a Ray-managed disaggregated mode (LoCoJAX-DA) overlaps rollout and training on separate hosts, and placement experiments establish when this overlap compensates for dividing devices and transferring weights: only when rollout dominates the step and receives most of the hosts. On NuminaMath-CoT at 256 chips, giving rollout 75% of hosts cuts step time by about 15% compared with LoCoJAX-CO.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-13
The authors' main claim, as the classifier picked it out 0.90
“In controlled comparisons against a tuned and optimized Pathways baseline on Qwen3-0.6B across three workloads and 8-64 TPU v7x chips, its colocated mode (LoCoJAX-CO) is 1.3-3.6x faster per step; on GSM8K, it completes a step faster on 8 chips than Pathways does on 64.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.89 · human participants 0.02 · controlled comparison 0.97 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Reinforcement learning (RL) increasingly targets small language models, and executing these workloads efficiently across multiple accelerators requires effective coordination. | background 1.00 | 0.05 | 0.03 | asserted 0.97 |
| On TPUs, Pathways-based RL uses a single controller to coordinate computation across hosts. | background 1.00 | 0.05 | 0.02 | asserted 1.00 |
| This execution model is effective for large models, but its coordination computational costs are harder to amortize when small models repeatedly invoke short computations. | background 0.67 | 0.18 | 0.05 | asserted 0.94 |
| To characterize these computational costs, we sweep dense models from 0.6B to 405B parameters on TPU v7x under a decode-heavy workload and measure how host and network computational overhead vary with model size. | method 0.85 | 0.96 | 0.03 | asserted 0.95 |
| At 0.6B, host orchestration occupies 36% of step time and network communication 2%, whereas for large models, estimated network computational costs dominate the non-compute time. | result 1.00 | 0.94 | 0.07 | asserted 1.00 |
| Motivated by these findings, we present Local Coordination JAX (LoCoJAX), which implements local rollout and reward processing on each TPU host with synchronous global training. | contribution 0.62 | 0.83 | 0.17 | asserted 0.99 |
| In controlled comparisons against a tuned and optimized Pathways baseline on Qwen3-0.6B across three workloads and 8-64 TPU v7x chips, its colocated mode (LoCoJAX-CO) is 1.3-3.6x faster per step; on GSM8K, it completes a step faster on 8 chips than Pathways does on 64. | result 1.00 | 0.96 | 0.42 | asserted 1.00 |
| Finally, a Ray-managed disaggregated mode (LoCoJAX-DA) overlaps rollout and training on separate hosts, and placement experiments establish when this overlap compensates for dividing devices and transferring weights: only when rollout dominates the step and receives most of the hosts. | method 0.45 | 0.94 | 0.13 | asserted 0.98 |
| On NuminaMath-CoT at 256 chips, giving rollout 75% of hosts cuts step time by about 15% compared with LoCoJAX-CO. | result 1.00 | 0.96 | 0.11 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 7,253 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (8)
Showing 6 of the paper's 8 figures. Figures are the authors', reproduced from the paper.
References (74)
- [1] Arora, D. and Zanette, A. , title =. NeurIPS , year =. 2502.04463 , archivePrefix = on PitharXiv
- [2] Fu, W. and Gao, J. and Shen, X. and Zhu, C. and Mei, Z. and He, C. and others , title =. NeurIPS , year =. 2505.24298 , archivePrefix = on PithPith's reviewarXiv
- [3] Tu, S. and Lin, J. and Zhang, Q. and Tian, X. and Li, L. and Lan, X. and others , title =. NeurIPS , year =. 2505.10832 , archivePrefix = on PitharXiv
- [4] Shao, Z. and Srivatsa, V. and Srivastava, S. and Wu, Q. and Ariyak, A. and others , title =. MLSys , year =. 2511.13841 , archivePrefix = on PitharXiv
- [5] Xu, Y. and Tian, K. and Tian, Y. and Yang, Z. and Yu, Y. and Li, Y. and others , title =. ICML , note =. 2026 , eprint = on Pith
- [6] Dang, Q.-A. and Ngo, C. , title =. 2025 , eprint = on Pith
- [7] Zhong, Y. and Liu, S. and Chen, J. and Hu, J. and Zhu, Y. and Liu, X. and others , title =. OSDI , year =. 2401.09670 , archivePrefix = on PithPith's reviewarXiv
- [8] Li, G. and Chen, Y. and Lin, M. and Yang, T. , title =. ICLR , year =. 2510.04474 , archivePrefix = on PitharXiv
- [9] Kim, M. and Lee, M. and Oh, S. and Galim, K. and Kim, D. and Hooper, C. and others , title =. 2026 , eprint = on Pith
- [10] Hu, Q. and Yang, S. and Guo, J. and Yao, X. and Lin, Y. and Gu, Y. and others , title =. ASPLOS , year =. 2511.16665 , archivePrefix = on PitharXiv
- [11] Xu, Y. and Lee, H. and Chen, D. and Hechtman, B. and Huang, Y. and Joshi, R. and others , title =. 2021 , eprint = on Pith
- [12] He, Y. and Zhang, S. and Gai, J. and Zhang, X. and Han, B. and Wang, B. and Rangwala, H. and Karypis, G. , title =. MLSys , year =. 2512.12476 , archivePrefix = on PithPith's reviewarXiv
- [13] Sheng, G. and Zhang, C. and Ye, Z. and Wu, X. and Zhang, W. and Zhang, R. and others , title =. EuroSys , year =. 2409.19256 , archivePrefix = on PithPith's reviewarXiv
- [14] He, B. and Qu, Z. and Liu, Z. and Chen, Y. and Zuo, Y. and Qian, C. and others , title =. 2025 , eprint = on Pith
- [15] Bauer, J. and Walshe, T. and Pham, D. and Vishwakarma, H. and Parchami, A. and others , title =. MLSys , year =. 2604.18381 , archivePrefix = on PithPith's reviewarXiv
- [16] Li, Y. and Yue, X. and Xu, Z. and Jiang, F. and Niu, L. and Lin, B. Y. and Ramasubramanian, B. and Poovendran, R. , title =. Findings of ACL , year =. 2502.12143 , archivePrefix = on PitharXiv
- [17] Li, X. and Zou, H. and Liu, P. , title =. 2025 , eprint = on Pith
- [18] Wang, Y. and Yang, Q. and Zeng, Z. and Ren, L. and Liu, L. and Peng, B. and others , title =. NeurIPS , note =. 2025 , eprint = on Pith
- [19] Barham, P. and Chowdhery, A. and Dean, J. and Ghemawat, S. and Hand, S. and Hurt, D. and others , title =. MLSys , year =. 2203.12533 , archivePrefix = on PithPith's reviewarXiv
- [20] Khan, Z. and Tafveez, O. and Bhatti, Z. H. , title =. EACL Student Research Workshop , year =. 2601.06677 , archivePrefix = on PitharXiv
show 54 more
- [21] Hessel, M. and Kroiss, M. and Clark, A. and Kemaev, I. and Quan, J. and Keck, T. and others , title =. 2021 , eprint = on Pith
- [22] Huang, L. and Anupam, S. and Lee, I. and Li, S. and Bastani, O. , title =. 2025 , eprint = on Pith
- [23] Moritz, P. and Nishihara, R. and Wang, S. and Tumanov, A. and Liaw, R. and Liang, E. and others , title =. OSDI , year =. 1712.05889 , archivePrefix = on PithPith's reviewarXiv
- [24] Mei, Z. and Fu, W. and Li, K. and Wang, G. and Zhang, H. and Wu, Y. , title =. MLSys , year =. 2406.14088 , archivePrefix = on PithPith's reviewarXiv
- [25] Chen, Q. and Liu, Z. and Sun, P. and Li, S. and Wang, G. and others , title =. MLSys , year =. 2510.26475 , archivePrefix = on PitharXiv
- [26] Zhou, J. and Hu, Q. and Jin, Y. and Wang, Z. and Sun, P. and Gu, Y. and Zhang, W. and Zhai, M. and Zhang, X. and Zhang, W. , title =. 2025 , eprint = on Pith
- [27] Zhou, R. and Zhang, L. and Wu, S.-C. and Zou, K. and Zhou, H. and Ye, K. and others , title =. 2025 , eprint = on Pith
- [28] Zhong, Y. and Zhang, Z. and Wu, B. and Liu, S. and Chen, Y. and Wan, C. and others , title =. NSDI , year =. 2409.13221 , archivePrefix = on PithPith's reviewarXiv
- [29] Yu, C. and Wang, Y. and Guo, Z. and Lin, H. and Xu, S. and Zang, H. and others , title =. OSDI , year =. 2509.15965 , archivePrefix = on PitharXiv
- [30] Jiang, J. and Chen, Y. and Hechtman, B. A. and Zhang, F. and Mu, Y. , title =. 2026 , eprint = on Pith
- [31] Agrawal, A. and Kedia, N. and Panwar, A. and Mohan, J. and Kwatra, N. and Gulavani, B. S. and others , title =. OSDI , year =. 2403.02310 , archivePrefix = on PithPith's reviewarXiv
- [32] Espeholt, L. and Marinier, R. and Stanczyk, P. and Wang, K. and Michalski, M. , title =. ICLR , year =. 1910.06591 , archivePrefix = on PithPith's reviewarXiv
- [33] Zeng, W. and Huang, Y. and Liu, Q. and Liu, W. and He, K. and Ma, Z. and others , title =. COLM , year =. 2503.18892 , archivePrefix = on PithPith's reviewarXiv
- [34] Zhong, Q. and Zheng, M. and Song, M. and Lin, X. and Sun, J. and Jiang, H. and others , title =. 2026 , eprint = on Pith
- [35] Zhong, Y. and Zhang, Z. and Song, X. and Hu, H. and Jin, C. and Wu, B. and others , title =. 2025 , eprint = on Pith
- [36] Fang, G. and Ma, X. and Wang, X. , title =. NeurIPS , year =. 2505.13379 , archivePrefix = on PithPith's reviewarXiv
- [37] Wang, S. and Asilis, J. and Akg. ICLR , year =. 2504.15777 , archivePrefix = on PithPith's reviewarXiv
- [38] Zhang, X. and Huang, Z. and Ni, C. and Xiong, Z. and Chen, J. and Oymak, S. , title =. 2025 , eprint = on Pith
- [39] Bao, T. and Carpenter, J. and Chai, L. and Gao, H. and Jiang, Y. and Noghabi, S. and others , title =. 2025 , url = on Pith
- [40] Kwon, W. and Li, Z. and Zhuang, S. and Sheng, Y. and Zheng, L. and Yu, C. H. and others , title =. SOSP , year =. 2309.06180 , archivePrefix = on PithPith's reviewarXiv
- [41] Wan, X. and Qi, P. and Huang, G. and Ruan, C. and Lin, M. and Li, J. , title =. ICLR , year =. 2601.19362 , archivePrefix = on PitharXiv
- [42] Zhuang, X. and Ma, P. and Jia, Z. and Cao, Z. and Liu, S. , title =. CPAL , year =. 2506.13404 , archivePrefix = on PitharXiv
- [43] Shao, Z. and Wang, P. and Zhu, Q. and Xu, R. and Song, J. and Bi, X. and others , title =. 2024 , eprint = on Pith
- [44] Hu, W. and Wadsworth, D. and Siddens, S. and Winata, S. and Fu, D. Y. and Swann, R. and others , title =. MLSys , year =. 2511.08083 , archivePrefix = on PitharXiv
- [45] Christiano, P. F. and Leike, J. and Brown, T. and Martic, M. and Legg, S. and Amodei, D. , title =. NeurIPS , year =. 1706.03741 , archivePrefix = on PithPith's reviewarXiv
- [46] Ouyang, L. and Wu, J. and Jiang, X. and Almeida, D. and Wainwright, C. L. and Mishkin, P. and others , title =. NeurIPS , year =. 2203.02155 , archivePrefix = on PithPith's reviewarXiv
- [47] Lambert, N. and Morrison, J. and Pyatkin, V. and Huang, S. and Ivison, H. and Brahman, F. and others , title =. 2024 , eprint = on Pith
- [48] Scaling Reinforcement Learning for Content Moderation with Large Language Models , author=. arXiv preprint arXiv:2512.20061 , year= on PitharXiv
- [49] Wang, L. , title =. 2025 , url = on Pith
- [50] Gao, Haoyu and Wang, Lance and Noghabi, Shadi and Bao, Tianshu and Yu, Weiren , title =. 2026 , url = on Pith
- [51] Zhong, T. and Ling, N. and Pi, Y. and Wei, Z. and others , title =. 2026 , eprint = on Pith
- [52] Belcak, P. and Heinrich, G. and Diao, S. and Fu, Y. and Dong, X. and Muralidharan, S. and Lin, Y. C. and Molchanov, P. , title =. 2025 , eprint = on Pith
- [53] Liu, Z. and Zhao, C. and Iandola, F. and Lai, C. and Tian, Y. and Fedorov, I. and Xiong, Y. and Chang, E. and Shi, Y. and Krishnamoorthi, R. and Lai, L. and Chandra, V. , title =. ICML , year =. 2402.14905 , archivePrefix = on PithPith's reviewarXiv
- [54] Yang, An and others , title =. 2025 , howpublished = on Pith
- [55] Grattafiori, Aaron and others , title =. 2024 , howpublished = on Pith
- [56] 2026 , howpublished = on Pith
- [57] Cobbe, Karl and Kosaraju, Vineet and Bavarian, Mohammad and Chen, Mark and Jun, Heewoo and Kaiser, Lukasz and Plappert, Matthias and Tworek, Jerry and Hilton, Jacob and Nakano, Reiichiro and Hesse, Christopher and Schulman, John , title =. 2021 , howpublished = on Pith
- [58] Li, Jia and Beeching, Edward and Tunstall, Lewis and Lipkin, Ben and Soletskyi, Roman and Huang, Shengyi Costa and Rasul, Kashif and Yu, Longhui and Jiang, Albert and Shen, Ziju and Qin, Zihan and Dong, Bin and Zhou, Li and Fleureau, Yann and Lample, Guillaume and Polu, Stanis... on Pith
- [59] Toshniwal, Shubham and Du, Wei and Moshkov, Ivan and Kisacanin, Branislav and Ayrapetyan, Alexan and Gitman, Igor , title =. 2024 , howpublished = on Pith
- [60] Yang, John and Jimenez, Carlos E. and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , title =. Advances in Neural Information Processing Systems (NeurIPS) , year = on Pith
- [61] 2024 , howpublished = on Pith
- [62] Bradbury, James and Frostig, Roy and Hawkins, Peter and Johnson, Matthew James and Leary, Chris and Maclaurin, Dougal and Necula, George and Paszke, Adam and VanderPlas, Jake and Wanderman-Milne, Skye and Zhang, Qiao , title =. 2018 , howpublished = on Pith
- [63] Fernandez, Jared and Kahn, Jacob and Na, Clara and Bisk, Yonatan and Strubell, Emma , title =. EMNLP , year = on Pith
- [64] Spector, Benjamin and Juravsky, Jordan and Sul, Stuart and Dugan, Owen and Lim, Dylan and Fu, Dan and Arora, Simran and R. 2025 , url = on Pith
- [65] Nrusimha, Aniruddha and Brandon, William and Mishra, Mayank and Shen, Yikang and Panda, Rameswar and Ragan-Kelley, Jonathan and Kim, Yoon , title = on Pith
- [66] Jouppi, Norman P. and Kurian, George and Li, Sheng and Ma, Peter and Nagarajan, Rahul and Nai, Lifeng and Patil, Nishant and Subramanian, Suvinay and Swing, Andy and Towles, Brian and Young, Cliff and Zhou, Xiang and Zhou, Zongwei and Patterson, David , title =. ISCA , year = on Pith
- [67] Lee, Mark and others , title = on Pith
- [68] Kumar, Sameer and Bradbury, James and Young, Cliff and Wang, Yu Emma and Levskaya, Anselm and Hechtman, Blake and Chen, Dehao and Lee, HyoukJoong and Deveci, Mehmet and Kumar, Naveen and Kanwar, Pankaj and Wang, Shibo and Wanderman-Milne, Skye and Lacy, Steve and Wang, Tao and... on Pith
- [69] Vellaisamy, Prabhu and Labonte, Thomas and Chakraborty, Sourav and Turner, Matt and Sury, Samantika and Shen, John Paul , title =. ISPASS , pages = on Pith
- [70] Shen, Gerald and Wang, Zhilin and Delalleau, Olivier and Zeng, Jiaqi and Dong, Yi and Egert, Daniel and others , title =. COLM , year = on Pith
- [71] Yao, Zhewei and Aminabadi, Reza Yazdani and Ruwase, Olatunji and Rajbhandari, Samyam and Wu, Xiaoxia and Awan, Ammar Ahmad and others , title = on Pith
- [72] Basile, C. and Voznesensky, M. and Park, G. and Wilson, B. and Holloway, P. and Wan, Z. and others , title = on Pith
- [73] Hendrycks, D. and Burns, C. and Kadavath, S. and Arora, A. and Basart, S. and Tang, E. and Song, D. and Steinhardt, J. , title =. NeurIPS Datasets and Benchmarks Track , year =. 2103.03874 , archivePrefix = on PithPith's reviewarXiv
- [74] Lightman, H. and Kosaraju, V. and Burda, Y. and Edwards, H. and Baker, B. and Lee, T. and Leike, J. and Schulman, J. and Sutskever, I. and Cobbe, K. , title =. ICLR , year =. 2305.20050 , archivePrefix = on PithPith's reviewarXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment