{ "release": "electronics-submission-2026-10-23-r2", "documents": { "RLVRAMBench_MDPI": { "sha256": "e561df8e88ae52c6d091afecb4146df10b17825d27866f5db10698c185b4ba00", "pages": 43, "figures": 7, "tables": 21, "captions": [ { "kind": "Table", "number": "1", "page": 4, "x": 166.098, "y": 363.774058, "first_line": "Table 1. Different questions addressed by related memory analyses. The last column describes the" }, { "kind": "Table", "number": "2", "page": 6, "x": 166.098, "y": 74.421058, "first_line": "Table 2. Key terms used in this article." }, { "kind": "Table", "number": "3", "page": 10, "x": 166.098, "y": 74.421058, "first_line": "Table 3. How the experiments support the research questions." }, { "kind": "Table", "number": "4", "page": 11, "x": 166.098, "y": 74.421058, "first_line": "Table 4. Workload settings for the boundary and batch/reservation experiments." }, { "kind": "Table", "number": "5", "page": 11, "x": 166.098, "y": 335.939058, "first_line": "Table 5. Configuration levels used in the boundary experiment. Each level bundles three settings." }, { "kind": "Table", "number": "6", "page": 12, "x": 166.098, "y": 74.421058, "first_line": "Table 6. Recorded attempts for the original 588 planned configuration\u2013seed slots. Completed runs" }, { "kind": "Table", "number": "7", "page": 17, "x": 166.098, "y": 113.076058, "first_line": "Table 7. Additional RL panels. Counts are planned slots. The external LLMem validation is a separate" }, { "kind": "Table", "number": "8", "page": 20, "x": 166.098, "y": 184.528058, "first_line": "Table 8. Outcomes of the independent-control experiment on four GPUs. Each cell contains twelve" }, { "kind": "Figure", "number": "1", "page": 21, "x": 166.394, "y": 346.024058, "first_line": "Figure 1. Reservation sensitivity with actor micro-batch 8 and generation-state release enabled. Cells" }, { "kind": "Table", "number": "9", "page": 22, "x": 166.098, "y": 74.421058, "first_line": "Table 9. All base configurations in the release experiment. Prompt/response limits are tokens; their" }, { "kind": "Figure", "number": "2", "page": 23, "x": 166.394, "y": 384.413058, "first_line": "Figure 2. Micro-batch effects in the same-device replication at each reservation. Each point is a seed\u2019s" }, { "kind": "Table", "number": "10", "page": 23, "x": 166.098, "y": 441.678058, "first_line": "Table 10. Same-device micro-batch/reservation replication with fixed full instrumentation. Entries are" }, { "kind": "Figure", "number": "3", "page": 24, "x": 166.394, "y": 350.004058, "first_line": "Figure 3. Targeted batch-by-logging control. Each connected pair compares the actor-stage batch-" }, { "kind": "Figure", "number": "4", "page": 26, "x": 166.394, "y": 346.972058, "first_line": "Figure 4. Full minus LoRA sampled memory peaks in the native-KL panel. Small points are seed" }, { "kind": "Table", "number": "11", "page": 26, "x": 166.098, "y": 490.050058, "first_line": "Table 11. Alternative minus anchor in the twenty-update iteration-cost panel, three same-device" }, { "kind": "Table", "number": "12", "page": 27, "x": 166.098, "y": 430.803058, "first_line": "Table 12. Frozen approval decisions for the 48 configurations of the independent-control experiment." }, { "kind": "Table", "number": "13", "page": 28, "x": 166.098, "y": 168.787058, "first_line": "Table 13. Frozen decisions on the complete three-repeat labels of the reservation-sensitivity panel." }, { "kind": "Figure", "number": "5", "page": 29, "x": 166.394, "y": 372.136058, "first_line": "Figure 5. Three-state labels across the boundary experiment. Each cell combines three fresh processes." }, { "kind": "Table", "number": "14", "page": 29, "x": 166.098, "y": 416.579058, "first_line": "Table 14. Copying labels from a source model to a target model at the primary margin. Each direction" }, { "kind": "Table", "number": "15", "page": 30, "x": 166.098, "y": 157.227058, "first_line": "Table 15. Outcomes at the same levels on two and four GPUs. \u201cWithin\u201d and \u201cabove\u201d are completed" }, { "kind": "Table", "number": "16", "page": 30, "x": 166.098, "y": 602.011058, "first_line": "Table 16. Admission results for four prospective cases. Entries under the rules give approved /" }, { "kind": "Table", "number": "17", "page": 32, "x": 166.098, "y": 168.253058, "first_line": "Table 17. LLMem native-training comparison, three pairs per cell. Peaks are in MiB; relative error is" }, { "kind": "Table", "number": "18", "page": 33, "x": 166.098, "y": 74.421058, "first_line": "Table 18. One-step and 100-step outcomes, with memory measurements from completed long runs." }, { "kind": "Table", "number": "19", "page": 33, "x": 166.098, "y": 559.808058, "first_line": "Table 19. Repeated labels of the 90 replayed A100 configurations on MI250X. Each label combines" }, { "kind": "Table", "number": "20", "page": 35, "x": 166.098, "y": 74.421058, "first_line": "Table 20. Additional MI250X panels. Labels combine three fresh processes; \u201cwake-up\u201d denotes" }, { "kind": "Figure", "number": "6", "page": 36, "x": 166.394, "y": 330.827058, "first_line": "Figure 6. Diagnosed memory failures by experiment. In the plot labels, \u201cState release\u201d denotes the" }, { "kind": "Figure", "number": "7", "page": 37, "x": 166.394, "y": 668.457058, "first_line": "Figure 7. Agreement of individual within-margin approval decisions across run lengths, by protocol," }, { "kind": "Table", "number": "21", "page": 38, "x": 166.098, "y": 476.639058, "first_line": "Table 21. Limits on extrapolating the measured memory boundaries." } ], "substantial_word_overlap_candidates": [ { "page": 32, "words": [ "M", "\u02c6\ufe01" ], "fraction": 0.818 } ], "caption_left_edge_range_pt": [ 166.098, 166.394 ], "unresolved_overlap_candidates": [], "overlap_review": "The hat accent over M in the estimation-error equation is intentional mathematical typesetting." }, "RLVRAMBench_supplement": { "sha256": "35fcae54892f4ff27a272fd30dae269a932ddf3453f0ae230ef5ad1101f07e03", "pages": 38, "figures": 3, "tables": 46, "captions": [ { "kind": "Table", "number": "S1", "page": 2, "x": 62.362, "y": 63.75752, "first_line": "Table S1: Completion and margin compliance in the original boundary experiment. The last column gives" }, { "kind": "Table", "number": "S2", "page": 3, "x": 62.362, "y": 63.75752, "first_line": "Table S2: Stage contrasts in external memory (MiB) in the instrumentation control. Each entry is full" }, { "kind": "Table", "number": "S3", "page": 3, "x": 62.362, "y": 564.17452, "first_line": "Table S3: Factorial effects per workload (MiB), from nine blocks per workload. All runs use full instrumentation." }, { "kind": "Figure", "number": "S1", "page": 4, "x": 62.362, "y": 474.95152, "first_line": "Figure S1: Supporting summaries of the memory mechanisms. The panel for the release experiment distinguishes" }, { "kind": "Table", "number": "S4", "page": 5, "x": 62.362, "y": 63.75752, "first_line": "Table S4: Realized sequence lengths in the 100-step duration tests, in tokens. \u201cCap batches\u201d counts batches with" }, { "kind": "Table", "number": "S5", "page": 5, "x": 62.362, "y": 516.66352, "first_line": "Table S5: Effects in the earlier 100-step duration test, by case (MiB). Brackets give observed ranges across runs," }, { "kind": "Figure", "number": "S2", "page": 6, "x": 62.362, "y": 299.19852, "first_line": "Figure S2: First-step and maximum peaks for the twelve completed runs of the earlier 100-step duration test." }, { "kind": "Table", "number": "S6", "page": 7, "x": 62.362, "y": 63.75752, "first_line": "Table S6: All usable control blocks. Entries are actor-stage effects on the external peak, in MiB. An asterisk marks" }, { "kind": "Table", "number": "S7", "page": 7, "x": 62.362, "y": 261.64752, "first_line": "Table S7: Realized response lengths for the blocks in Table S6. Entries are mean / maximum tokens, rounded to" }, { "kind": "Table", "number": "S8", "page": 8, "x": 62.362, "y": 63.75752, "first_line": "Table S8: Descriptive aggregate effects with 95% resampling intervals. Memory effects are in MiB unless stated" }, { "kind": "Table", "number": "S9", "page": 8, "x": 62.362, "y": 493.53052, "first_line": "Table S9: Calibration configurations. Reservation is the generation reservation, the fraction of GPU memory" }, { "kind": "Table", "number": "S10", "page": 9, "x": 62.362, "y": 263.58652, "first_line": "Table S10: How missing admission evidence is treated. A recorded invocation is charged even when it does not" }, { "kind": "Table", "number": "S11", "page": 10, "x": 62.362, "y": 63.75752, "first_line": "Table S11: All prospective candidates. Peaks are in MiB and refer to completed processes; the evaluation peak is" }, { "kind": "Table", "number": "S12", "page": 11, "x": 62.362, "y": 63.75752, "first_line": "Table S12: Fixed training controls. BF16 is 16-bit bfloat16, and FP32 is 32-bit floating point. Fully Sharded Data" }, { "kind": "Table", "number": "S13", "page": 11, "x": 62.362, "y": 489.89352, "first_line": "Table S13: Realized work in the boundary experiment. Qwen, Phi, and Granite denote Qwen2.5-3B, Phi-4-mini," }, { "kind": "Table", "number": "S14", "page": 12, "x": 62.362, "y": 312.67052, "first_line": "Table S14: Diagnostic rules on the 36 GPU-count transfer queries. \u201cAlways reject\u201d predicts memory failure for" }, { "kind": "Table", "number": "S15", "page": 13, "x": 62.362, "y": 63.75752, "first_line": "Table S15: Duration settings. Caps are in tokens; every final-step operation listed here is required for completion." }, { "kind": "Table", "number": "S16", "page": 13, "x": 62.362, "y": 306.13952, "first_line": "Table S16: Retrospective sensitivity to the margin. \u201cCapacity allowed\u201d is the share of device capacity used as the" }, { "kind": "Table", "number": "S17", "page": 13, "x": 62.362, "y": 508.36352, "first_line": "Table S17: Instrumentation control. Count cells give external-only; full instrumentation, each out of three. Peak" }, { "kind": "Table", "number": "S18", "page": 16, "x": 62.362, "y": 63.75752, "first_line": "Table S18: Fitting costs and regression peak errors with one model family held out. Each fitting pool has sixty" }, { "kind": "Table", "number": "S19", "page": 16, "x": 62.362, "y": 439.43752, "first_line": "Table S19: Predictions with one model family held out. Each row evaluates thirty configurations, and the other" }, { "kind": "Table", "number": "S20", "page": 17, "x": 62.362, "y": 344.64852, "first_line": "Table S20: Original larger-model panel (7B). Each entry gives the number of eligible seeds out of three and the" }, { "kind": "Table", "number": "S21", "page": 20, "x": 62.362, "y": 63.75752, "first_line": "Table S21: Configuration peaks, in MiB, at four GPUs for levels c1 and c2 in the larger-model replication. Every" }, { "kind": "Table", "number": "S22", "page": 20, "x": 62.362, "y": 400.12152, "first_line": "Table S22: Predictions by model and actor-offload group (c1 offloads actor parameters; c2/c3 keep them on the" }, { "kind": "Table", "number": "S23", "page": 21, "x": 62.362, "y": 111.52252, "first_line": "Table S23: Frozen predictions in the larger-model replication. Qwen has eighteen resolved configurations (four" }, { "kind": "Table", "number": "S24", "page": 22, "x": 62.362, "y": 63.75752, "first_line": "Table S24: Prediction results with independently varied controls, shown separately by model. Each model" }, { "kind": "Table", "number": "S25", "page": 23, "x": 62.362, "y": 417.33452, "first_line": "Table S25: Observed contrast ranges within blocks, in MiB. Brackets give the minimum and maximum over n" }, { "kind": "Table", "number": "S26", "page": 25, "x": 62.362, "y": 243.07752, "first_line": "Table S26: Effect of actor offload (enabled minus disabled) at generation reservation 0.45. Each entry is the" }, { "kind": "Table", "number": "S27", "page": 25, "x": 62.362, "y": 388.91952, "first_line": "Table S27: Whole-run offload contrasts by case at generation reservation 0.45. Each entry is the three-pair mean" }, { "kind": "Table", "number": "S28", "page": 28, "x": 62.362, "y": 63.75752, "first_line": "Table S28: Verified actor parameter inventories before FSDP wrapping. Stored parameter tensors are FP32" }, { "kind": "Table", "number": "S29", "page": 28, "x": 62.362, "y": 199.88252, "first_line": "Table S29: Full minus LoRA peak differences, native KL. Values are mean [minimum, maximum] across jointly" }, { "kind": "Table", "number": "S30", "page": 28, "x": 62.362, "y": 451.05052, "first_line": "Table S30: Actor offload enabled minus disabled, native KL. Each comparison keeps adaptation mode and" }, { "kind": "Table", "number": "S31", "page": 29, "x": 62.362, "y": 63.00752, "first_line": "Table S31: Full minus LoRA peak differences, no-KL control. Values are mean [minimum, maximum] across jointly" }, { "kind": "Table", "number": "S32", "page": 29, "x": 62.362, "y": 172.75652, "first_line": "Table S32: Actor offload enabled minus disabled, no-KL control. Each comparison keeps adaptation mode and" }, { "kind": "Table", "number": "S33", "page": 29, "x": 62.362, "y": 403.44852, "first_line": "Table S33: Reservation sensitivity. Each cell reports within margin/above margin/memory failure/unresolved" }, { "kind": "Table", "number": "S34", "page": 30, "x": 62.362, "y": 63.00752, "first_line": "Table S34: Frozen prediction decisions by reservation scope. Scored/planned counts retain unresolved targets;" }, { "kind": "Table", "number": "S35", "page": 30, "x": 62.362, "y": 457.26352, "first_line": "Table S35: State-release sensitivity at the two GSM8K anchors. Outcome counts are within margin/above" }, { "kind": "Table", "number": "S36", "page": 31, "x": 62.362, "y": 63.00752, "first_line": "Table S36: Iteration costs for Qwen-7B/GSM8K and Mistral-7B/longer-prompt code. Counts are within" }, { "kind": "Table", "number": "S37", "page": 31, "x": 62.362, "y": 249.02052, "first_line": "Table S37: Observed work and host memory for the same iteration-cost runs. Token rates use prompt plus" }, { "kind": "Table", "number": "S38", "page": 31, "x": 62.362, "y": 435.03352, "first_line": "Table S38: Same-device timing contrasts against the offload-on, reservation-0.45, release-on anchor. Time" }, { "kind": "Figure", "number": "S3", "page": 32, "x": 62.362, "y": 330.64552, "first_line": "Figure S3: Mean iteration time of each run over updates 3\u201320. Points are seeds, diamonds are condition means," }, { "kind": "Table", "number": "S39", "page": 32, "x": 62.362, "y": 501.46352, "first_line": "Table S39: GPU-hours used by the follow-up experiments. The allocation and task columns overlap and must not" }, { "kind": "Table", "number": "S40", "page": 33, "x": 62.362, "y": 97.57952, "first_line": "Table S40: MI250X slots by panel. Executions count every started execution, including amendment reruns and one Slurm restart after a node failure; the latest attempt gives" }, { "kind": "Table", "number": "S41", "page": 33, "x": 62.362, "y": 236.73652, "first_line": "Table S41: Every replayed A100 configuration. Peaks are maxima over completed repetitions (MiB); failure stages count failing processes (A100 / MI250X). Off = actor offload" }, { "kind": "Table", "number": "S41", "page": 34, "x": 62.362, "y": 62.67552, "first_line": "Table S41: Continued." }, { "kind": "Table", "number": "S41", "page": 35, "x": 62.362, "y": 62.67552, "first_line": "Table S41: Continued." }, { "kind": "Table", "number": "S42", "page": 36, "x": 62.362, "y": 63.00752, "first_line": "Table S42: 7B configurations: A100 and MI250X repeated labels." }, { "kind": "Table", "number": "S43", "page": 36, "x": 62.362, "y": 390.28952, "first_line": "Table S43: Reservation sweep at c3 settings (GSM8K, two GCDs, LoRA-GRPO, one step). Times are per" }, { "kind": "Table", "number": "S44", "page": 37, "x": 62.362, "y": 63.00752, "first_line": "Table S44: GRPO-LoRA, PPO with a full-parameter critic, and full-parameter GRPO (GSM8K, two GCDs, one" }, { "kind": "Table", "number": "S45", "page": 37, "x": 62.362, "y": 285.56252, "first_line": "Table S45: One-step screens and 40-step runs of the A100 temporal configurations (LoRA-GRPO). Times are per" }, { "kind": "Table", "number": "S46", "page": 38, "x": 62.362, "y": 377.49752, "first_line": "Table S46: Estimator decisions on all benchmark targets. Recall is within-margin recall; approval means a" } ], "substantial_word_overlap_candidates": [], "unresolved_overlap_candidates": [] }, "RLVRAMBench_reviewer_response": { "sha256": "124c501e804fef58220812cabdf3bb61a0073044551c023b35c637b2f23953af", "pages": 6, "figures": 0, "tables": 0, "captions": [], "substantial_word_overlap_candidates": [], "unresolved_overlap_candidates": [] } }, "new_gpu_experiments": false, "status": "passed", "reviewed_at_utc": "2026-10-23T12:59:11.490431+00:00", "visual_review": { "all_pages_inspected": true, "main_figures": 7, "main_tables": 21, "supplement_figures": 3, "supplement_tables": 46, "main_figure_width": "manuscript text column", "main_captions_left_aligned": true, "table_headers_and_numeric_cells": "aligned; no clipped or overlapping cells observed", "continued_inventory": "MI250X replay uses three landscape pages with repeated headers", "short_tables_kept_together": [ "S34", "S44" ], "all_main_float_labels_referenced": true }, "material_build_warnings": [], "source_zip_standalone_rebuild": true, "data_preservation_report": "layout-data-preservation.json", "figure1_clarification": "Dash marks the one invalid-telemetry run; caption identifies its configuration and exclusion from the counts.", "float_page_changes_since_previous_release": [ [ "RLVRAMBench_MDPI", [ "Table", "9" ], 21, 22 ], [ "RLVRAMBench_MDPI", [ "Figure", "4" ], 25, 26 ], [ "RLVRAMBench_MDPI", [ "Table", "18" ], 32, 33 ] ] }