{
"paper": {
"title": "Language Models Learn to Improve Language Models",
"method": "LM2LM",
"arxiv": "2610.23344",
"version": "v2",
"revisionDate": "2026-11-02",
"url": "https://arxiv.org/abs/2610.23344",
"pdf": "https://arxiv.org/pdf/2610.23344v2",
"authorsSource": {
"label": "Title page",
"url": "https://arxiv.org/pdf/2610.23344v2#page=1",
"page": 1,
"tex": "amazonpaper.tex / author and affiliation declarations"
}
},
"roles": {
"proposer": {
"model": "Qwen3.8-27B",
"weightsUpdated": true,
"method": "DAPO reinforcement learning"
},
"rwm": {
"model": "GPT-6-sol",
"weightsUpdated": false,
"reasoning": "high",
"recordsUpdated": true,
"description": "Predicts experimental gain using real records supplied in context."
},
"builder": {
"model": "GPT-6-astra",
"reasoning": "high",
"description": "Implements proposals for real-execution evaluation and calibration."
},
"source": {
"label": "Section 4.1; Appendix E",
"url": "https://arxiv.org/pdf/2610.23344v2#page=7",
"page": 7,
"tex": "5_experiments_v2.tex / Implementation Details; 7_appendix_v2.tex / sec:exp-impl-details"
}
},
"protocol": {
"trainingTasks": 40,
"heldoutTasks": 10,
"heldoutStates": 78,
"arStates": 60,
"diffusionStates": 18,
"familiesInTraining": 8,
"familiesInHeldoutEvaluation": 7,
"proposalsPerState": 4,
"defaultRlSteps": 96,
"initialRealRecords": 100,
"calibrationSteps": [
16,
40,
64
],
"calibrationRecordsAdded": [
31,
38,
31
],
"defaultTotalRealRuns": 200,
"baselineIclRecords": 300,
"sftPairs": 300,
"rftPositivePairs": 114,
"definitionOfTask": "Model family, model size, training budget, and reference recipe. Each real experiment trains the target model from scratch.",
"definitions": {
"gain": "Reference validation loss minus proposal validation loss; positive is better.",
"avg4": "Mean of four real proposal gains per state, then equal-weight average over states.",
"best4": "Maximum of four real proposal gains per state, then equal-weight average over states.",
"negativePct": "Percentage of proposals with negative real gain.",
"winRate": "Percentage of states with higher Avg@4; ties count as 0.5.",
"invalidOrUnchanged": "Assigned zero gain; negative gains are retained."
},
"isolation": "No held-out experimental records are used for proposer training, RWM grounding, or calibration.",
"calibrationRecordUse": "Calibration outcomes enter the RWM context only; they do not enter training-state histories.",
"source": {
"label": "Section 4.1; Appendices C-E",
"url": "https://arxiv.org/pdf/2610.23344v2#page=7",
"page": 7,
"tex": "5_experiments_v2.tex / ssec:exp-env; 7_appendix_v2.tex / app:data-construction, app:metrics, app:impl-cal"
}
},
"main": {
"source": {
"label": "Table 1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=6",
"page": 6,
"tex": "5_experiments_v2.tex / tab:main"
},
"ar": {
"unit": "10^-3 BPB",
"unitHtml": "10-3 BPB",
"states": 60,
"rows": [
{
"id": "base",
"label": "Base",
"avg4": -7.94,
"best4": 11.39,
"negativePct": 70.4,
"winVsBasePct": null,
"lm2lmWinVsRowPct": 93.3
},
{
"id": "base_icl",
"label": "Base-ICL",
"avg4": 4.04,
"best4": 16.54,
"negativePct": 40.8,
"winVsBasePct": 81.7,
"lm2lmWinVsRowPct": 66.7
},
{
"id": "gpt6_sol",
"label": "GPT-6-sol",
"avg4": 6.51,
"best4": 8.79,
"negativePct": 52.5,
"winVsBasePct": 83.3,
"lm2lmWinVsRowPct": 56.7
},
{
"id": "gpt6_sol_icl",
"label": "GPT-6-sol-ICL",
"avg4": 6.15,
"best4": 11.27,
"negativePct": 29.6,
"winVsBasePct": 78.3,
"lm2lmWinVsRowPct": 51.7
},
{
"id": "sft",
"label": "SFT",
"avg4": -12.19,
"best4": 5.51,
"negativePct": 71.5,
"winVsBasePct": 55.0,
"lm2lmWinVsRowPct": 93.3
},
{
"id": "rft_real",
"label": "RFT-Real",
"avg4": -22.95,
"best4": 9.69,
"negativePct": 66.5,
"winVsBasePct": 45.0,
"lm2lmWinVsRowPct": 95.0
},
{
"id": "lm2lm",
"label": "LM2LM",
"avg4": 17.06,
"best4": 23.4,
"negativePct": 39.2,
"winVsBasePct": 93.3,
"lm2lmWinVsRowPct": null
}
]
},
"diffusion": {
"unit": "10^-2 nats per token",
"unitHtml": "10-2 nats / token",
"states": 18,
"rows": [
{
"id": "base",
"label": "Base",
"avg4": -4.18,
"best4": 5.89,
"negativePct": 59.7,
"winVsBasePct": null,
"lm2lmWinVsRowPct": 100
},
{
"id": "base_icl",
"label": "Base-ICL",
"avg4": 4.03,
"best4": 8.5,
"negativePct": 13.9,
"winVsBasePct": 83.3,
"lm2lmWinVsRowPct": 72.2
},
{
"id": "gpt6_sol",
"label": "GPT-6-sol",
"avg4": 1.98,
"best4": 4.43,
"negativePct": 38.9,
"winVsBasePct": 83.3,
"lm2lmWinVsRowPct": 83.3
},
{
"id": "gpt6_sol_icl",
"label": "GPT-6-sol-ICL",
"avg4": 4.94,
"best4": 5.93,
"negativePct": 0,
"winVsBasePct": 94.4,
"lm2lmWinVsRowPct": 72.2
},
{
"id": "sft",
"label": "SFT",
"avg4": -3.09,
"best4": 3.73,
"negativePct": 54.2,
"winVsBasePct": 61.1,
"lm2lmWinVsRowPct": 100
},
{
"id": "rft_real",
"label": "RFT-Real",
"avg4": -3.72,
"best4": 3.37,
"negativePct": 52.8,
"winVsBasePct": 50,
"lm2lmWinVsRowPct": 88.9
},
{
"id": "lm2lm",
"label": "LM2LM",
"avg4": 7.1,
"best4": 9.93,
"negativePct": 7.6,
"winVsBasePct": 100,
"lm2lmWinVsRowPct": null
}
]
},
"baseModel": "Qwen3.8-27B",
"iclRecords": 300,
"metricsAreRealExecution": true,
"aggregateCaution": "Highest aggregate Avg@4 and Best@4 in both domains; not a win in every family or the lowest negative-gain rate."
},
"transfer": {
"source": {
"label": "Table 2; Section 4.3",
"url": "https://arxiv.org/pdf/2610.23344v2#page=8",
"page": 8,
"tex": "5_experiments_v2.tex / tab:rq3-transfer-l2, ssec:RQ3"
},
"excludedFamily": "Qwen3",
"unit": "10^-3 BPB",
"rows": [
{
"id": "base",
"label": "Base",
"avg4": 3.53,
"best4": 31.39,
"winVsBasePct": null,
"lm2lmWinVsRowPct": 87.5
},
{
"id": "base_icl",
"label": "Base-ICL",
"avg4": 2.97,
"best4": 25.22,
"winVsBasePct": 58.3,
"lm2lmWinVsRowPct": 100
},
{
"id": "gpt6_sol",
"label": "GPT-6-sol",
"avg4": 20.19,
"best4": 22.75,
"winVsBasePct": 66.7,
"lm2lmWinVsRowPct": 62.5
},
{
"id": "gpt6_sol_icl",
"label": "GPT-6-sol-ICL",
"avg4": 21.63,
"best4": 25.27,
"winVsBasePct": 75,
"lm2lmWinVsRowPct": 62.5
},
{
"id": "lm2lm",
"label": "LM2LM",
"avg4": 53.72,
"best4": 86.69,
"winVsBasePct": 87.5,
"lm2lmWinVsRowPct": null
}
],
"excludedFrom": [
"proposer training tasks",
"RWM grounding records"
],
"description": "A separate experiment holds out the entire Qwen3 target family, trains on the remaining families, and evaluates real proposal gains on Qwen3.",
"avgGainRatioVsGptIcl": 2.483587609801202,
"ratioIsDerived": true,
"importantDistinction": "Qwen3 is the target model family being improved. The proposing agent is Qwen3.8-27B. The family exclusion concerns this study's experimental records and training tasks, not the base agent's pretraining knowledge."
},
"costs": {
"source": {
"label": "Section 4.2; Tables 10 and 11",
"url": "https://arxiv.org/pdf/2610.23344v2#page=24",
"page": 24,
"tex": "5_experiments_v2.tex / Cost Efficiency; 7_appendix_v2.tex / tab:total-cost, tab:rq1-cost"
},
"comparison": {
"lm2lm": {
"checkpoint": "Step 96 / M3",
"realRuns": 200,
"dataGenerationUsd": 40050,
"dataGenerationUsdRounded": 40100,
"dataGenerationHours": 343.2,
"proposerTrainingUsdRounded": 16100,
"totalUsdRounded": 56100
},
"icl300": {
"realRuns": 300,
"dataGenerationUsd": 60001,
"dataGenerationUsdRounded": 60000,
"dataGenerationHours": 510,
"proposerTrainingUsdRounded": 0,
"totalUsdRounded": 60000
},
"dataGenerationReductionPct": 33.25111248145865,
"headlineDataGenerationReductionPct": 33,
"totalReductionPctFromRoundedCosts": 6.5,
"dataGenerationTimeReductionPct": 32.705882352941174
},
"definition": "Data generation includes GPU cost of real experimental runs (including failed runs) plus RWM prediction token costs.",
"excludedFromDataGeneration": [
"proposer inference",
"builder inference",
"proposer fine-tuning",
"shared research-state construction"
],
"totalDefinition": "The paper's total adds proposer-training GPU cost to data-generation cost; it is not an all-inclusive project or deployment bill.",
"priceBasis": {
"gpuModel": "NVIDIA B200",
"usdPerGpuHour": 14.208,
"note": "Paper accounting prices, not a current price quote. Table 10 uses B200-equivalent hours with H200 converted at 1.198."
},
"timeDefinition": "Sum of real-execution durations assuming sequential execution, plus RWM prediction time; not observed end-to-end elapsed time.",
"arCheckpoints": [
{
"recordSet": "M0",
"rlStep": 16,
"realRuns": 100,
"gpuHours": 1665,
"rwmTokensMillions": 140.6,
"dataGenerationUsd": 23901,
"realExecutionHours": 202.0,
"rwmHours": 1.4,
"avg4": 1.81,
"best4": 18.91
},
{
"recordSet": "M1",
"rlStep": 40,
"realRuns": 131,
"gpuHours": 1960,
"rwmTokensMillions": 294.1,
"dataGenerationUsd": 28330,
"realExecutionHours": 239.0,
"rwmHours": 2.5,
"avg4": 10.85,
"best4": 21.28
},
{
"recordSet": "M2",
"rlStep": 64,
"realRuns": 169,
"gpuHours": 2500,
"rwmTokensMillions": 459.4,
"dataGenerationUsd": 36183,
"realExecutionHours": 306.6,
"rwmHours": 3.4,
"avg4": 18.01,
"best4": 23.91
},
{
"recordSet": "M3",
"rlStep": 96,
"realRuns": 200,
"gpuHours": 2758,
"rwmTokensMillions": 664.7,
"dataGenerationUsd": 40050,
"realExecutionHours": 338.9,
"rwmHours": 4.3,
"avg4": 17.06,
"best4": 23.4
},
{
"recordSet": "M4",
"rlStep": 128,
"realRuns": 236,
"gpuHours": 3206,
"rwmTokensMillions": 889.3,
"dataGenerationUsd": 46626,
"realExecutionHours": 395.0,
"rwmHours": 5.3,
"avg4": 19.4,
"best4": 24.69
}
],
"arCheckpointsSource": {
"label": "Table 11",
"url": "https://arxiv.org/pdf/2610.23344v2#page=24",
"page": 24,
"tex": "7_appendix_v2.tex / tab:rq1-cost"
},
"totalCheckpoints": [
{
"recordSet": "M0",
"dataGenerationUsdThousands": 23.9,
"proposerTrainingGpuHours": 152,
"proposerTrainingUsdThousands": 2.2,
"totalUsdThousands": 26.1
},
{
"recordSet": "M1",
"dataGenerationUsdThousands": 28.3,
"proposerTrainingGpuHours": 434,
"proposerTrainingUsdThousands": 6.2,
"totalUsdThousands": 34.5
},
{
"recordSet": "M2",
"dataGenerationUsdThousands": 36.2,
"proposerTrainingGpuHours": 898,
"proposerTrainingUsdThousands": 12.8,
"totalUsdThousands": 48.9
},
{
"recordSet": "M3",
"dataGenerationUsdThousands": 40.1,
"proposerTrainingGpuHours": 1131,
"proposerTrainingUsdThousands": 16.1,
"totalUsdThousands": 56.1
},
{
"recordSet": "M4",
"dataGenerationUsdThousands": 46.6,
"proposerTrainingGpuHours": 1376,
"proposerTrainingUsdThousands": 19.6,
"totalUsdThousands": 66.2
}
],
"totalSource": {
"label": "Table 10",
"url": "https://arxiv.org/pdf/2610.23344v2#page=23",
"page": 23,
"tex": "7_appendix_v2.tex / tab:total-cost"
},
"roundingNote": "Individual displayed costs are rounded independently; 40.1 + 16.1 need not equal displayed total 56.1."
},
"calibration": {
"source": {
"label": "Section 4.4; Figure 4",
"url": "https://arxiv.org/pdf/2610.23344v2#page=8",
"page": 8,
"tex": "5_experiments_v2.tex / ssec:ablation"
},
"ablation": {
"domain": "ar",
"unit": "10^-3 BPB",
"rlStep": 96,
"realRuns": 200,
"rows": [
{
"id": "static200",
"label": "Static records (200)",
"avg4": 14.76,
"best4": null,
"note": "100 initial records plus 100 random records, fixed throughout RL; Best@4 not transcribed because no exact value is printed in the prose."
},
{
"id": "random_calibration",
"label": "Random calibration",
"avg4": 10.84,
"best4": 18.6
},
{
"id": "top_rated",
"label": "Top-rated calibration (LM2LM)",
"avg4": 17.06,
"best4": 23.4
}
]
},
"reliability": {
"source": {
"label": "Section 4.5; Figure 5",
"url": "https://arxiv.org/pdf/2610.23344v2#page=9",
"page": 9,
"tex": "5_experiments_v2.tex / ssec:RQ2"
},
"extendedRlSteps": 128,
"extendedRealRuns": 236,
"comparison": "Each variant is scored by its own training RWM on real-executed held-out AR proposals.",
"static": {
"pairwiseAccuracy": [
{
"step": 16,
"value": 0.72
},
{
"step": 128,
"value": 0.37
}
],
"advantageCosine": [
{
"step": 16,
"value": 0.69
},
{
"step": 128,
"value": -0.42
}
],
"rewardGap": [
{
"step": 96,
"value": 2.0,
"approximate": true
}
]
},
"calibrated": {
"pairwiseAccuracyAt128": 0.61,
"pairwiseAccuracyRange": [
0.61,
0.69
],
"advantageCosineRange": [
0.5,
0.71
],
"rewardGap": [
{
"step": 96,
"value": 0.85
},
{
"step": 128,
"value": 0.17
}
]
},
"chancePairwiseAccuracy": 0.5,
"identicalProposalsAt96": {
"staticRecordsPairwiseAccuracy": 0.46,
"calibratedRecordsPairwiseAccuracy": 0.61
},
"plottingNote": "Only stated endpoints and ranges are transcribed. Do not invent intermediate learning-curve samples."
}
},
"environments": [
{
"family": "OLMo-3",
"domain": "ar",
"trainingTasks": [
"100m-4c",
"100m-8c",
"190m-1c",
"190m-2c",
"370m-1c",
"370m-2c",
"760m-1c"
],
"trainingTaskCount": 7,
"heldoutTasks": [
{
"id": "100m-1c",
"states": 17,
"totalParametersMillions": 153,
"budget": "2.0B tokens",
"referenceLoss": 1.375,
"referenceSeedStd": 0.0011
},
{
"id": "100m-2c",
"states": 8,
"totalParametersMillions": 153,
"budget": "4.0B tokens",
"referenceLoss": 1.2939,
"referenceSeedStd": 0.0056
},
{
"id": "190m-4c",
"states": 4,
"totalParametersMillions": 267,
"budget": "15.2B tokens",
"referenceLoss": 1.1275,
"referenceSeedStd": 0.0017
}
],
"heldoutTaskCount": 3,
"heldoutStates": 29,
"initialRecords": 16,
"baselinePoolRecords": 47,
"trainingData": "Dolma 3",
"metric": "Mean BPB over ten downstream tasks",
"description": "Autoregressive transformer",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
},
{
"family": "Qwen3",
"domain": "ar",
"trainingTasks": [
"q06b-1b",
"q06b-2b",
"q17b-1b",
"q17b-2b",
"q152m-8b",
"q267m-1b",
"q267m-2b"
],
"trainingTaskCount": 7,
"heldoutTasks": [
{
"id": "q152m-2b",
"states": 8,
"totalParametersMillions": 153,
"budget": "2.0B tokens",
"referenceLoss": 1.3641,
"referenceSeedStd": 0.0025
},
{
"id": "q267m-4b",
"states": 4,
"totalParametersMillions": 267,
"budget": "4.0B tokens",
"referenceLoss": 1.1683,
"referenceSeedStd": 0.0042
}
],
"heldoutTaskCount": 2,
"heldoutStates": 12,
"initialRecords": 25,
"baselinePoolRecords": 62,
"trainingData": "Dolma 3",
"metric": "Mean BPB over ten downstream tasks",
"description": "Official Qwen3 architecture and tokenizer",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
},
{
"family": "NanoChat",
"domain": "ar",
"trainingTasks": [
"d8-8k-0p82b",
"d8-8k-3p28b",
"d12-8k-8p59b",
"d16-8k-4p58b",
"d20-8k-8p49b"
],
"trainingTaskCount": 5,
"heldoutTasks": [
{
"id": "d12-8k-2p15b",
"states": 17,
"totalParametersMillions": 286,
"budget": "2.15B tokens",
"referenceLoss": 0.8135,
"referenceSeedStd": 0.0007
}
],
"heldoutTaskCount": 1,
"heldoutStates": 17,
"initialRecords": 15,
"baselinePoolRecords": 79,
"trainingData": "ClimbMix",
"metric": "BPB on a held-out ClimbMix validation shard",
"description": "Depth-parameterized transformer",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
},
{
"family": "Marin",
"domain": "ar",
"trainingTasks": [
"153m-2p9b",
"153m-8b",
"267m-2p9b",
"267m-8b",
"486m-2p9b"
],
"trainingTaskCount": 5,
"heldoutTasks": [
{
"id": "486m-8b",
"states": 2,
"totalParametersMillions": 486,
"budget": "8.0B tokens",
"referenceLoss": 0.9959,
"referenceSeedStd": 0.0008
}
],
"heldoutTaskCount": 1,
"heldoutStates": 2,
"initialRecords": 11,
"baselinePoolRecords": 30,
"trainingData": "Dolma 3",
"metric": "Mean BPB over ten downstream tasks",
"description": "Marin architecture",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
},
{
"family": "E2D2",
"domain": "diffusion",
"trainingTasks": [
"34m-6e18",
"113m-3e18",
"113m-6e18",
"231m-6e18"
],
"trainingTaskCount": 4,
"heldoutTasks": [
{
"id": "66m-1e18",
"states": 6,
"totalParametersMillions": 115,
"budget": "1e18 FLOPs / 1.55B tokens",
"referenceLoss": 3.5508,
"referenceSeedStd": 0.0037
}
],
"heldoutTaskCount": 1,
"heldoutStates": 6,
"initialRecords": 11,
"baselinePoolRecords": 31,
"trainingData": "SlimPajama",
"metric": "NLL in nats per token",
"description": "Encoder-decoder masked diffusion; block-autoregressive",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
},
{
"family": "BD3LM",
"domain": "diffusion",
"trainingTasks": [
"34m-6e18",
"113m-3e18",
"113m-1e19",
"231m-6e18"
],
"trainingTaskCount": 4,
"heldoutTasks": [
{
"id": "113m-6e18",
"states": 6,
"totalParametersMillions": 162,
"budget": "6e18 FLOPs / 3.46B tokens",
"referenceLoss": 3.1218,
"referenceSeedStd": 0.0005
}
],
"heldoutTaskCount": 1,
"heldoutStates": 6,
"initialRecords": 7,
"baselinePoolRecords": 19,
"trainingData": "SlimPajama",
"metric": "NLL in nats per token",
"description": "Block diffusion; block-structured attention",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
},
{
"family": "MDM",
"domain": "diffusion",
"trainingTasks": [
"34m-6e18",
"85m-6e18",
"85m-1e19",
"231m-6e18"
],
"trainingTaskCount": 4,
"heldoutTasks": [],
"heldoutTaskCount": 0,
"heldoutStates": 0,
"initialRecords": 6,
"baselinePoolRecords": 14,
"trainingData": "SlimPajama",
"metric": "NLL in nats per token",
"description": "Full-sequence bidirectional masked diffusion; training only",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
},
{
"family": "E2D2-SW",
"domain": "diffusion",
"trainingTasks": [
"34m-6e18",
"113m-3e18",
"113m-1e19",
"206m-6e18"
],
"trainingTaskCount": 4,
"heldoutTasks": [
{
"id": "113m-6e18",
"states": 6,
"totalParametersMillions": 196,
"budget": "6e18 FLOPs / 2.44B tokens",
"referenceLoss": 3.2846,
"referenceSeedStd": 0.0026
}
],
"heldoutTaskCount": 1,
"heldoutStates": 6,
"initialRecords": 9,
"baselinePoolRecords": 18,
"trainingData": "SlimPajama",
"metric": "NLL in nats per token",
"description": "Token-causal encoder with sliding-window decoder training",
"source": {
"label": "Tables 3 and 4; Appendix C.1",
"url": "https://arxiv.org/pdf/2610.23344v2#page=17",
"page": 17,
"tex": "7_appendix_v2.tex / tab:data, tab:heldout, app:env"
}
}
],
"authors": [
{
"name": "Minhua Lin",
"affiliations": [
1
]
},
{
"name": "Zijun Wang",
"affiliations": [
2
]
},
{
"name": "Zewen Liu",
"affiliations": [
3
]
},
{
"name": "Zhaotian Weng",
"affiliations": [
4
]
},
{
"name": "Bing He",
"affiliations": [
5
]
},
{
"name": "Zhan Shi",
"affiliations": [
5
]
},
{
"name": "Benoit Dumoulin",
"affiliations": [
5
]
},
{
"name": "Xiang Zhang",
"affiliations": [
1
]
},
{
"name": "Suhang Wang",
"affiliations": [
1
]
},
{
"name": "Hanqing Lu",
"affiliations": [
5
]
}
],
"affiliations": [
{
"id": 1,
"name": "The Pennsylvania State University"
},
{
"id": 2,
"name": "UC Santa Cruz"
},
{
"id": 3,
"name": "Emory University"
},
{
"id": 4,
"name": "UC Santa Barbara"
},
{
"id": 5,
"name": "Amazon"
}
],
"caveats": [
"This is proposal-level improvement in controlled LM development, not a demonstration of end-to-end recursive self-improvement.",
"Eight model families are used in training; real held-out evaluation covers seven because MDM has no held-out task.",
"AR and diffusion gains use different units and must not share an unlabeled numerical scale.",
"Main results average equally over states, not equally over families. There are 29 OLMo-3, 12 Qwen3, 17 NanoChat, 2 Marin, and 6 states for each evaluated diffusion family.",
"LM2LM has the best aggregate Avg@4 and Best@4, but GPT-6-sol-ICL has lower negative-gain rates; individual-family wins are not universal.",
"The 33% reduction is for data-generation cost against ICL with 300 records. Including proposer RL training gives about 6.5% lower reported total cost using rounded figures.",
"The family-held-out Qwen3 experiment is separate from the main task-held-out benchmark.",
"The RWM is frozen; on-policy calibration updates its context records. The proposer is the model trained with RL.",
"The paper's cost accounting omits proposer and builder inference and shared state construction.",
"Calibration corrects selected high-rated proposals using real outcomes; surrogate predictions remain imperfect."
],
"verification": {
"sourceVersion": "arXiv v2, 2 November 2026",
"method": "Non-comment LaTeX table/prose checked against text extracted from the published v2 PDF.",
"staleCommentValuesExcluded": [
"AR/diffusion main Avg@4 15.07/6.88",
"static pairwise accuracy 0.34; calibrated 0.59 at step 128",
"static advantage cosine -0.43",
"calibrated reward gap 0.35 at step 128"
],
"publishedValuesUsed": [
"main Avg@4 17.06/7.10",
"static pairwise accuracy 0.37; calibrated 0.61 at step 128",
"static advantage cosine -0.42",
"calibrated reward gap 0.17 at step 128"
]
},
"display": {
"providers": {
"alibaba": {
"name": "Alibaba",
"color": "#FF7018",
"logo": "assets/model-logos/alibaba.svg"
},
"openai": {
"name": "OpenAI",
"color": "#1F1F1F",
"logo": "assets/model-logos/openai.svg"
}
},
"methods": {
"base": {
"label": "Base",
"model": "Qwen3.8 27B",
"provider": "alibaba",
"description": "Untuned proposer"
},
"base_icl": {
"label": "Base + records",
"model": "Qwen3.8 27B",
"provider": "alibaba",
"description": "300-record ICL"
},
"gpt6_sol": {
"label": "GPT-6 Sol",
"model": "GPT-6 Sol",
"provider": "openai",
"description": "Zero-shot proposer"
},
"gpt6_sol_icl": {
"label": "GPT-6 Sol + records",
"model": "GPT-6 Sol",
"provider": "openai",
"description": "300-record ICL"
},
"sft": {
"label": "SFT",
"model": "Qwen3.8 27B",
"provider": "alibaba",
"description": "Supervised fine-tuning"
},
"rft_real": {
"label": "RFT-Real",
"model": "Qwen3.8 27B",
"provider": "alibaba",
"description": "Rejection-sampling fine-tuning"
},
"lm2lm": {
"label": "LM2LM",
"model": "Qwen3.8 27B",
"provider": "alibaba",
"description": "Trained proposer"
}
}
}
}