File size: 2,732 Bytes
f791e67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
{
  "schema_version": 1,
  "claim_id": 5,
  "verdict": "BLOCKED",
  "paper_statement": "Table 5 reports Spearman rho 0.800 for the decoder-head formulation versus 0.717 for a normalized regression head and 0.478 for a standard regression head; Table 6 reports that scaling T5Gemma prefix-LM from 300M to 600M parameters improves Spearman rho from 0.744 to 0.782.",
  "source": {
    "paper_id": "2509.26476",
    "html_url": "https://ar5iv.labs.arxiv.org/html/2509.26476",
    "retrieved_utc_date": "2026-07-27",
    "html_sha256": "5947f4512cc86850a63409adf52af25ac1f40b15dcc797348fd8ae91a2740913",
    "anchors": [
      "S6.SS2",
      "S6.T5",
      "S6.T6"
    ]
  },
  "table5": {
    "metric": "Spearman rank correlation",
    "evaluation_domain": "512 NASBench101 validation samples",
    "training_domains": [
      "NASBench101",
      "SNAS",
      "OFA ResNet",
      "OFA ProxylessNAS",
      "OFA MobileNet"
    ],
    "reported": {
      "standard_regression_head": 0.478,
      "normalized_regression_head": 0.717,
      "decoder_head": 0.800
    },
    "acceptance": "Reproduce all three formulations under the paper's same data split, optimization, checkpoint-selection, input, and evaluation protocol, with independently recomputed Spearman correlations and negative controls."
  },
  "table6": {
    "metric": "Spearman rank correlation",
    "evaluation_domain": "1024 CodeNet samples",
    "training_domain": "a smaller subset of CodeNet, APPS, and KernelBook",
    "reported": {
      "t5gemma_s_s_prefixlm_300m": 0.744,
      "t5gemma_b_b_prefixlm_600m": 0.782,
      "difference": 0.038
    },
    "acceptance": "Reproduce both sizes with the exact same settings and data identities, independently recompute correlations and uncertainty, and show a positive 600M-minus-300M difference."
  },
  "required_evidence": [
    "Exact Table 5 implementations and hyperparameters for all three heads.",
    "Exact Table 5 training and validation row identities and deterministic seeds.",
    "Exact Table 5 checkpoints or a complete reproducible training recipe.",
    "Exact Table 6 300M and 600M RLM checkpoints or a complete reproducible training recipe.",
    "Exact Table 6 smaller-subset and evaluation row identities and seeds.",
    "Raw predictions, independent metric checker, and shuffled-target controls."
  ],
  "non_substitutions": [
    "A generic sequence-classification regressor is not the missing Table 5 experiment.",
    "Public base-model parameter metadata is not the Table 6 scaling result.",
    "The released 181.5M RLM is not a counterexample unless it is identified as an exact Table 6 checkpoint.",
    "A failure caused by gated access or absent artifacts is not falsification."
  ]
}