Spaces:
Configuration error
Configuration error
File size: 4,151 Bytes
9670323 | 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 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 | {
"attempt_id": "1e84c33a-e5bd-4a24-b551-de7b4d675054",
"claims": [
{
"challenge_claim_sha256": "333c510d8a8d69cc59827726bb86dd399983b01e7a253af8887d0f2251cda61b",
"evidence": "Repository mechanism audit plus deterministic masked-MDP transition checks.",
"status": "verified",
"target_claim": "The paper formalizes masked diffusion sampling as an MDP in which the diffusion language model is the environment and the policy chooses which tokens to unmask (Section 3)."
},
{
"challenge_claim_sha256": "bf6aebbeea700b651067e91333f97aef0e4fffec15565daf27c4cf0e89b06056",
"evidence": "Repository policy/config audit plus local confidence-to-action checks.",
"status": "verified",
"target_claim": "The learned unmasking policy is a lightweight single-layer transformer mapping token confidences to unmasking decisions (Section 3.2)."
},
{
"challenge_claim_sha256": "28fdce6dd76e8df860c9149960ceaf5f1edf8398eb3da00be9d4844911be16f5",
"evidence": "No trained checkpoint or raw Figure 4 evaluation output is bundled for CPU recomputation.",
"status": "inconclusive",
"target_claim": "Policy sampling matches state-of-the-art heuristic samplers in semi-autoregressive block generation settings (Figure 4)."
},
{
"challenge_claim_sha256": "d2a26b39ada20dfa224d97c827be80a8f15fe7a06bdc9811c6a53f086fc2e607",
"evidence": "No trained checkpoint or raw Figure 5 evaluation output is bundled for CPU recomputation.",
"status": "inconclusive",
"target_claim": "Learned policies outperform heuristic unmasking strategies in the full-diffusion generation setting (Figure 5)."
},
{
"challenge_claim_sha256": "1969db18a5c2b4252edee8f22c1e947bb7d185870e1016c0f8a491141c5415ac",
"evidence": "Source/config audit plus deterministic expert left-to-right unmasking simulation.",
"status": "toy",
"target_claim": "Visualization of learned full-diffusion policies shows expert-steered policies recovering a left-to-right unmasking order on GSM8K samples (Figure 7)."
}
],
"computed_outputs": {
"left_to_right": [
0,
2,
3
],
"mdp_step": {
"action": [
3,
0
],
"done": false,
"next_state": [
"A",
"[MASK]",
"fixed",
"D"
],
"state": [
"[MASK]",
"[MASK]",
"fixed",
"[MASK]"
],
"unmasked": 2
},
"semi_ar_blocks": [
[
0,
1,
2,
3
],
[
4,
5,
6,
7
],
[
8,
9
]
]
},
"fencing_token": 1,
"generated_at": "2026-08-01T00:00:00+00:00",
"limitations": [
"Benchmark claims are inconclusive without released raw evaluation outputs or CPU-feasible trained policy checkpoints.",
"Local simulations verify mechanism contracts and do not substitute for paper-scale model evaluation."
],
"owner": "codex-paper-owner-05",
"paper_id": "F9NDKf5oPy",
"repository_audit": {
"file_sha256": {
"README.md": "2e29199c7bce1da1cb199bf693fb753e9fe9e473c34357633a6f560e6178346d",
"configs/experiment_configs/llada_8b_instruct_dit_confidence_BL32_mixture.yaml": "eb3b223982d15564e358d0ba6e174cc25e4f6094bb42a32630952672829d8580",
"eval/pipeline.py": "7294c782b17d91116ee67c7a1fab347c277f9416f3797a381f465010c52e2dc6"
},
"terms_found": {
"confidence": true,
"environment": true,
"evaluation": true,
"mdp": true,
"single_block_transformer": true
},
"upstream_revision": "35e4830485f1821d57f9ac3f1a303f3d4531fb82"
},
"snapshot_id": "9e9d22e53a0f5eba83916747aebd400e61cf28e84e57cf5219a34f0c7a3b00dd",
"title": "Learning Unmasking Policies for Diffusion Language Models",
"upstream": {
"arxiv": "2512.09106",
"huggingface_paper": "https://huggingface.co/papers/2512.09106",
"openreview": "https://openreview.net/forum?id=F9NDKf5oPy",
"repository": "https://github.com/apple/ml-rl-dllm",
"revision": "35e4830485f1821d57f9ac3f1a303f3d4531fb82"
}
}
|