Pith Number

pith:BYFFEMDJ

pith:2026:BYFFEMDJN5AAXFGZ5K2OHHRJ44

not attested not anchored not stored refs resolved

Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models

Guangtao Zhai, Haoxing Chen, Huayu Zheng, Jianghan Shen, Junjun He, Siqi Luo, Xiaohong Liu, Yan Tai, Yihao Liu, Yi Xin, Yue Li, Yuewen Cao

Hierarchical reinforcement learning with staged sketch-then-paint updates improves how diffusion multimodal models assign credit during image generation.

arxiv:2605.16842 v1 · 2026-05-16 · cs.AI

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{BYFFEMDJN5AAXFGZ5K2OHHRJ44}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest claim

HT-GRPO achieves substantial gains on the GenEval and DPG benchmarks. Evaluations across six additional metrics confirm significant improvements in image quality, aesthetics, and human preference.

C2weakest assumption

The prompt-conditioned estimator correctly computes importance ratios from a fully masked state and the hierarchical credit assignment accurately prioritizes structural tokens without introducing new biases or instabilities in the policy optimization.

C3one line summary

Proposes HT-GRPO with sketch-then-paint staged updates, prompt-conditioned importance ratios, and hierarchical credit assignment for dMLLMs, reporting gains on GenEval and DPG plus quality metrics.

References

49 extracted · 49 resolved · 18 Pith anchors

[1] Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer 2025 · arXiv:2511.22699

[2] Lumina-image 2.0: A unified and efficient image generative framework 2025

[3] Qwen-Image Technical Report 2025 · arXiv:2508.02324

[4] LongCat-Image Technical Report 2025 · arXiv:2512.07584

[5] Emu3: Next-Token Prediction is All You Need 2024 · arXiv:2409.18869

Formal links

2 machine-checked theorem links

Receipt and verification

First computed	2026-05-20T00:03:25.755287Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

0e0a5230696f400b94d9eab4e39e29e71158dca69d8cdf346aab34f3a1460273

Aliases

arxiv: 2605.16842 · arxiv_version: 2605.16842v1 · doi: 10.48550/arxiv.2605.16842 · pith_short_12: BYFFEMDJN5AA · pith_short_16: BYFFEMDJN5AAXFGZ · pith_short_8: BYFFEMDJ

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/BYFFEMDJN5AAXFGZ5K2OHHRJ44 \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 0e0a5230696f400b94d9eab4e39e29e71158dca69d8cdf346aab34f3a1460273

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "5b1c3ab3f01a8eda0f7c556aeac57b2b5fe7414e73ae723e5628244892489f00",
    "cross_cats_sorted": [],
    "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
    "primary_cat": "cs.AI",
    "submitted_at": "2026-05-16T06:59:54Z",
    "title_canon_sha256": "58a4888ac1e2fff49fc3abc3084a79e22f0b0c689af7db46ae3d44292e3f65a7"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2605.16842",
    "kind": "arxiv",
    "version": 1
  }
}