P2TPO: Structure-Aligned Page-to-Token Credit Assignment for Full-Page Document Parsing
We introduce Page-to-Token Policy Optimization (P2TPO), a structure-aligned credit-assignment framework that operates on full-page policy trajectories.
ProofPaper ↗
Key points
- End-to-end full-page document parsing must preserve the content and structure of heterogeneous elements, yet otherwise accurate outputs can contain local errors and omissions.
- Reference-element alignment first associates block-level quality contrasts with generated spans.
- Within matched blocks, Span Credit Redistribution (SCR) redistributes existing block credit according to aligned correct and erroneous spans, using boundary proxies for within-block omissions while preserving the total block-derived advantage across its tokens.
- Under a fixed-weight additive page utility, we show that element-indexed contrasts retain local quality differences that page-level aggregation can cancel out.
Sources (1)
- [1]P2TPO: Structure-Aligned Page-to-Token Credit Assignment for Full-Page Document ParsingarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 02:32 PM
We introduce Page-to-Token Policy Optimization (P2TPO), a structure-aligned credit-assignment framework that operates on full-page policy trajectories.
End-to-end full-page document parsing must preserve the content and structure of heterogeneous elements, yet otherwise accurate outputs can contain local errors and omissions.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026Velox: Spatial Memory Makes World Models Efficient
- Oct 14, 2026mistralai/Timestral-209M
- Oct 14, 2026Learning to Compute Is Not Learning to Retain or Reuse in Looped Transformers
- Oct 14, 2026How Block is building AI-powered security with OpenAI technology
- Oct 14, 2026LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation
- Oct 14, 2026Your Transformer Is Secretly Elastic