Can iterative revision cycles match how humans actually write?

Existing deep research agents combine test-time scaling techniques (CoT, best-of-n, MCTS, debate, self-refinement) without a deliberate cognitive design. Most public agents employ a linear or parallelized pipeline of planning → searching → generation, which loses global context and misses critical dependencies. Cognitive studies of human writing (Flower and Hayes, 1981) show that people do not write linearly — they establish a high-level plan, draft based on the plan, and then engage in multiple revision cycles that interleave further information gathering with rewriting.

TTD-DR observes a structural similarity between this human pattern and retrieval-augmented diffusion sampling: a noisy initial draft is iteratively denoised toward higher-quality outputs, with each denoising step informed by retrieved external information. The framework operationalizes this as report-level diffusion — a preliminary draft serves as an updatable skeleton that evolves through iterative refinement, with each step augmented by targeted retrieval. The draft is a global anchor that maintains coherence across iterations, addressing the information-loss problem of linear pipelines.

Two mechanisms make the analogy useful in practice. Denoising with retrieval drives report-level evolution: the draft and research plan jointly steer the next retrieval, and retrieved content drives the next denoising step. Self-evolution operates at the component level: each unit agent (plan generator, question generator, answer searcher, report generator) undergoes its own optimization, mitigating per-component information loss across long agentic trajectories. The interplay is essential — without component-level self-evolution, the draft-level diffusion lacks high-quality context to refine on.

The conceptual yield is that diffusion is not just a generation technique but a process model for cognitively-inspired iterative work. Any task that humans approach as draft-and-revise rather than write-once-correctly — research reports, design documents, complex prose — is a candidate for the same draft-centric, retrieval-augmented diffusion treatment. The draft becomes the persistent state that the agentic system refines, rather than a final output produced by a feed-forward pipeline — analogous to how Why does vanilla RAG produce shallow and redundant results? argues iterative loops are required for depth.