Doubao Work's RSI Loop: Plan & Goal Modes Enable Self-Correcting Deep Research

Plan Mode: Align Before Execution

Traditional AI deep research often hallucinates or mixes promotional content with fabricated citations. The root cause is the lack of a self-correction loop based on feedback — a core idea behind Recursive Self-Improvement (RSI), a key 2026 AI trend. Doubao Work implements this via two modes: Plan Mode and Goal Mode .

In Plan Mode, the user defines the research domain and goal (e.g., "systematically evaluate Jev's technical positioning"). The AI does not start searching immediately; instead it produces a reviewable, editable Markdown plan ( plan_v1.md) that the user must approve before execution. The plan includes:

Core Research Question (RQ): Whether Jev, as the first System One decision model, holds up under independently verifiable evidence.

Sub-questions: RQ1 capability boundaries, RQ2 performance-cost positioning, RQ3 category and ecosystem positioning.

Falsifiable hypotheses (H1–H3): Each with explicit test methods, rejection criteria, and required data.

The user can then modify the plan. In the case study, the author expanded the literature search strategy from arXiv + Semantic Scholar to also include Google Scholar, OpenReview, ACL Anthology, and IEEE Xplore for broader coverage, and reframed "implications for research" to "implications for research agent architecture" for practitioner focus.

Goal Mode & Agent Parallelism: Execute with Acceptance Criteria

After plan approval, the user switches to Goal Mode and issues an execution command. Doubao Work automatically splits the task into four mutually exclusive, collectively exhaustive parallel shards that run simultaneously:

A | Fact Verification: Cross-check official pricing pages and OpenRouter model pages, producing ≥20 verified facts.

B | arXiv Verification: Fetch and verify 7 papers against JevResearch reproductions, outputting an evidence comparison table.

C | Literature Screening: Multi-query parallel search yielding ≥50 papers (preprints/peer-reviewed ≥30), with search logs recorded. Actual intake reached 71 papers, meeting the Plan Mode acceptance standard.

D | Pre-registration & Evaluation: Write a pre-registration document and evaluate hypotheses H1–H3 with positioning analysis.

After each shard returns, the system performs cross-validation and acceptance aggregation . If any criterion fails, the relevant agent automatically reworks until all standards are met. This "self-volving" loop continues until a satisfactory deliverable is produced.

Recursive Self-Improvement Loop in Practice

The combination of Plan Mode (search-strategy self-alignment), Goal Mode (quantified acceptance criteria with auto-rework), and Agent Parallelism (decomposing a large search tree into concurrent sub-tasks) realizes a minimal, runnable RSI cycle: Plan → Execute → Evaluate → Rework if not met .

Concrete metrics from the Jev evaluation: - Literature intake: 71 papers (target ≥50). - Fact checklist: ≥20 verified items. - arXiv verification: 7 papers cross-checked. - Hypotheses H1–H3 evaluated with explicit rejection conditions.

From Prompt Engineering to System Architect

The author reflects that the role of the human shifts from "quality inspector chasing AI hallucinations" to "architect of the verification system." The user only needs to set direction at the start (Plan) and define the floor at the end (Goal); the intermediate grind and iterative trial-and-error are handled by the closed loop.