self-correction-guided-refinement

name

self-correction-guided-refinement

description

Use this skill when the user wants to ensure the highest degree of accuracy through a 'think-twice' approach or self-critique. Trigger it for requests like 'double check your answer against the text', 'reflect on your initial guess', 'make sure you didn't miss any contradictions between the files and your internal logic', or 'correct your reasoning if the documents say something counter-intuitive'. This skill is specifically for RAG agents that follow an iterative 'Refine' loop to improve reasoning steps by critiquing a preliminary response against retrieved evidence.

Skill: self-correction-guided-refinement

1. Capability Definition & Real Case

• Professional Definition: The ability of an agent to iteratively evaluate and improve a preliminary response by identifying discrepancies between its internal parametric biases and the retrieved external evidence. This involves a 'reflection' step where the agent critiques its initial reasoning (R0) for hallucinations, omissions, or over-simplifications, then synthesizes a final refined response that better incorporates technical referential details and resolves reasoning-level errors.
• Dimension Hierarchy: Grounded Response Reliability->Evidence Faithfulness->self-correction-guided-refinement

Real Case

[Case 1]

• Initial Environment: A RAG context containing a PubMed research abstract (PubMedQA) describing a study on the effects of a specific drug stimulation. The abstract contains a nuanced 'No' or 'Maybe' result that contradicts common medical stereotypes or broad generalizations.
• Real Question: Does the protein glycoRNA show an increase in M0 macrophages upon stimulation with LPS?
• Real Trajectory: 1. Analyze the initial question. 2. Fetch internal knowledge which suggests that immune markers generally increase under LPS stimulation (initial hypothesis R0: Yes). 3. Retrieve the specific abstract for the targeted protein (U8 glycoRNA). 4. Perform a 'Refine' step: examine the retrieved 'Results' section which explicitly states 'U8 did not show an increase'. 5. Critique the initial hypothesis: 'The initial assumption of an increase is corrected by the specific experimental data in the document regarding the U8 variant.' 6. Synthesize the final corrected answer.
• Real Answer: No
• Why this demonstrates the capability: The case demonstrates the 'COT-Refine' logic where a model must reflect on a potentially biased or generalized initial guess (that LPS always increases markers) and correct it based on the specific, counter-intuitive evidence found in the document. Success requires the agent to prioritize the refined 'reflection' over its first reaction.

---

[Case 2]

• Initial Environment: A RAG environment with an online medical consultation query (LFQA) and a retrieved set from StatPearls. The retrieved text is highly technical and verbatim, while the user is a layperson asking for advice.
• Real Question: What are the primary goals of the SPADEX mission and how does it affect future space docking?
• Real Trajectory: 1. Retrieve technical documents describing SPADEX autonomous docking and refueling. 2. Generate a preliminary answer (R0) that is highly technical and uses verbatim medical/technical jargon. 3. Perform a 'Readability and Usefulness Refine' step: analyze if the response is too formal for a layperson. 4. Adjust the reasoning: 'The technical terms must be preserved for accuracy but explained or re-organized for the user's reading level.' 5. Synthesize a response that balances formal precision with structured usefulness.
• Real Answer: The primary goal is the demonstration of autonomous docking and refueling. This is a critical building block for assembling the planned space station, as it allows spacecraft to join together without manual intervention.
• Why this demonstrates the capability: This demonstrates the 'Refine' aspect where an agent critiques its initial 'verbatim-heavy' RAG output to balance 'Readability' with 'Knowledge,' as identified in the MRAG benchmark. It shows the model iteratively improving the organization of its reasoning to be both grounded and useful.

Pipeline Execution Instructions

To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:

1. Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds: references/EXPLORATION.md

2. Phase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: references/SELECTION.md

3. Phase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: references/SYNTHESIS.md