paper-to-code-implementation-fidelity

# Skill: paper-to-code-implementation-fidelity ## 1. Capability Definition & Real Case * **Professional Definition**: The ability to convert a research paper’s methodological description into faithful source code, preserving key modeling choices, data processing steps, and implementation dependencies required by the paper’s intended experiments. * **Dimension Hierarchy**: Research Reproduction Engineering->Paper-Grounded Implementation->paper-to-code-implementation-fidelity ### Real Case **[Case 1]** * **Initial Environment**: A workspace contains a research paper in PDF and markdown form, plus an addendum clarifying a few implementation details. No original author code is available. The paper specifies a feature-extraction pipeline using a named encoder and a particular grading procedure. * **Real Question**: Implement the paper’s described feature-construction and evaluation pipeline from scratch in the local repository. * **Real Trajectory**: Read the methods and experimental sections carefully, identify required datasets, model components, and output artifacts, scaffold the repository structure, implement the data and modeling code, and run a small smoke test to confirm the pipeline executes. * **Real Answer**: A new codebase implements the paper’s stated pipeline in a way that reflects the described model components, preprocessing, and evaluation procedure. * **Why this demonstrates the capability**: This capability focuses on faithful methodology translation from paper text to executable code. The core challenge is reading underspecified scientific prose and making correct implementation decisions without access to original source code. --- **[Case 2]** * **Initial Environment**: Another paper workspace describes an experiment comparing two RL variants under a dense-reward setup and specifies the main training loop, metrics, and ablation dimensions. The repository starts empty except for the paper and instructions. * **Real Question**: Build the code needed to reproduce the described RL comparison and logging setup from the paper text. * **Real Trajectory**: Extract the core experimental components from the paper, design a code skeleton for environment interaction, policy training, logging, and evaluation, implement the baseline and comparison variant, and verify that the code can run end-to-end on a toy smoke test. * **Real Answer**: The repository now contains the core implementation needed to run the paper’s comparison under the intended experimental setup. * **Why this demonstrates the capability**: The task measures fidelity in turning research prose into source code structures and experimental machinery. It is not enough to write vaguely similar RL code; the implementation must preserve the paper’s central design choices. ## 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`