Load model + tokenizer on first use.

Normalize OpenAlex work data to common format (matching S2 format).

Return the experiment Python path. Resolution order: 1. config.experiment_python (user-managed environment) 2. experiment/.ve

Find \\cite keys in LaTeX that are missing from the bib, and fill them. Strategy: 1. Extract all cited keys from LaTeX. 2. Ex

Review the paper section-by-section to avoid JSON truncation.

Fix common LLM output issues that break LaTeX compilation. Applies, in order: 1. Unicode replacement (dashes, quotes) 2. Perc

Write deterministic runner assets for a generated experiment project.

Generate a reflection-augmentation response. Parameters ---------- task_type : str One of ``"method_gen"``, ``"co

Get list of available template formats. Returns: List of available format names (e.g., ['arxiv', 'neurips', 'icml'])

Normalize a model-generated train command into runnable tokens + env vars.

Restore code from a snapshot. Returns True on success.

Execute all checks and return an aggregated report.

Check that all metric values in a list are finite numbers.

LLM modifies specific files using search-replace edits (OpenClaw style). Uses precise search-replace blocks instead of full file rewrites to:

Fallback: when search-replace fails, ask LLM to rewrite the target file entirely.

Build context block from comparison_with_baselines analysis data.

Build an evidence summary for chart generation prompts. Priority: real experiment results > literature numbers > empty.

Build a GroundingPacket from all available evidence sources.

Compress historical rounds into a compact summary (~100 chars each).

Build a runner-backed shell command for legacy experiment execution.

Build the concrete argv for a legacy experiment subprocess.

Build a must-cite instruction block for writing prompts. Maps must-cite titles to their actual cite_keys so the LLM knows exactly whi

Build a context string from reference GitHub repos.

Check if a Python file has valid syntax.

Close open string/bracket state for a truncated JSON fragment.

Return finite numeric values from *source*, coercing string floats.

Append a pipeline mode flag if it is not already present.

Compile LaTeX to PDF with automatic error-fix loop. If compilation fails, feed the error back to the LLM, apply the fix, and retry up

Compile LaTeX to PDF with automatic error-fix loop. If compilation fails, feed the error back to the LLM, apply the fix, and retry up

Build an artifact-grounded Experiments section with text/table/figure interleaving. The composer is intentionally deterministic: it reports o

Return contract violations in the generated unified experiment runner.

Derive feedback string from stage output for SDPO data collection.

Determine BibTeX entry type from venue name.

Regenerate the unified runner when static contract checks fail.

Ensure Figure 1 is an image2-generated method schematic.

Create ~/.nanoresearch and its subdirectories if they don't exist.

Return a readable method/variant name from real artifact fields.

Execute LLM-written trim code in a subprocess. Pre-defines INPUT_PATH and OUTPUT_PATH variables for the code. Returns True if executi

Execute main.py --dry-run with up to 5 batch-fix cycles. Each cycle: run -> collect all errors -> fix ALL affected files in one LLM c

r"""Move figure/figure* blocks out of itemize/enumerate environments. Figures inside list environments cause severe formatting issues,

Extract high-value sections from a SKILL.md file. Keeps sections whose heading (lowered) matches *_EXTRACT_HEADINGS*. Returns concatenated se

Pull top-level numeric metrics from the experiment results.

Generate LaTeX without templates as a fallback. Uses the NeurIPS 2025 document class by default. The neurips_2025.sty file is copied

Return a compact, artifact-grounded section when the LLM endpoint times out.

Fetch the referenced_works list for a single OpenAlex work. Returns list of short OpenAlex IDs (e.g. ["W123", "W456"]).