llm-finetuning

May 4, 2026 · View on GitHub

Code release for the paper "Beyond Inference-Time Search: Reinforcement Learning Synthesizes Reusable Solvers".

This repository packages the paper-backed experiments as first-class reproducibility surfaces. The main benchmark is Synergistic Dependency Selection (SDS), with additional-domain evidence on the Job Shop Scheduling Problem (JSSP) inventoried in the release manifest.

Quick start

# Local environment
./setup_dev.sh --dev
conda activate llm-finetuning

# Validate everything that is honestly checkable from a local checkout
./scripts/validate_paper_release.sh

# Recreate the main paper bundles from the canonical public manifest
./scripts/generate_paper_results.sh

# Validate appendix/supporting evidence and regenerate soft-gate plots when
# the corresponding frozen SDS result roots are available locally
./scripts/generate_paper_appendix_results.sh

Canonical public entrypoints

  • Main paper manifest:
    • experiments/report_sets/paper_public_main_v1.json
  • Appendix / supporting-evidence manifest:
    • experiments/report_sets/paper_public_appendix_v1.json
  • Human-readable release manifest:
    • docs/release_manifest.md
  • Machine-readable artifact inventory:
    • docs/release_artifact_inventory.json

The checked-in paper bundles currently live at:

  • SDS:
    • evaluation/sds/aggregated_report_batches/paper_public_main_v1/
  • BigCode:
    • evaluation/bigcode/aggregated_report_batches/paper_public_main_v1/
  • Fixed-code / runtime audit bundle:
    • evaluation/sds/aggregated_report_batches/20260326_baseline-eval-v1/

What is included in this release

The public release path now includes:

  • refreshed SDS baseline package with the neutral-prompt ShinkaEvolve rerun
  • fixed-code SDS evaluation support for frozen-solver validation
  • manually specified constraint-aware simulated annealing baseline
  • soft-gate SDS ablation support
  • reward-normalization ablation support
  • feasibility-sparsity logging + summary artifacts
  • timeout analysis report
  • paper-aligned manifests, figures, tables, and release docs

CVRP remains intentionally out of scope for this release.

Reproducibility model

This repository now has two complementary reproducibility modes:

  1. Paper-bundle verification

    • inspect the checked-in aggregated outputs that match the final manuscript
    • use docs/release_manifest.md to map every paper-facing number to its source bundle

  2. Frozen-artifact regeneration

    • regenerate the paper figures/tables from the canonical manifests once the frozen private result roots have been synced locally
    • regenerate the appendix-only SDS analyses through the dedicated appendix manifest and helper script

The repo intentionally keeps the main SDS comparison frame separate from the late diagnostic ablations so the virtual-best-solver denominator for the headline figures remains stable.

Local vs cluster validation

Most of the public-release surface can be validated on a MacBook:

  • checked-in bundle presence
  • manifest and inventory integrity
  • shell syntax
  • SDS / BigCode / open-r1 tests
  • appendix/supporting-evidence validation

The remaining cluster-only checks are:

  • sbatch / srun launcher smoke tests
  • EDF environment resolution
  • Capstor checkpoint and dataset path assumptions
  • any retraining or reevaluation that depends on GH200 resources or private frozen roots

Use ./scripts/validate_paper_release.sh for the local portion first, then use docs/REPRODUCTION.md for the remaining cluster-only checks.

If you want the validator itself to exercise the full main-paper regeneration path, run:

./scripts/validate_paper_release.sh --run-main-regen

Repository structure

llm-finetuning/
├── evaluation/                    # SDS + BigCode evaluation and aggregation
├── analysis/feasibility_sparsity/ # Checked-in feasibility-density summaries
├── experiments/report_sets/       # Canonical public manifests
├── docs/                          # Release manifest, reproduction guide, technical reports
├── scripts/                       # Training/evaluation launchers and reproducibility helpers
├── deps/                          # Pinned companion dependency trees
└── tests/                         # Top-level validation tests

Default documentation path

If you are trying to reproduce the paper, start here:

  1. docs/release_manifest.md
  2. docs/REPRODUCTION.md
  3. docs/technical-reports/README.md

Internal review-response archives are intentionally omitted from this standalone code release.