IDEALLab/Neural-Solver-Synthesis-Final-Evidence-v1

Neural Solver Synthesis Final Evidence v1

This repository is the immutable large-artifact companion to the public code release for "Beyond Inference-Time Search: Reinforcement Learning Synthesizes Reusable Solvers."

Public code anchor: `07a798e7d7eca736cd1ef13a15209d402d401ef6`

Current public release: Neural Solver Synthesis evidence snapshot v1.0.2

Interactive experiment companion: Neural Solver Synthesis

Aggregate evidence snapshot: Final Evidence

Public policy checkpoints: Neural Solver Synthesis collection

The v1.0.2 patch updates public documentation pointers only. The checksummed scientific evidence in this dataset is unchanged.

Contents

• —certified full and content-clean SDS summaries;
• —duplicate-safe Base Best-of-64 selection audit;
• —same-model and hosted adaptive-repair controls;
• —end-to-end cost and break-even accounting;
• —input-disjoint universal search;
• —inference-time and training-time Hypothesize sensitivity;
• —compile-once JSSP summaries;
• —bounded TSP RL-training results, corrected trial rows, and selected programs.

The human-readable map is docs/EVIDENCE_MAP.md. Machine-readable claims, uncertainty conventions, sample counts, and file checksums are in docs/final_evidence_index.json and artifacts/neurips2026/checksums.sha256.

Verification

After downloading the repository, run the validator from the matching public Git release with this repository's artifacts/neurips2026 directory in place:

python scripts/validate_neurips2026_public_evidence.py

Limitations and correction history

• —The same-model adaptive-repair result evaluates one 64-completion execution-feedback controller and does not exhaust non-RL search.
• —The hosted medium-reasoning control is not token-, dollar-, latency-, or training-compute-matched to the smaller open model. Its final allocation followed an earlier truncated attempt.
• —JSSP is a within-family deployment test using JSSP-trained policies.
• —TSP is direct-from-base RL training without an SFT stage. Its predeclared quality/stability gate failed, and native 2-opt and OR-Tools remained stronger.
• —The TSP parser correction was applied after outcomes were observed, so TSP is boundary evidence rather than blind confirmation.
• —The Qwen same-model adaptive-repair selected programs are identified by frozen SHA-256 values; their historical source files were unavailable in the compact publication workspace and were not reconstructed from test outcomes.

The evidence does not establish that RL is uniquely necessary, that transfer is tuning-free, or that generated solvers dominate native solvers.

Licensing

See docs/LICENSING.md. This evidence repository does not add a blanket license to code, models, datasets, or standard benchmark instances.