拓展阅读 / NanoJev：小型决策模型

知识库收录材料 · 原始文件 ↗ · 来源、版本与许可

Development results: fixed maze and Snake pilot

The primary pipeline combines atomic judgments + code planning: focused model propositions, exploration memory, and code that composes actions. This page preserves the separate whole-map direct-action stress test. See atomic training and model-guided exploration for the implemented local pipeline. Planner success and event calibration have separate measurements.

All six cases were selected before the new rollouts, with a shared 128-step limit and identical initial states. See the frozen protocol, machine-readable audit, and environment/runbook. This is a bounded integration pilot, not a full-size completion benchmark or a population estimate.

Systems and roles

Initial NanoJev is the earlier trained checkpoint, not raw untrained Qwen. The two game-training arms begin from that same checkpoint. Reference-code wins are due to its explicit algorithm; they are not credited to a learned judgment model. Uniform random samples; the other five systems use greedy actions. This pilot does not compare both controllers for every system.

Maze: success / steps

A success is reaching the goal. Every failure remains in the table. The three exact shortest-path lengths are 16, 24, and 96, so all three are feasible within 128 moves.

Snake: food score / steps / outcome

Food score counts food eaten. Surviving to the time limit is distinct from filling the board, and is not a win. wall means wall collision; body means self-collision; limit means the snake remained alive at step 128.

The learned direct-action models and Jev do not solve any of these three mazes within the shared 128-step limit. Local Snake behavior varies substantially, and no system fills a board. These results motivated the separate local-question training and code-exploration pipeline linked above.

Independent question and probability measurements

The probability report and RLCD experiment measure fixed question outputs separately from game trajectories. Atomic clear_*/safe_* truths, planning truth questions, action preferences, and stochastic events have different targets. Better one-step probability scores do not establish long-horizon game success.

The aggregate atomic accuracies of these full-map game-training arms do not exceed the fixed always-true control. The OOD Snake subset contains only true safety labels. The subsequent local-maze experiment uses matching local training inputs and reports true/false cases separately from these full-map runs.

The small JSON audit retains confusion counts, separate planning-truth and action-support diagnostics; they are not pooled into atomic accuracy. Maze optimal-action mass and Snake local reference-action mass are separate metrics.

Verification and reproduction

The local audit replays 36 episodes / 3201 transitions, checks every saved before/after hash, final outcome, offered probability support, greedy choice, cohort identity, and horizon. Source-file hashes are retained. This does not rerun model inference, verify private API receipts, or reconstruct random draws absent from the compact artifacts.

python scripts/summarize_development_results.py

This command reads the frozen compact reports and rebuilds the small JSON audit and this document without API/GPU calls. Keep all six source reports and the preselected episode file unchanged.

快照未包含部分工程附件或配图；请通过上游链接或原始材料查看，收录范围见来源清单。