From 3a6bab75c03056d8d27d2c34657e757cd8e036a9 Mon Sep 17 00:00:00 2001
From: Guo Cheng <224264187+GuoCheng24@users.noreply.github.com>
Date: Wed, 2 Sep 2026 08:49:57 +0800
Subject: [PATCH] AutoTPPR: minimise Test Top 20 DE MSE, do not maximise it

tasks/metric_config.json marks Gears/"Test Top 20 DE MSE" as maximize=true.
The value comes from gears_model.test_metrics['mse_de'] (experiment.py:1409),
a mean squared error, and the flag drives both node selection and reward in
mcts_experiments_utils_claude.py, so a worse test MSE is scored as progress.
The sibling error metrics in the same file (IEEE39 rmse, AutoMolecule3D MAE,
AutoScale RMSE) all use false, and the module's own fallback heuristic would
also infer false from the metric name.
---
 tasks/metric_config.json | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tasks/metric_config.json b/tasks/metric_config.json
index 7ac3d9e..0008495 100644
--- a/tasks/metric_config.json
+++ b/tasks/metric_config.json
@@ -25,7 +25,7 @@
       "mIoU": true
     },
     "Gears": {
-      "Test Top 20 DE MSE": true
+      "Test Top 20 DE MSE": false
     },
     "Geometry QA":{
       "Accuracy": true