synthetic-disentangled-Mar19 Results

Collection State

  • planned jobs: 63
  • completed jobs: 62
  • missing job: sorting_algorithms + beta0p1_base

Across the collection, controlled_test_final/mean_overall_final_answer_accuracy stays in a narrow 0.9853 to 1.0 range, while controlled_test_final/alignment_one_to_one ranges from 0.1624 to 0.6930.

Best Results By Task

grid_pathfinding

  • best setting: beta0p1_twr2
  • alignment_one_to_one = 0.6930
  • local_strategy_coverage_mean = 1.0000
  • local_strategy_full_coverage_rate = 1.0000
  • mean_overall_final_answer_accuracy = 1.0000

Representative non-token-weighted comparison:

  • beta0p5_base
  • alignment_one_to_one = 0.224233
  • local_strategy_coverage_mean = 0.442500
  • mean_overall_final_answer_accuracy = 1.0000

sorting_algorithms

  • best setting: beta0p1_twr2
  • alignment_one_to_one = 0.270660
  • local_strategy_coverage_mean = 0.397800
  • mean_overall_final_answer_accuracy = 0.993040

Representative non-token-weighted comparison:

  • beta0p1_pe0p5_rmk0p5_rsw0p1
  • alignment_one_to_one = 0.175460
  • local_strategy_coverage_mean = 0.223340
  • mean_overall_final_answer_accuracy = 0.999720

multidigit_addition

  • best setting: beta0p1_rsw1
  • alignment_one_to_one = 0.332367
  • local_strategy_coverage_mean = 0.332400
  • mean_overall_final_answer_accuracy = 0.996600

Comparison with token-weighted reconstruction:

  • beta0p1_twr2
  • alignment_one_to_one = 0.3280
  • local_strategy_coverage_mean = 0.3281
  • mean_overall_final_answer_accuracy = 0.991267

Figures

Best alignment by task family and block. Caption: Token-weighted reconstruction is the only block that changes the best achievable alignment on grid_pathfinding and sorting_algorithms.

Alignment versus token-weighted reconstruction weight. Caption: Alignment varies sharply with token-weighted reconstruction weight compared with the flatter occupancy-only trends seen elsewhere in the aggregate bundle.

Strategy heatmap for the strongest grid_pathfinding token-weighted run. Caption: The standout grid_pathfinding token-weighted run reaches full local strategy coverage in a way the non-token-weighted baselines do not.

Built with LogoFlowershow