BRlkl/xl-1024-reasoning-1

Samantha Recursive Latent Reasoning Validation

Answer-only supervised validation of silent latent recurrence with adaptive stopping.

• Base model: google/t5gemma-l-l-prefixlm-it
• Dataset mixture: BRlkl/samantha-r01-recursive-reasoning-corpus
• Latent slots: 1024
• Maximum training depth: 16
• Credit assignment: hrm_stable_credit
• Stop objective: hazard
• Silent ticks use a zero attention mask and no explicit clock features.

Validation metrics

• validation/active_train_depth: 8.000000
• validation/adaptive_0.4_accuracy: 0.250000
• validation/adaptive_0.4_forced_max_rate: 0.000000
• validation/adaptive_0.4_mean_ticks: 1.250000
• validation/adaptive_0.5_accuracy: 0.250000
• validation/adaptive_0.5_forced_max_rate: 0.000000
• validation/adaptive_0.5_mean_ticks: 1.312500
• validation/adaptive_0.6_accuracy: 0.125000
• validation/adaptive_0.6_forced_max_rate: 0.250000
• validation/adaptive_0.6_mean_ticks: 9.125000
• validation/adaptive_0.7_accuracy: 0.062500
• validation/adaptive_0.7_forced_max_rate: 0.875000
• validation/adaptive_0.7_mean_ticks: 28.500000
• validation/adaptive_0.8_accuracy: 0.062500
• validation/adaptive_0.8_forced_max_rate: 0.875000
• validation/adaptive_0.8_mean_ticks: 28.625000
• validation/evaluation_max_depth: 32.000000
• validation/example_count: 16.000000
• validation/fixed_depth_16_accuracy: 0.000000
• validation/fixed_depth_16_nll: 5.377218
• validation/fixed_depth_1_accuracy: 0.250000
• validation/fixed_depth_1_nll: 0.408200
• validation/fixed_depth_2_accuracy: 0.250000
• validation/fixed_depth_2_nll: 0.412210
• validation/fixed_depth_32_accuracy: 0.000000
• validation/fixed_depth_32_nll: 8.305442
• validation/fixed_depth_4_accuracy: 0.250000
• validation/fixed_depth_4_nll: 0.419760
• validation/fixed_depth_8_accuracy: 0.000000
• validation/fixed_depth_8_nll: 1.681179
• validation/fixed_depth_gain: -0.250000