original_paper/paper_statistics.md

Paper Statistics

Results Directory: results_temp_1.1

1. QC Pass Rates

• Base: 5/100 (5.0%)
• SFT: 10/100 (10.0%)
• RL: 77/100 (77.0%)

2. Held-Out Continuation Task (Log-Probability)

Model	Mean Log-Prob	Std	N
Base	-12.4492	6.1440	85
SFT	-12.4492	6.1440	85
RL	-10.9660	2.7415	85

Paired T-Tests (Completion)

• Base vs SFT: t=3.945, p=1.65e-04 *** (n=85 pairs)
• Base vs RL: t=-2.474, p=1.54e-02 * (n=85 pairs)
• SFT vs RL: t=-2.474, p=1.54e-02 * (n=85 pairs)

Alignment Tax Analysis

• Base vs RL (paired t-test)
  • N pairs: 85
  • Base mean: -12.4492
  • RL mean: -10.9660
  • Difference (RL - Base): +1.4831
  • t-statistic: -2.4741
  • p-value: 1.54e-02 *
  • Cohen's d: 0.268 (small effect)
  • Interpretation: No alignment tax (RL performs better)

3. Surprisal Benchmark

Model	Mean Log-Prob	Std	N
Base	-13.8384	2.5110	28
SFT	-13.8384	2.5110	28
RL	-11.6637	1.8210	28

Paired T-Tests (Surprisal)

• Base vs SFT: t=7.408, p=5.72e-08 *** (n=28 pairs)
• Base vs RL: t=-5.128, p=2.16e-05 *** (n=28 pairs)
• SFT vs RL: t=-5.128, p=2.16e-05 *** (n=28 pairs)

4. Distribution Metrics

3-mer KL Divergence from Real Plasmids

| Model | KL(Model||Real) | Mean JS Divergence | |-------|-----------------|--------------------| | Base | 0.0142 | 0.1037 | | SFT | 0.0084 | 0.1074 | | RL | 0.0106 | 0.0866 |

GC Content

Model	Mean GC	Std
Real	0.5172	0.0291
Base	0.4779	0.0336
SFT	0.4953	0.0501
RL	0.5185	0.0293

Sequence Length

Model	Mean Length	Std	Median
Real	7273	2924	6690
Base	7482	3829	7168
SFT	5812	4202	4409
RL	7107	1489	6668

MFE Density (Thermodynamic Stability)

Model	Mean MFE/nt	Std
Real	-0.3643	0.0267
Base	-0.3597	0.0652
SFT	-0.3577	0.0728
RL	-0.3622	0.0338

5. Diversity Metrics

Model	Pass Rate (%)	Diversity Score
Base	5.0	0.9150
SFT	10.0	0.8964
GRPO	77.0	0.5878

6. Sample Counts

Category	Count
Real	254
Base	53
SFT	100
RL	100
Total	507