Add GenSelect plugin for solution selection

[codelion]

Introduces the GenSelect plugin implementing generative solution selection based on the AIMO-2 winning solution paper. Updates README with plugin documentation and reference, bumps version to 0.1.23, and adds a GenSelect math test case.

[codelion]

Test-Time Compute Evaluation Results

Model Comparison

Model	Approach	Overall Accuracy	GSM8K	MMLU_MATH	BOOLQ	AQUA_RAT	Avg Time (s)
Gemini Flash	avg@5	60.80%	72.14%	64.55%	96.52%	15.56%	0.89
	pass@5	70.00% (+15.1%)	75.00%	68.18%	95.65%	44.44%	0.79
	maj@5	68.00% (+11.8%)	78.57%	68.18%	100.00%	29.63%	0.84
	genselect@5	67.00% (+10.2%)	82.14%	72.73%	95.65%	22.22%	3.20
Qwen3-0.6B-8bit	avg@5	51.49%	60.71%	50.91%	86.09%	14.81%	5.84
	pass@5	59.59% (+15.7%)	67.86%	69.09%	86.96%	22.22%	5.64
	maj@5	53.00% (+2.9%)	67.86%	54.55%	86.96%	3.70%	5.62
	genselect@5	55.00% (+6.8%)	64.29%	58.18%	86.96%	11.11%	18.78
Qwen3-0.6B-4bit	avg@5	39.20%	64.29%	29.09%	62.61%	1.48%	334.55
	pass@5	54.00% (+37.8%)	75.00%	59.09%	82.61%	3.70%	30.92
	maj@5	42.00% (+7.1%)	71.43%	31.82%	60.87%	3.70%	30.55
	genselect@5	41.00% (+4.6%)	60.71%	27.27%	69.57%	7.41%	68.27

Key Insights

Model Performance

• Gemini Flash: Best overall performance (60.80% baseline), excellent on BOOLQ (96-100%)
• Qwen3-0.6B-8bit: Moderate performance (51.49% baseline), consistent across categories
• Qwen3-0.6B-4bit: Significant degradation (39.20% baseline), severe timing issues

Test-Time Compute Benefits

• High-quality models (Gemini Flash): Moderate gains (10-15%)
• Medium-quality models (Qwen 8-bit): Similar gains (3-16%)
• Low-quality models (Qwen 4-bit): Dramatic gains (up to 38%) due to high variance

Category-Specific Patterns

• GSM8K: GenSelect excels with larger models, Pass@5 best for quantized
• MMLU_MATH: Pass@5 consistently strong across all models
• BOOLQ: Majority voting achieves perfect accuracy on Gemini Flash
• AQUA_RAT: Pass@5 dramatically improves results (3x on Gemini Flash)

Timing Considerations

• Gemini Flash: Sub-second for most approaches
• Qwen 8-bit: 5-19 seconds, reasonable for local model
• Qwen 4-bit: Extreme delays on avg@5 (334s), suggesting implementation issues

Response Variance Indicator (pass@5 - avg@5 gap)

• Gemini Flash: 15.1% (low variance, high quality)
• Qwen 8-bit: 15.7% (low variance, moderate quality)
• Qwen 4-bit: 37.8% (high variance, low quality)

Recommendations

1. For production: Use high-quality models with maj@5 for balanced accuracy/speed
2. For best accuracy: Use genselect@5 with sufficient compute budget
3. For uncertain domains: Use pass@5 to maximize coverage
4. For quantized models: Pass@5 provides significant accuracy recovery