## Bar Charts: Decoupled Fine-Tuning vs Evolutionary Reasoning RL
### Overview
The image contains two side-by-side bar charts comparing performance scores under two pre-training conditions: "Pre-training w/o CoT Data" (blue) and "Pre-training w/ CoT Data" (orange). The left chart focuses on "Decoupled Fine-Tuning" across multiple tasks, while the right chart compares "Evolutionary Reasoning RL" across two years of AIME benchmarks.
### Components/Axes
**Left Chart (Decoupled Fine-Tuning):**
- **X-axis (Categories):** gpqa, aider, LCBench, aime2024, aime2025, minerva math, bbh, bbeh
- **Y-axis (Score):** 0–80 (linear scale)
- **Legend:**
- Blue: Pre-training w/o CoT Data
- Orange: Pre-training w/ CoT Data
**Right Chart (Evolutionary Reasoning RL):**
- **X-axis (Years):** AIME 2024, AIME 2025
- **Y-axis (Score):** 0–45 (linear scale)
- **Legend:** Same as left chart (blue/orange for pre-training conditions)
### Detailed Analysis
**Left Chart Values:**
- **gpqa:** Blue 34.5 | Orange 41.9
- **aider:** Blue 33.8 | Orange 38.4
- **LCBench:** Blue 59.6 | Orange 60.7
- **aime2024:** Blue 14.6 | Orange 16.0
- **aime2025:** Blue 11.6 | Orange 13.3
- **minerva math:** Blue 60.3 | Orange 62.1
- **bbh:** Blue 72.8 | Orange 78.3
- **bbeh:** Blue 8.5 | Orange 9.4
**Right Chart Values:**
- **AIME 2024:** Blue 19.7 | Orange 26.5 (+6.8% increase)
- **AIME 2025:** Blue 14.2 | Orange 20.2 (+6.0% increase)
### Key Observations
1. **Consistent CoT Advantage:** Across all tasks in the left chart, pre-training with CoT data (orange) outperforms the baseline (blue), with the largest gap in `bbh` (78.3 vs 72.8).
2. **Task-Specific Variability:**
- `aime2025` and `bbeh` show minimal improvement with CoT (1.7 and 0.9 point gains, respectively).
- `LCBench` and `minerva math` demonstrate near-par performance between methods (0.4–1.8 point differences).
3. **Yearly Decline in AIME Scores:** Both methods show reduced scores from 2024 to 2025, but CoT maintains a consistent 6–7 point advantage.
4. **Percentage Annotations:** The right chart explicitly highlights relative improvements (+6.8% and +6.0%) for CoT pre-training between years.
### Interpretation
The data suggests that Chain-of-Thought (CoT) pre-training consistently enhances performance across diverse reasoning tasks, with particularly strong effects in complex benchmarks like `bbh`. However, the diminishing returns in `aime2025` and `bbeh` imply potential task-specific limitations. The AIME results reveal a concerning trend: while CoT pre-training remains superior, both methods experience score declines from 2024 to 2025, raising questions about dataset evolution or model generalization. The near-identical improvement percentages (+6.8% vs +6.0%) between years suggest CoT's benefits are stable but not transformative in later benchmarks.
