## Bar Charts: Comparison of Training Methods for Final-round Accuracy
### Overview
The image displays three grouped bar charts (labeled a, b, and c) comparing the performance of different machine learning training approaches. The primary metric is "Final-round Accuracy (%)" on the y-axis. Each chart explores a different dimension of the training process: objectives, fine-tuning methods, and training signals. A shared legend at the top defines the bar patterns and reference lines.
### Components/Axes
* **Shared Y-Axis:** "Final-round Accuracy (%)", scale from 0 to 100 in increments of 20.
* **Shared Legend (Top Center):**
* **Bar Patterns:**
* `Direct`: Diagonal striped pattern (blue, orange, green).
* `Beliefs`: Solid fill (blue, orange, green).
* **Reference Lines:**
* `Bayesian Assistant`: Dashed light brown line.
* `Random`: Dashed dark gray line.
* **Subplot Titles & X-Axis Categories:**
* **a. Training Objectives:** Categories are "SFT" and "DPO".
* **b. Fine-tuning Methods:** Categories are "Full" and "LoRA".
* **c. Training Signals:** Categories are "Interaction", "Preferences", and "Both".
### Detailed Analysis
**Subplot a. Training Objectives**
* **SFT Category:**
* `Direct` (Blue, striped): ~76%
* `Beliefs` (Blue, solid): ~72%
* **DPO Category:**
* `Direct` (Orange, striped): ~66% (with a small error bar)
* `Beliefs` (Orange, solid): ~70%
* **Reference Lines:**
* `Bayesian Assistant` (Dashed line): Positioned at approximately 80%.
* `Random` (Dashed line): Positioned at approximately 30%.
**Subplot b. Fine-tuning Methods**
* **Full Category:**
* `Direct` (Blue, striped): ~76%
* `Beliefs` (Blue, solid): ~72%
* **LoRA Category:**
* `Direct` (Orange, striped): ~70%
* `Beliefs` (Orange, solid): ~68%
* **Reference Lines:** Same as in subplot a (Bayesian Assistant ~80%, Random ~30%).
**Subplot c. Training Signals**
* **Interaction Category:**
* `Direct` (Blue, striped): ~76%
* `Beliefs` (Blue, solid): ~72%
* **Preferences Category:**
* `Direct` (Green, striped): ~55%
* `Beliefs` (Green, solid): ~79% (with a small error bar)
* **Both Category:**
* `Direct` (Orange, striped): ~78%
* `Beliefs` (Orange, solid): ~79%
* **Reference Lines:** Same as in subplots a and b.
### Key Observations
1. **Performance vs. Baselines:** All reported model performances (bars) are significantly above the `Random` baseline (~30%) but generally below the `Bayesian Assistant` baseline (~80%).
2. **Direct vs. Beliefs Trend:** The relationship between `Direct` and `Beliefs` methods is not consistent.
* In **SFT, Full, and Interaction** settings, `Direct` outperforms `Beliefs` by 4-5 percentage points.
* In **DPO and LoRA** settings, the gap narrows or reverses, with `Beliefs` performing slightly better or comparably.
* In the **Preferences** signal setting, there is a dramatic reversal: `Beliefs` (~79%) vastly outperforms `Direct` (~55%).
3. **Highest Performers:** The highest accuracy values (~79%) are achieved by the `Beliefs` method using either the `Preferences` signal alone or the `Both` signal combination.
4. **Lowest Performer:** The `Direct` method with the `Preferences` signal is the clear outlier, performing at ~55%, which is notably lower than all other configurations.
### Interpretation
This data suggests that the optimal training strategy is highly dependent on the specific context (objective, fine-tuning method, and signal type). There is no universally superior approach between `Direct` and `Beliefs`.
* The `Direct` method appears more effective when the training signal is based on **Interaction** or when using standard **SFT** objectives and **Full** fine-tuning.
* The `Beliefs` method shows a critical advantage when the training signal is derived from **Preferences**. This indicates that modeling or incorporating "beliefs" may be particularly beneficial for learning from preference-based data, potentially by better capturing the underlying rationale behind human choices.
* The fact that combining signals (`Both`) yields high performance for both methods suggests complementarity between interaction and preference data.
* The consistent gap below the `Bayesian Assistant` baseline indicates that these training methods, while effective, have not yet reached the theoretical performance ceiling represented by that benchmark. The `Random` baseline confirms the tasks are non-trivial and the models are learning meaningful patterns.
**Language Declaration:** All text in the image is in English.
