contrastive-probing-mac / summary_plot.py
ch-min's picture
Upload summary_plot.py with huggingface_hub
cb9715e verified
import os
import argparse
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from matplotlib.lines import Line2D
import warnings
# 경고 메시지 숨김
warnings.simplefilter("ignore", UserWarning)
def main():
# 1. Argument parsing
parser = argparse.ArgumentParser(description='Generate Spatial Representation Figures')
parser.add_argument('--data_path', type=str, required=True, help='Path to the input CSV file')
args = parser.parse_args()
# 2. Load Data
df = pd.read_csv(args.data_path)
# 3. Preprocessing
rename_dict = {
'EmbSpatial (con)': 'EmbSpatial Acc (con)',
'EmbSpatial (ctr)': 'EmbSpatial Acc (ctr)',
'CVBench3D (con)': 'CVBench3D Acc (con)',
'CVBench3D (ctr)': 'CVBench3D Acc (ctr)',
'Layer Horiz SC Layer': 'Layer Horiz SC',
'Layer Vert SC Layer': 'Layer Vert SC',
'Layer Dist SC Layer': 'Layer Dist SC',
'Entanglement': 'VD-Entanglement'
}
df.rename(columns=rename_dict, inplace=True)
cols_to_convert = ['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)', 'CVBench3D Acc (con)', 'CVBench3D Acc (ctr)']
for col in cols_to_convert:
if col in df.columns:
df[col] = df[col].astype(str).str.replace('%', '', regex=False).replace('N/A', np.nan).astype(float)
# Add group and scale labels
def get_base(name):
name_upper = str(name).upper()
if 'MOLMO' in name_upper: return 'Molmo'
if 'NVILA-ST' in name_upper or 'NVILA ST' in name_upper: return 'NVILA-ST'
if 'ROBOREFER' in name_upper: return 'RoboRefer'
if 'NVILA' in name_upper: return 'NVILA'
if 'QWEN3' in name_upper: return 'Qwen3'
if 'QWEN' in name_upper: return 'Qwen'
return 'Other'
def get_scale(name):
name_lower = str(name).lower()
if 'vanilla' in name_lower: return 'vanilla'
if '80k' in name_lower: return '80k'
if '400k' in name_lower: return '400k'
if '800k' in name_lower: return '800k'
if '2m' in name_lower: return '2M'
return 'Special'
df['Base'] = df['Model'].apply(get_base)
df['Scale'] = df['Model'].apply(get_scale)
# 4. Filter out Qwen3
df = df[df['Base'] != 'Qwen3']
# 5. Output directory & Settings
output_dir = "evaluation_figures_final"
os.makedirs(output_dir, exist_ok=True)
sns.set_theme(style="whitegrid")
# Colors & Orders
colors = {
'Molmo': '#1f77b4',
'NVILA': '#ff7f0e',
'NVILA-ST': '#cc5500',
'Qwen': '#2ca02c',
'RoboRefer': '#d62728'
}
scale_order = ['vanilla', '80k', '400k', '800k', '2M']
st_scale_order = ['80k', '400k', '800k']
custom_lines_global = [
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Molmo'], markersize=10, label='Molmo'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA'], markersize=10, label='NVILA-Lite-2B'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA-ST'], markeredgecolor='black', markersize=11, label='NVILA-ST'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Qwen'], markersize=10, label='Qwen'),
Line2D([0], [0], marker='*', color='w', markerfacecolor=colors['RoboRefer'], markeredgecolor='black', markersize=18, label='RoboRefer-2B-SFT'),
]
# 공통 사용 RoboRefer 데이터프레임
df_robo = df[df['Base'] == 'RoboRefer']
# -----------------------------------------------------------------------------
# Figure 1: The Two Paths (Scatter with Trajectories and Labels)
# -----------------------------------------------------------------------------
plt.figure(figsize=(12, 9))
df_bases = df[df['Base'] != 'RoboRefer']
sns.scatterplot(data=df_bases, x='VD-Entanglement', y='Layer Dist SC', hue='Base', palette=colors, s=100, alpha=0.8, edgecolor='black', legend=False)
for i, row in df_bases.iterrows():
model_name = str(row['Model'])
is_st = row['Base'] == 'NVILA-ST'
is_pct = 'pct' in model_name.lower()
if is_st or is_pct:
plt.annotate(model_name, (row['VD-Entanglement'], row['Layer Dist SC']),
xytext=(6, 6), textcoords='offset points',
fontsize=9, color=colors[row['Base']], fontweight='bold', alpha=0.9)
for base in ['Molmo', 'NVILA', 'Qwen', 'NVILA-ST']:
current_order = st_scale_order if base == 'NVILA-ST' else scale_order
base_df = df[(df['Base'] == base) & (df['Scale'].isin(current_order))]
base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(current_order).dropna(subset=['VD-Entanglement', 'Layer Dist SC'])
if len(base_df) > 1:
for i in range(len(base_df)-1):
x1, y1 = base_df.iloc[i]['VD-Entanglement'], base_df.iloc[i]['Layer Dist SC']
x2, y2 = base_df.iloc[i+1]['VD-Entanglement'], base_df.iloc[i+1]['Layer Dist SC']
ls = '--' if base == 'NVILA-ST' else '-'
plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
arrowprops=dict(arrowstyle="->", color=colors[base], alpha=0.6, lw=1.8, ls=ls))
if not df_robo.empty:
sns.scatterplot(data=df_robo, x='VD-Entanglement', y='Layer Dist SC', hue='Base',
palette=colors, s=400, marker='*', edgecolor='black', zorder=5, legend=False)
for i, row in df_robo.iterrows():
plt.annotate('RoboRefer-2B-SFT', (row['VD-Entanglement'], row['Layer Dist SC']),
xytext=(10, -5), textcoords='offset points',
fontsize=11, color=colors['RoboRefer'], fontweight='bold')
plt.legend(handles=custom_lines_global, title='Model Group', loc='upper right')
plt.title('Figure 1: The Two Paths of Spatial Representation\n(Naive Scaling vs. Genuine Understanding)', fontsize=14, fontweight='bold')
plt.xlabel('VD-Entanglement Shortcut [Lower is Better]', fontsize=12)
plt.ylabel('Distance Representation (Layer Dist SC) [Higher is Better]', fontsize=12)
plt.text(0.6, 0.05, "Mechanism 1:\nNaive Scaling\n(More Data = More Entanglement)",
fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
plt.text(0.1, 0.20, "Mechanism 2:\nGenuine Understanding\n(Break the Shortcut)",
fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure1_two_paths_trajectory.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 1-1: NVILA and RoboRefer Only
# -----------------------------------------------------------------------------
plt.figure(figsize=(10, 8))
df_nvila_fig1_1 = df[(df['Base'] == 'NVILA') & (df['Scale'].isin(scale_order))]
sns.scatterplot(data=df_nvila_fig1_1, x='VD-Entanglement', y='Layer Dist SC', color=colors['NVILA'],
s=150, alpha=0.9, edgecolor='black', zorder=4, legend=False)
base_df = df_nvila_fig1_1.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(scale_order).dropna(subset=['VD-Entanglement', 'Layer Dist SC'])
if len(base_df) > 1:
for i in range(len(base_df)-1):
x1, y1 = base_df.iloc[i]['VD-Entanglement'], base_df.iloc[i]['Layer Dist SC']
x2, y2 = base_df.iloc[i+1]['VD-Entanglement'], base_df.iloc[i+1]['Layer Dist SC']
plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
arrowprops=dict(arrowstyle="-|>", color=colors['NVILA'], alpha=0.7, lw=2.5, mutation_scale=15))
for scale_val, row in base_df.iterrows():
# if scale_val == 'vanilla':
# plt.annotate('base model', (row['VD-Entanglement'], row['Layer Dist SC']),
# xytext=(10, 10), textcoords='offset points',
# fontsize=15, color=colors['NVILA'], fontweight='bold', alpha=1.0)
if scale_val in ['80k', '400k', '800k', '2M']:
plt.annotate(scale_val, (row['VD-Entanglement'], row['Layer Dist SC']),
xytext=(10, 10), textcoords='offset points',
fontsize=15, color=colors['NVILA'], fontweight='bold', alpha=1.0)
if not df_robo.empty:
sns.scatterplot(data=df_robo, x='VD-Entanglement', y='Layer Dist SC', color=colors['RoboRefer'],
s=200, marker='o', edgecolor='black', zorder=5, legend=False)
# for i, row in df_robo.iterrows():
# plt.annotate('RoboRefer-2B-SFT', (row['VD-Entanglement'], row['Layer Dist SC']),
# xytext=(10, 10), textcoords='offset points',
# fontsize=15, color=colors['RoboRefer'], fontweight='bold', alpha=1.0)
custom_lines_fig1_1 = [
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA'], markeredgecolor='black', markersize=12, label='NVILA-Lite-2B'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['RoboRefer'], markeredgecolor='black', markersize=12, label='RoboRefer-2B-SFT')
]
plt.legend(handles=custom_lines_fig1_1, title='Model Group', loc='upper right', fontsize=14, title_fontsize=16)
plt.xlabel('VD-Entanglement Index (VD-EI)', fontsize=16, fontweight='bold')
plt.ylabel(r'Distance Coherence ($\mathrm{Coh}_{\mathrm{D}}$)', fontsize=16, fontweight='bold')
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure1_1_nvila_roborefer_only.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 1-2: NVILA and RoboRefer Only (1x3 Subplots: Horiz, Vert, Dist)
# -----------------------------------------------------------------------------
fig, axes = plt.subplots(1, 3, figsize=(24, 8))
coherence_metrics = [
('Layer Horiz SC', 'Horizontal Coherence'),
('Layer Vert SC', 'Vertical Coherence'),
('Layer Dist SC', r'Distance Coherence ($\mathrm{Coh}_{\mathrm{D}}$)')
]
for ax, (col, ylabel) in zip(axes, coherence_metrics):
if col not in df.columns:
continue
df_nvila_fig1_2 = df[(df['Base'] == 'NVILA') & (df['Scale'].isin(scale_order))]
sns.scatterplot(data=df_nvila_fig1_2, x='VD-Entanglement', y=col, color=colors['NVILA'],
s=150, alpha=0.9, edgecolor='black', zorder=4, ax=ax, legend=False)
base_df = df_nvila_fig1_2.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(scale_order).dropna(subset=['VD-Entanglement', col])
if len(base_df) > 1:
for i in range(len(base_df)-1):
x1, y1 = base_df.iloc[i]['VD-Entanglement'], base_df.iloc[i][col]
x2, y2 = base_df.iloc[i+1]['VD-Entanglement'], base_df.iloc[i+1][col]
ax.annotate('', xy=(x2, y2), xytext=(x1, y1),
arrowprops=dict(arrowstyle="-|>", color=colors['NVILA'], alpha=0.7, lw=2.5, mutation_scale=15))
for scale_val, row in base_df.iterrows():
if scale_val == 'vanilla':
ax.annotate('NVILA-Lite-2B', (row['VD-Entanglement'], row[col]),
xytext=(10, 10), textcoords='offset points',
fontsize=15, color=colors['NVILA'], fontweight='bold', alpha=1.0)
elif scale_val in ['80k', '400k', '800k', '2M']:
ax.annotate(scale_val, (row['VD-Entanglement'], row[col]),
xytext=(10, 10), textcoords='offset points',
fontsize=15, color=colors['NVILA'], fontweight='bold', alpha=1.0)
if not df_robo.empty and col in df_robo.columns:
sns.scatterplot(data=df_robo, x='VD-Entanglement', y=col, color=colors['RoboRefer'],
s=200, marker='o', edgecolor='black', zorder=5, ax=ax, legend=False)
for i, row in df_robo.iterrows():
ax.annotate('RoboRefer-2B-SFT', (row['VD-Entanglement'], row[col]),
xytext=(10, 10), textcoords='offset points',
fontsize=15, color=colors['RoboRefer'], fontweight='bold', alpha=1.0)
ax.set_xlabel('VD-Entanglement Index (VD-EI)', fontsize=16, fontweight='bold')
ax.set_ylabel(ylabel, fontsize=16, fontweight='bold')
ax.tick_params(axis='both', which='major', labelsize=13)
custom_lines_fig1_2 = [
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA'], markeredgecolor='black', markersize=12, label='NVILA-Lite-2B'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['RoboRefer'], markeredgecolor='black', markersize=12, label='RoboRefer-2B-SFT')
]
plt.tight_layout()
fig.subplots_adjust(bottom=0.2)
fig.legend(handles=custom_lines_fig1_2, title='Model Group', loc='upper center', ncol=2,
fontsize=14, title_fontsize=16, bbox_to_anchor=(0.5, 0.12))
plt.savefig(os.path.join(output_dir, 'figure1_2_nvila_roborefer_coherence_1x3.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 2: Entanglement Paradox (Line Plot)
# -----------------------------------------------------------------------------
df_scale = df[df['Base'] != 'RoboRefer'].copy()
df_scale = df_scale[df_scale['Scale'].isin(scale_order)]
df_scale = df_scale.drop_duplicates(subset=['Base', 'Scale'])
df_scale['Scale'] = pd.Categorical(df_scale['Scale'], categories=scale_order, ordered=True)
plt.figure(figsize=(10, 6))
sns.lineplot(data=df_scale, x='Scale', y='VD-Entanglement', hue='Base', style='Base',
palette=colors, dashes={'Molmo':'', 'NVILA':'', 'Qwen':'', 'NVILA-ST':(4,2)},
marker='o', linewidth=2.5, markersize=8)
if not df_robo.empty:
robo_ent = df_robo['VD-Entanglement'].values[0]
plt.axhline(y=robo_ent, color=colors['RoboRefer'], linestyle='--', label=f'RoboRefer-2B-SFT ({robo_ent:.4f})')
plt.title('Figure 2: The Entanglement Paradox during Naive Scaling', fontsize=14, fontweight='bold')
plt.ylabel('VD-Entanglement Index (VD-EI)', fontsize=12)
plt.xlabel('Fine-tuning Data Scale', fontsize=12)
plt.legend(title='Model Group', bbox_to_anchor=(1.05, 1), loc='upper left')
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure2_entanglement_paradox.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 3: Depth Independence Ratio (Bar Chart)
# -----------------------------------------------------------------------------
if 'Layer Dist SC' in df.columns and 'Layer Vert SC' in df.columns:
df['Depth Independence Ratio'] = df['Layer Dist SC'] / df['Layer Vert SC']
rep_models = ['Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer']
available_rep = [m for m in rep_models if m in df['Model'].values]
df_rep = df.set_index('Model').loc[available_rep]
plt.figure(figsize=(12, 6))
bar_colors = ['#aec7e8', '#1f77b4', '#ffbb78', '#ff7f0e', '#98df8a', '#2ca02c', '#d62728']
ax = df_rep['Depth Independence Ratio'].plot(kind='bar', color=bar_colors[:len(available_rep)], edgecolor='black')
plt.title('Figure 3: Depth Independence Ratio', fontsize=14, fontweight='bold')
plt.ylabel('Ratio', fontsize=12)
plt.xticks(rotation=45, ha='right')
plt.axhline(y=0.2, color='gray', linestyle=':', alpha=0.7)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure3_depth_independence_ratio.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 4: Behavioral Consequence - The Gap (Grouped Bar Chart)
# -----------------------------------------------------------------------------
df_acc = df.dropna(subset=['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)']).copy()
available_acc_models = df_acc['Model'].tolist()
fig_width = max(12, len(available_acc_models) * 0.7)
x = np.arange(len(available_acc_models))
width = 0.35
fig, ax = plt.subplots(figsize=(fig_width, 6))
rects1 = ax.bar(x - width/2, df_acc['EmbSpatial Acc (con)'], width, label='Consistent (Shortcut Works)', color='#2b83ba')
rects2 = ax.bar(x + width/2, df_acc['EmbSpatial Acc (ctr)'], width, label='Counter (Shortcut Fails)', color='#d7191c')
for i in range(len(available_acc_models)):
con_val = df_acc['EmbSpatial Acc (con)'].iloc[i]
ctr_val = df_acc['EmbSpatial Acc (ctr)'].iloc[i]
gap = con_val - ctr_val
ax.plot([i - width/2, i + width/2], [con_val, ctr_val], color='black', linestyle='-', linewidth=1.5, alpha=0.5)
ax.text(i, max(con_val, ctr_val) + 2, f'Gap: {gap:.1f}%p', ha='center', fontsize=9, fontweight='bold')
ax.set_ylabel('Accuracy (%)', fontsize=12)
ax.set_title('Figure 4: The Consistent vs. Counter Performance Gap (All Valid Models)', fontsize=14, fontweight='bold')
ax.set_xticks(x)
display_acc_models = ['RoboRefer-2B-SFT' if m == 'RoboRefer' else m for m in available_acc_models]
ax.set_xticklabels(display_acc_models, rotation=45, ha='right')
ax.legend()
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure4_accuracy_gap.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 5: Distance Coherence vs Counter Accuracy
# -----------------------------------------------------------------------------
custom_lines_fig5 = [
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Molmo'], markeredgecolor='black', markersize=12, label='Molmo-7B-O-0924'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA'], markeredgecolor='black', markersize=12, label='NVILA-Lite-2B'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Qwen'], markeredgecolor='black', markersize=12, label='Qwen2.5-VL-3B-Instruct'),
Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['RoboRefer'], markeredgecolor='black', markersize=12, label='RoboRefer-2B-SFT'),
]
if 'Layer Dist SC' in df.columns and 'EmbSpatial Acc (ctr)' in df.columns:
plt.figure(figsize=(10, 8))
df_valid_acc = df.dropna(subset=['EmbSpatial Acc (ctr)', 'Layer Dist SC'])
# NVILA-ST 제외
df_bases_fig5 = df_valid_acc[(df_valid_acc['Base'] != 'RoboRefer') & (df_valid_acc['Base'] != 'NVILA-ST')]
sns.scatterplot(data=df_bases_fig5, x='Layer Dist SC', y='EmbSpatial Acc (ctr)', hue='Base',
palette=colors, s=200, edgecolor='black', alpha=0.9, legend=False)
# 궤적 그리기
for base in ['Molmo', 'NVILA', 'Qwen']:
base_df = df_valid_acc[(df_valid_acc['Base'] == base) & (df_valid_acc['Scale'].isin(scale_order))]
base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(scale_order).dropna(subset=['Layer Dist SC', 'EmbSpatial Acc (ctr)'])
if len(base_df) > 1:
for i in range(len(base_df)-1):
x1, y1 = base_df.iloc[i]['Layer Dist SC'], base_df.iloc[i]['EmbSpatial Acc (ctr)']
x2, y2 = base_df.iloc[i+1]['Layer Dist SC'], base_df.iloc[i+1]['EmbSpatial Acc (ctr)']
plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
arrowprops=dict(arrowstyle="-|>", color=colors[base], alpha=0.8, lw=2.5, mutation_scale=20))
# 텍스트 라벨링 추가 (80k, 400k, 800k, 2M)
for scale_val, row in base_df.iterrows():
if scale_val in ['80k', '400k', '800k', '2M']:
plt.annotate(scale_val, (row['Layer Dist SC'], row['EmbSpatial Acc (ctr)']),
xytext=(10, 10), textcoords='offset points',
fontsize=15, color=colors[base], fontweight='bold', alpha=1.0)
# RoboRefer를 빨간색 큰 원으로 표시
df_robo_fig5 = df_valid_acc[df_valid_acc['Base'] == 'RoboRefer']
if not df_robo_fig5.empty:
sns.scatterplot(data=df_robo_fig5, x='Layer Dist SC', y='EmbSpatial Acc (ctr)', color=colors['RoboRefer'],
s=250, marker='o', edgecolor='black', zorder=5, legend=False)
plt.legend(handles=custom_lines_fig5, title='Model Group', loc='lower right', fontsize=14, title_fontsize=16)
plt.xlabel(r'Distance Coherence ($\mathrm{Coh}_{\mathrm{D}}$)', fontsize=16, fontweight='bold')
plt.ylabel('Embspatial-Bench (Counter) Accuracy (%)', fontsize=16, fontweight='bold')
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.grid(True, linestyle='--', alpha=0.7)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure5_distSC_vs_counterAcc.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 5-1: Dist/Vert Coherence Ratio vs Counter Accuracy
# -----------------------------------------------------------------------------
if 'Layer Dist SC' in df.columns and 'Layer Vert SC' in df.columns and 'EmbSpatial Acc (ctr)' in df.columns:
df['Dist_Vert_Ratio'] = df['Layer Dist SC'] / df['Layer Vert SC']
plt.figure(figsize=(10, 8))
df_valid_ratio = df.dropna(subset=['EmbSpatial Acc (ctr)', 'Dist_Vert_Ratio'])
# NVILA-ST 제외
df_bases_ratio = df_valid_ratio[(df_valid_ratio['Base'] != 'RoboRefer') & (df_valid_ratio['Base'] != 'NVILA-ST')]
sns.scatterplot(data=df_bases_ratio, x='Dist_Vert_Ratio', y='EmbSpatial Acc (ctr)', hue='Base',
palette=colors, s=200, edgecolor='black', alpha=0.9, legend=False)
# 궤적 그리기
for base in ['Molmo', 'NVILA', 'Qwen']:
base_df = df_valid_ratio[(df_valid_ratio['Base'] == base) & (df_valid_ratio['Scale'].isin(scale_order))]
base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(scale_order).dropna(subset=['Dist_Vert_Ratio', 'EmbSpatial Acc (ctr)'])
if len(base_df) > 1:
for i in range(len(base_df)-1):
x1, y1 = base_df.iloc[i]['Dist_Vert_Ratio'], base_df.iloc[i]['EmbSpatial Acc (ctr)']
x2, y2 = base_df.iloc[i+1]['Dist_Vert_Ratio'], base_df.iloc[i+1]['EmbSpatial Acc (ctr)']
plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
arrowprops=dict(arrowstyle="-|>", color=colors[base], alpha=0.8, lw=2.5, mutation_scale=20))
# 텍스트 라벨링 추가
for scale_val, row in base_df.iterrows():
if scale_val in ['80k', '400k', '800k', '2M']:
plt.annotate(scale_val, (row['Dist_Vert_Ratio'], row['EmbSpatial Acc (ctr)']),
xytext=(10, 10), textcoords='offset points',
fontsize=15, color=colors[base], fontweight='bold', alpha=1.0)
# RoboRefer
df_robo_ratio = df_valid_ratio[df_valid_ratio['Base'] == 'RoboRefer']
if not df_robo_ratio.empty:
sns.scatterplot(data=df_robo_ratio, x='Dist_Vert_Ratio', y='EmbSpatial Acc (ctr)', color=colors['RoboRefer'],
s=250, marker='o', edgecolor='black', zorder=5, legend=False)
plt.legend(handles=custom_lines_fig5, title='Model Group', loc='lower right', fontsize=14, title_fontsize=16)
plt.xlabel(r'Distance Coherence / Vertical Coherence ($\mathrm{Coh}_{\mathrm{D}} / \mathrm{Coh}_{\mathrm{V}}$)', fontsize=16, fontweight='bold')
plt.ylabel('Embspatial-Bench (Counter) Accuracy (%)', fontsize=16, fontweight='bold')
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.grid(True, linestyle='--', alpha=0.7)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure5_1_dist_vert_ratio_vs_counterAcc.png'), dpi=300)
plt.close()
# -----------------------------------------------------------------------------
# Figure 6: 3D Scatter (Layer Dist SC, Counter Acc, VD-Entanglement)
# -----------------------------------------------------------------------------
if 'Layer Dist SC' in df.columns and 'EmbSpatial Acc (ctr)' in df.columns:
fig = plt.figure(figsize=(12, 10))
ax = fig.add_subplot(111, projection='3d')
for base_name in df_valid_acc['Base'].unique():
subset = df_valid_acc[df_valid_acc['Base'] == base_name]
c = colors.get(base_name, 'gray')
if base_name == 'RoboRefer':
ax.scatter(subset['Layer Dist SC'], subset['VD-Entanglement'], subset['EmbSpatial Acc (ctr)'],
c=c, s=300, marker='*', edgecolors='k', alpha=0.9, zorder=5)
else:
ax.scatter(subset['Layer Dist SC'], subset['VD-Entanglement'], subset['EmbSpatial Acc (ctr)'],
c=c, s=120, edgecolors='k', alpha=0.8)
for base in ['Molmo', 'NVILA', 'Qwen', 'NVILA-ST']:
current_order = st_scale_order if base == 'NVILA-ST' else scale_order
base_df = df_valid_acc[(df_valid_acc['Base'] == base) & (df_valid_acc['Scale'].isin(current_order))]
base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(current_order).dropna(subset=['Layer Dist SC', 'VD-Entanglement', 'EmbSpatial Acc (ctr)'])
if len(base_df) > 1:
for i in range(len(base_df)-1):
x1, y1, z1 = base_df.iloc[i]['Layer Dist SC'], base_df.iloc[i]['VD-Entanglement'], base_df.iloc[i]['EmbSpatial Acc (ctr)']
x2, y2, z2 = base_df.iloc[i+1]['Layer Dist SC'], base_df.iloc[i+1]['VD-Entanglement'], base_df.iloc[i+1]['EmbSpatial Acc (ctr)']
ls = '--' if base == 'NVILA-ST' else '-'
ax.plot([x1, x2], [y1, y2], [z1, z2], color=colors[base], linestyle=ls, linewidth=2.5, alpha=0.7)
ax.set_xlabel('Layer Dist SC', labelpad=10, fontsize=11)
ax.set_ylabel('VD-Entanglement Index (VD-EI)', labelpad=10, fontsize=11)
ax.set_zlabel('Counter Accuracy (%)', labelpad=10, fontsize=11)
ax.set_title('Figure 6: Navigating the Representation Space', fontweight='bold', fontsize=14)
ax.view_init(elev=20, azim=135)
plt.legend(handles=custom_lines_global, title='Model Group', loc='upper left', bbox_to_anchor=(1.05, 1))
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'figure6_3d_distSC_ent_counterAcc.png'), dpi=300, bbox_inches='tight')
plt.close()
# -----------------------------------------------------------------------------
# Figure 7: Consistency vs. Norm (Scatter per dimension)
# -----------------------------------------------------------------------------
norm_cols = ['Norm Horiz', 'Norm Vert', 'Norm Dist']
sc_cols = ['Layer Horiz SC', 'Layer Vert SC', 'Layer Dist SC']
if all(c in df.columns for c in norm_cols + sc_cols):
fig, axes = plt.subplots(1, 3, figsize=(18, 6))
dims = [
('Layer Horiz SC', 'Norm Horiz', 'Horizontal Dimension'),
('Layer Vert SC', 'Norm Vert', 'Vertical Dimension'),
('Layer Dist SC', 'Norm Dist', 'Distance Dimension')
]
for ax, (sc_col, norm_col, title) in zip(axes, dims):
df_dim = df.dropna(subset=[sc_col, norm_col])
df_bases_dim = df_dim[df_dim['Base'] != 'RoboRefer']
sns.scatterplot(data=df_bases_dim, x=sc_col, y=norm_col, hue='Base', palette=colors,
s=100, edgecolor='black', alpha=0.8, ax=ax, legend=False)
for base in ['Molmo', 'NVILA', 'Qwen', 'NVILA-ST']:
current_order = st_scale_order if base == 'NVILA-ST' else scale_order
base_df = df_dim[(df_dim['Base'] == base) & (df_dim['Scale'].isin(current_order))]
base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(current_order).dropna(subset=[sc_col, norm_col])
if len(base_df) > 1:
for i in range(len(base_df)-1):
x1, y1 = base_df.iloc[i][sc_col], base_df.iloc[i][norm_col]
x2, y2 = base_df.iloc[i+1][sc_col], base_df.iloc[i+1][norm_col]
ls = '--' if base == 'NVILA-ST' else '-'
ax.annotate('', xy=(x2, y2), xytext=(x1, y1),
arrowprops=dict(arrowstyle="->", color=colors.get(base, 'gray'), alpha=0.6, lw=1.8, ls=ls))
df_robo_dim = df_dim[df_dim['Base'] == 'RoboRefer']
if not df_robo_dim.empty:
sns.scatterplot(data=df_robo_dim, x=sc_col, y=norm_col, hue='Base',
palette=colors, s=400, marker='*', edgecolor='black', zorder=5, ax=ax, legend=False)
for i, row in df_robo_dim.iterrows():
ax.annotate('RoboRefer-2B-SFT', (row[sc_col], row[norm_col]),
xytext=(10, -5), textcoords='offset points',
fontsize=11, color=colors['RoboRefer'], fontweight='bold')
for i, row in df_bases_dim.iterrows():
model_name = str(row['Model'])
if row['Base'] == 'NVILA-ST' or 'pct' in model_name.lower():
ax.annotate(model_name, (row[sc_col], row[norm_col]),
xytext=(6, 6), textcoords='offset points',
fontsize=9, color=colors[row['Base']], fontweight='bold', alpha=0.9)
if len(df_dim) > 1:
corr = np.corrcoef(df_dim[sc_col], df_dim[norm_col])[0, 1]
ax.set_title(f"{title}\n(r={corr:.2f})", fontsize=14, fontweight='bold')
else:
ax.set_title(title, fontsize=14, fontweight='bold')
ax.set_xlabel(f'Consistency ({sc_col})', fontsize=12)
ax.set_ylabel(f'Magnitude ({norm_col})', fontsize=12)
ax.grid(True, linestyle='--', alpha=0.7)
fig.legend(handles=custom_lines_global, title='Model Group', loc='center right', bbox_to_anchor=(0.99, 0.5))
fig.suptitle('Figure 7: Sign-Corrected Consistency vs. Vector Norm', fontsize=18, fontweight='bold', y=1.05)
plt.tight_layout(rect=[0, 0, 0.88, 1])
plt.savefig(os.path.join(output_dir, 'figure7_consistency_vs_norm.png'), dpi=300, bbox_inches='tight')
plt.close()
# -----------------------------------------------------------------------------
# Figure 8: Consistency Trajectory Across Scales (1x3 Line plots)
# -----------------------------------------------------------------------------
if all(c in df.columns for c in sc_cols):
fig, axes = plt.subplots(1, 3, figsize=(18, 6))
dims = [
('Layer Horiz SC', 'Horizontal Consistency'),
('Layer Vert SC', 'Vertical Consistency'),
('Layer Dist SC', 'Distance Consistency')
]
df_scale_fig8 = df[df['Base'] != 'RoboRefer'].copy()
df_scale_fig8 = df_scale_fig8[df_scale_fig8['Scale'].isin(scale_order)]
df_scale_fig8 = df_scale_fig8.drop_duplicates(subset=['Base', 'Scale'])
df_scale_fig8['Scale'] = pd.Categorical(df_scale_fig8['Scale'], categories=scale_order, ordered=True)
for ax, (col, title) in zip(axes, dims):
sns.lineplot(data=df_scale_fig8, x='Scale', y=col, hue='Base', style='Base',
palette=colors, dashes={'Molmo':'', 'NVILA':'', 'Qwen':'', 'NVILA-ST':(4,2)},
marker='o', linewidth=2.5, markersize=8, ax=ax, legend=False)
if not df_robo.empty and col in df_robo.columns:
robo_val = df_robo[col].values[0]
ax.axhline(y=robo_val, color=colors['RoboRefer'], linestyle='--', alpha=0.8, label='RoboRefer-2B-SFT')
ax.set_title(title, fontsize=14, fontweight='bold')
ax.set_xlabel('Fine-tuning Data Scale', fontsize=12)
ax.set_ylabel(col, fontsize=12)
ax.grid(True, linestyle='--', alpha=0.7)
fig.legend(handles=custom_lines_global, title='Model Group', loc='center right', bbox_to_anchor=(0.99, 0.5))
fig.suptitle('Figure 8: Sign-Corrected Consistency Trajectory Across Scales', fontsize=18, fontweight='bold', y=1.05)
plt.tight_layout(rect=[0, 0, 0.88, 1])
plt.savefig(os.path.join(output_dir, 'figure8_consistency_trajectory.png'), dpi=300, bbox_inches='tight')
plt.close()
print(f"Success! Evaluation figures have been saved to ./{output_dir}/")
if __name__ == "__main__":
main()
# import os
# import argparse
# import pandas as pd
# import numpy as np
# import matplotlib.pyplot as plt
# import seaborn as sns
# from matplotlib.lines import Line2D
# import warnings
# # 경고 메시지 숨김
# warnings.simplefilter("ignore", UserWarning)
# def main():
# # 1. Argument parsing
# parser = argparse.ArgumentParser(description='Generate Spatial Representation Figures')
# parser.add_argument('--data_path', type=str, required=True, help='Path to the input CSV file')
# args = parser.parse_args()
# # 2. Load Data
# df = pd.read_csv(args.data_path)
# # 3. Preprocessing
# rename_dict = {
# 'EmbSpatial (con)': 'EmbSpatial Acc (con)',
# 'EmbSpatial (ctr)': 'EmbSpatial Acc (ctr)',
# 'CVBench3D (con)': 'CVBench3D Acc (con)',
# 'CVBench3D (ctr)': 'CVBench3D Acc (ctr)'
# }
# df.rename(columns=rename_dict, inplace=True)
# cols_to_convert = ['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)', 'CVBench3D Acc (con)', 'CVBench3D Acc (ctr)']
# for col in cols_to_convert:
# if col in df.columns:
# df[col] = df[col].astype(str).str.replace('%', '', regex=False).replace('N/A', np.nan).astype(float)
# # Add group and scale labels
# def get_base(name):
# name_upper = str(name).upper()
# if 'MOLMO' in name_upper: return 'Molmo'
# if 'NVILA-ST' in name_upper or 'NVILA ST' in name_upper: return 'NVILA-ST'
# if 'ROBOREFER' in name_upper: return 'RoboRefer'
# if 'NVILA' in name_upper: return 'NVILA'
# if 'QWEN3' in name_upper: return 'Qwen3'
# if 'QWEN' in name_upper: return 'Qwen'
# return 'Other'
# def get_scale(name):
# name_lower = str(name).lower()
# if 'vanilla' in name_lower: return 'vanilla'
# if '80k' in name_lower: return '80k'
# if '400k' in name_lower: return '400k'
# if '800k' in name_lower: return '800k'
# if '2m' in name_lower: return '2M'
# return 'Special'
# df['Base'] = df['Model'].apply(get_base)
# df['Scale'] = df['Model'].apply(get_scale)
# # 4. Filter out Qwen3
# df = df[df['Base'] != 'Qwen3']
# # 5. Output directory & Settings
# output_dir = "evaluation_figures_final"
# os.makedirs(output_dir, exist_ok=True)
# sns.set_theme(style="whitegrid")
# # Colors & Orders
# colors = {
# 'Molmo': '#1f77b4',
# 'NVILA': '#ff7f0e',
# 'NVILA-ST': '#cc5500',
# 'Qwen': '#2ca02c',
# 'RoboRefer': '#d62728'
# }
# scale_order = ['vanilla', '80k', '400k', '800k', '2M']
# st_scale_order = ['80k', '400k', '800k']
# custom_lines = [
# Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Molmo'], markersize=10, label='Molmo'),
# Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA'], markersize=10, label='NVILA'),
# Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA-ST'], markeredgecolor='black', markersize=11, label='NVILA-ST'),
# Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Qwen'], markersize=10, label='Qwen'),
# Line2D([0], [0], marker='*', color='w', markerfacecolor=colors['RoboRefer'], markeredgecolor='black', markersize=18, label='RoboRefer'),
# ]
# # -----------------------------------------------------------------------------
# # Figure 1: The Two Paths (Scatter with Trajectories and Labels)
# # -----------------------------------------------------------------------------
# plt.figure(figsize=(12, 9))
# # Scatter: RoboRefer를 제외한 모든 모델 (pct 모델들도 점으로 그려지게 포함됨)
# df_bases = df[df['Base'] != 'RoboRefer']
# sns.scatterplot(data=df_bases, x='Entanglement', y='Peak Dist', hue='Base', palette=colors, s=100, alpha=0.8, edgecolor='black', legend=False)
# # 텍스트 라벨링: NVILA-ST와 이름에 'pct'가 들어간 모델들만 표시
# for i, row in df_bases.iterrows():
# model_name = str(row['Model'])
# is_st = row['Base'] == 'NVILA Syn'
# is_pct = '%' in model_name.lower()
# if is_st or is_pct:
# plt.annotate(model_name, (row['Entanglement'], row['Peak Dist']),
# xytext=(6, 6), textcoords='offset points',
# fontsize=9, color=colors[row['Base']], fontweight='bold', alpha=0.9)
# # Trajectories (Arrows)
# for base in ['Molmo', 'NVILA', 'Qwen', 'NVILA-ST']:
# current_order = st_scale_order if base == 'NVILA-ST' else scale_order
# base_df = df[(df['Base'] == base) & (df['Scale'].isin(current_order))]
# base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(current_order).dropna(subset=['Entanglement', 'Peak Dist'])
# if len(base_df) > 1:
# for i in range(len(base_df)-1):
# x1, y1 = base_df.iloc[i]['Entanglement'], base_df.iloc[i]['Peak Dist']
# x2, y2 = base_df.iloc[i+1]['Entanglement'], base_df.iloc[i+1]['Peak Dist']
# ls = '--' if base == 'NVILA-ST' else '-'
# plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
# arrowprops=dict(arrowstyle="->", color=colors[base], alpha=0.6, lw=1.8, ls=ls))
# # Scatter & Label: RoboRefer
# df_robo = df[df['Base'] == 'RoboRefer']
# if not df_robo.empty:
# sns.scatterplot(data=df_robo, x='Entanglement', y='Peak Dist', hue='Base',
# palette=colors, s=400, marker='*', edgecolor='black', zorder=5, legend=False)
# for i, row in df_robo.iterrows():
# plt.annotate(row['Model'], (row['Entanglement'], row['Peak Dist']),
# xytext=(10, -5), textcoords='offset points',
# fontsize=11, color=colors['RoboRefer'], fontweight='bold')
# plt.legend(handles=custom_lines, title='Model Group', loc='upper right')
# plt.title('Figure 1: The Two Paths of Spatial Representation', fontsize=14, fontweight='bold')
# plt.xlabel('Entanglement Shortcut [Lower is Better]', fontsize=12)
# plt.ylabel('Distance Representation (Dist SC) [Higher is Better]', fontsize=12)
# plt.text(0.6, 0.05, "Mechanism 1:\nNaive Scaling\n(More Data = More Entanglement)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.text(0.1, 0.20, "Mechanism 2:\nGenuine Understanding\n(Break the Shortcut)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure1_two_paths_trajectory.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 2: Entanglement Paradox (Line Plot)
# # -----------------------------------------------------------------------------
# df_scale = df[df['Base'] != 'RoboRefer'].copy()
# df_scale = df_scale[df_scale['Scale'].isin(scale_order)]
# df_scale = df_scale.drop_duplicates(subset=['Base', 'Scale'])
# df_scale['Scale'] = pd.Categorical(df_scale['Scale'], categories=scale_order, ordered=True)
# plt.figure(figsize=(10, 6))
# sns.lineplot(data=df_scale, x='Scale', y='Entanglement', hue='Base', style='Base',
# palette=colors, dashes={'Molmo':'', 'NVILA':'', 'Qwen':'', 'NVILA-ST':(4,2)},
# marker='o', linewidth=2.5, markersize=8)
# if not df_robo.empty:
# robo_ent = df_robo['Entanglement'].values[0]
# plt.axhline(y=robo_ent, color=colors['RoboRefer'], linestyle='--', label=f'RoboRefer ({robo_ent:.4f})')
# plt.title('Figure 2: The Entanglement Paradox during Naive Scaling', fontsize=14, fontweight='bold')
# plt.ylabel('Entanglement', fontsize=12)
# plt.xlabel('Fine-tuning Data Scale', fontsize=12)
# plt.legend(title='Models', bbox_to_anchor=(1.05, 1), loc='upper left')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure2_entanglement_paradox.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 4: Behavioral Consequence - The Gap (Grouped Bar Chart)
# # -----------------------------------------------------------------------------
# df_acc = df.dropna(subset=['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)']).copy()
# available_acc_models = df_acc['Model'].tolist()
# fig_width = max(12, len(available_acc_models) * 0.7)
# x = np.arange(len(available_acc_models))
# width = 0.35
# fig, ax = plt.subplots(figsize=(fig_width, 6))
# rects1 = ax.bar(x - width/2, df_acc['EmbSpatial Acc (con)'], width, label='Consistent (Shortcut Works)', color='#2b83ba')
# rects2 = ax.bar(x + width/2, df_acc['EmbSpatial Acc (ctr)'], width, label='Counter (Shortcut Fails)', color='#d7191c')
# for i in range(len(available_acc_models)):
# con_val = df_acc['EmbSpatial Acc (con)'].iloc[i]
# ctr_val = df_acc['EmbSpatial Acc (ctr)'].iloc[i]
# gap = con_val - ctr_val
# ax.plot([i - width/2, i + width/2], [con_val, ctr_val], color='black', linestyle='-', linewidth=1.5, alpha=0.5)
# ax.text(i, max(con_val, ctr_val) + 2, f'Gap: {gap:.1f}%p', ha='center', fontsize=9, fontweight='bold')
# ax.set_ylabel('Accuracy (%)', fontsize=12)
# ax.set_title('Figure 4: The Consistent vs. Counter Performance Gap (All Valid Models)', fontsize=14, fontweight='bold')
# ax.set_xticks(x)
# ax.set_xticklabels(available_acc_models, rotation=45, ha='right')
# ax.legend()
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure4_accuracy_gap.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 5: Distance Consistency vs Counter Accuracy
# # -----------------------------------------------------------------------------
# plt.figure(figsize=(10, 8))
# df_valid_acc = df.dropna(subset=['EmbSpatial Acc (ctr)', 'Peak Dist'])
# df_bases_fig5 = df_valid_acc[df_valid_acc['Base'] != 'RoboRefer']
# sns.scatterplot(data=df_bases_fig5, x='Peak Dist', y='EmbSpatial Acc (ctr)', hue='Base', palette=colors, s=150, edgecolor='black', alpha=0.8, legend=False)
# for base in ['Molmo', 'NVILA', 'Qwen', 'NVILA-ST']:
# current_order = st_scale_order if base == 'NVILA-ST' else scale_order
# base_df = df_valid_acc[(df_valid_acc['Base'] == base) & (df_valid_acc['Scale'].isin(current_order))]
# base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(current_order).dropna(subset=['Peak Dist', 'EmbSpatial Acc (ctr)'])
# if len(base_df) > 1:
# for i in range(len(base_df)-1):
# x1, y1 = base_df.iloc[i]['Peak Dist'], base_df.iloc[i]['EmbSpatial Acc (ctr)']
# x2, y2 = base_df.iloc[i+1]['Peak Dist'], base_df.iloc[i+1]['EmbSpatial Acc (ctr)']
# ls = '--' if base == 'NVILA-ST' else '-'
# plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
# arrowprops=dict(arrowstyle="->", color=colors[base], alpha=0.6, lw=1.8, ls=ls))
# df_robo_fig5 = df_valid_acc[df_valid_acc['Base'] == 'RoboRefer']
# if not df_robo_fig5.empty:
# sns.scatterplot(data=df_robo_fig5, x='Peak Dist', y='EmbSpatial Acc (ctr)', hue='Base',
# palette=colors, s=400, marker='*', edgecolor='black', zorder=5, legend=False)
# if len(df_valid_acc) > 1:
# corr = np.corrcoef(df_valid_acc['Peak Dist'], df_valid_acc['EmbSpatial Acc (ctr)'])[0, 1]
# plt.text(0.05, 0.95, f"Pearson r = {corr:.2f}", transform=plt.gca().transAxes,
# fontsize=14, fontweight='bold', bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.legend(handles=custom_lines, title='Model Group', loc='lower right')
# plt.title('Figure 5: Distance SC is the strongest predictor of Counter Accuracy', fontsize=14, fontweight='bold')
# plt.xlabel('Distance Sign-Corrected Consistency (Peak Dist)', fontsize=12)
# plt.ylabel('EmbSpatial Counter Accuracy (%)', fontsize=12)
# plt.grid(True, linestyle='--', alpha=0.7)
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure5_distSC_vs_counterAcc.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 6: 3D Scatter (Peak Dist, Counter Acc, Entanglement)
# # -----------------------------------------------------------------------------
# fig = plt.figure(figsize=(12, 10))
# ax = fig.add_subplot(111, projection='3d')
# for base_name in df_valid_acc['Base'].unique():
# subset = df_valid_acc[df_valid_acc['Base'] == base_name]
# c = colors.get(base_name, 'gray')
# if base_name == 'RoboRefer':
# ax.scatter(subset['Peak Dist'], subset['Entanglement'], subset['EmbSpatial Acc (ctr)'],
# c=c, s=300, marker='*', edgecolors='k', alpha=0.9, zorder=5)
# else:
# ax.scatter(subset['Peak Dist'], subset['Entanglement'], subset['EmbSpatial Acc (ctr)'],
# c=c, s=120, edgecolors='k', alpha=0.8)
# for base in ['Molmo', 'NVILA', 'Qwen', 'NVILA-ST']:
# current_order = st_scale_order if base == 'NVILA-ST' else scale_order
# base_df = df_valid_acc[(df_valid_acc['Base'] == base) & (df_valid_acc['Scale'].isin(current_order))]
# base_df = base_df.drop_duplicates(subset=['Scale']).set_index('Scale').reindex(current_order).dropna(subset=['Peak Dist', 'Entanglement', 'EmbSpatial Acc (ctr)'])
# if len(base_df) > 1:
# for i in range(len(base_df)-1):
# x1, y1, z1 = base_df.iloc[i]['Peak Dist'], base_df.iloc[i]['Entanglement'], base_df.iloc[i]['EmbSpatial Acc (ctr)']
# x2, y2, z2 = base_df.iloc[i+1]['Peak Dist'], base_df.iloc[i+1]['Entanglement'], base_df.iloc[i+1]['EmbSpatial Acc (ctr)']
# ls = '--' if base == 'NVILA-ST' else '-'
# ax.plot([x1, x2], [y1, y2], [z1, z2], color=colors[base], linestyle=ls, linewidth=2.5, alpha=0.7)
# ax.set_xlabel('Peak Dist SC', labelpad=10, fontsize=11)
# ax.set_ylabel('Entanglement', labelpad=10, fontsize=11)
# ax.set_zlabel('Counter Accuracy (%)', labelpad=10, fontsize=11)
# ax.set_title('Figure 6: Navigating the Representation Space\n(Dist SC vs. Entanglement vs. Counter Acc)', fontweight='bold', fontsize=14)
# ax.view_init(elev=20, azim=135)
# plt.legend(handles=custom_lines, title='Model Group', loc='upper left', bbox_to_anchor=(1.05, 1))
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure6_3d_distSC_ent_counterAcc.png'), dpi=300, bbox_inches='tight')
# plt.close()
# print(f"Success! Evaluation figures have been saved to ./{output_dir}/")
# if __name__ == "__main__":
# main()
# import os
# import pandas as pd
# import numpy as np
# import matplotlib.pyplot as plt
# import seaborn as sns
# from io import StringIO
# import matplotlib.patches as patches
# # 1. Input data
# data = """Model,Peak Horiz,Peak Vert,Peak Dist,Peak cos(dv dd),Last cos(dv dd),EmbSpatial Acc (con),EmbSpatial Acc (ctr),CVBench3D Acc (con),CVBench3D Acc (ctr)
# Molmo vanilla,0.1427,0.2279,0.0749,0.3486,0.2367,63.5%,34.9%,93.1%,75.4%
# Molmo 80k,0.1224,0.3317,0.0718,0.5760,0.4397,60.6%,29.5%,80.2%,56.9%
# Molmo 400k,0.2355,0.5971,0.0961,0.7113,0.6151,62.7%,27.1%,89.5%,56.9%
# Molmo 800k,0.2468,0.5586,0.1073,0.8009,0.6150,65.2%,34.1%,88.7%,70.8%
# Molmo 2M,0.2390,0.5743,0.1122,0.7986,0.6543,65.3%,39.5%,90.6%,72.3%
# NVILA vanilla,0.3232,0.2890,0.0521,0.5609,0.2887,49.0%,27.1%,74.4%,40.0%
# RoboRefer,0.6490,0.8302,0.1824,0.5725,0.1117,87.0%,59.7%,98.9%,95.4%
# NVILA 80k,0.2946,0.4967,0.0701,0.7094,0.4125,57.7%,15.5%,71.6%,50.8%
# NVILA 400k,0.2415,0.5741,0.0949,0.8056,0.4197,61.1%,34.1%,81.3%,58.5%
# NVILA 800k,0.2779,0.4984,0.0885,0.8377,0.4133,63.2%,38.8%,84.6%,67.7%
# NVILA 2M,0.2409,0.5531,0.1039,0.8371,0.4267,60.7%,41.1%,97.2%,93.8%
# Qwen vanilla,0.3674,0.2925,0.0428,0.5047,0.2043,54.7%,32.6%,75.5%,55.4%
# Qwen 80k,0.3862,0.3148,0.0404,0.4553,0.1897,50.6%,30.2%,69.7%,60.0%
# Qwen 400k,0.4496,0.4522,0.0418,0.5112,0.2041,52.6%,27.1%,65.8%,58.5%
# Qwen 800k,0.4730,0.5376,0.0454,0.5429,0.2177,55.8%,26.4%,61.2%,58.5%
# Qwen 2M,0.4845,0.5858,0.0519,0.6049,0.2426,60.9%,24.0%,62.0%,53.8%
# Qwen3 235b,0.6610,0.6187,0.2285,0.5126,0.1783,N/A,N/A,N/A,N/A"""
# # Preprocessing
# df = pd.read_csv(StringIO(data))
# cols_to_convert = ['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)', 'CVBench3D Acc (con)', 'CVBench3D Acc (ctr)']
# for col in cols_to_convert:
# df[col] = df[col].astype(str).str.replace('%', '').replace('N/A', np.nan).astype(float)
# # Add group and scale labels
# def get_base(name):
# if 'Molmo' in name: return 'Molmo'
# if 'NVILA' in name: return 'NVILA'
# if 'Qwen3' in name: return 'Genuine (Qwen3)'
# if 'Qwen' in name: return 'Qwen'
# if 'RoboRefer' in name: return 'Genuine (RoboRefer)'
# return 'Other'
# def get_scale(name):
# for s in ['vanilla', '80k', '400k', '800k', '2M']:
# if s in name: return s
# return 'Special'
# df['Base'] = df['Model'].apply(get_base)
# df['Scale'] = df['Model'].apply(get_scale)
# # Output directory
# output_dir = "evaluation_figures_v2"
# os.makedirs(output_dir, exist_ok=True)
# sns.set_theme(style="whitegrid")
# # -----------------------------------------------------------------------------
# # Figure 1: The Two Paths (Scatter with Trajectories)
# # -----------------------------------------------------------------------------
# plt.figure(figsize=(10, 8))
# # Define colors
# colors = {'Molmo': '#1f77b4', 'NVILA': '#ff7f0e', 'Qwen': '#2ca02c',
# 'Genuine (RoboRefer)': '#d62728', 'Genuine (Qwen3)': '#9467bd'}
# # Plot naive scaling models (with smaller dots)
# df_naive = df[df['Scale'] != 'Special']
# sns.scatterplot(data=df_naive, x='Last cos(dv dd)', y='Peak Dist', hue='Base', palette=colors, s=80, alpha=0.7)
# # Draw arrows to show trajectories (vanilla -> 2M)
# scale_order = ['vanilla', '80k', '400k', '800k', '2M']
# for base in ['Molmo', 'NVILA', 'Qwen']:
# base_df = df_naive[df_naive['Base'] == base].set_index('Scale').loc[scale_order]
# for i in range(len(scale_order)-1):
# x1, y1 = base_df.iloc[i]['Last cos(dv dd)'], base_df.iloc[i]['Peak Dist']
# x2, y2 = base_df.iloc[i+1]['Last cos(dv dd)'], base_df.iloc[i+1]['Peak Dist']
# plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
# arrowprops=dict(arrowstyle="->", color=colors[base], alpha=0.5, lw=1.5))
# # Plot Genuine models (with big stars)
# df_genuine = df[df['Scale'] == 'Special']
# sns.scatterplot(data=df_genuine, x='Last cos(dv dd)', y='Peak Dist', hue='Base',
# palette=colors, s=400, marker='*', edgecolor='black', zorder=5)
# plt.title('Figure 1: The Two Paths of Spatial Representation\n(Naive Scaling vs. Genuine Understanding)', fontsize=14, fontweight='bold')
# plt.xlabel('Entanglement Shortcut: Last Layer cos(dv, dd) [Lower is Better]', fontsize=12)
# plt.ylabel('Distance Representation: Peak Dist [Higher is Better]', fontsize=12)
# # Add text annotations for clarity
# plt.text(0.6, 0.05, "Mechanism 1:\nNaive Scaling\n(More Data = More Entanglement)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.text(0.1, 0.20, "Mechanism 2:\nGenuine Understanding\n(Break the Shortcut)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.legend(title='Model Group')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure1_two_paths_trajectory.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 2: Entanglement Paradox (Line Plot)
# # -----------------------------------------------------------------------------
# df_scale = df[df['Scale'] != 'Special'].copy()
# df_scale['Scale'] = pd.Categorical(df_scale['Scale'], categories=scale_order, ordered=True)
# plt.figure(figsize=(10, 6))
# sns.lineplot(data=df_scale, x='Scale', y='Last cos(dv dd)', hue='Base', marker='o', linewidth=2.5, markersize=8)
# # Add baselines for Genuine Models
# robo_ent = df[df['Model'] == 'RoboRefer']['Last cos(dv dd)'].values[0]
# qwen3_ent = df[df['Model'] == 'Qwen3 235b']['Last cos(dv dd)'].values[0]
# plt.axhline(y=qwen3_ent, color=colors['Genuine (Qwen3)'], linestyle='--', label=f'Qwen3 235b ({qwen3_ent:.4f})')
# plt.axhline(y=robo_ent, color=colors['Genuine (RoboRefer)'], linestyle='--', label=f'RoboRefer ({robo_ent:.4f})')
# plt.title('Figure 2: The Entanglement Paradox during Naive Scaling', fontsize=14, fontweight='bold')
# plt.ylabel('Entanglement (Last Layer cos(dv, dd))', fontsize=12)
# plt.xlabel('Fine-tuning Data Scale', fontsize=12)
# plt.legend(title='Models', bbox_to_anchor=(1.05, 1), loc='upper left')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure2_entanglement_paradox.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 3: Depth Independence Ratio (Bar Chart)
# # -----------------------------------------------------------------------------
# # Calculate ratio: Peak Dist / Peak Vert
# df['Depth Independence Ratio'] = df['Peak Dist'] / df['Peak Vert']
# # Select representative models
# rep_models = ['Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer', 'Qwen3 235b']
# df_rep = df.set_index('Model').loc[rep_models]
# plt.figure(figsize=(12, 6))
# # Color scheme: Naive (blue/orange/green) vs Genuine (Red/Purple)
# bar_colors = ['#aec7e8', '#1f77b4', '#ffbb78', '#ff7f0e', '#98df8a', '#2ca02c', '#d62728', '#9467bd']
# ax = df_rep['Depth Independence Ratio'].plot(kind='bar', color=bar_colors, edgecolor='black')
# plt.title('Figure 3: Depth Independence Ratio (Peak Dist / Peak Vert)', fontsize=14, fontweight='bold')
# plt.ylabel('Ratio (Higher = Less dependent on vertical shortcut)', fontsize=12)
# plt.xticks(rotation=45, ha='right')
# # Add horizontal line to show the "Genuine" threshold
# plt.axhline(y=0.2, color='gray', linestyle=':', alpha=0.7)
# plt.text(-0.5, 0.22, "Genuine Understanding Threshold", color='gray', fontstyle='italic')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure3_depth_independence_ratio.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 4: Behavioral Consequence - The Gap (Grouped Bar Chart)
# # -----------------------------------------------------------------------------
# acc_models = ['Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer']
# df_acc = df.set_index('Model').loc[acc_models]
# # We focus on EmbSpatial as it clearly shows the counter gap
# x = np.arange(len(acc_models))
# width = 0.35
# fig, ax = plt.subplots(figsize=(12, 6))
# rects1 = ax.bar(x - width/2, df_acc['EmbSpatial Acc (con)'], width, label='Consistent (Shortcut Works)', color='#2b83ba')
# rects2 = ax.bar(x + width/2, df_acc['EmbSpatial Acc (ctr)'], width, label='Counter (Shortcut Fails)', color='#d7191c')
# # Draw gap lines
# for i in range(len(acc_models)):
# con_val = df_acc['EmbSpatial Acc (con)'].iloc[i]
# ctr_val = df_acc['EmbSpatial Acc (ctr)'].iloc[i]
# gap = con_val - ctr_val
# # Plot an error bar-like line connecting the two tops
# ax.plot([i - width/2, i + width/2], [con_val, ctr_val], color='black', linestyle='-', linewidth=1.5, alpha=0.5)
# ax.text(i, max(con_val, ctr_val) + 2, f'Gap: {gap:.1f}%p', ha='center', fontsize=9, fontweight='bold')
# ax.set_ylabel('Accuracy (%)', fontsize=12)
# ax.set_title('Figure 4: The Consistent vs. Counter Performance Gap', fontsize=14, fontweight='bold')
# ax.set_xticks(x)
# ax.set_xticklabels(acc_models, rotation=45, ha='right')
# ax.legend()
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure4_accuracy_gap.png'), dpi=300)
# plt.close()
# print(f"Success! 4 figures supporting 'Naive Scaling vs Genuine Understanding' have been saved to ./{output_dir}/")
# import os
# import pandas as pd
# import numpy as np
# import matplotlib.pyplot as plt
# import seaborn as sns
# from io import StringIO
# import matplotlib.patches as patches
# # 1. Input data (Updated with new Entanglement values)
# data = """Model,Peak Horiz,Peak Vert,Peak Dist,Entanglement Layer,Entanglement,EmbSpatial Acc (con),EmbSpatial Acc (ctr),CVBench3D Acc (con),CVBench3D Acc (ctr)
# Molmo vanilla,0.1427,0.2279,0.0749,23,0.2791,63.5%,34.9%,93.1%,75.4%
# Molmo 80k,0.1224,0.3317,0.0718,23,0.3879,60.6%,29.5%,80.2%,56.9%
# Molmo 400k,0.2355,0.5971,0.0961,23,0.4586,62.7%,27.1%,89.5%,56.9%
# Molmo 800k,0.2468,0.5586,0.1073,23,0.5142,65.2%,34.1%,88.7%,70.8%
# Molmo 2M,0.2390,0.5743,0.1122,23,0.4744,65.3%,39.5%,90.6%,72.3%
# NVILA vanilla,0.3232,0.2890,0.0521,20,0.5387,15.3%,7.8%,74.4%,40.0%
# RoboRefer,0.6490,0.8302,0.1824,20,0.3619,87.0%,59.7%,98.9%,95.4%
# NVILA 80k,0.2946,0.4967,0.0701,20,0.6061,57.7%,15.5%,71.6%,50.8%
# NVILA 400k,0.2415,0.5741,0.0949,20,0.5889,61.1%,34.1%,81.3%,58.5%
# NVILA 800k,0.2779,0.4984,0.0885,20,0.5914,63.2%,38.8%,84.6%,67.7%
# NVILA 2M,0.2409,0.5531,0.1039,20,0.5501,60.7%,41.1%,97.2%,93.8%
# NVILA 10pct,0.4141,0.8469,0.1106,20,0.6028,N/A,N/A,N/A,N/A
# NVILA 20pct,0.4329,0.8179,0.0881,20,0.6871,N/A,N/A,N/A,N/A
# NVILA 30pct,0.4687,0.7497,0.0670,20,0.6669,N/A,N/A,N/A,N/A
# Qwen vanilla,0.3674,0.2925,0.0428,27,0.4568,54.7%,32.6%,75.5%,55.4%
# Qwen 80k,0.3862,0.3148,0.0404,27,0.4563,50.6%,30.2%,69.7%,60.0%
# Qwen 400k,0.4496,0.4522,0.0418,27,0.4514,52.6%,27.1%,65.8%,58.5%
# Qwen 800k,0.4730,0.5376,0.0454,27,0.4287,55.8%,26.4%,61.2%,58.5%
# Qwen 2M,0.4845,0.5858,0.0519,27,0.4720,60.9%,24.0%,62.0%,53.8%
# Qwen3-235B,0.6610,0.6187,0.2285,87,0.1169,70.5%,39.5%,N/A,N/A"""
# # Preprocessing
# df = pd.read_csv(StringIO(data))
# cols_to_convert = ['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)', 'CVBench3D Acc (con)', 'CVBench3D Acc (ctr)']
# for col in cols_to_convert:
# df[col] = df[col].astype(str).str.replace('%', '').replace('N/A', np.nan).astype(float)
# # Add group and scale labels
# def get_base(name):
# if 'Molmo' in name: return 'Molmo'
# if 'NVILA' in name: return 'NVILA'
# if 'Qwen3' in name: return 'Qwen3' # Genuine 텍스트 제거
# if 'Qwen' in name: return 'Qwen'
# if 'RoboRefer' in name: return 'RoboRefer' # Genuine 텍스트 제거
# return 'Other'
# def get_scale(name):
# for s in ['vanilla', '80k', '400k', '800k', '2M']:
# if s in name: return s
# return 'Special'
# df['Base'] = df['Model'].apply(get_base)
# df['Scale'] = df['Model'].apply(get_scale)
# # Output directory
# output_dir = "evaluation_figures_v2_updated"
# os.makedirs(output_dir, exist_ok=True)
# sns.set_theme(style="whitegrid")
# # -----------------------------------------------------------------------------
# # Figure 1: The Two Paths (Scatter with Trajectories)
# # -----------------------------------------------------------------------------
# plt.figure(figsize=(10, 8))
# # Define colors
# colors = {'Molmo': '#1f77b4', 'NVILA': '#ff7f0e', 'Qwen': '#2ca02c',
# 'RoboRefer': '#d62728', 'Qwen3': '#9467bd'} # Genuine 키 수정
# # Plot naive scaling models (with smaller dots)
# df_naive = df[df['Scale'] != 'Special']
# sns.scatterplot(data=df_naive, x='Entanglement', y='Peak Dist', hue='Base', palette=colors, s=80, alpha=0.7)
# # Draw arrows to show trajectories (vanilla -> 2M)
# scale_order = ['vanilla', '80k', '400k', '800k', '2M']
# for base in ['Molmo', 'NVILA', 'Qwen']:
# base_df = df_naive[df_naive['Base'] == base].set_index('Scale').loc[scale_order]
# for i in range(len(scale_order)-1):
# x1, y1 = base_df.iloc[i]['Entanglement'], base_df.iloc[i]['Peak Dist']
# x2, y2 = base_df.iloc[i+1]['Entanglement'], base_df.iloc[i+1]['Peak Dist']
# plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
# arrowprops=dict(arrowstyle="->", color=colors[base], alpha=0.5, lw=1.5))
# # Plot RoboRefer and Qwen3 models (with big stars) - NVILA 제외
# df_genuine = df[df['Scale'] == 'Special']
# df_genuine_filtered = df_genuine[df_genuine['Base'] != 'NVILA'] # NVILA Special 모델 제외
# sns.scatterplot(data=df_genuine_filtered, x='Entanglement', y='Peak Dist', hue='Base',
# palette=colors, s=400, marker='*', edgecolor='black', zorder=5)
# plt.title('Figure 1: The Two Paths of Spatial Representation\n(Naive Scaling vs. Genuine Understanding)', fontsize=14, fontweight='bold')
# plt.xlabel('Entanglement Shortcut [Lower is Better]', fontsize=12)
# plt.ylabel('Distance Representation: Peak Dist [Higher is Better]', fontsize=12)
# # Add text annotations for clarity
# plt.text(0.6, 0.05, "Mechanism 1:\nNaive Scaling\n(More Data = More Entanglement)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.text(0.1, 0.20, "Mechanism 2:\nGenuine Understanding\n(Break the Shortcut)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.legend(title='Model Group')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure1_two_paths_trajectory_updated.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 2: Entanglement Paradox (Line Plot)
# # -----------------------------------------------------------------------------
# df_scale = df[df['Scale'] != 'Special'].copy()
# df_scale['Scale'] = pd.Categorical(df_scale['Scale'], categories=scale_order, ordered=True)
# plt.figure(figsize=(10, 6))
# sns.lineplot(data=df_scale, x='Scale', y='Entanglement', hue='Base', marker='o', linewidth=2.5, markersize=8)
# # Add baselines for Genuine Models
# robo_ent = df[df['Model'] == 'RoboRefer']['Entanglement'].values[0]
# qwen3_ent = df[df['Model'] == 'Qwen3-235B']['Entanglement'].values[0]
# plt.axhline(y=qwen3_ent, color=colors['Qwen3'], linestyle='--', label=f'Qwen3-235B ({qwen3_ent:.4f})')
# plt.axhline(y=robo_ent, color=colors['RoboRefer'], linestyle='--', label=f'RoboRefer ({robo_ent:.4f})')
# plt.title('Figure 2: The Entanglement Paradox during Naive Scaling', fontsize=14, fontweight='bold')
# plt.ylabel('Entanglement', fontsize=12)
# plt.xlabel('Fine-tuning Data Scale', fontsize=12)
# plt.legend(title='Models', bbox_to_anchor=(1.05, 1), loc='upper left')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure2_entanglement_paradox_updated.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 3: Depth Independence Ratio (Bar Chart)
# # -----------------------------------------------------------------------------
# # Calculate ratio: Peak Dist / Peak Vert
# df['Depth Independence Ratio'] = df['Peak Dist'] / df['Peak Vert']
# # Select representative models
# rep_models = ['Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer', 'Qwen3-235B']
# df_rep = df.set_index('Model').loc[rep_models]
# plt.figure(figsize=(12, 6))
# # Color scheme: Naive (blue/orange/green) vs Genuine (Red/Purple)
# bar_colors = ['#aec7e8', '#1f77b4', '#ffbb78', '#ff7f0e', '#98df8a', '#2ca02c', '#d62728', '#9467bd']
# ax = df_rep['Depth Independence Ratio'].plot(kind='bar', color=bar_colors, edgecolor='black')
# plt.title('Figure 3: Depth Independence Ratio (Peak Dist / Peak Vert)', fontsize=14, fontweight='bold')
# plt.ylabel('Ratio (Higher = Less dependent on vertical shortcut)', fontsize=12)
# plt.xticks(rotation=45, ha='right')
# # Add horizontal line to show the "Genuine" threshold
# plt.axhline(y=0.2, color='gray', linestyle=':', alpha=0.7)
# plt.text(-0.5, 0.22, "Genuine Understanding Threshold", color='gray', fontstyle='italic')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure3_depth_independence_ratio_updated.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 4: Behavioral Consequence - The Gap (Grouped Bar Chart)
# # -----------------------------------------------------------------------------
# acc_models = ['Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer', 'Qwen3-235B'] # Qwen3 추가
# df_acc = df.set_index('Model').loc[acc_models]
# # We focus on EmbSpatial as it clearly shows the counter gap
# x = np.arange(len(acc_models))
# width = 0.35
# fig, ax = plt.subplots(figsize=(12, 6))
# rects1 = ax.bar(x - width/2, df_acc['EmbSpatial Acc (con)'], width, label='Consistent (Shortcut Works)', color='#2b83ba')
# rects2 = ax.bar(x + width/2, df_acc['EmbSpatial Acc (ctr)'], width, label='Counter (Shortcut Fails)', color='#d7191c')
# # Draw gap lines
# for i in range(len(acc_models)):
# con_val = df_acc['EmbSpatial Acc (con)'].iloc[i]
# ctr_val = df_acc['EmbSpatial Acc (ctr)'].iloc[i]
# gap = con_val - ctr_val
# # Plot an error bar-like line connecting the two tops
# ax.plot([i - width/2, i + width/2], [con_val, ctr_val], color='black', linestyle='-', linewidth=1.5, alpha=0.5)
# ax.text(i, max(con_val, ctr_val) + 2, f'Gap: {gap:.1f}%p', ha='center', fontsize=9, fontweight='bold')
# ax.set_ylabel('Accuracy (%)', fontsize=12)
# ax.set_title('Figure 4: The Consistent vs. Counter Performance Gap', fontsize=14, fontweight='bold')
# ax.set_xticks(x)
# ax.set_xticklabels(acc_models, rotation=45, ha='right')
# ax.legend()
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure4_accuracy_gap_updated.png'), dpi=300)
# plt.close()
# print(f"Success! 4 figures with requested updates have been saved to ./{output_dir}/")
# import os
# import pandas as pd
# import numpy as np
# import matplotlib.pyplot as plt
# import seaborn as sns
# from io import StringIO
# from matplotlib.lines import Line2D
# # 1. Input data
# data = """Model,Peak Horiz,Peak Vert,Peak Dist,Entanglement Layer,Entanglement,EmbSpatial Acc (con),EmbSpatial Acc (ctr),CVBench3D Acc (con),CVBench3D Acc (ctr)
# Molmo vanilla,0.1427,0.2279,0.0749,23,0.2791,63.5%,34.9%,93.1%,75.4%
# Molmo 80k,0.1224,0.3317,0.0718,23,0.3879,60.6%,29.5%,80.2%,56.9%
# Molmo 400k,0.2355,0.5971,0.0961,23,0.4586,62.7%,27.1%,89.5%,56.9%
# Molmo 800k,0.2468,0.5586,0.1073,23,0.5142,65.2%,34.1%,88.7%,70.8%
# Molmo 2M,0.2390,0.5743,0.1122,23,0.4744,65.3%,39.5%,90.6%,72.3%
# NVILA vanilla,0.3232,0.2890,0.0521,20,0.5387,15.3%,7.8%,74.4%,40.0%
# RoboRefer,0.6490,0.8302,0.1824,20,0.3619,87.0%,59.7%,98.9%,95.4%
# NVILA 80k,0.2946,0.4967,0.0701,20,0.6061,57.7%,15.5%,71.6%,50.8%
# NVILA 400k,0.2415,0.5741,0.0949,20,0.5889,61.1%,34.1%,81.3%,58.5%
# NVILA 800k,0.2779,0.4984,0.0885,20,0.5914,63.2%,38.8%,84.6%,67.7%
# NVILA 2M,0.2409,0.5531,0.1039,20,0.5501,60.7%,41.1%,97.2%,93.8%
# NVILA 10pct,0.4141,0.8469,0.1106,20,0.6028,N/A,N/A,N/A,N/A
# NVILA 20pct,0.4329,0.8179,0.0881,20,0.6871,N/A,N/A,N/A,N/A
# NVILA 30pct,0.4687,0.7497,0.0670,20,0.6669,N/A,N/A,N/A,N/A
# Qwen vanilla,0.3674,0.2925,0.0428,27,0.4568,54.7%,32.6%,75.5%,55.4%
# Qwen 80k,0.3862,0.3148,0.0404,27,0.4563,50.6%,30.2%,69.7%,60.0%
# Qwen 400k,0.4496,0.4522,0.0418,27,0.4514,52.6%,27.1%,65.8%,58.5%
# Qwen 800k,0.4730,0.5376,0.0454,27,0.4287,55.8%,26.4%,61.2%,58.5%
# Qwen 2M,0.4845,0.5858,0.0519,27,0.4720,60.9%,24.0%,62.0%,53.8%
# Qwen3-235B,0.6610,0.6187,0.2285,87,0.1169,70.5%,39.5%,N/A,N/A"""
# # Preprocessing
# df = pd.read_csv(StringIO(data))
# cols_to_convert = ['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)', 'CVBench3D Acc (con)', 'CVBench3D Acc (ctr)']
# for col in cols_to_convert:
# df[col] = df[col].astype(str).str.replace('%', '').replace('N/A', np.nan).astype(float)
# # Add group and scale labels
# def get_base(name):
# if 'Molmo' in name: return 'Molmo'
# if 'NVILA' in name: return 'NVILA'
# if 'Qwen3' in name: return 'Qwen3'
# if 'Qwen' in name: return 'Qwen'
# if 'RoboRefer' in name: return 'RoboRefer'
# return 'Other'
# def get_scale(name):
# for s in ['vanilla', '80k', '400k', '800k', '2M']:
# if s in name: return s
# return 'Special'
# df['Base'] = df['Model'].apply(get_base)
# df['Scale'] = df['Model'].apply(get_scale)
# # Output directory
# output_dir = "evaluation_figures_v4"
# os.makedirs(output_dir, exist_ok=True)
# sns.set_theme(style="whitegrid")
# # Colors
# colors = {'Molmo': '#1f77b4', 'NVILA': '#ff7f0e', 'Qwen': '#2ca02c',
# 'RoboRefer': '#d62728', 'Qwen3': '#9467bd'}
# synth_color = '#cc5500' # 진한 주황색 (Darker Orange)
# # -----------------------------------------------------------------------------
# # Figure 1: The Two Paths (Scatter with Trajectories)
# # -----------------------------------------------------------------------------
# plt.figure(figsize=(10, 8))
# # 1. Plot naive scaling models
# df_naive = df[df['Scale'] != 'Special']
# sns.scatterplot(data=df_naive, x='Entanglement', y='Peak Dist', hue='Base', palette=colors, s=80, alpha=0.7, legend=False)
# # 2. Draw arrows for naive scaling
# scale_order = ['vanilla', '80k', '400k', '800k', '2M']
# for base in ['Molmo', 'NVILA', 'Qwen']:
# base_df = df_naive[df_naive['Base'] == base].set_index('Scale').loc[scale_order]
# for i in range(len(scale_order)-1):
# x1, y1 = base_df.iloc[i]['Entanglement'], base_df.iloc[i]['Peak Dist']
# x2, y2 = base_df.iloc[i+1]['Entanglement'], base_df.iloc[i+1]['Peak Dist']
# plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
# arrowprops=dict(arrowstyle="->", color=colors[base], alpha=0.5, lw=1.5))
# # 3. Plot RoboRefer and Qwen3 (Stars)
# df_genuine = df[(df['Scale'] == 'Special') & (df['Base'] != 'NVILA')]
# sns.scatterplot(data=df_genuine, x='Entanglement', y='Peak Dist', hue='Base',
# palette=colors, s=400, marker='*', edgecolor='black', zorder=5, legend=False)
# # 4. Plot NVILA Synthetic Mix (Darker Orange Circles)
# df_synth = df[(df['Scale'] == 'Special') & (df['Base'] == 'NVILA')]
# plt.scatter(df_synth['Entanglement'], df_synth['Peak Dist'],
# color=synth_color, marker='o', s=130, edgecolor='black', zorder=4, label='_nolegend_')
# # Draw dashed arrows for Synthetic Mix trajectory (10 -> 20 -> 30)
# synth_order = ['NVILA 10pct', 'NVILA 20pct', 'NVILA 30pct']
# df_synth_sorted = df_synth.set_index('Model').loc[synth_order]
# for i in range(len(synth_order)-1):
# x1, y1 = df_synth_sorted.iloc[i]['Entanglement'], df_synth_sorted.iloc[i]['Peak Dist']
# x2, y2 = df_synth_sorted.iloc[i+1]['Entanglement'], df_synth_sorted.iloc[i+1]['Peak Dist']
# plt.annotate('', xy=(x2, y2), xytext=(x1, y1),
# arrowprops=dict(arrowstyle="->", color=synth_color, alpha=0.8, lw=1.5, ls='--'))
# # 5. Custom Legend
# custom_lines = [
# Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Molmo'], markersize=10, label='Molmo'),
# Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['NVILA'], markersize=10, label='NVILA'),
# Line2D([0], [0], marker='o', color='w', markerfacecolor=synth_color, markeredgecolor='black', markersize=11, label='NVILA (Synthetic)'),
# Line2D([0], [0], marker='o', color='w', markerfacecolor=colors['Qwen'], markersize=10, label='Qwen'),
# Line2D([0], [0], marker='*', color='w', markerfacecolor=colors['RoboRefer'], markeredgecolor='black', markersize=18, label='RoboRefer'),
# Line2D([0], [0], marker='*', color='w', markerfacecolor=colors['Qwen3'], markeredgecolor='black', markersize=18, label='Qwen3')
# ]
# plt.legend(handles=custom_lines, title='Model Group', loc='upper right')
# plt.title('Figure 1: The Two Paths of Spatial Representation\n(Naive Scaling vs. Genuine Understanding)', fontsize=14, fontweight='bold')
# plt.xlabel('Entanglement Shortcut [Lower is Better]', fontsize=12)
# plt.ylabel('Distance Representation: Peak Dist [Higher is Better]', fontsize=12)
# plt.text(0.6, 0.05, "Mechanism 1:\nNaive Scaling\n(More Data = More Entanglement)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.text(0.1, 0.20, "Mechanism 2:\nGenuine Understanding\n(Break the Shortcut)",
# fontsize=12, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray'))
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure1_two_paths_trajectory.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 2: Entanglement Paradox (Line Plot)
# # -----------------------------------------------------------------------------
# df_scale = df[df['Scale'] != 'Special'].copy()
# df_scale['Scale'] = pd.Categorical(df_scale['Scale'], categories=scale_order, ordered=True)
# plt.figure(figsize=(10, 6))
# sns.lineplot(data=df_scale, x='Scale', y='Entanglement', hue='Base', marker='o', linewidth=2.5, markersize=8)
# robo_ent = df[df['Model'] == 'RoboRefer']['Entanglement'].values[0]
# qwen3_ent = df[df['Model'] == 'Qwen3-235B']['Entanglement'].values[0]
# plt.axhline(y=qwen3_ent, color=colors['Qwen3'], linestyle='--', label=f'Qwen3-235B ({qwen3_ent:.4f})')
# plt.axhline(y=robo_ent, color=colors['RoboRefer'], linestyle='--', label=f'RoboRefer ({robo_ent:.4f})')
# plt.title('Figure 2: The Entanglement Paradox during Naive Scaling', fontsize=14, fontweight='bold')
# plt.ylabel('Entanglement', fontsize=12)
# plt.xlabel('Fine-tuning Data Scale', fontsize=12)
# plt.legend(title='Models', bbox_to_anchor=(1.05, 1), loc='upper left')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure2_entanglement_paradox.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 3: Depth Independence Ratio (Bar Chart)
# # -----------------------------------------------------------------------------
# df['Depth Independence Ratio'] = df['Peak Dist'] / df['Peak Vert']
# rep_models = ['Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer', 'Qwen3-235B']
# df_rep = df.set_index('Model').loc[rep_models]
# plt.figure(figsize=(12, 6))
# bar_colors = ['#aec7e8', '#1f77b4', '#ffbb78', '#ff7f0e', '#98df8a', '#2ca02c', '#d62728', '#9467bd']
# ax = df_rep['Depth Independence Ratio'].plot(kind='bar', color=bar_colors, edgecolor='black')
# plt.title('Figure 3: Depth Independence Ratio (Peak Dist / Peak Vert)', fontsize=14, fontweight='bold')
# plt.ylabel('Ratio (Higher = Less dependent on vertical shortcut)', fontsize=12)
# plt.xticks(rotation=45, ha='right')
# plt.axhline(y=0.2, color='gray', linestyle=':', alpha=0.7)
# plt.text(-0.5, 0.22, "Genuine Understanding Threshold", color='gray', fontstyle='italic')
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure3_depth_independence_ratio.png'), dpi=300)
# plt.close()
# # -----------------------------------------------------------------------------
# # Figure 4: Behavioral Consequence - The Gap (Grouped Bar Chart)
# # -----------------------------------------------------------------------------
# acc_models = ['Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer', 'Qwen3-235B']
# df_acc = df.set_index('Model').loc[acc_models]
# x = np.arange(len(acc_models))
# width = 0.35
# fig, ax = plt.subplots(figsize=(12, 6))
# rects1 = ax.bar(x - width/2, df_acc['EmbSpatial Acc (con)'], width, label='Consistent (Shortcut Works)', color='#2b83ba')
# rects2 = ax.bar(x + width/2, df_acc['EmbSpatial Acc (ctr)'], width, label='Counter (Shortcut Fails)', color='#d7191c')
# for i in range(len(acc_models)):
# con_val = df_acc['EmbSpatial Acc (con)'].iloc[i]
# ctr_val = df_acc['EmbSpatial Acc (ctr)'].iloc[i]
# if pd.isna(con_val) or pd.isna(ctr_val):
# continue
# gap = con_val - ctr_val
# ax.plot([i - width/2, i + width/2], [con_val, ctr_val], color='black', linestyle='-', linewidth=1.5, alpha=0.5)
# ax.text(i, max(con_val, ctr_val) + 2, f'Gap: {gap:.1f}%p', ha='center', fontsize=9, fontweight='bold')
# ax.set_ylabel('Accuracy (%)', fontsize=12)
# ax.set_title('Figure 4: The Consistent vs. Counter Performance Gap', fontsize=14, fontweight='bold')
# ax.set_xticks(x)
# ax.set_xticklabels(acc_models, rotation=45, ha='right')
# ax.legend()
# plt.tight_layout()
# plt.savefig(os.path.join(output_dir, 'figure4_accuracy_gap.png'), dpi=300)
# plt.close()