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 warnings.simplefilter("ignore", UserWarning) SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__)) OUTPUT_DIR = os.path.join(SCRIPT_DIR, "summarize_metrics", "plots") 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'), ] def load_and_preprocess(path: str) -> pd.DataFrame: df = pd.read_csv(path) df.rename(columns={ 'EmbSpatial (con)': 'EmbSpatial Acc (con)', 'EmbSpatial (ctr)': 'EmbSpatial Acc (ctr)', 'CVBench3D (con)': 'CVBench3D Acc (con)', 'CVBench3D (ctr)': 'CVBench3D Acc (ctr)', }, inplace=True) for col in ['EmbSpatial Acc (con)', 'EmbSpatial Acc (ctr)', 'CVBench3D Acc (con)', 'CVBench3D Acc (ctr)']: if col in df.columns: df[col] = (df[col].astype(str) .str.replace('%', '', regex=False) .replace('N/A', np.nan) .astype(float)) def get_base(name): if 'Molmo' in name: return 'Molmo' if 'NVILA-ST'in name: return 'NVILA-ST' 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) df = df[df['Base'] != 'Qwen3'] return df # ── helpers ────────────────────────────────────────────────────────────────── def _draw_2d_arrows(df_src, base_list, x_col, y_col): """Draw annotate arrows for a 2-D figure.""" for base in base_list: order = ST_SCALE_ORDER if base == 'NVILA-ST' else SCALE_ORDER base_df = (df_src[(df_src['Base'] == base) & (df_src['Scale'].isin(order))] .set_index('Scale').reindex(order) .dropna(subset=[x_col, y_col])) if len(base_df) < 2: continue ls = '--' if base == 'NVILA-ST' else '-' for i in range(len(base_df) - 1): x1, y1 = base_df.iloc[i][[x_col, y_col]] x2, y2 = base_df.iloc[i + 1][[x_col, y_col]] plt.annotate('', xy=(x2, y2), xytext=(x1, y1), arrowprops=dict(arrowstyle="->", color=COLORS[base], alpha=0.6, lw=1.8, ls=ls)) def _draw_3d_arrows(ax, df_src, base_list, x_col, y_col, z_col): """Draw arrows in a 3-D axes using quiver + dashed line for NVILA-ST.""" for base in base_list: order = ST_SCALE_ORDER if base == 'NVILA-ST' else SCALE_ORDER base_df = (df_src[(df_src['Base'] == base) & (df_src['Scale'].isin(order))] .set_index('Scale').reindex(order) .dropna(subset=[x_col, y_col, z_col])) if len(base_df) < 2: continue ls = '--' if base == 'NVILA-ST' else '-' for i in range(len(base_df) - 1): x1, y1, z1 = base_df.iloc[i][[x_col, y_col, z_col]] x2, y2, z2 = base_df.iloc[i + 1][[x_col, y_col, z_col]] dx, dy, dz = x2 - x1, y2 - y1, z2 - z1 # Draw the shaft as a line (supports dashes for NVILA-ST) ax.plot([x1, x2], [y1, y2], [z1, z2], color=COLORS[base], linestyle=ls, linewidth=2.0, alpha=0.6) # Draw an arrowhead at the end using quiver tip_frac = 0.25 # arrowhead covers 25 % of segment ax.quiver(x2 - tip_frac * dx, y2 - tip_frac * dy, z2 - tip_frac * dz, tip_frac * dx, tip_frac * dy, tip_frac * dz, color=COLORS[base], alpha=0.8, arrow_length_ratio=1.0, linewidth=0) # ── figures ────────────────────────────────────────────────────────────────── def figure1(df: pd.DataFrame, outdir: str): """Figure 1: Two Paths of Spatial Representation (scatter + arrows).""" plt.figure(figsize=(10, 8)) 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) # Arrows for all non-RoboRefer groups including NVILA-ST _draw_2d_arrows(df, ['Molmo', 'NVILA', 'NVILA-ST', 'Qwen'], 'Entanglement', 'Peak Dist') 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) 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 (Dist SC) [Higher is Better]', fontsize=12) plt.text(0.60, 0.05, "Mechanism 1:\nNaive Scaling\n(More Data = More Entanglement)", fontsize=12, transform=plt.gca().transAxes, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray')) plt.text(0.02, 0.72, "Mechanism 2:\nGenuine Understanding\n(Break the Shortcut)", fontsize=12, transform=plt.gca().transAxes, bbox=dict(facecolor='white', alpha=0.8, edgecolor='gray')) plt.tight_layout() plt.savefig(os.path.join(outdir, 'figure1_two_paths_trajectory.png'), dpi=300) plt.close() def figure2(df: pd.DataFrame, outdir: str): """Figure 2: Entanglement Paradox (line plot, NVILA-ST included).""" df_scale = df[df['Base'] != 'RoboRefer'].copy() df_scale = df_scale[df_scale['Scale'].isin(SCALE_ORDER)] df_scale['Scale'] = pd.Categorical(df_scale['Scale'], categories=SCALE_ORDER, ordered=True) plt.figure(figsize=(10, 6)) # Separate NVILA-ST for manual plotting to support dashed style df_main = df_scale[df_scale['Base'] != 'NVILA-ST'] df_st = df_scale[df_scale['Base'] == 'NVILA-ST'].sort_values('Scale') sns.lineplot(data=df_main, x='Scale', y='Entanglement', hue='Base', style='Base', palette={k: v for k, v in COLORS.items() if k != 'NVILA-ST'}, marker='o', linewidth=2.5, markersize=8, legend='auto') # Overlay NVILA-ST as dashed line if not df_st.empty: ax = plt.gca() ax.plot(df_st['Scale'].cat.codes if hasattr(df_st['Scale'], 'cat') else [SCALE_ORDER.index(s) for s in df_st['Scale']], df_st['Entanglement'], color=COLORS['NVILA-ST'], linestyle='--', linewidth=2.5, marker='o', markersize=8, label='NVILA-ST') df_robo = df[df['Base'] == 'RoboRefer'] 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(outdir, 'figure2_entanglement_paradox.png'), dpi=300) plt.close() def figure4(df: pd.DataFrame, outdir: str): """Figure 4: Consistent vs Counter Performance Gap (grouped bar).""" acc_models = [ 'Molmo vanilla', 'Molmo 2M', 'NVILA vanilla', 'NVILA 2M', 'NVILA-ST 80k', 'NVILA-ST 800k', 'Qwen vanilla', 'Qwen 2M', 'RoboRefer', ] available = [m for m in acc_models if m in df['Model'].values] df_acc = df.set_index('Model').loc[available] x = np.arange(len(available)) width = 0.35 fig, ax = plt.subplots(figsize=(14, 6)) ax.bar(x - width / 2, df_acc['EmbSpatial Acc (con)'], width, label='Consistent (Shortcut Works)', color='#2b83ba') ax.bar(x + width / 2, df_acc['EmbSpatial Acc (ctr)'], width, label='Counter (Shortcut Fails)', color='#d7191c') for i in range(len(available)): 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(available, rotation=45, ha='right') ax.legend() plt.tight_layout() plt.savefig(os.path.join(outdir, 'figure4_accuracy_gap.png'), dpi=300) plt.close() def figure5(df: pd.DataFrame, outdir: str): """Figure 5: Peak Dist vs EmbSpatial Counter Accuracy (scatter + arrows). NVILA-ST has N/A accuracy in the default CSV; their points/arrows will appear automatically if the data file contains numeric values for them. """ df_valid = df.dropna(subset=['EmbSpatial Acc (ctr)', 'Peak Dist']) plt.figure(figsize=(10, 8)) df_bases = df_valid[df_valid['Base'] != 'RoboRefer'] sns.scatterplot(data=df_bases, x='Peak Dist', y='EmbSpatial Acc (ctr)', hue='Base', palette=COLORS, s=150, edgecolor='black', alpha=0.8, legend=False) # Arrows for all groups including NVILA-ST _draw_2d_arrows(df_valid, ['Molmo', 'NVILA', 'NVILA-ST', 'Qwen'], 'Peak Dist', 'EmbSpatial Acc (ctr)') df_robo = df_valid[df_valid['Base'] == 'RoboRefer'] if not df_robo.empty: sns.scatterplot(data=df_robo, x='Peak Dist', y='EmbSpatial Acc (ctr)', hue='Base', palette=COLORS, s=400, marker='*', edgecolor='black', zorder=5, legend=False) if len(df_valid) > 1: corr = np.corrcoef(df_valid['Peak Dist'], df_valid['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(outdir, 'figure5_distSC_vs_counterAcc.png'), dpi=300) plt.close() def figure6(df: pd.DataFrame, outdir: str): """Figure 6: 3-D scatter (Peak Dist × Entanglement × Counter Acc) with arrows.""" df_valid = df.dropna(subset=['EmbSpatial Acc (ctr)', 'Peak Dist']) fig = plt.figure(figsize=(12, 10)) ax = fig.add_subplot(111, projection='3d') for base_name in df_valid['Base'].unique(): subset = df_valid[df_valid['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) # Arrows (shaft + quiver arrowhead) for trajectory _draw_3d_arrows(ax, df_valid, ['Molmo', 'NVILA', 'NVILA-ST', 'Qwen'], 'Peak Dist', 'Entanglement', 'EmbSpatial Acc (ctr)') 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(outdir, 'figure6_3d_distSC_ent_counterAcc.png'), dpi=300, bbox_inches='tight') plt.close() # ── main ───────────────────────────────────────────────────────────────────── def main(): 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() df = load_and_preprocess(args.data_path) os.makedirs(OUTPUT_DIR, exist_ok=True) sns.set_theme(style='whitegrid') figure1(df, OUTPUT_DIR) figure2(df, OUTPUT_DIR) figure4(df, OUTPUT_DIR) figure5(df, OUTPUT_DIR) figure6(df, OUTPUT_DIR) print(f"Saved figures to {OUTPUT_DIR}/") if __name__ == '__main__': main()