Datasets:
File size: 49,497 Bytes
3c4b9d3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 | # -*- coding: utf-8 -*-
"""
Plot evaluation result charts
1. Joint success rate by difficulty (LaTeX table)
2. Joint success rate by scenario (model-grouped and scenario-grouped bar charts)
3. Error reason pie charts (by model and by scenario)
4. Average tokens per trajectory vs joint success rate
5. Average conversation rounds per trajectory vs joint success rate
6. Average tool calls per trajectory vs joint success rate
"""
import os
import re
import json
import math
from collections import defaultdict
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
from matplotlib.offsetbox import OffsetImage, AnnotationBbox
from adjustText import adjust_text
# =========================
# Path Configuration
# =========================
CURRENT_DIR = os.path.dirname(os.path.abspath(__file__))
PROJECT_ROOT = os.path.dirname(CURRENT_DIR)
EVAL_RESULT_DIR = os.path.join(PROJECT_ROOT, "eval_result")
LOGO_DIR = os.path.join(CURRENT_DIR, "logo")
ERROR_ANALYSIS_DIR = os.path.join(PROJECT_ROOT, "error_analysis")
FIGURES_DIR = os.path.join(EVAL_RESULT_DIR, "figures")
os.makedirs(FIGURES_DIR, exist_ok=True)
def save_figure_pair(png_path, *, dpi=200, bbox_inches="tight"):
"""Save the current Matplotlib figure as both PNG and vector PDF."""
stem, extension = os.path.splitext(png_path)
if extension.lower() != ".png":
raise ValueError(f"Expected a .png output path, got: {png_path}")
plt.savefig(png_path, dpi=dpi, bbox_inches=bbox_inches)
plt.savefig(f"{stem}.pdf", bbox_inches=bbox_inches)
# =========================
# Global Plot Configuration
# =========================
plt.rcParams["font.sans-serif"] = [
"SimHei", "Microsoft YaHei", "Arial Unicode MS", "DejaVu Sans"
]
plt.rcParams["axes.unicode_minus"] = False
plt.rcParams["font.size"] = 18
# =========================
# Model Configuration
# =========================
MODEL_NAMES = {
"glm-5v-turbo": "GLM-5V-Turbo",
"qwen3-vl-235b": "Qwen3-VL-235B",
"qwen3.6-plus": "Qwen3.6-Plus",
"Qwen3.5-397B-A17B": "Qwen3.5-397B-A17B",
"gemini-3.1-pro-preview": "Gemini-3.1-Pro",
"kimi-k2.6": "Kimi-K2.6",
"mimo-v2.5-omni": "MiMo-V2.5-Omni",
"doubao-seed-2-0-pro-260215": "Doubao-seed-2-0-pro"
}
LOGO_FILES = {
"glm-5v-turbo": "chatglm-color.png",
"qwen3-vl-235b": "qwen-color.png",
"qwen3.6-plus": "qwen-color.png",
"Qwen3.5-397B-A17B": "qwen-color.png",
"gemini-3.1-pro-preview": "gemini-color.png",
"kimi-k2.6": "kimi.png",
"mimo-v2.5-omni": "xiaomimimo.png",
"doubao-seed-2-0-pro-260215": "doubao-color.png"
}
MODEL_COLORS = {
"glm-5v-turbo": "#E63946",
"qwen3-vl-235b": "#457B9D",
"qwen3.6-plus": "#2A9D8F",
"Qwen3.5-397B-A17B": "#E9C46A",
"gemini-3.1-pro-preview": "#9B59B6",
"kimi-k2.6": "#3498DB",
"mimo-v2.5-omni": "#FF6FA5",
"doubao-seed-2-0-pro-260215": "#1ABC9C"
}
LIGHT_COLORS = {
"glm-5v-turbo": "#F5A5A8",
"qwen3-vl-235b": "#8CBFD6",
"qwen3.6-plus": "#72D4CC",
"Qwen3.5-397B-A17B": "#F5DFA8",
"gemini-3.1-pro-preview": "#C9A3D4",
"kimi-k2.6": "#8DC6E8",
"mimo-v2.5-omni": "#F5A593",
"doubao-seed-2-0-pro-260215": "#6ED9CB"
}
MODEL_ORDER = list(MODEL_NAMES.keys())
SCENARIOS = ["retail", "restaurant", "kitchen", "warehouse", "household"]
DIFFICULTIES = ["easy", "hard", "static"]
# =========================
# Utility Functions
# =========================
def safe_read_json(path):
if not os.path.exists(path):
return None
try:
with open(path, "r", encoding="utf-8") as f:
return json.load(f)
except Exception as e:
print(f"[WARN] Failed to read JSON: {path}, error={e}")
return None
def list_model_dirs():
"""List model directories under eval_result that contain summary.json"""
models = []
if not os.path.exists(EVAL_RESULT_DIR):
return models
for name in os.listdir(EVAL_RESULT_DIR):
model_dir = os.path.join(EVAL_RESULT_DIR, name)
if os.path.isdir(model_dir) and os.path.exists(os.path.join(model_dir, "summary.json")):
models.append(name)
# Keep only models defined in MODEL_NAMES, sorted by predefined order, extras at the end
models = [m for m in models if m in MODEL_NAMES]
ordered = [m for m in MODEL_ORDER if m in models]
extras = [m for m in models if m not in ordered]
return ordered + sorted(extras)
def get_model_summary(model_dir):
path = os.path.join(EVAL_RESULT_DIR, model_dir, "summary.json")
return safe_read_json(path)
def get_model_eval_files(model_dir):
model_path = os.path.join(EVAL_RESULT_DIR, model_dir)
if not os.path.exists(model_path):
return []
files = []
for fn in os.listdir(model_path):
if fn.endswith("_eval.json"):
files.append(os.path.join(model_path, fn))
return sorted(files)
def mean_or_zero(values):
values = [v for v in values if v is not None]
return float(np.mean(values)) if values else 0.0
def normalize_rate(x):
"""
Some fields may be 0~1 or already 0~100.
Normalize to percentage values.
"""
if x is None:
return 0.0
try:
x = float(x)
except Exception:
return 0.0
if x <= 1.0:
return x * 100
return x
def is_complete_summary_item(item):
"""Exclude partial result files from every chart-level accuracy aggregate."""
if item.get("error"):
return False
if "included_in_summary" in item:
return bool(item["included_in_summary"])
total = item.get("total_scenarios", 0)
valid = item.get("valid_scenarios", 0)
return total > 0 and valid == total
def load_logo_image(model, zoom=0.12):
logo_file = LOGO_FILES.get(model)
if not logo_file:
return None
logo_path = os.path.join(LOGO_DIR, logo_file)
if not os.path.exists(logo_path):
return None
try:
img = Image.open(logo_path).convert("RGBA")
return OffsetImage(img, zoom=zoom)
except Exception as e:
print(f"[WARN] Failed to load logo: model={model}, error={e}")
return None
# =========================
# Metric Extraction
# =========================
def get_four_rates_by_difficulty(summary, model=None):
"""
Extract four metrics by difficulty: micro_accuracy, tool_based_success_rate, result_based_success_rate, joint_success_rate
Weighted average using valid_scenarios. Incomplete files are excluded first,
so valid_scenarios equals the selected ground-truth task count.
Returns: {difficulty: {"micro": x, "tool": x, "result": x, "joint": x}}
The mode labels come directly from the evaluated result filenames; no
model-specific reordering is applied.
"""
result = {d: {"micro": 0.0, "tool": 0.0, "result": 0.0, "joint": 0.0} for d in DIFFICULTIES}
if not summary:
return result
all_results = summary.get("all_results", [])
# Collect (rate, weight) pairs from complete files only.
buckets = {d: {"micro": [], "tool": [], "result": [], "joint": [], "weights": []} for d in DIFFICULTIES}
for item in all_results:
if not is_complete_summary_item(item):
continue
mode = item.get("mode")
if mode in DIFFICULTIES:
weight = item.get("valid_scenarios", 0)
buckets[mode]["weights"].append(weight)
buckets[mode]["micro"].append(normalize_rate(item.get("micro_accuracy", 0)))
buckets[mode]["tool"].append(normalize_rate(item.get("tool_based_success_rate", 0)))
buckets[mode]["result"].append(normalize_rate(item.get("result_based_success_rate", 0)))
buckets[mode]["joint"].append(normalize_rate(item.get("joint_success_rate", 0)))
for d in DIFFICULTIES:
total_weight = sum(buckets[d]["weights"])
if total_weight > 0:
for key in ["micro", "tool", "result", "joint"]:
result[d][key] = sum(r * w for r, w in zip(buckets[d][key], buckets[d]["weights"])) / total_weight
else:
for key in ["micro", "tool", "result", "joint"]:
result[d][key] = mean_or_zero(buckets[d][key])
return result
def get_joint_success_rate_by_difficulty(summary, model=None):
result = {d: 0.0 for d in DIFFICULTIES}
if not summary:
return result
rates = get_four_rates_by_difficulty(summary, model=model)
for d in DIFFICULTIES:
result[d] = rates[d]["joint"]
return result
def get_joint_success_rate_by_scenario(summary):
result = {s: 0.0 for s in SCENARIOS}
if not summary:
return result
all_results = summary.get("all_results", [])
# Collect (rate, weight) pairs from complete files only.
bucket = defaultdict(list)
for item in all_results:
if not is_complete_summary_item(item):
continue
scenario = item.get("scenario")
if scenario in SCENARIOS:
rate = normalize_rate(item.get("joint_success_rate", 0))
weight = item.get("valid_scenarios", 0)
bucket[scenario].append((rate, weight))
for s in SCENARIOS:
if bucket[s]:
total_weight = sum(w for _, w in bucket[s])
if total_weight > 0:
result[s] = sum(r * w for r, w in bucket[s]) / total_weight
else:
result[s] = mean_or_zero([r for r, _ in bucket[s]])
return result
def get_overall_joint_success_rate(summary):
if not summary:
return 0.0
complete_items = [
item for item in summary.get("all_results", [])
if is_complete_summary_item(item)
]
total = sum(item.get("valid_scenarios", 0) for item in complete_items)
if total <= 0:
return 0.0
return sum(
normalize_rate(item.get("joint_success_rate", 0))
* item.get("valid_scenarios", 0)
for item in complete_items
) / total
def get_metrics_from_summary(summary):
"""
Extract metrics from summary.json summary field
Returns: avg_tokens, avg_input_tokens, avg_output_tokens, avg_rounds, avg_tool_calls
"""
if not summary:
return {"avg_tokens": 0.0, "avg_input_tokens": 0.0, "avg_output_tokens": 0.0, "avg_rounds": 0.0, "avg_tool_calls": 0.0}
complete_items = [
item for item in summary.get("all_results", [])
if is_complete_summary_item(item)
]
if summary.get("all_results"):
total = sum(item.get("valid_scenarios", 0) for item in complete_items)
if total <= 0:
return {"avg_tokens": 0.0, "avg_input_tokens": 0.0, "avg_output_tokens": 0.0, "avg_rounds": 0.0, "avg_tool_calls": 0.0}
def weighted(key):
return sum(
(item.get(key, 0) or 0) * item.get("valid_scenarios", 0)
for item in complete_items
) / total
avg_input_tokens = weighted("avg_input_tokens")
avg_output_tokens = weighted("avg_output_tokens")
return {
"avg_tokens": float(avg_input_tokens + avg_output_tokens),
"avg_input_tokens": float(avg_input_tokens),
"avg_output_tokens": float(avg_output_tokens),
"avg_rounds": float(weighted("avg_rounds_count")),
"avg_tool_calls": float(weighted("avg_tool_calls_count")),
}
s = summary.get("summary", {})
# Token consumption = input + output
avg_input_tokens = s.get("avg_input_tokens", 0) or 0
avg_output_tokens = s.get("avg_output_tokens", 0) or 0
avg_tokens = avg_input_tokens + avg_output_tokens
# Conversation rounds
avg_rounds = s.get("avg_rounds_count", 0) or 0
# Tool call count
avg_tool_calls = s.get("avg_tool_calls_count", 0) or 0
return {
"avg_tokens": float(avg_tokens),
"avg_input_tokens": float(avg_input_tokens),
"avg_output_tokens": float(avg_output_tokens),
"avg_rounds": float(avg_rounds),
"avg_tool_calls": float(avg_tool_calls)
}
# =========================
# Axis Model Logo Labels
# =========================
def add_model_logos_below_axis(ax, models, y_offset_axes=-0.16, zoom=0.065, fontsize=10):
"""
Add logo + model name below x-axis
Use coordinate transform for stable layout
"""
xticks = ax.get_xticks()
if len(xticks) < len(models):
return
for i, model in enumerate(models):
x = xticks[i]
# logo
img = load_logo_image(model, zoom=zoom)
if img is not None:
ab = AnnotationBbox(
img,
(x, y_offset_axes + 0.035),
xycoords=ax.get_xaxis_transform(),
frameon=False,
box_alignment=(0.5, 0.5),
pad=0
)
ax.add_artist(ab)
# text
ax.text(
x, y_offset_axes - 0.02,
MODEL_NAMES.get(model, model),
transform=ax.get_xaxis_transform(),
ha="center", va="top",
fontsize=fontsize, fontweight="bold"
)
ax.set_xticklabels([])
plt.subplots_adjust(bottom=0.30)
# =========================
# Chart 1: LaTeX Table by Difficulty
# =========================
def _find_best_and_second(values):
"""Return (best_idx, second_idx), best is max value index, second is second-max value index"""
if not values:
return None, None
sorted_indices = sorted(range(len(values)), key=lambda i: values[i], reverse=True)
best = sorted_indices[0] if len(sorted_indices) > 0 else None
second = sorted_indices[1] if len(sorted_indices) > 1 else None
return best, second
def _fmt_val(val, is_best, is_second):
s = f"{val:.2f}"
if is_best:
return f"\\textbf{{{s}}}"
if is_second:
return f"\\underline{{{s}}}"
return s
def plot_chart1_joint_success_by_difficulty(models):
print("Generating Table 1: Four accuracy metrics by difficulty (LaTeX)")
# Collect four metrics per model per difficulty
model_data = {}
for model in models:
summary = get_model_summary(model)
model_data[model] = get_four_rates_by_difficulty(summary, model=model)
metrics = ["micro", "tool", "result", "joint"]
metric_labels = ["Micro", "Tool", "Result", "Joint"]
# Pre-compute best / second best per column
col_best = {} # (difficulty, metric) -> best_model_idx
col_second = {}
for d in DIFFICULTIES:
for m in metrics:
vals = [model_data[model][d][m] for model in models]
best, second = _find_best_and_second(vals)
col_best[(d, m)] = best
col_second[(d, m)] = second
# Build LaTeX
lines = []
lines.append(r"\begin{table*}[t]")
lines.append(r"\centering")
lines.append(r"% \caption{Performance by Difficulty on EgoBench.}")
lines.append(r"\label{tab:difficulty_results}")
lines.append(r"\large")
lines.append(r"\setlength{\tabcolsep}{4pt}")
lines.append(r"\begin{tabular}{clcccccccccccc}")
lines.append(r"\toprule")
lines.append(r" & & \multicolumn{4}{c}{\textbf{Easy}} & \multicolumn{4}{c}{\textbf{Hard}} & \multicolumn{4}{c}{\textbf{Static}} \\")
lines.append(r"\cmidrule(lr){3-6}\cmidrule(lr){7-10}\cmidrule(lr){11-14}")
lines.append(r" & \textbf{Model} & \textbf{Micro} & \textbf{Tool} & \textbf{Result} & \textbf{Joint} & \textbf{Micro} & \textbf{Tool} & \textbf{Result} & \textbf{Joint} & \textbf{Micro} & \textbf{Tool} & \textbf{Result} & \textbf{Joint} \\")
lines.append(r"\midrule")
for i, model in enumerate(models):
logo_file = LOGO_FILES.get(model, "")
logo_name = logo_file.replace(".png", "") if logo_file else ""
display_name = MODEL_NAMES.get(model, model)
row_vals = []
for d in DIFFICULTIES:
for j, m in enumerate(metrics):
val = model_data[model][d][m]
is_best = (col_best[(d, m)] == i)
is_second = (col_second[(d, m)] == i)
row_vals.append(_fmt_val(val, is_best, is_second))
logo_part = f"\\includegraphics[height=0.3cm]{{logo/{logo_name}}}" if logo_name else ""
line = f"{logo_part} & {display_name} & " + " & ".join(row_vals) + r" \\"
lines.append(line)
lines.append(r"\bottomrule")
lines.append(r"\end{tabular}")
lines.append(r"\end{table*}")
latex_str = "\n".join(lines)
save_path = os.path.join(FIGURES_DIR, "table1_difficulty_results.tex")
with open(save_path, "w", encoding="utf-8") as f:
f.write(latex_str)
print(f"LaTeX table saved to: {save_path}")
# =========================
# Chart 2: Bar Chart by Scenario
# =========================
def plot_chart2_joint_success_by_scenario(models):
print("Plotting Chart 2: Joint success rate by scenario (bar chart)")
x = np.arange(len(models))
width = 0.8 / len(SCENARIOS)
scenario_data = {s: [] for s in SCENARIOS}
for model in models:
summary = get_model_summary(model)
rates = get_joint_success_rate_by_scenario(summary)
for s in SCENARIOS:
scenario_data[s].append(rates[s])
fig, ax = plt.subplots(figsize=(15, 8))
colors = {
"retail": "#2E86AB",
"restaurant": "#E55934",
"kitchen": "#9BC53D",
"warehouse": "#7B61A8",
"household": "#F2B134",
}
for idx, scenario in enumerate(SCENARIOS):
ax.bar(
x + (idx - (len(SCENARIOS) - 1) / 2) * width,
scenario_data[scenario],
width=width,
label=scenario.capitalize(),
color=colors[scenario],
edgecolor="white",
linewidth=1.2
)
ax.set_ylabel("Joint Success Rate (%)", fontsize=27)
# No title
ax.set_xticks(x)
ax.grid(True, axis="y", linestyle="--", alpha=0.3)
ax.tick_params(axis="y", labelsize=22)
ax.legend(fontsize=21)
ymax = max([max(v) if v else 0 for v in scenario_data.values()] + [5])
ax.set_ylim(0, min(100, ymax + 10))
add_model_logos_below_axis(ax, models, y_offset_axes=-0.14, zoom=0.06, fontsize=10)
save_path = os.path.join(FIGURES_DIR, "chart2_joint_success_by_scenario.png")
save_figure_pair(save_path)
plt.close()
def plot_chart2b_joint_success_grouped_by_scenario(models):
"""Plot five scenario groups, with one bar per model in each group."""
print("Plotting Chart 2b: Joint success rate grouped by scenario (bar chart)")
x = np.arange(len(SCENARIOS))
width = 0.8 / len(models)
model_data = {}
for model in models:
summary = get_model_summary(model)
rates = get_joint_success_rate_by_scenario(summary)
model_data[model] = [rates[scenario] for scenario in SCENARIOS]
fig, ax = plt.subplots(figsize=(15, 8))
for idx, model in enumerate(models):
ax.bar(
x + (idx - (len(models) - 1) / 2) * width,
model_data[model],
width=width,
label=MODEL_NAMES.get(model, model),
color=MODEL_COLORS.get(model),
edgecolor="white",
linewidth=1.2,
)
ax.set_ylabel("Joint Success Rate (%)", fontsize=27)
ax.set_xticks(x)
ax.set_xticklabels([scenario.capitalize() for scenario in SCENARIOS], fontsize=22)
ax.grid(True, axis="y", linestyle="--", alpha=0.3)
ax.tick_params(axis="y", labelsize=22)
ax.legend(fontsize=15, ncol=2, loc="upper left")
ymax = max([max(values) if values else 0 for values in model_data.values()] + [5])
ax.set_ylim(0, min(100, ymax + 10))
save_path = os.path.join(
FIGURES_DIR, "chart2b_joint_success_grouped_by_scenario.png"
)
save_figure_pair(save_path)
plt.close()
# =========================
# Error Analysis Data Loading
# =========================
PIE_LABELS = ["Structural Non-Compliance", "Multimodal Perceptual Misinterpretations", "Hallucination", "Logical Fallacies", "Risky Operations", "Correct"]
PIE_KEYS = ["syntax", "multimodal", "hallucination", "logic", "over_action", "correct"]
# High-contrast color scheme; Correct: green
PIE_COLORS = ["#C62828", "#E64A19", "#FFA726", "#FDD835", "#7CB342", "#2E7D32"]
def _compute_pie_proportions(counts):
"""
Calculate proportions (percentages) from error type counts, summing to 100%.
Memory errors are merged into Logic.
counts: {"total": N, "syntax_error": n, ...}
Returns: {"syntax": %, "multimodal": %, "hallucination": %, "logic": %, "over_action": %, "correct": %}
"""
total = counts.get("total", 0)
if total == 0:
return {}
syntax = counts.get("syntax_error", 0)
multimodal = counts.get("multimodal_error", 0)
hallucination = counts.get("hallucination_error", 0)
logic = counts.get("logic_error", 0) + counts.get("memory_error", 0) # memory merged into logic
over_op = counts.get("over_operation_error", 0)
correct = counts.get("correct", 0)
props = {}
props["syntax"] = syntax / total * 100
props["multimodal"] = multimodal / total * 100
props["hallucination"] = hallucination / total * 100
props["logic"] = logic / total * 100
props["over_action"] = over_op / total * 100
props["correct"] = correct / total * 100
return props
def load_all_error_analysis():
"""
Read detailed results from error_analysis/all_scenarios_error_analysis_details.json,
aggregate error counts by scenario and model, then calculate pie chart proportions.
Returns:
(per_scenario, overall)
per_scenario: {scenario: {model: {"syntax": %, ...}}}
overall: {model: {"syntax": %, ...}}
"""
per_scenario = {s: defaultdict(dict) for s in SCENARIOS}
overall_counts = {}
json_path = os.path.join(ERROR_ANALYSIS_DIR, "all_scenarios_error_analysis_details.json")
if not os.path.exists(json_path):
print(f"[WARN] Error analysis JSON not found: {json_path}")
return per_scenario, {}
details = safe_read_json(json_path)
if not details:
return per_scenario, {}
# Aggregate error type counts by (scenario_type, model)
_EMPTY = lambda: {"total": 0, "syntax_error": 0, "multimodal_error": 0,
"hallucination_error": 0, "logic_error": 0, "memory_error": 0,
"over_operation_error": 0, "correct": 0}
counts = defaultdict(lambda: defaultdict(_EMPTY))
for entry in details:
if "error" in entry:
continue
scenario_type = entry.get("scenario_type")
model = entry.get("model")
if not scenario_type or not model:
continue
for r in entry.get("results", []):
error_type = r.get("error_type")
if error_type in ("missing_gt", "missing_eval"):
continue
counts[scenario_type][model]["total"] += 1
if error_type in counts[scenario_type][model]:
counts[scenario_type][model][error_type] += 1
# Calculate pie proportions per scenario per model
for scenario_type in SCENARIOS:
for model, cnt in counts[scenario_type].items():
per_scenario[scenario_type][model] = _compute_pie_proportions(cnt)
# Calculate pie proportions per model across scenarios
all_models = set()
for s in SCENARIOS:
all_models.update(counts[s].keys())
for model in all_models:
merged = _EMPTY()
for s in SCENARIOS:
if model in counts[s]:
for k in merged:
merged[k] += counts[s][model][k]
overall_counts[model] = _compute_pie_proportions(merged)
return per_scenario, overall_counts
# =========================
# Chart 3: Error Reason Pie Charts
# =========================
def _draw_single_pie(ax, props, model, logo_zoom=0.10,
name_fontsize=15, pct_fontsize=11):
"""
Draw a solid pie chart with the model logo and name below it.
props: {"syntax": %, "multimodal": %, ...} summing to 100%
"""
values = [props.get(k, 0) for k in PIE_KEYS]
# Filter out zero-value items
filtered_labels = []
filtered_values = []
filtered_colors = []
for i, v in enumerate(values):
if v > 0:
filtered_labels.append(PIE_LABELS[i])
filtered_values.append(v)
filtered_colors.append(PIE_COLORS[i])
if filtered_values:
_, _, autotexts = ax.pie(
filtered_values,
colors=filtered_colors,
autopct=lambda pct: f"{pct:.1f}" if pct >= 4 else "",
startangle=90,
pctdistance=0.70,
radius=0.78,
center=(0, 0.22),
wedgeprops=dict(edgecolor="white", linewidth=1.5),
textprops=dict(fontsize=pct_fontsize, fontweight="bold"),
)
for text in autotexts:
text.set_fontsize(pct_fontsize)
text.set_color("#333333")
else:
ax.text(0.5, 0.58, "No Data", transform=ax.transAxes,
ha="center", va="center", fontsize=name_fontsize)
# Keep the logo and model name outside the pie so the center remains solid.
img = load_logo_image(model, zoom=logo_zoom)
if img is not None:
ab = AnnotationBbox(
img, (0.5, 0.085), xycoords=ax.transAxes,
frameon=False, box_alignment=(0.5, 0.5), pad=0,
)
ax.add_artist(ab)
ax.text(0.5, -0.025, MODEL_NAMES.get(model, model),
transform=ax.transAxes, ha="center", va="top",
fontsize=name_fontsize, fontweight="bold", clip_on=False)
# Remove the generous default margins added by ``Axes.pie``. Keeping a
# little extra room at the bottom leaves space for the logo and name.
ax.set_xlim(-0.92, 0.92)
ax.set_ylim(-1.00, 1.15)
ax.set_aspect("equal")
ax.axis("off")
def plot_chart3a_pie_by_model(models):
"""Chart 3a: Overall error reason pies, with up to eight models per row."""
print(f"Plotting Chart 3a: Overall error reason pies for {len(models)} models")
_, overall_data = load_all_error_analysis()
ncols = min(8, max(1, len(models)))
nrows = math.ceil(len(models) / ncols)
fig, axes = plt.subplots(
nrows, ncols, figsize=(ncols * 1.85, nrows * 2.8),
gridspec_kw={"wspace": 0.0, "hspace": 0.06},
)
axes = np.asarray(axes, dtype=object).reshape(nrows, ncols)
for idx, model in enumerate(models):
r, c = divmod(idx, ncols)
if r >= nrows:
break
ax = axes[r][c]
props = overall_data.get(model, {})
_draw_single_pie(
ax, props, model,
logo_zoom=0.055, name_fontsize=11, pct_fontsize=10,
)
# Hide extra subplots
for idx in range(len(models), nrows * ncols):
r, c = divmod(idx, ncols)
axes[r][c].axis("off")
# Legend
legend_elements = [plt.matplotlib.patches.Patch(
facecolor=PIE_COLORS[i], edgecolor="white", label=PIE_LABELS[i]
) for i in range(len(PIE_LABELS))]
fig.legend(handles=legend_elements, loc="lower center", ncol=len(PIE_LABELS),
fontsize=11, frameon=True, bbox_to_anchor=(0.5, -0.02))
plt.subplots_adjust(left=0.005, right=0.995, top=0.995, bottom=0.20,
wspace=0.0, hspace=0.06)
save_path = os.path.join(FIGURES_DIR, "chart3a_error_pie_by_model.png")
save_figure_pair(save_path)
plt.close()
print(f"Saved: {save_path}")
def plot_chart3b_pie_by_model_scenario(models):
"""Chart 3b: Error reason pies by scenario and model."""
per_scenario, _ = load_all_error_analysis()
plotted_scenarios = [
scenario for scenario in SCENARIOS
if any(per_scenario[scenario].get(model) for model in models)
] or SCENARIOS
print(f"Plotting Chart 3b: Error reason pies for "
f"{len(plotted_scenarios)} scenarios")
scenario_labels = {
"retail": "Retail", "restaurant": "Restaurant",
"kitchen": "Kitchen", "warehouse": "Warehouse",
"household": "Household",
}
nrows = len(plotted_scenarios)
ncols = len(models)
fig, axes = plt.subplots(nrows, ncols,
figsize=(ncols * 1.9, nrows * 2.5))
for row, scenario in enumerate(plotted_scenarios):
for col, model in enumerate(models):
ax = axes[row][col] if nrows > 1 else axes[col]
props = per_scenario[scenario].get(model, {})
_draw_single_pie(ax, props, model,
logo_zoom=0.045, name_fontsize=8,
pct_fontsize=10)
# Add scenario label to first column
if col == 0:
ax.text(-0.25, 0.5, scenario_labels.get(scenario, scenario),
transform=ax.transAxes, ha="center", va="center",
fontsize=21, fontweight="bold", rotation=90)
# Legend
legend_elements = [plt.matplotlib.patches.Patch(
facecolor=PIE_COLORS[i], edgecolor="white", label=PIE_LABELS[i]
) for i in range(len(PIE_LABELS))]
fig.legend(handles=legend_elements, loc="lower center", ncol=len(PIE_LABELS),
fontsize=12, frameon=True, bbox_to_anchor=(0.5, -0.01))
plt.subplots_adjust(left=0.045, right=0.995, top=0.995, bottom=0.10,
wspace=0.0, hspace=0.04)
save_path = os.path.join(FIGURES_DIR, "chart3b_error_pie_by_model_scenario.png")
save_figure_pair(save_path)
plt.close()
print(f"Saved: {save_path}")
def plot_chart3_error_pie(models):
"""Plot all error reason pie charts."""
plot_chart3a_pie_by_model(models)
plot_chart3b_pie_by_model_scenario(models)
# =========================
# Scatter Plot Utility Functions
# =========================
def annotate_scatter_with_logo(ax, x, y, model, text_offset=(0, -12), zoom=0.11, fontsize=10):
img = load_logo_image(model, zoom=zoom)
if img is not None:
ab = AnnotationBbox(
img, (x, y),
frameon=False,
box_alignment=(0.5, 0.5),
pad=0
)
ax.add_artist(ab)
ax.annotate(
MODEL_NAMES.get(model, model),
(x, y),
textcoords="offset points",
xytext=text_offset,
ha="center",
va="top",
fontsize=fontsize,
fontweight="bold"
)
def annotate_scatter_points(ax, models, xs, ys, fontsize=16, top_only=True,
top_strict=False, color_overrides=None):
"""Deterministically place each model's label ABOVE its point with no overlaps.
color_overrides: optional {model: hex_color} to override the label text color
for specific models (e.g. render a single model's label in pink).
top_strict=True tightens top_only mode so the label sits DIRECTLY above its own
point and is never allowed to drift below any other point; crowded points trade
horizontal offset rather than dropping below.
Labels must sit above their own point (model name over the scatter dot) and not
cover any other point or label. Because several charts cluster points tightly
(e.g. conversation rounds all land near 4.5), apply a deterministic screen-space
placement instead of a force-directed packer:
- Isolated points keep their label close; crowded points push farther out.
- Candidate anchors are tried at increasing screen radii around each point along a
ring of angles; the nearest legal anchor that avoids point- and label-overlap is
chosen, with the search ordered so sparser points grab the near slots first.
- top_only=True restricts the candidate ring to the upper semicircle so every label
sits strictly above its own point (still nudged left/right to dodge overlaps).
- A final objective-driven local search rejects any label that ends up closer to
another point than that point's own label, and clears residual point/label hits.
- Thin leader lines tie each label back to its point; a white text halo keeps the
colored name readable over the grid.
"""
import matplotlib.patheffects as pe
fig = ax.figure
renderer = fig.canvas.get_renderer()
xs = np.asarray(xs, dtype=float)
ys = np.asarray(ys, dtype=float)
n = len(xs)
names = [MODEL_NAMES.get(m, m) for m in models]
colors = [MODEL_COLORS.get(m, "#000000") for m in models]
if color_overrides:
for i, m in enumerate(models):
if m in color_overrides:
colors[i] = color_overrides[m]
# Point centers in pixels.
def _to_pixels(xy):
return ax.transData.transform(np.asarray(xy, dtype=float))
pt_px = _to_pixels(list(zip(xs, ys)))
# Measure each label's size in pixels with a hidden text artist.
def _text_extent(txt):
bb = txt.get_window_extent(renderer=renderer)
return ((bb.x0 + bb.x1) / 2.0, (bb.y0 + bb.y1) / 2.0,
bb.x1 - bb.x0, bb.y1 - bb.y0)
label_sizes = []
tmp = ax.text(0, 0, "", fontsize=fontsize, fontweight="bold")
for nm in names:
tmp.set_text(nm)
_, _, w, h = _text_extent(tmp)
label_sizes.append((w, h))
tmp.remove()
# Axes size in pixels, for scaling candidate radii.
xlim = ax.get_xlim(); ylim = ax.get_ylim()
x0, y0 = _to_pixels([(xlim[0], ylim[0])])[0]
x1, y1 = _to_pixels([(xlim[1], ylim[1])])[0]
axes_w, axes_h = abs(x1 - x0), abs(y1 - y0)
axes_min = min(axes_w, axes_h)
pt_marker_r = 11.0 # half-extent of an s=150 marker on a 14x9in fig at 200 dpi
# Angles to try around each point. top_only keeps every candidate in the upper
# semicircle (screen-up), straight up first then fanning into the upper diagonals.
if top_only:
a = np.pi / 2
angles = [a,
a - np.pi / 8, a + np.pi / 8,
a - np.pi / 4, a + np.pi / 4,
a - 3 * np.pi / 8, a + 3 * np.pi / 8,
a - (np.pi / 2 - 0.18), a + (np.pi / 2 - 0.18)]
else:
angles = [0.0, np.pi, np.pi / 2, -np.pi / 2,
np.pi / 4, -np.pi / 4, 3 * np.pi / 4, -3 * np.pi / 4]
# Nearest-neighbor isolation in pixels (bigger = more isolated).
nn_dist = np.full(n, np.inf)
for i in range(n):
for j in range(n):
if i != j:
nn_dist[i] = min(nn_dist[i], np.hypot(pt_px[i, 0] - pt_px[j, 0],
pt_px[i, 1] - pt_px[j, 1]))
cand_radii = [axes_min * r for r in (0.05, 0.08, 0.11, 0.15, 0.20, 0.26, 0.33, 0.42, 0.52)]
# Axes pixel bounds (allow labels inside the axes rectangle, with small margin).
margin = 6.0
ax_bb = ax.get_window_extent(renderer=renderer)
ax_left = ax_bb.x0 + margin; ax_right = ax_bb.x1 - margin
ax_bottom = ax_bb.y0 + margin; ax_top = ax_bb.y1 - margin
placed = [None] * n # (cx, cy, w, h)
def overlaps_rect(cx, cy, w, h, other):
ox, oy, ow, oh = other
pad = 5.0
return (abs(cx - ox) < (w + ow) / 2 + pad and abs(cy - oy) < (h + oh) / 2 + pad)
def overlaps_any_point(cx, cy, w, h):
pad = 6.0
for j in range(n):
if (abs(cx - pt_px[j, 0]) < w / 2 + pt_marker_r + pad and
abs(cy - pt_px[j, 1]) < h / 2 + pt_marker_r + pad):
return True
return False
def out_of_bounds(cx, cy, w, h):
return (cx - w / 2 < ax_left or cx + w / 2 > ax_right or
cy - h / 2 < ax_bottom or cy + h / 2 > ax_top)
def label_label_ok(cx, cy, w, h, skip=None):
for k in range(n):
if k == skip or placed[k] is None:
continue
if overlaps_rect(cx, cy, w, h, placed[k]):
return False
return True
# Isolated points first so they claim the near slots; crowded points take the
# farther rings their neighbors haven't used.
order = sorted(range(n), key=lambda i: -nn_dist[i])
for i in order:
w, h = label_sizes[i]
chosen = None
fallback = None # in-bounds-but-overlapping last resort
for r in cand_radii:
for ang in angles:
cx = pt_px[i, 0] + r * np.cos(ang)
cy = pt_px[i, 1] + r * np.sin(ang)
ob = out_of_bounds(cx, cy, w, h)
if overlaps_any_point(cx, cy, w, h):
continue
lo = label_label_ok(cx, cy, w, h, skip=i)
if ob:
if lo and fallback is None:
fallback = (cx, cy, w, h)
continue
if not lo:
continue
chosen = (cx, cy, w, h)
break
if chosen is not None:
break
if chosen is None:
if fallback is not None:
chosen = fallback
else:
r = cand_radii[-1]
# last resort: directly above the point
chosen = (pt_px[i, 0], pt_px[i, 1] + r, w, h)
placed[i] = chosen
# Objective-driven local search: minimize (nearest-violations, point-overlaps,
# label-overlaps, out-of-bounds) so a label never lands closer to a neighbor's
# point than that neighbor's own label.
fine_angles = np.linspace(0, 2 * np.pi, 24, endpoint=False)
if top_only:
fine_angles = fine_angles[np.sin(fine_angles) > 0.25]
def score(layout):
s_viol = 0
for j in range(n):
cx, cy, w, h = layout[j]
dj = np.hypot(cx - pt_px[j, 0], cy - pt_px[j, 1])
best_other = min(
np.hypot(layout[i][0] - pt_px[j, 0], layout[i][1] - pt_px[j, 1])
for i in range(n) if i != j)
if best_other + 14 < dj:
s_viol += 1
s_pt = sum(1 for i in range(n) if overlaps_any_point(*layout[i]))
s_ll = sum(1 for i in range(n) for k in range(n)
if k < i and overlaps_rect(*layout[i], layout[k]))
s_oob = sum(1 for i in range(n) if out_of_bounds(*layout[i]))
return (s_viol, s_pt, s_ll, s_oob)
cur = list(placed)
cur_score = score(cur)
improved = True
for _ in range(40):
if not improved or cur_score[:3] == (0, 0, 0):
break
improved = False
for i in range(n):
w, h = label_sizes[i]
best = cur[i]; best_s = cur_score
for r in cand_radii:
for ang in fine_angles:
nx = pt_px[i, 0] + r * np.cos(ang)
ny = pt_px[i, 1] + r * np.sin(ang)
cand = (nx, ny, w, h)
trial = list(cur); trial[i] = cand
s = score(trial)
if s < best_s or (s == best_s and
np.hypot(nx - pt_px[i, 0], ny - pt_px[i, 1]) <
np.hypot(best[0] - pt_px[i, 0], best[1] - pt_px[i, 1])):
best = cand; best_s = s
if best is not cur[i]:
cur[i] = best; cur_score = best_s; improved = True
placed = cur
# Pixel centers back to data coords; draw labels + leader lines.
inv = ax.transData.inverted()
for i in range(n):
cx, cy, w, h = placed[i]
dx, dy = inv.transform((cx, cy))
text = ax.text(
dx, dy, names[i],
fontsize=fontsize, fontweight="bold", color=colors[i],
ha="center", va="center", zorder=12,
path_effects=[pe.withStroke(linewidth=3.5, foreground="white")],
)
ax.plot([xs[i], dx], [ys[i], dy],
color="gray", lw=0.6, alpha=0.5, zorder=6,
solid_capstyle="round")
def get_axis_padding(vals, ratio=0.08):
vals = [v for v in vals if v is not None]
if not vals:
return 0, 1
vmin = min(vals)
vmax = max(vals)
if math.isclose(vmin, vmax):
pad = vmax * ratio if vmax != 0 else 1
else:
pad = (vmax - vmin) * ratio
return vmin - pad, vmax + pad
# =========================
# Chart 4: Tokens vs Success
# =========================
def plot_chart4_tokens_vs_success(models):
print("Plotting Chart 4: Avg token consumption vs joint success rate")
xs, ys = [], []
for model in models:
summary = get_model_summary(model)
metrics = get_metrics_from_summary(summary)
xs.append(metrics["avg_tokens"])
ys.append(get_overall_joint_success_rate(summary))
fig, ax = plt.subplots(figsize=(14, 9))
# MiMo (mimo-v2.5-omni) is highlighted in pink per request.
MIMO_PINK = "#FF6FA5"
# Draw with large points
for model, x, y in zip(models, xs, ys):
color = MODEL_COLORS.get(model, "#000000")
if model == "mimo-v2.5-omni":
color = MIMO_PINK
ax.scatter(x, y, c=color, s=150, edgecolors="white", linewidths=1.5, zorder=5)
ax.set_xlabel("Average Tokens per Trajectory (Input + Output)", fontsize=28)
ax.set_ylabel("Joint Success Rate (%)", fontsize=28)
ax.tick_params(labelsize=22)
ax.grid(True, linestyle="--", alpha=0.3)
# Set x-axis to k units
ax.xaxis.set_major_formatter(plt.FuncFormatter(lambda x, p: f"{int(round(x/1000))}k"))
xmin, xmax = get_axis_padding(xs)
ymin, ymax = get_axis_padding(ys)
ax.set_xlim(max(0, xmin), xmax)
# Extra headroom on top so labels placed strictly above their points never clip.
ax.set_ylim(max(0, ymin - 2), min(100, ymax + 12))
annotate_scatter_points(
ax, models, xs, ys, fontsize=16,
top_strict=True,
color_overrides={"mimo-v2.5-omni": MIMO_PINK},
)
save_path = os.path.join(FIGURES_DIR, "chart4_tokens_vs_success.png")
save_figure_pair(save_path)
plt.close()
# =========================
# Chart 4a: Input Tokens vs Success
# =========================
def plot_chart4a_input_tokens_vs_success(models):
print("Plotting Chart 4a: Avg input token consumption vs joint success rate")
xs, ys = [], []
for model in models:
summary = get_model_summary(model)
metrics = get_metrics_from_summary(summary)
xs.append(metrics["avg_input_tokens"])
ys.append(get_overall_joint_success_rate(summary))
fig, ax = plt.subplots(figsize=(14, 9))
# Draw with large points
for model, x, y in zip(models, xs, ys):
color = MODEL_COLORS.get(model, "#000000")
ax.scatter(x, y, c=color, s=150, edgecolors="white", linewidths=1.5, zorder=5)
ax.set_xlabel("Average Input Tokens per Trajectory", fontsize=28)
ax.set_ylabel("Joint Success Rate (%)", fontsize=28)
ax.tick_params(labelsize=22)
ax.grid(True, linestyle="--", alpha=0.3)
# Set x-axis to k units
ax.xaxis.set_major_formatter(plt.FuncFormatter(lambda x, p: f"{int(round(x/1000))}k"))
xmin, xmax = get_axis_padding(xs)
ymin, ymax = get_axis_padding(ys)
ax.set_xlim(max(0, xmin), xmax)
ax.set_ylim(max(0, ymin - 2), min(100, ymax + 9))
annotate_scatter_points(ax, models, xs, ys, fontsize=16)
save_path = os.path.join(FIGURES_DIR, "chart4a_input_tokens_vs_success.png")
save_figure_pair(save_path)
plt.close()
# =========================
# Chart 4b: Output Tokens vs Success
# =========================
def plot_chart4b_output_tokens_vs_success(models):
print("Plotting Chart 4b: Avg output token consumption vs joint success rate")
xs, ys = [], []
for model in models:
summary = get_model_summary(model)
metrics = get_metrics_from_summary(summary)
xs.append(metrics["avg_output_tokens"])
ys.append(get_overall_joint_success_rate(summary))
fig, ax = plt.subplots(figsize=(14, 9))
# Draw with large points
for model, x, y in zip(models, xs, ys):
color = MODEL_COLORS.get(model, "#000000")
ax.scatter(x, y, c=color, s=150, edgecolors="white", linewidths=1.5, zorder=5)
ax.set_xlabel("Average Output Tokens per Trajectory", fontsize=28)
ax.set_ylabel("Joint Success Rate (%)", fontsize=28)
ax.tick_params(labelsize=22)
ax.grid(True, linestyle="--", alpha=0.3)
# Set x-axis to k units
ax.xaxis.set_major_formatter(plt.FuncFormatter(lambda x, p: f"{int(round(x/1000))}k"))
xmin, xmax = get_axis_padding(xs)
ymin, ymax = get_axis_padding(ys)
ax.set_xlim(max(0, xmin), xmax)
ax.set_ylim(max(0, ymin - 2), min(100, ymax + 9))
annotate_scatter_points(ax, models, xs, ys, fontsize=16)
save_path = os.path.join(FIGURES_DIR, "chart4b_output_tokens_vs_success.png")
save_figure_pair(save_path)
plt.close()
# =========================
# Chart 5: Rounds vs Success
# =========================
def plot_chart5_rounds_vs_success(models):
print("Plotting Chart 5: Avg conversation rounds vs joint success rate")
xs, ys = [], []
for model in models:
summary = get_model_summary(model)
metrics = get_metrics_from_summary(summary)
xs.append(metrics["avg_rounds"])
ys.append(get_overall_joint_success_rate(summary))
fig, ax = plt.subplots(figsize=(14, 9))
# Draw with large points
for model, x, y in zip(models, xs, ys):
color = MODEL_COLORS.get(model, "#000000")
ax.scatter(x, y, c=color, s=150, edgecolors="white", linewidths=1.5, zorder=5)
ax.set_xlabel("Average User-Agent Conversation Rounds per Trajectory", fontsize=28)
ax.set_ylabel("Joint Success Rate (%)", fontsize=28)
ax.tick_params(labelsize=22)
ax.grid(True, linestyle="--", alpha=0.3)
xmin, xmax = get_axis_padding(xs)
ymin, ymax = get_axis_padding(ys)
ax.set_xlim(max(0, xmin), xmax)
ax.set_ylim(max(0, ymin - 2), min(100, ymax + 9))
annotate_scatter_points(ax, models, xs, ys, fontsize=16)
save_path = os.path.join(FIGURES_DIR, "chart5_rounds_vs_success.png")
save_figure_pair(save_path)
plt.close()
# =========================
# Chart 6: Tool Calls vs Success
# =========================
def plot_chart6_tool_calls_vs_success(models):
print("Plotting Chart 6: Avg tool calls vs joint success rate")
xs, ys = [], []
for model in models:
summary = get_model_summary(model)
metrics = get_metrics_from_summary(summary)
xs.append(metrics["avg_tool_calls"])
ys.append(get_overall_joint_success_rate(summary))
fig, ax = plt.subplots(figsize=(14, 9))
# Draw with large points
for model, x, y in zip(models, xs, ys):
color = MODEL_COLORS.get(model, "#000000")
ax.scatter(x, y, c=color, s=150, edgecolors="white", linewidths=1.5, zorder=5)
ax.set_xlabel("Average Tool Calls per Trajectory", fontsize=28)
ax.set_ylabel("Joint Success Rate (%)", fontsize=28)
ax.tick_params(labelsize=22)
ax.grid(True, linestyle="--", alpha=0.3)
xmin, xmax = get_axis_padding(xs)
ymin, ymax = get_axis_padding(ys)
ax.set_xlim(max(0, xmin), xmax)
ax.set_ylim(max(0, ymin - 2), min(100, ymax + 9))
annotate_scatter_points(ax, models, xs, ys, fontsize=16)
save_path = os.path.join(FIGURES_DIR, "chart6_tool_calls_vs_success.png")
save_figure_pair(save_path)
plt.close()
# =========================
# Main Function
# =========================
def main():
import argparse
parser = argparse.ArgumentParser(description="Plot evaluation result charts")
parser.add_argument("--eval_root", type=str, default=None,
help="Root dir of per-model eval output (default: eval_result). "
"Set to GPT_user_eval_result to plot the GPT-user run; figures "
"are written under <eval_root>/figures.")
parser.add_argument("--error_analysis_root", type=str, default=None,
help="Root dir of error-analysis details JSON (default: error_analysis). "
"Must match the --output_root used by analyze_error_reasons.py.")
args = parser.parse_args()
global EVAL_RESULT_DIR, ERROR_ANALYSIS_DIR, FIGURES_DIR
if args.eval_root is not None:
EVAL_RESULT_DIR = os.path.join(PROJECT_ROOT, args.eval_root)
FIGURES_DIR = os.path.join(EVAL_RESULT_DIR, "figures")
if args.error_analysis_root is not None:
ERROR_ANALYSIS_DIR = os.path.join(PROJECT_ROOT, args.error_analysis_root)
os.makedirs(FIGURES_DIR, exist_ok=True)
print("=" * 60)
print("Starting evaluation result chart generation")
print(f"EVAL_RESULT_DIR : {EVAL_RESULT_DIR}")
print(f"LOGO_DIR : {LOGO_DIR}")
print(f"ERROR_ANALYSIS_DIR: {ERROR_ANALYSIS_DIR}")
print(f"FIGURES_DIR : {FIGURES_DIR}")
print("=" * 60)
models = list_model_dirs()
if not models:
print("[ERROR] No valid model directories found under eval_result (containing summary.json)")
return
print("Detected models:")
for m in models:
print(f" - {m}")
plot_chart1_joint_success_by_difficulty(models)
plot_chart2_joint_success_by_scenario(models)
plot_chart2b_joint_success_grouped_by_scenario(models)
plot_chart3_error_pie(models)
plot_chart4_tokens_vs_success(models)
plot_chart4a_input_tokens_vs_success(models)
plot_chart4b_output_tokens_vs_success(models)
plot_chart5_rounds_vs_success(models)
plot_chart6_tool_calls_vs_success(models)
print("=" * 60)
print("All charts generated.")
print(f"Output directory: {FIGURES_DIR}")
print("=" * 60)
if __name__ == "__main__":
main()
|