| |
| """Generate the figures of the paper from results.json (build_data.py) into <output>/figures/.""" |
| import json |
| import os |
|
|
| import matplotlib |
| matplotlib.use('Agg') |
| import matplotlib.pyplot as plt |
| from matplotlib.lines import Line2D |
| from matplotlib.patches import Patch |
| import numpy as np |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| RES_DIR = OUT_ROOT = os.path.join(HERE, 'output') |
| FIG = os.path.join(OUT_ROOT, 'figures') |
| os.makedirs(FIG, exist_ok=True) |
| R = json.load(open(os.path.join(RES_DIR, 'results.json'))) |
| CF = R['configs'] |
|
|
| OPEN = ['qwen-3.5-27b', 'qwen-3.5-122b-a10b', 'gemma-4-31b', 'glm-4.7-flash', |
| 'gpt-oss-20b', 'gpt-oss-120b', 'minimax-m2.5'] |
| FRONTIER = ['seed-2.0', 'gpt-5.4'] |
| DISP = {'qwen-3.5-27b': 'Qwen3.5-27B', 'qwen-3.5-122b-a10b': 'Qwen3.5-122B-A10B', 'gemma-4-31b': 'Gemma4-31B', |
| 'glm-4.7-flash': 'GLM-4.7-Flash', 'gpt-oss-20b': 'GPT-OSS-20B', 'gpt-oss-120b': 'GPT-OSS-120B', |
| 'minimax-m2.5': 'MiniMax-M2.5', 'seed-2.0': 'Seed-2.0', 'gpt-5.4': 'GPT-5.4'} |
| METHODS = ['rag-nothink', 'rag-think', 'miniswe', 'openhands', 'claude'] |
| AGENTS = ['miniswe', 'openhands', 'claude'] |
| MDISP = {'rag-nothink': 'RAG-NoThink', 'rag-think': 'RAG-Think', 'miniswe': 'Mini-SWE-Agent', |
| 'openhands': 'OpenHands', 'claude': 'Claude-Code', 'rag': 'RAG'} |
| |
| |
| MCOL = {'rag-nothink': '#2a78d6', 'rag-think': '#4a3aa7', 'miniswe': '#eb6834', |
| 'openhands': '#1baf7a', 'claude': '#e34948', 'rag': '#2a78d6'} |
| MMARK = {'rag-nothink': 'o', 'rag-think': 's', 'miniswe': '^', 'openhands': 'D', 'claude': 'v', 'rag': 'o'} |
| CATS = ['Image', 'Video/Audio', 'Documents', 'Fonts', 'Data Formats', 'Compilers/Interpreters', |
| 'Binaries', 'Network', 'Others'] |
| CAT_SHORT = {'Image': 'Image', 'Video/Audio': 'A/V', 'Documents': 'Doc', 'Fonts': 'Font', |
| 'Data Formats': 'Data', 'Compilers/Interpreters': 'Code', 'Binaries': 'Bin', |
| 'Network': 'Net', 'Others': 'Other'} |
| INK, INK2, GRID = '#0b0b0b', '#52514e', '#e4e3df' |
|
|
| plt.rcParams.update({ |
| 'font.size': 10, 'axes.edgecolor': '#b5b4ae', 'axes.labelcolor': INK, 'xtick.color': INK2, |
| 'ytick.color': INK2, 'axes.spines.top': False, 'axes.spines.right': False, |
| 'axes.grid': True, 'grid.color': GRID, 'grid.linewidth': 0.6, 'axes.axisbelow': True, |
| 'legend.frameon': False, 'pdf.fonttype': 42, 'ps.fonttype': 42, |
| }) |
|
|
|
|
| def cid(model, mk): |
| return f'{mk}|{model}' |
|
|
|
|
| def best_open(mk): |
| """best open-weight LLM for a method, by pass@k_max.""" |
| return max(OPEN, key=lambda m: (CF[cid(m, mk)]['pass'][str(CF[cid(m, mk)]['N'])], |
| CF[cid(m, mk)]['pass']['1'])) |
|
|
|
|
| def best_any(mk): |
| models = OPEN if mk.startswith('rag') else OPEN + FRONTIER |
| return max(models, key=lambda m: CF[cid(m, mk)]['pass']['1']) |
|
|
|
|
| def curve(c, field, cat=None): |
| N = c['N'] |
| if cat is None: |
| return [c[field][str(k)] for k in range(1, N + 1)] |
| return [c['per_category_pass'][cat][str(k)] for k in range(1, N + 1)] |
|
|
|
|
| def draw_kcurves(ax, field='pass', cat=None, legend=True, small=False, open_mode='best'): |
| """solid: open-weight curve per method ('best' LLM, or 'mean' over the 7 LLMs); |
| dashed: Seed-2.0 per agent; star: GPT-5.4 per agent (k=1 only).""" |
| ms = 5 if small else 6 |
| for mk in METHODS: |
| if open_mode == 'best': |
| m = best_open(mk) |
| y = curve(CF[cid(m, mk)], field, cat) |
| lab = f'{MDISP[mk]} ({DISP[m]})' |
| else: |
| ys = [curve(CF[cid(m, mk)], field, cat) for m in OPEN] |
| y = list(np.mean(ys, axis=0)) |
| lab = MDISP[mk] |
| ax.plot(range(1, len(y) + 1), y, color=MCOL[mk], marker=MMARK[mk], ms=ms, lw=1.6, label=lab) |
| for mk in AGENTS: |
| y = curve(CF[cid('seed-2.0', mk)], field, cat) |
| ax.plot(range(1, len(y) + 1), y, color=MCOL[mk], marker=MMARK[mk], ms=ms, lw=1.4, |
| ls='--', mfc='white') |
| y = curve(CF[cid('gpt-5.4', mk)], field, cat) |
| ax.plot([1], y, color=MCOL[mk], marker='*', ms=ms + 6, ls='none', mec='white', mew=0.6) |
| ax.set_xticks(range(1, 6)) |
| if legend: |
| h, l = ax.get_legend_handles_labels() |
| style = [Line2D([], [], color=INK2, lw=1.6, |
| label='best open-weight LLM' if open_mode == 'best' else 'mean of 7 open-weight LLMs'), |
| Line2D([], [], color=INK2, lw=1.4, ls='--', marker='o', mfc='white', label='Seed-2.0'), |
| Line2D([], [], color=INK2, ls='none', marker='*', ms=12, label='GPT-5.4 (1 sample)')] |
| return h + style |
| return None |
|
|
|
|
| def fig_kcurves(field, fname, ylabel, loc='upper right'): |
| fig, ax = plt.subplots(figsize=(6.3, 3.9)) |
| hs = draw_kcurves(ax, field) |
| ax.set_xlabel('k') |
| ax.set_ylabel(ylabel) |
| ax.set_xlim(0.8, 5.2) |
| ax.legend(handles=hs, loc=loc, fontsize=8, handlelength=2.0, labelspacing=0.3) |
| fig.tight_layout() |
| fig.savefig(os.path.join(FIG, fname), bbox_inches='tight') |
| plt.close(fig) |
|
|
|
|
| def fig_funnel(): |
| rows = [('openhands', 'gpt-oss-120b'), ('miniswe', 'minimax-m2.5'), ('claude', 'gpt-oss-120b'), |
| ('claude', 'minimax-m2.5'), ('openhands', 'seed-2.0'), ('openhands', 'gpt-5.4'), |
| ('claude', 'gpt-5.4')] |
| bands = ['pass', 'reach \\ pass', 'sfunc \\ reach', 'miss function', 'no output'] |
| cols = ['#1c5cab', '#5598e7', '#9ec5f4', '#d9d8d3', '#7a7974'] |
| fig, ax = plt.subplots(figsize=(8.3, 2.35)) |
| ys = np.arange(len(rows))[::-1] |
| for y, (mk, m) in zip(ys, rows): |
| f = CF[cid(m, mk)]['funnel'] |
| left = 0.0 |
| for b, c in zip(bands, cols): |
| ax.barh(y, f[b], left=left, color=c, height=0.72, edgecolor='white', linewidth=1.0) |
| left += f[b] |
| ax.text(f['pass'] / 2, y, f"{f['pass']:.1f}", ha='center', va='center', color='white', fontsize=8) |
| ax.axhline(ys[3] - 0.5, color=INK2, lw=0.8, ls=':') |
| ax.set_yticks(ys) |
| ax.set_yticklabels([f'{MDISP[mk]} | {DISP[m]}' for mk, m in rows], fontsize=9) |
| ax.set_xlim(0, 100) |
| ax.grid(axis='y', visible=False) |
| ax.legend(handles=[Patch(color=c, label=b) for b, c in zip(bands, cols)], ncol=5, |
| loc='upper center', bbox_to_anchor=(0.5, -0.16), fontsize=9) |
| fig.tight_layout() |
| fig.savefig(os.path.join(FIG, 'f2_funnel.pdf'), bbox_inches='tight') |
| plt.close(fig) |
|
|
|
|
| def fig_method_cat(): |
| rows = [(mk, best_any(mk)) for mk in METHODS] |
| M = np.array([[CF[cid(m, mk)]['per_category_pass'][c]['1'] for c in CATS] + [CF[cid(m, mk)]['pass']['1']] |
| for mk, m in rows]) |
| fig, ax = plt.subplots(figsize=(9.2, 2.45)) |
| im = ax.imshow(M, cmap='Blues', vmin=0, vmax=max(70, M.max()), aspect='auto') |
| for i in range(M.shape[0]): |
| for j in range(M.shape[1]): |
| v = M[i, j] |
| ax.text(j, i, f'{v:.1f}', ha='center', va='center', fontsize=8.5, |
| color='white' if v > 38 else INK) |
| ax.set_xticks(range(len(CATS) + 1)) |
| ax.set_xticklabels([CAT_SHORT[c] for c in CATS] + ['Overall']) |
| ax.set_yticks(range(len(rows))) |
| ax.set_yticklabels([f'{MDISP[mk]}\n({DISP[m]})' for mk, m in rows], fontsize=8.5) |
| ax.grid(False) |
| ax.axvline(len(CATS) - 0.5, color='white', lw=2) |
| cb = fig.colorbar(im, ax=ax, fraction=0.025, pad=0.01) |
| cb.set_label('pass@1 (%)') |
| fig.tight_layout() |
| fig.savefig(os.path.join(FIG, 'f3_method_cat.pdf'), bbox_inches='tight') |
| plt.close(fig) |
|
|
|
|
| def fig_upset(key, set_order, set_labels, fname): |
| inter = R[key] |
| items = [(k, v) for k, v in inter.items() if k != '(none)'] |
| items.sort(key=lambda x: -x[1]) |
| items.append(('(none)', inter.get('(none)', 0))) |
| fig = plt.figure(figsize=(10.9, 3.0)) |
| gs = fig.add_gridspec(2, 1, height_ratios=[2.3, 1.1], hspace=0.05) |
| ax = fig.add_subplot(gs[0]) |
| ax2 = fig.add_subplot(gs[1], sharex=ax) |
| x = np.arange(len(items)) |
| cols = ['#3d3c39'] * (len(items) - 1) + ['#e34948'] |
| ax.bar(x, [v for _, v in items], color=cols, width=0.72) |
| for xi, (_, v) in zip(x, items): |
| ax.text(xi, v + 3, str(v), ha='center', va='bottom', fontsize=8.5, color=INK) |
| ax.set_ylabel('# tasks') |
| ax.tick_params(axis='x', bottom=False, labelbottom=False) |
| ax.set_ylim(0, max(v for _, v in items) * 1.15) |
| ax.grid(axis='x', visible=False) |
| for yi, s in enumerate(set_order): |
| yy = len(set_order) - 1 - yi |
| for xi, (k, _) in zip(x, items): |
| on = k != '(none)' and s in k.split('|') |
| ax2.scatter(xi, yy, s=55, color=MCOL[s] if on else '#dcdbd6', zorder=3) |
| for xi, (k, _) in zip(x, items): |
| if k == '(none)': |
| continue |
| ys = [len(set_order) - 1 - set_order.index(s) for s in k.split('|')] |
| if len(ys) > 1: |
| ax2.plot([xi, xi], [min(ys), max(ys)], color='#3d3c39', lw=1.6, zorder=2) |
| ax2.set_yticks(range(len(set_order))) |
| ax2.set_yticklabels(set_labels[::-1], fontsize=9) |
| ax2.set_ylim(-0.6, len(set_order) - 0.4) |
| ax2.tick_params(axis='x', bottom=False, labelbottom=False) |
| ax2.grid(False) |
| for sp in ('left', 'bottom'): |
| ax2.spines[sp].set_visible(False) |
| fig.savefig(os.path.join(FIG, fname), bbox_inches='tight') |
| plt.close(fig) |
|
|
|
|
| def jaccard(a, b): |
| a, b = set(a), set(b) |
| return len(a & b) / len(a | b) if a | b else 0.0 |
|
|
|
|
| def fig_jaccard(): |
| solved = {k: set(v['solved']) for k, v in CF.items()} |
| msets = {'rag': set().union(*[solved[cid(m, mk)] for m in OPEN for mk in ('rag-think', 'rag-nothink')])} |
| for mk in AGENTS: |
| msets[mk] = set().union(*[solved[cid(m, mk)] for m in OPEN + FRONTIER]) |
| mo = ['rag'] + AGENTS |
| A = np.array([[jaccard(msets[a], msets[b]) for b in mo] for a in mo]) |
| models = OPEN + FRONTIER |
| best = {} |
| for m in models: |
| mks = AGENTS if m in FRONTIER else METHODS |
| mk = max(mks, key=lambda x: (CF[cid(m, x)]['pass'][str(CF[cid(m, x)]['N'])], CF[cid(m, x)]['pass']['1'])) |
| best[m] = solved[cid(m, mk)] |
| B = np.array([[jaccard(best[a], best[b]) for b in models] for a in models]) |
| fig, (a1, a2) = plt.subplots(1, 2, figsize=(11.5, 4.6), gridspec_kw={'width_ratios': [1, 1.45]}) |
| for ax, M, labs, title in ((a1, A, [MDISP[m] for m in mo], 'Method-level Jaccard'), |
| (a2, B, [DISP[m] for m in models], 'Model-level Jaccard (best method per model)')): |
| im = ax.imshow(M, cmap='Blues', vmin=0, vmax=1) |
| for i in range(M.shape[0]): |
| for j in range(M.shape[1]): |
| ax.text(j, i, f'{M[i, j]:.2f}', ha='center', va='center', fontsize=8 if M.shape[0] > 5 else 9.5, |
| color='white' if M[i, j] > 0.6 else INK) |
| ax.set_xticks(range(len(labs))) |
| ax.set_xticklabels(labs, rotation=25, ha='right', fontsize=9) |
| ax.set_yticks(range(len(labs))) |
| ax.set_yticklabels(labs, fontsize=9) |
| ax.grid(False) |
| ax.set_title(title, fontsize=11) |
| fig.colorbar(im, ax=ax, fraction=0.046, pad=0.03) |
| fig.tight_layout() |
| fig.savefig(os.path.join(FIG, 'fa3b_jaccard.pdf'), bbox_inches='tight') |
| plt.close(fig) |
|
|
|
|
| def fig_anatomy(): |
| labels = ['Pass', 'Exec success (cov-miss)', 'Code error', 'Missing/invalid output', 'Timeout', |
| 'Resource exceeded', 'No output'] |
| cols = ['#1c5cab', '#9ec5f4', '#e34948', '#eb6834', '#eda100', '#e87ba4', '#b5b4ae'] |
| panels = ['rag-think'] + AGENTS |
| models = OPEN + FRONTIER |
| fig, axes = plt.subplots(1, 4, figsize=(14.3, 5.3), sharey=True) |
| ys = np.arange(len(models))[::-1] |
| for ax, mk in zip(axes, panels): |
| for y, m in zip(ys, models): |
| c = CF.get(cid(m, mk)) |
| if c is None: |
| ax.text(50, y, 'not evaluated', ha='center', va='center', fontsize=8.5, color=INK2) |
| continue |
| left = 0.0 |
| for l, col in zip(labels, cols): |
| v = c['anatomy'][l] |
| ax.barh(y, v, left=left, color=col, height=0.72, edgecolor='white', linewidth=0.8) |
| left += v |
| ax.axhline(ys[len(OPEN) - 1] - 0.5, color=INK2, lw=0.8, ls=':') |
| ax.set_title(MDISP[mk], fontsize=11) |
| ax.set_xlim(0, 100) |
| ax.grid(axis='y', visible=False) |
| axes[0].set_yticks(ys) |
| axes[0].set_yticklabels([DISP[m] for m in models]) |
| fig.legend(handles=[Patch(color=c, label=l) for l, c in zip(labels, cols)], ncol=4, |
| loc='lower center', bbox_to_anchor=(0.5, -0.02), fontsize=10) |
| fig.tight_layout(rect=(0, 0.1, 1, 1)) |
| fig.savefig(os.path.join(FIG, 'fa4_anatomy.pdf'), bbox_inches='tight') |
| plt.close(fig) |
|
|
|
|
| def fig_percat(): |
| fig, axes = plt.subplots(3, 3, figsize=(10.9, 9.2), sharey=True) |
| hs = None |
| for ax, cat in zip(axes.flat, CATS): |
| h = draw_kcurves(ax, 'pass', cat, legend=True, small=True, open_mode='mean') |
| hs = hs or h |
| ax.set_title(cat, fontsize=10) |
| ax.set_ylim(-2, 80) |
| for ax in axes[:, 0]: |
| ax.set_ylabel('pass@k (%)') |
| for ax in axes[-1]: |
| ax.set_xlabel('k') |
| fig.legend(handles=hs, ncol=4, loc='lower center', bbox_to_anchor=(0.5, -0.06), fontsize=9) |
| fig.tight_layout() |
| fig.savefig(os.path.join(FIG, 'fa5_per_cat_passk.pdf'), bbox_inches='tight') |
| plt.close(fig) |
|
|
|
|
| def main(): |
| fig_kcurves('pass', 'f1_passk.pdf', 'pass@k (%)') |
| fig_kcurves('reach', 'fa1_reachk.pdf', 'reach@k (%)', loc='lower right') |
| fig_kcurves('sfunc', 'fa1_sfunck.pdf', 'sfunc@k (%)', loc='lower right') |
| fig_funnel() |
| fig_method_cat() |
| fig_upset('upset_final', ['rag'] + AGENTS, ['RAG', 'Mini-SWE-Agent', 'OpenHands', 'Claude-Code'], |
| 'f4_method_upset.pdf') |
| fig_upset('upset_open', ['rag'] + AGENTS, ['RAG', 'Mini-SWE-Agent', 'OpenHands', 'Claude-Code'], |
| 'fa3a_upset_detailed.pdf') |
| fig_jaccard() |
| fig_anatomy() |
| fig_percat() |
| print('figures written to', os.path.abspath(FIG)) |
|
|
|
|
| if __name__ == '__main__': |
| main() |
|
|
