make_figures.py

#!/usr/bin/env python3
"""Generate publication figures from the actual experimental result files."""
import json, os
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import Patch
import numpy as np

FIG = os.path.expanduser("~/vulnllm-bench/paper/figures")
os.makedirs(FIG, exist_ok=True)
plt.rcParams.update({
    "font.family": "DejaVu Sans", "font.size": 10,
    "axes.spines.top": False, "axes.spines.right": False, "figure.dpi": 150,
    "axes.titlesize": 11, "axes.titleweight": "bold", "savefig.bbox": "tight",
})
# colorblind-safe palette (Okabe-Ito)
BLUE="#0072B2"; ORANGE="#E69F00"; GREEN="#009E73"; RED="#D55E00"; GREY="#999999"; PURPLE="#CC79A7"; SKY="#56B4E9"

# ---------------- Figure 1: the ladder — every lever hits the same wall ----------------
def fig1():
    labels = ["Base\n(+prompt)", "LoRA v1\nverdict", "LoRA v2\nSTaR", "LoRA v3\nteacher", "LoRA v3\n×7 scaled", "Qwen3.6-27B\n(3.8× base)"]
    pc = [17.4, 17.2, 17.1, 20.0, 14.2, 0.0]   # pair-correct %
    colors = [GREY, SKY, SKY, GREEN, ORANGE, RED]
    fig, ax = plt.subplots(figsize=(7.2, 3.7))
    x = np.arange(len(labels))
    bars = ax.bar(x, pc, color=colors, width=0.66, edgecolor="white", linewidth=0.6)
    ax.axhline(17.4, ls="--", lw=1.1, color=GREY, zorder=0)
    ax.text(len(labels)-0.5, 18.1, "base ceiling 17.4%", ha="right", va="bottom", fontsize=8.5, color="#555")
    # random-pairing reference (both-correct by chance ~ 0.5*0.5=25% only if independent&balanced -> but empirical near-chance)
    for b, v in zip(bars, pc):
        ax.text(b.get_x()+b.get_width()/2, v+0.5, f"{v:.1f}", ha="center", va="bottom", fontsize=9, fontweight="bold")
    ax.set_xticks(x); ax.set_xticklabels(labels, fontsize=8.4)
    ax.set_ylabel("Pair-correct (%)"); ax.set_ylim(0, 24)
    ax.set_title("Figure 1.  No fine-tuning recipe or larger base beats the single-function wall")
    fig.text(0.5, -0.06, "PrimeVul test_paired, blind, matched harness. Pair-correct = both the vulnerable function and its patched twin judged correctly.",
             ha="center", fontsize=7.6, color="#555")
    fig.savefig(f"{FIG}/fig1_ladder.png"); plt.close(fig); print("fig1 done")

# ---------------- Figure 2: in-distribution vs OOD, and the generalization boundary ----------------
def fig2():
    fig, (axA, axB) = plt.subplots(1, 2, figsize=(7.6, 3.5), gridspec_kw={"width_ratios":[1,1.15]})
    # Panel A: distribution shift
    groups = ["In-distribution\n(authors' setup)", "Out-of-distribution\n(PrimeVul blind)"]
    acc = [81.0, 53.0]; f1 = [80.0, 44.7]
    x = np.arange(len(groups)); w=0.36
    axA.bar(x-w/2, acc, w, label="Accuracy", color=BLUE, edgecolor="white")
    axA.bar(x+w/2, f1, w, label="F1", color=ORANGE, edgecolor="white")
    for xi,a,f in zip(x,acc,f1):
        axA.text(xi-w/2, a+1, f"{a:.0f}", ha="center", fontsize=8.5, fontweight="bold")
        axA.text(xi+w/2, f+1, f"{f:.0f}", ha="center", fontsize=8.5, fontweight="bold")
    axA.axhline(50, ls=":", color=GREY, lw=1); axA.text(1.4, 51, "chance", fontsize=7.5, color="#666", ha="right")
    axA.set_xticks(x); axA.set_xticklabels(groups, fontsize=8.2); axA.set_ylabel("Score (%)"); axA.set_ylim(0,95)
    axA.set_title("A. Distribution shift", fontsize=9.5, loc="left")
    axA.legend(frameon=False, fontsize=8, loc="upper right")
    # Panel B: trained vs novel mechanism (recall is the collapse)
    cats = ["Accuracy", "F1", "Recall"]
    trained = [85.0, 87.0, 100.0]; novel = [70.0, 64.0, 53.3]
    x2 = np.arange(len(cats))
    axB.bar(x2-w/2, trained, w, label="Trained mechanisms", color=GREEN, edgecolor="white")
    axB.bar(x2+w/2, novel, w, label="Novel mechanisms", color=PURPLE, edgecolor="white")
    for xi,t,n in zip(x2,trained,novel):
        axB.text(xi-w/2, t+1.5, f"{t:.0f}", ha="center", fontsize=8.5, fontweight="bold")
        axB.text(xi+w/2, n+1.5, f"{n:.0f}", ha="center", fontsize=8.5, fontweight="bold")
    axB.set_xticks(x2); axB.set_xticklabels(cats, fontsize=8.6); axB.set_ylim(0,112); axB.set_ylabel("Score (%)")
    axB.set_title("B. Novel mechanisms: recall halves", fontsize=9.5, loc="left")
    axB.legend(frameon=False, fontsize=8, loc="lower left")
    fig.suptitle("Figure 2.  The model learned transferable reasoning, but recall is fragile off-distribution", fontsize=10.5, fontweight="bold", y=1.02)
    fig.savefig(f"{FIG}/fig2_boundary.png"); plt.close(fig); print("fig2 done")

# ---------------- Figure 3: the core finding — where the signal lives ----------------
def fig3():
    dec = [json.loads(l) for l in open(os.path.expanduser("~/vulnllm-bench/phase0_decide_results.jsonl"))]
    intra = sum(1 for r in dec if r.get("cls")=="INTRA"); inter = sum(1 for r in dec if r.get("cls")=="INTER")
    tot = intra+inter
    diff = [json.loads(l) for l in open(os.path.expanduser("~/vulnllm-bench/phase0_diff_results.jsonl"))]
    cc = [r for r in diff if r.get("lang")=="c" and not r.get("skipped")]
    directional = sum(1 for r in cc if r.get("directional")); both = sum(1 for r in cc if r.get("both_same_flag"))
    neither = sum(1 for r in cc if r.get("neither_flagged")); other = len(cc)-directional-both-neither

    fig, (axA, axB) = plt.subplots(1, 2, figsize=(7.8, 3.6), gridspec_kw={"width_ratios":[1,1]})
    # Panel A: decidability decomposition (stacked horizontal bar)
    ip = intra/tot*100; xp = inter/tot*100
    axA.barh([0], [ip], color=GREEN, edgecolor="white", label=f"INTRA-procedural  {ip:.0f}%")
    axA.barh([0], [xp], left=[ip], color=RED, edgecolor="white", label=f"INTER-procedural  {xp:.0f}%")
    axA.text(ip/2, 0, f"{ip:.0f}%", ha="center", va="center", color="white", fontweight="bold", fontsize=11)
    axA.text(ip+xp/2, 0, f"{xp:.0f}%", ha="center", va="center", color="white", fontweight="bold", fontsize=11)
    axA.set_xlim(0,100); axA.set_ylim(-0.6,0.6); axA.set_yticks([]); axA.set_xlabel("Share of 99 PrimeVul pairs")
    axA.set_title("A. Where the vuln↔patch signal lives\n(frontier oracle judgment)", fontsize=9.3)
    axA.legend(frameon=False, fontsize=8.2, loc="upper center", bbox_to_anchor=(0.5,-0.28), ncol=1)
    axA.text(50, 0.48, "29% is the ceiling for ANY\nlocal single-function method", ha="center", fontsize=8, color="#444")

    # Panel B: what cppcheck actually delivers on the same pairs
    segs = [("Directional\n(usable signal)", directional, GREEN),
            ("Same finding\nboth twins", both, GREY),
            ("Tool blind\n(neither)", neither, "#BBBBBB"),
            ("Other\ndiff", other, ORANGE)]
    names=[s[0] for s in segs]; vals=[s[1] for s in segs]; cols=[s[2] for s in segs]
    y=np.arange(len(segs))[::-1]
    axB.barh(y, vals, color=cols, edgecolor="white")
    for yi,v in zip(y,vals): axB.text(v+0.8, yi, f"{v}%", va="center", fontsize=9, fontweight="bold")
    axB.set_yticks(y); axB.set_yticklabels(names, fontsize=8.2); axB.set_xlim(0,72); axB.set_xlabel("Share of 100 C pairs")
    axB.set_title("B. cppcheck recovers ~2% of it\n(intra-procedural linter, isolated fns)", fontsize=9.3)
    fig.suptitle("Figure 3.  The wall is interprocedural: 71% of the signal is outside the function, and the cheap extractor captures almost none of the rest",
                 fontsize=9.6, fontweight="bold", y=1.06)
    fig.savefig(f"{FIG}/fig3_ceiling.png"); plt.close(fig); print("fig3 done", intra, inter, directional, both, neither, other)

fig1(); fig2(); fig3()
print("all figures in", FIG)