"""Engram-Figure-4-style summary figure from exp/analysis/repr_summary.json. Panel layout mirrors the Engram paper's Fig.4 (LogitLens convergence + CKA alignment with best-match trajectory), plus our shuffled-delta control and the draft-vs-target alignment: (a) LogitLens: own-slot accuracy vs draft layer, all-mask / soft / shuf. Dashed line = all-mask final-layer accuracy ("soft L1 即达到基线终点"). (b) CKA(soft layer i, all-mask layer j) with per-row best-match trajectory (white) and diagonal reference — the engram-style "+1 layer" shift. (c) CKA(shuf, all-mask), same trajectory — identical shift with WRONG content: the geometric shift alone is not evidence of effective depth. (d) CKA(draft, target 36 layers) heatmap with best-match trajectories for all-mask (white) and soft (red): the soft INPUT aligns to target L22. No GPU needed; reads only the JSON. """ from __future__ import annotations import argparse import json from pathlib import Path import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import numpy as np TENSOR_NAMES = ["emb_in", "L0", "L1", "L2", "L3", "L4", "norm_out"] def cka_panel(ax, mat, title, traj_color="white", extra_traj=None): n = mat.shape[0] im = ax.imshow(mat, vmin=0, vmax=1, cmap="viridis", aspect="auto") best = mat.argmax(axis=1) ax.plot(best, np.arange(n), color=traj_color, marker="o", ms=5, lw=2, label="best match (row argmax)") ax.plot(np.arange(n), np.arange(n), "w--", lw=1, alpha=0.6, label="diagonal") if extra_traj is not None: ax.plot(extra_traj, np.arange(n), color="red", marker="s", ms=5, lw=2, label="soft best match") ax.set_xticks(range(n), TENSOR_NAMES, rotation=45, fontsize=8) ax.set_yticks(range(n), TENSOR_NAMES, fontsize=8) ax.set_xlabel("all-mask draft layer") ax.set_title(title, fontsize=10) for i in range(n): for j in range(n): ax.text(j, i, f"{mat[i, j]:.2f}", ha="center", va="center", fontsize=6, color="w" if mat[i, j] < 0.55 else "k") ax.legend(fontsize=7, loc="upper left") return im def main() -> None: ap = argparse.ArgumentParser() ap.add_argument("--summary", default=str(Path(__file__).parent / "analysis" / "repr_summary.json")) ap.add_argument("--fig-dir", default=str(Path(__file__).parent / "figs")) args = ap.parse_args() s = json.load(open(args.summary)) fig_dir = Path(args.fig_dir) fig_dir.mkdir(parents=True, exist_ok=True) lens = {c: np.array(s["lens_acc"][c]) for c in ("allmask", "soft", "shuf", "half")} cross_soft = np.array(s["cross_cka_soft"]) cross_shuf = np.array(s["cross_cka_shuf"]) d2t_all = np.array(s["d2t_allmask"]) d2t_soft = np.array(s["d2t_soft"]) fig, axes = plt.subplots(1, 4, figsize=(21, 4.6)) # (a) LogitLens convergence ax = axes[0] styles = {"allmask": ("k--", "all-mask"), "soft": ("r-", "soft (oracle, slots 1-15)"), "shuf": ("g:", "shuffled control"), "half": ("b-.", "half (oracle, slots 1-8)")} for cond, (st, lab) in styles.items(): acc = lens[cond][:, 1:].mean(axis=1) ax.plot(range(len(TENSOR_NAMES)), acc, st, marker="o", ms=5, label=lab) final_all = lens["allmask"][-1, 1:].mean() ax.axhline(final_all, color="k", ls=":", lw=1, alpha=0.6) ax.annotate(f"all-mask final = {final_all:.3f}", (0.4, final_all + 0.002), fontsize=8) ax.set_xticks(range(len(TENSOR_NAMES)), TENSOR_NAMES, rotation=45, fontsize=8) ax.set_xlabel("draft layer") ax.set_ylabel("own-slot top-1 acc (slots 1-15)") ax.set_title("(a) LogitLens: faster convergence", fontsize=10) ax.grid(alpha=0.3) ax.legend(fontsize=8) # (b)(c) draft-draft CKA with best-match trajectory cka_panel(axes[1], cross_soft, "(b) CKA: soft → all-mask (rows=soft)") cka_panel(axes[2], cross_shuf, "(c) CKA: shuffled → all-mask (control)") # (d) draft-target CKA with two trajectories ax = axes[3] im = ax.imshow(d2t_soft, vmin=0, vmax=0.6, cmap="viridis", aspect="auto") ax.plot(d2t_all.argmax(axis=1), np.arange(7), "w-o", ms=5, lw=2, label="all-mask best match") ax.plot(d2t_soft.argmax(axis=1), np.arange(7), "r-s", ms=5, lw=2, label="soft best match") ax.set_xticks(range(0, 37, 3), [str(i) for i in range(0, 37, 3)], fontsize=8) ax.set_yticks(range(7), TENSOR_NAMES, fontsize=8) ax.set_xlabel("target layer") ax.set_title("(d) CKA: draft → target layers (heatmap = soft)", fontsize=10) ax.legend(fontsize=7, loc="upper left") fig.tight_layout() out = fig_dir / "fig4_engram_style.png" fig.savefig(out, dpi=200) plt.close(fig) print(f"saved {out}") if __name__ == "__main__": main()