File size: 18,108 Bytes
6c5f29f | 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 | """Budget sweep + ablations + significance + hybrid controller.
Run on local GPU. Handles the 4 most critical reviewer concerns:
1. Budget sweep at 5 budget levels
2. Ablations (no-cache, no-consolidate) at each level
3. Paired bootstrap significance tests
4. Hybrid heuristic+utility controller
"""
from __future__ import annotations
import time, json, numpy as np
from collections import Counter, defaultdict
from pathlib import Path
from sklearn.neural_network import MLPRegressor
from sklearn.pipeline import Pipeline as SKPipeline
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import accuracy_score, f1_score
import sys
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from llm_memory_validation.bsc_longmemeval import (
load_dataset, build_bsc, build_replay_only_router, build_fifo_replay,
build_uniform_replay, classify_action, count_words, session_text, tail_snippet,
extract_fact_lines, full_budget_words, MemoryEntry, QUESTION_TYPES,
)
from llm_memory_validation.counterfactual_dense_bsc import (
POSITIVE_ACTIONS, ACTION_TO_ID, build_context, candidate_gain,
action_utilities_for_session, feature_vector, decisions_from_utilities,
oversample_keep_rows, counterfactual_oracle_select, split_examples,
build_learned_selection, dense_predict_ids_from_candidates,
)
from llm_memory_validation.paper_competitor_suite import (
DenseEmbedder, DenseItem, dense_rag_retrieve, dense_items_from_entries,
memorybank_retrieve, ld_agent_retrieve,
)
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
OUT = Path("llm_memory_validation/neurips_full_results")
OUT.mkdir(parents=True, exist_ok=True)
TOPK = 5
BUDGET_FRACS = [0.10, 0.15, 0.20, 0.30, 0.40]
SEEDS = [0, 1, 2]
print("[1/7] Loading dataset and building embeddings...")
t0 = time.time()
examples = load_dataset()
embedder = DenseEmbedder(model_name="intfloat/e5-base-v2")
train_ex, val_ex, test_ex = split_examples(examples, seed=11)
print(f" Done in {time.time()-t0:.1f}s: {len(examples)} examples, {len(train_ex)}/{len(val_ex)}/{len(test_ex)} split")
print("[2/7] Building contexts for all budget levels...")
all_contexts = {}
for bf in BUDGET_FRACS:
t1 = time.time()
all_contexts[bf] = {ex["question_id"]: build_context(ex, bf, embedder) for ex in examples}
print(f" Budget {bf:.0%}: {time.time()-t1:.1f}s")
print("[3/7] Running budget sweep with all methods + ablations...")
sweep = {}
for bf in BUDGET_FRACS:
print(f"\n === Budget {bf:.0%} ===")
t1 = time.time()
contexts = all_contexts[bf]
def eval_fn(name, fn, examples_list):
recalls, mrrs, per_type = [], [], defaultdict(list)
for ex in examples_list:
ctx = contexts[ex["question_id"]]
gold = set(ex["answer_session_ids"])
ids = fn(ex, ctx)
hits = [r for r, sid in enumerate(ids, 1) if sid in gold]
recalls.append(len(set(ids) & gold) / max(len(gold), 1))
mrrs.append(0.0 if not hits else 1.0 / min(hits))
per_type[ex["question_type"]].append(recalls[-1])
return {
"recall_at_5": float(np.mean(recalls)),
"mrr_at_5": float(np.mean(mrrs)),
"per_type_recall_at_5": {qt: float(np.mean(v)) for qt, v in per_type.items()},
"n": len(recalls),
}
budget_words_default = max(256, int(full_budget_words(examples[0]) * bf))
# 1. FIFO replay
def fifo_fn(ex, ctx):
entries = build_fifo_replay(ex, bf)
items = dense_items_from_entries(ex, entries, embedder, TOPK)
return [item.session_id for item in items]
# 2. Dense RAG
def rag_fn(ex, ctx):
items = dense_rag_retrieve(ex, embedder, TOPK)
return [item.session_id for item in items]
# 3. Replay-only router
def replay_fn(ex, ctx):
entries = build_replay_only_router(ex, bf)
items = dense_items_from_entries(ex, entries, embedder, TOPK)
return [item.session_id for item in items]
# 4. Heuristic BSC
def heur_fn(ex, ctx):
entries = build_bsc(ex, bf)
items = dense_items_from_entries(ex, entries, embedder, TOPK)
return [item.session_id for item in items]
# 5. Oracle BSC
def oracle_fn(ex, ctx):
cands, _, _ = counterfactual_oracle_select(ctx, TOPK)
return dense_predict_ids_from_candidates(ctx, cands, TOPK)
# 6. No-cache ablation (oracle: only replay + consolidate)
def no_cache_fn(ex, ctx):
candidates = []
for si in range(len(ex["haystack_sessions"])):
best_action, best_util = "discard", -999.0
for a in ["replay", "consolidate"]:
cand = ctx.candidates_by_session.get(si, {}).get(a)
if cand is None: continue
g = candidate_gain([], ctx, cand, TOPK)
if g > best_util: best_util, best_action = g, a
if best_util > 0.01 and best_action != "discard":
candidates.append(ctx.candidates_by_session[si][best_action])
sorted_c = sorted(candidates, key=lambda c: (c.similarity - 0.25 * c.cost_words / max(ctx.budget_words, 1)), reverse=True)
budget_c, used = [], 0
for c in sorted_c:
if used + c.cost_words <= ctx.budget_words:
budget_c.append(c); used += c.cost_words
return dense_predict_ids_from_candidates(ctx, budget_c, TOPK)
# 7. No-consolidate ablation (oracle: only replay + cache)
def no_consolidate_fn(ex, ctx):
candidates = []
for si in range(len(ex["haystack_sessions"])):
best_action, best_util = "discard", -999.0
for a in ["replay", "cache"]:
cand = ctx.candidates_by_session.get(si, {}).get(a)
if cand is None: continue
g = candidate_gain([], ctx, cand, TOPK)
if g > best_util: best_util, best_action = g, a
if best_util > 0.01 and best_action != "discard":
candidates.append(ctx.candidates_by_session[si][best_action])
sorted_c = sorted(candidates, key=lambda c: (c.similarity - 0.25 * c.cost_words / max(ctx.budget_words, 1)), reverse=True)
budget_c, used = [], 0
for c in sorted_c:
if used + c.cost_words <= ctx.budget_words:
budget_c.append(c); used += c.cost_words
return dense_predict_ids_from_candidates(ctx, budget_c, TOPK)
# 8. MemoryBank proxy
def memorybank_fn(ex, ctx):
items = memorybank_retrieve(ex, embedder, TOPK)
return [item.session_id for item in items]
# 9. LD-Agent proxy
def ldagent_fn(ex, ctx):
items = ld_agent_retrieve(ex, embedder, TOPK)
return [item.session_id for item in items]
methods = {
"fifo_replay": fifo_fn,
"dense_rag_e5": rag_fn,
"replay_only_router": replay_fn,
"heuristic_bsc": heur_fn,
"oracle_bsc": oracle_fn,
"no_cache_oracle": no_cache_fn,
"no_consolidate_oracle": no_consolidate_fn,
"memorybank": memorybank_fn,
"ld_agent": ldagent_fn,
}
ret = {}
for name, fn in methods.items():
r = eval_fn(name, fn, test_ex)
ret[name] = r
print(f" {name:30s} R@5={r['recall_at_5']:.4f} MRR@5={r['mrr_at_5']:.4f}")
# 10. Train learned controller at this budget
print(f" Training learned controller at {bf:.0%}...")
train_x, train_y, train_ora = [], [], []
for ex in train_ex:
ctx_ = contexts[ex["question_id"]]
_, decs, _ = counterfactual_oracle_select(ctx_, TOPK)
for si in range(len(ex["haystack_sessions"])):
train_x.append(feature_vector(ex, ctx_, si))
train_y.append(action_utilities_for_session(ctx_, si, TOPK))
train_ora.append(ACTION_TO_ID[decs[si]])
train_x = np.array(train_x, dtype=np.float32)
train_y = np.array(train_y, dtype=np.float32)
train_ora = np.array(train_ora, dtype=np.int64)
val_x, val_y, val_ora = [], [], []
for ex in val_ex:
ctx_ = contexts[ex["question_id"]]
_, decs, _ = counterfactual_oracle_select(ctx_, TOPK)
for si in range(len(ex["haystack_sessions"])):
val_x.append(feature_vector(ex, ctx_, si))
val_y.append(action_utilities_for_session(ctx_, si, TOPK))
val_ora.append(ACTION_TO_ID[decs[si]])
val_x = np.array(val_x, dtype=np.float32)
val_y = np.array(val_y, dtype=np.float32)
val_ora = np.array(val_ora, dtype=np.int64)
best_pipe, best_thresh, best_f1 = None, 0.0, -1.0
for seed in SEEDS:
sx, sy = oversample_keep_rows(train_x, train_y, seed)
pipe = SKPipeline([
("s", StandardScaler()),
("m", MLPRegressor(hidden_layer_sizes=(128, 128), activation="relu", solver="adam",
alpha=1e-4, learning_rate_init=1e-3, batch_size=256, max_iter=250,
random_state=seed, early_stopping=True, validation_fraction=0.1,
n_iter_no_change=15)),
])
pipe.fit(sx, sy)
vp = pipe.predict(val_x)
for th in [-0.05, 0.0, 0.01, 0.02, 0.03, 0.05]:
vd = decisions_from_utilities(vp, float(th))
f = f1_score(val_ora, vd, average="macro")
a = accuracy_score(val_ora, vd)
if (f, a) > (best_f1, 0):
best_pipe, best_thresh, best_f1 = pipe, float(th), f
controller = {"pipeline": best_pipe, "threshold": best_thresh}
def learned_fn(ex, ctx):
cands, _, _ = build_learned_selection(ex, ctx, controller)
return dense_predict_ids_from_candidates(ctx, cands, TOPK)
ret["learned_bsc"] = eval_fn("learned_bsc", learned_fn, test_ex)
print(f" {'learned_bsc':30s} R@5={ret['learned_bsc']['recall_at_5']:.4f} MRR@5={ret['learned_bsc']['mrr_at_5']:.4f}")
# 11. Hybrid: heuristic action selection + utility-based discard threshold
def hybrid_fn(ex, ctx):
heuristic_entries = build_bsc(ex, bf)
filtered = []
for entry in heuristic_entries:
si_idx = None
for si, sid in enumerate(ex["haystack_session_ids"]):
if sid == entry.session_id:
si_idx = si
break
if si_idx is not None:
feat = feature_vector(ex, ctx, si_idx)
pred_utils = best_pipe.predict(feat.reshape(1, -1))[0]
max_util = float(max(pred_utils))
if max_util > best_thresh:
filtered.append(entry)
else:
filtered.append(entry) # keep if we can't find the session
if not filtered:
heuristic_entries.sort(key=lambda e: e.priority if hasattr(e, 'priority') and e.priority else 0, reverse=True)
filtered = heuristic_entries[:max(1, int(len(heuristic_entries) * 0.5))]
items = dense_items_from_entries(ex, filtered, embedder, TOPK)
return [item.session_id for item in items]
ret["hybrid_bsc"] = eval_fn("hybrid_bsc", hybrid_fn, test_ex)
print(f" {'hybrid_bsc':30s} R@5={ret['hybrid_bsc']['recall_at_5']:.4f} MRR@5={ret['hybrid_bsc']['mrr_at_5']:.4f}")
sweep[f"budget_{bf:.2f}"] = {"budget_frac": bf, "retrieval": ret}
print(f" Budget {bf:.0%} done in {time.time()-t1:.1f}s")
print("\n[4/7] Paired bootstrap significance tests (budget=0.20)...")
ref_idx = "budget_0.20"
if ref_idx in sweep:
ref = sweep[ref_idx]["retrieval"]
pairs = [
("oracle_bsc", "replay_only_router"),
("heuristic_bsc", "replay_only_router"),
("heuristic_bsc", "dense_rag_e5"),
("learned_bsc", "replay_only_router"),
("hybrid_bsc", "heuristic_bsc"),
("oracle_bsc", "heuristic_bsc"),
]
sig_results = {}
for ma, mb in pairs:
if ma in ref and mb in ref:
diff = ref[ma]["recall_at_5"] - ref[mb]["recall_at_5"]
sig_results[f"{ma}_vs_{mb}"] = {
"recall_diff": diff,
"method_a": ref[ma]["recall_at_5"],
"method_b": ref[mb]["recall_at_5"],
"note": "Paired bootstrap CI requires per-example scores; aggregate diff reported here",
}
print(f" {ma} vs {mb}: diff={diff:+.4f}")
else:
sig_results = {}
print("\n[5/7] Computing heuristic action distribution by budget...")
action_dist = {}
for bf in BUDGET_FRACS:
actions = Counter()
for ex in examples:
total = len(ex["haystack_sessions"])
for i, session in enumerate(ex["haystack_sessions"]):
a = classify_action(session, i, total)
actions[a] += 1
total_dec = sum(actions.values())
action_dist[bf] = {a: actions[a] / total_dec for a in ["discard", "replay", "cache", "consolidate"]}
action_dist[bf]["_total"] = total_dec
action_dist[bf]["_counts"] = dict(actions)
print("\n[6/7] Per-example significance between heuristic and RAG at 20%...")
if ref_idx in sweep:
heuristic_recalls = []
rag_recalls = []
for ex in test_ex:
ctx = all_contexts[0.20][ex["question_id"]]
gold = set(ex["answer_session_ids"])
h_entries = build_bsc(ex, 0.20)
h_items = dense_items_from_entries(ex, h_entries, embedder, TOPK)
h_ids = [item.session_id for item in h_items]
h_recall = len(set(h_ids) & gold) / max(len(gold), 1)
heuristic_recalls.append(h_recall)
r_items = dense_rag_retrieve(ex, embedder, TOPK)
r_ids = [item.session_id for item in r_items]
r_recall = len(set(r_ids) & gold) / max(len(gold), 1)
rag_recalls.append(r_recall)
heuristic_recalls = np.array(heuristic_recalls)
rag_recalls = np.array(rag_recalls)
diffs = heuristic_recalls - rag_recalls
observed_diff = float(np.mean(diffs))
rng = np.random.default_rng(42)
n = len(diffs)
bootstrap_diffs = np.array([float(np.mean(diffs[rng.integers(0, n, size=n)])) for _ in range(10000)])
ci_lower = float(np.percentile(bootstrap_diffs, 2.5))
ci_upper = float(np.percentile(bootstrap_diffs, 97.5))
p_value = float(min(np.mean(bootstrap_diffs <= 0) * 2, 1.0))
sig_results["heuristic_vs_rag_bootstrap"] = {
"observed_diff": observed_diff,
"ci_95": [ci_lower, ci_upper],
"p_value": p_value,
"significant_at_005": p_value < 0.05,
"n_examples": n,
"heuristic_mean": float(np.mean(heuristic_recalls)),
"rag_mean": float(np.mean(rag_recalls)),
}
print(f" Heuristic vs RAG: diff={observed_diff:+.4f}, 95% CI=[{ci_lower:.4f}, {ci_upper:.4f}], p={p_value:.6f}")
print(f" Significant at p<0.05: {p_value < 0.05}")
print("\n[7/7] Saving results and generating figures...")
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
method_labels = {
"replay_only_router": "Replay-only",
"dense_rag_e5": "Dense RAG",
"heuristic_bsc": "Heuristic BSC",
"oracle_bsc": "Oracle BSC",
"learned_bsc": "Learned BSC",
"hybrid_bsc": "Hybrid BSC",
"no_cache_oracle": "No-cache",
"no_consolidate_oracle": "No-consolidate",
"memorybank": "MemoryBank",
"ld_agent": "LD-Agent",
"fifo_replay": "FIFO",
}
colors = {
"replay_only_router": "gray", "dense_rag_e5": "purple", "heuristic_bsc": "steelblue",
"oracle_bsc": "green", "learned_bsc": "coral", "hybrid_bsc": "darkred",
"no_cache_oracle": "orange", "no_consolidate_oracle": "brown",
"memorybank": "pink", "ld_agent": "gold", "fifo_replay": "lightgray",
}
markers = {
"replay_only_router": "v", "dense_rag_e5": "D", "heuristic_bsc": "o",
"oracle_bsc": "*", "learned_bsc": "s", "hybrid_bsc": "P",
"no_cache_oracle": "^", "no_consolidate_oracle": "<",
}
for metric_key, metric_name, ax in [("recall_at_5", "Recall@5", axes[0]), ("mrr_at_5", "MRR@5", axes[1])]:
for mk, label in method_labels.items():
bvs, mvs = [], []
for bk in sorted(sweep.keys()):
if mk in sweep[bk]["retrieval"]:
bvs.append(sweep[bk]["budget_frac"])
mvs.append(sweep[bk]["retrieval"][mk][metric_key])
if bvs:
ax.plot(bvs, mvs, marker=markers.get(mk, "o"), label=label, color=colors.get(mk, "black"), linewidth=1.5)
ax.set_xlabel("Memory Budget (%)")
ax.set_ylabel(metric_name)
ax.set_title(f"{metric_name} vs Memory Budget")
ax.legend(fontsize=7, loc="lower right")
ax.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(OUT / "budget_sweep.png", dpi=200)
plt.close()
fig, ax = plt.subplots(figsize=(8, 5))
budgets = sorted(action_dist.keys())
for action, color in [("discard", "gray"), ("replay", "steelblue"), ("cache", "orange"), ("consolidate", "green")]:
vals = [action_dist[bf][action] for bf in budgets]
ax.plot(budgets, vals, marker="o", label=action, color=color)
ax.set_xlabel("Memory Budget (%)")
ax.set_ylabel("Fraction of sessions")
ax.set_title("Heuristic Action Distribution vs Budget")
ax.legend()
ax.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(OUT / "action_dist_vs_budget.png", dpi=200)
plt.close()
results = {
"budget_sweep": {k: {kk: vv for kk, vv in v.items() if kk != "retrieval" or isinstance(vv, dict)} for k, v in sweep.items()},
"action_distribution_by_budget": action_dist,
"significance": sig_results,
}
with open(OUT / "full_results.json", "w") as f:
json.dump(results, f, indent=2, default=str)
print("\n" + "="*70)
print("BUDGET SWEEP RESULTS")
print("="*70)
for bk in sorted(sweep.keys()):
bf = sweep[bk]["budget_frac"]
r = sweep[bk]["retrieval"]
print(f"\n Budget {bf:.0%}:")
for mk in ["fifo_replay", "replay_only_router", "dense_rag_e5", "memorybank", "ld_agent", "heuristic_bsc", "learned_bsc", "hybrid_bsc", "no_cache_oracle", "no_consolidate_oracle", "oracle_bsc"]:
if mk in r:
print(f" {mk:30s} R@5={r[mk]['recall_at_5']:.4f} MRR={r[mk]['mrr_at_5']:.4f}")
print(f"\nResults saved to {OUT}") |