File size: 14,891 Bytes
e53f10b | 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 | """
Stage 4: Steering sweep (Apr 2026 update).
Changes vs v1:
- Sweep restricted to α ∈ [0, 1] (no over-suppression / amplification)
- 2 versions (v1_raw, v_pca_subspace) instead of 4
- --save_texts default True (so we always have CoT texts for inspection)
- --joint flag enables anti-leak coupling steering
- Robust collapse detection (ngram-based, length-relative)
- Real-reflection vs filler distinction in monitoring counts
"""
import sys
import argparse
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
import torch
from tqdm import tqdm
from configs.paths import (
ensure_dirs, LOGS_DIR, TEST_MATH_PATH,
PLAN_V1_RAW, PLAN_V_PCA_SUBSPACE,
MON_V1_RAW, MON_V_PCA_SUBSPACE,
RESULTS_DIR,
)
from configs.model import (
MODEL_CONFIG, ALPHA_SWEEP, GEN_CONFIG_FAST,
ANTI_LEAK_BETA,
)
from src.utils import setup_logger, read_jsonl, append_jsonl, write_json, cleanup_memory
from src.model_io import load_model_and_tokenizer, build_thinking_prompt, generate
from src.detectors import (
BehaviorDetector, compute_rr,
count_real_monitoring, is_collapsed,
)
from src.planning_quality import compute_pqs
from src.steering import (
ResidualSteerer, JointResidualSteerer,
build_force_prompt, is_neutral_alpha,
)
from src.directions import load_directions
SWEEP_LOG = RESULTS_DIR / "sweep_log.jsonl"
def get_direction_paths():
return {
"planning": {"v1_raw": PLAN_V1_RAW,
"v_pca_subspace": PLAN_V_PCA_SUBSPACE},
"monitoring": {"v1_raw": MON_V1_RAW,
"v_pca_subspace": MON_V_PCA_SUBSPACE},
}
def make_config_key(dim, version, alpha, idx, joint=False):
"""Stable resume key."""
a_str = "NA" if alpha is None else f"{alpha}"
j_str = "_J" if joint else ""
return f"{dim}|{version}|alpha{a_str}|idx{idx}{j_str}"
def load_completed_keys(log_path: Path):
done = set()
if not log_path.exists():
return done
import json as _json
with open(log_path, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
try:
obj = _json.loads(line)
done.add(make_config_key(
obj.get("dim"), obj.get("version"),
obj.get("alpha"), obj.get("idx"),
joint=obj.get("joint", False),
))
except Exception:
pass
return done
def evaluate_cot(text, base_text, mon_det, plan_det):
"""Compute all metrics for a CoT."""
mon_cnt = mon_det.detect(text)["total"]
plan_cnt = plan_det.detect(text)["total"]
real_mon = count_real_monitoring(text)
pqs = compute_pqs(text)
coll = is_collapsed(text, base_text=base_text)
return {
"mon_count": mon_cnt,
"plan_count": plan_cnt,
"mon_real": real_mon["real_reflection"],
"mon_filler": real_mon["filler_only"],
"mon_ambiguous": real_mon["ambiguous"],
"pqs": pqs,
"collapsed": coll["collapsed"],
"collapse_reason": coll["reason"],
"ngram_repetition": coll["ngram_repetition"],
"length_ratio": coll["length_ratio"],
"len_chars": len(text),
}
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--n_test", type=int, default=30)
parser.add_argument("--resume", action="store_true")
parser.add_argument("--max_new_tokens", type=int, default=GEN_CONFIG_FAST["max_new_tokens"])
parser.add_argument("--skip_force_prompt", action="store_true")
parser.add_argument("--only_dim", choices=["planning", "monitoring", "both"], default="both")
parser.add_argument("--only_versions", nargs="+", default=None,
help="Subset of direction versions, e.g. v_pca_subspace")
parser.add_argument("--save_texts", action="store_true", default=True,
help="Save full CoT text in log (default: True)")
parser.add_argument("--no_save_texts", dest="save_texts", action="store_false")
parser.add_argument("--joint", action="store_true",
help="Enable anti-leak joint steering (suppress both dims together)")
parser.add_argument("--anti_leak_beta", type=float, default=ANTI_LEAK_BETA)
args = parser.parse_args()
ensure_dirs()
log = setup_logger("09_sweep", LOGS_DIR / "09_sweep.log")
problems = read_jsonl(TEST_MATH_PATH)[: args.n_test]
log.info(f"Test problems: {len(problems)}")
log.info(f"Joint steering (anti-leak): {args.joint}, beta={args.anti_leak_beta}")
log.info(f"Save texts: {args.save_texts}")
log.info("Loading model...")
model, tokenizer = load_model_and_tokenizer()
dir_paths = get_direction_paths()
loaded_dirs = {dim: {} for dim in dir_paths}
for dim in dir_paths:
for version, p in dir_paths[dim].items():
loaded_dirs[dim][version] = load_directions(p)
n_layers = len(loaded_dirs[dim][version])
n_nonzero = sum(
1 for w in loaded_dirs[dim][version].values()
if (w.dim() == 1 and w.norm() > 1e-8) or
(w.dim() == 2 and w.shape[0] > 0)
)
log.info(f"Loaded {dim}/{version}: {n_layers} layers, {n_nonzero} non-zero")
mon_det = BehaviorDetector("monitoring")
plan_det = BehaviorDetector("planning")
completed = load_completed_keys(SWEEP_LOG) if args.resume else set()
log.info(f"Resume: {len(completed)} experiments already logged")
# Baselines (alpha = 1.0)
log.info("Computing baselines (alpha=1, NEW semantics: no steering)...")
baselines = {}
for prob in tqdm(problems, desc="baselines"):
prompt = build_thinking_prompt(tokenizer, prob["problem"], enable_thinking=True)
try:
text = generate(model, tokenizer, prompt, max_new_tokens=args.max_new_tokens)
except Exception as e:
log.error(f"baseline idx={prob.get('idx')} failed: {e}")
continue
ev = evaluate_cot(text, base_text=text, mon_det=mon_det, plan_det=plan_det)
baselines[prob["idx"]] = {"text": text, "prompt": prompt, **ev}
cleanup_memory()
log.info(f"Baselines done. {len(baselines)} OK")
# Sweep
dimensions = ["planning", "monitoring"] if args.only_dim == "both" else [args.only_dim]
versions_to_use = args.only_versions or ["v1_raw", "v_pca_subspace"]
total_runs = len(dimensions) * len(versions_to_use) * len(ALPHA_SWEEP) * len(problems)
log.info(f"Total sweep runs: {total_runs}")
for dim in dimensions:
other_dim = "monitoring" if dim == "planning" else "planning"
for version in versions_to_use:
target_dirs = loaded_dirs[dim][version]
other_dirs = loaded_dirs[other_dim][version]
# Sanity check: skip if all directions zero
def _has_signal(dirs):
for w in dirs.values():
if w.dim() == 1 and w.norm() > 1e-6:
return True
if w.dim() == 2 and w.shape[0] > 0:
return True
return False
if not _has_signal(target_dirs):
log.warning(f"{dim}/{version}: all zero, skipping")
continue
for alpha in ALPHA_SWEEP:
desc = f"{dim[:4]}/{version}/α={alpha:.2f}{' J' if args.joint else ''}"
for prob in tqdm(problems, desc=desc, leave=False):
key = make_config_key(dim, version, alpha, prob["idx"], joint=args.joint)
if key in completed:
continue
base = baselines.get(prob["idx"])
if base is None:
continue
if is_neutral_alpha(alpha):
steered_text = base["text"]
else:
if args.joint:
steerer = JointResidualSteerer(
model, target_dirs, other_dirs,
alpha=alpha, beta=args.anti_leak_beta,
)
else:
steerer = ResidualSteerer(model, target_dirs, alpha=alpha)
steerer.start()
try:
steered_text = generate(
model, tokenizer, base["prompt"],
max_new_tokens=args.max_new_tokens,
)
except Exception as e:
log.error(f"{key}: generation failed: {e}")
steerer.stop()
continue
steerer.stop()
ev = evaluate_cot(steered_text, base_text=base["text"],
mon_det=mon_det, plan_det=plan_det)
rec = {
"dim": dim,
"version": version,
"alpha": alpha,
"joint": args.joint,
"beta": args.anti_leak_beta if args.joint else None,
"idx": prob["idx"],
"base_mon": base["mon_count"],
"base_plan": base["plan_count"],
"base_mon_real": base["mon_real"],
"base_pqs": base["pqs"]["pqs"],
"base_len": base["len_chars"],
"steered_mon": ev["mon_count"],
"steered_plan": ev["plan_count"],
"steered_mon_real": ev["mon_real"],
"steered_mon_filler": ev["mon_filler"],
"steered_pqs": ev["pqs"]["pqs"],
"steered_q1": ev["pqs"]["q1_structural_depth"],
"steered_q2": ev["pqs"]["q2_strategy_diversity"],
"steered_q3": ev["pqs"]["q3_long_range_coherence"],
"steered_q4": ev["pqs"]["q4_premature_execution"],
"steered_len": ev["len_chars"],
"collapsed": ev["collapsed"],
"collapse_reason": ev["collapse_reason"],
"ngram_repetition": ev["ngram_repetition"],
"length_ratio": ev["length_ratio"],
"steered_text": steered_text if args.save_texts else None,
}
# Use real_reflection for monitoring RR (excludes filler)
if dim == "monitoring":
rec["rr"] = compute_rr(rec["base_mon_real"], rec["steered_mon_real"])
rec["rr_total"] = compute_rr(rec["base_mon"], rec["steered_mon"])
else:
rec["rr"] = compute_rr(rec["base_plan"], rec["steered_plan"])
append_jsonl(rec, SWEEP_LOG)
cleanup_memory()
# Force-prompt baseline
if not args.skip_force_prompt:
log.info("Force-prompt baselines...")
for dim in dimensions:
for mode in ["suppress", "enhance"]:
desc = f"force_{mode}/{dim[:4]}"
version_name = f"force_{mode}"
for prob in tqdm(problems, desc=desc, leave=False):
key = make_config_key(dim, version_name, None, prob['idx'], joint=False)
if key in completed:
continue
sys_prompt = build_force_prompt(
MODEL_CONFIG["default_system_prompt"], dim, mode
)
prompt = build_thinking_prompt(
tokenizer, prob["problem"],
system_prompt=sys_prompt, enable_thinking=True,
)
try:
text = generate(model, tokenizer, prompt,
max_new_tokens=args.max_new_tokens)
except Exception as e:
log.error(f"{key}: failed: {e}")
continue
base = baselines.get(prob["idx"])
if base is None:
continue
ev = evaluate_cot(text, base_text=base["text"],
mon_det=mon_det, plan_det=plan_det)
rec = {
"dim": dim,
"version": version_name,
"alpha": None,
"joint": False,
"idx": prob["idx"],
"base_mon": base["mon_count"],
"base_plan": base["plan_count"],
"base_mon_real": base["mon_real"],
"base_pqs": base["pqs"]["pqs"],
"base_len": base["len_chars"],
"steered_mon": ev["mon_count"],
"steered_plan": ev["plan_count"],
"steered_mon_real": ev["mon_real"],
"steered_mon_filler": ev["mon_filler"],
"steered_pqs": ev["pqs"]["pqs"],
"steered_q1": ev["pqs"]["q1_structural_depth"],
"steered_q2": ev["pqs"]["q2_strategy_diversity"],
"steered_q3": ev["pqs"]["q3_long_range_coherence"],
"steered_q4": ev["pqs"]["q4_premature_execution"],
"steered_len": ev["len_chars"],
"collapsed": ev["collapsed"],
"collapse_reason": ev["collapse_reason"],
"ngram_repetition": ev["ngram_repetition"],
"length_ratio": ev["length_ratio"],
"steered_text": text if args.save_texts else None,
}
if dim == "monitoring":
rec["rr"] = compute_rr(rec["base_mon_real"], rec["steered_mon_real"])
else:
rec["rr"] = compute_rr(rec["base_plan"], rec["steered_plan"])
append_jsonl(rec, SWEEP_LOG)
cleanup_memory()
log.info(f"Sweep complete. Log: {SWEEP_LOG}")
if __name__ == "__main__":
main()
|