File size: 9,506 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
"""
Standalone inference tool (Apr 2026 update).

Given a math problem (or default), produce N outputs at different alphas.
Default alphas: 1.0, 0.5, 0.25, 0.0 (NEW SEMANTICS: 1=baseline, 0=full suppress).

Default direction version: v_pca_subspace (k-D subspace, more robust than v1).

Usage:
    # Single dim, multi-alpha
    python scripts/10_infer.py --dim planning --alphas 1.0 0.5 0.0

    # With anti-leak joint steering
    python scripts/10_infer.py --dim planning --joint --alphas 1.0 0.5 0.0

    # Use v1_raw for comparison
    python scripts/10_infer.py --dim planning --version v1_raw --alphas 1.0 0.0

This script is also called by runall.sh as a sanity-check on
2 sample problems × {planning, monitoring} × {alpha=1, alpha=0}.
"""
import sys
import argparse
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))

import torch

from configs.paths import (
    ensure_dirs, LOGS_DIR,
    PLAN_V1_RAW, PLAN_V_PCA_SUBSPACE,
    MON_V1_RAW, MON_V_PCA_SUBSPACE,
    RESULTS_DIR,
)
from configs.model import GEN_CONFIG_FAST, ANTI_LEAK_BETA
from src.utils import setup_logger, write_json, cleanup_memory
from src.model_io import load_model_and_tokenizer, build_thinking_prompt, generate
from src.detectors import BehaviorDetector, count_real_monitoring, is_collapsed
from src.planning_quality import compute_pqs
from src.steering import (
    ResidualSteerer, JointResidualSteerer,
    is_neutral_alpha,
)
from src.directions import load_directions


DIRECTION_PATHS = {
    "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,
    },
}

DEFAULT_PROBLEMS = [
    "Find the smallest positive integer n such that n^2 + n + 41 is composite.",
    "If $\\sin x + \\cos x = \\frac{1}{5}$ and $0 \\le x < \\pi$, find $\\tan x$.",
]


def run_inference(model, tokenizer, prompt, target_dirs, other_dirs,
                  alpha, max_new_tokens, joint=False, beta=ANTI_LEAK_BETA):
    if is_neutral_alpha(alpha):
        return generate(model, tokenizer, prompt, max_new_tokens=max_new_tokens)
    if joint and other_dirs is not None:
        steerer = JointResidualSteerer(model, target_dirs, other_dirs,
                                        alpha=alpha, beta=beta)
    else:
        steerer = ResidualSteerer(model, target_dirs, alpha=alpha)
    steerer.start()
    try:
        text = generate(model, tokenizer, prompt, max_new_tokens=max_new_tokens)
    finally:
        steerer.stop()
    return text


def format_report(problem, alpha, text, base_text, mon_det, plan_det):
    mon_full = mon_det.detect(text)
    plan_full = plan_det.detect(text)
    real_mon = count_real_monitoring(text)
    pqs = compute_pqs(text)
    coll = is_collapsed(text, base_text=base_text)

    if alpha is None:
        profile = "force-prompt baseline"
    elif abs(alpha - 1.0) < 1e-5:
        profile = "BASELINE (no steering, full ability)"
    elif abs(alpha - 0.0) < 1e-5:
        profile = "ZERO ABILITY (full suppression)"
    elif 0.0 < alpha < 1.0:
        profile = f"PARTIAL ABILITY ({alpha*100:.0f}% of native)"
    elif alpha > 1.0:
        profile = f"OVERTHINKER (amplified by {alpha-1.0:+.1f})"
    else:
        profile = f"OVER-SUPPRESSED ({alpha:.1f})"

    lines = [
        "=" * 70,
        f"[α = {alpha}]   profile: {profile}",
        "=" * 70,
        f"Length: {len(text)} chars   Length-ratio vs base: {coll['length_ratio']}",
        f"Collapsed: {coll['collapsed']}  reason={coll['reason']}  "
        f"ngram_rep={coll['ngram_repetition']:.3f}",
        "",
        f"Planning triggers (total):       {plan_full['total']}",
        f"  by_subtype: {plan_full['by_type']}",
        f"Monitoring triggers (total):     {mon_full['total']}",
        f"  REAL reflection:               {real_mon['real_reflection']}",
        f"  Filler ('wait, 5+3=...'):      {real_mon['filler_only']}",
        f"  Ambiguous:                     {real_mon['ambiguous']}",
        f"  by_subtype: {mon_full['by_type']}",
        "",
        f"Planning Quality Score (PQS): {pqs['pqs']:.3f}",
        f"  Q1 structural_depth:    {pqs['q1_structural_depth']:.3f}",
        f"  Q2 strategy_diversity:  {pqs['q2_strategy_diversity']:.3f}",
        f"  Q3 long_range_coherence: {pqs['q3_long_range_coherence']:.3f}",
        f"  Q4 premature_execution: {pqs['q4_premature_execution']:.3f}  (lower=better)",
        "",
        "=== Generated CoT (first 800 chars) ===",
        text[:800],
        "...",
        "=== last 200 chars ===",
        text[-200:],
        "",
    ]
    return "\n".join(lines), {
        "alpha": alpha,
        "len": len(text),
        "mon_total": mon_full["total"],
        "mon_real":  real_mon["real_reflection"],
        "plan_total": plan_full["total"],
        "pqs": pqs["pqs"],
        "collapsed": coll["collapsed"],
    }


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--dim", choices=["planning", "monitoring"], default="planning")
    parser.add_argument("--version", choices=["v1_raw", "v_pca_subspace"],
                        default="v_pca_subspace")
    parser.add_argument("--problem", type=str, default=None)
    parser.add_argument("--problem_file", type=str, default=None)
    parser.add_argument("--alphas", nargs="+", type=float,
                        default=[1.0, 0.5, 0.25, 0.0])
    parser.add_argument("--max_new_tokens", type=int, default=4096)
    parser.add_argument("--save_to", type=str, default=None)
    parser.add_argument("--joint", action="store_true",
                        help="Enable anti-leak coupling steering")
    parser.add_argument("--beta", type=float, default=ANTI_LEAK_BETA)
    parser.add_argument("--auto_problems", action="store_true",
                        help="Run on built-in default problems (used by runall sanity)")
    args = parser.parse_args()

    ensure_dirs()
    log = setup_logger("10_infer", LOGS_DIR / "10_infer.log")

    # Determine problems
    if args.auto_problems:
        problems_to_run = DEFAULT_PROBLEMS
    elif args.problem:
        problems_to_run = [args.problem]
    elif args.problem_file:
        problems_to_run = [Path(args.problem_file).read_text(encoding="utf-8").strip()]
    else:
        problems_to_run = [DEFAULT_PROBLEMS[0]]

    log.info(f"Dimension: {args.dim}  Version: {args.version}")
    log.info(f"Alphas: {args.alphas}")
    log.info(f"Joint anti-leak: {args.joint}  beta={args.beta}")
    log.info(f"Problems: {len(problems_to_run)}")

    log.info("Loading model...")
    model, tokenizer = load_model_and_tokenizer()
    log.info("Loading directions...")
    target_dirs = load_directions(DIRECTION_PATHS[args.dim][args.version])
    other_dim = "monitoring" if args.dim == "planning" else "planning"
    other_dirs = load_directions(DIRECTION_PATHS[other_dim][args.version]) if args.joint else None

    mon_det = BehaviorDetector("monitoring")
    plan_det = BehaviorDetector("planning")

    full_outputs = []
    for prob_idx, problem in enumerate(problems_to_run):
        log.info(f"\n========== Problem {prob_idx+1}/{len(problems_to_run)} ==========")
        log.info(f"Q: {problem[:120]}")
        prompt = build_thinking_prompt(tokenizer, problem, enable_thinking=True)

        # First pass at alpha=1 (baseline) for length comparison
        baseline_text = None
        prob_outputs = {}

        for a in args.alphas:
            log.info(f"-- α={a} --")
            text = run_inference(model, tokenizer, prompt,
                                  target_dirs, other_dirs,
                                  a, args.max_new_tokens,
                                  joint=args.joint, beta=args.beta)
            if is_neutral_alpha(a):
                baseline_text = text
            base_for_eval = baseline_text if baseline_text else text
            report, summary = format_report(problem, a, text, base_for_eval, mon_det, plan_det)
            print(report)
            prob_outputs[str(a)] = {
                "text": text,
                "summary": summary,
            }
            cleanup_memory()
        full_outputs.append({
            "problem": problem,
            "outputs": prob_outputs,
        })

    if args.save_to:
        write_json({
            "dim": args.dim, "version": args.version,
            "alphas": args.alphas, "joint": args.joint, "beta": args.beta,
            "problems": full_outputs,
        }, Path(args.save_to))
        log.info(f"Saved to {args.save_to}")

    # Print summary table
    print("\n" + "=" * 70)
    print("SUMMARY TABLE (across all problems)")
    print("=" * 70)
    print(f"{'α':>6} {'mon_total':>10} {'mon_real':>10} {'plan':>6} {'pqs':>6} {'len':>8} {'collapse':>10}")
    for a in args.alphas:
        # Aggregate over problems
        rows = []
        for po in full_outputs:
            rows.append(po["outputs"][str(a)]["summary"])
        mt = sum(r["mon_total"] for r in rows) / len(rows)
        mr = sum(r["mon_real"] for r in rows) / len(rows)
        pl = sum(r["plan_total"] for r in rows) / len(rows)
        pq = sum(r["pqs"] for r in rows) / len(rows)
        ln = sum(r["len"] for r in rows) / len(rows)
        coll_pct = sum(1 for r in rows if r["collapsed"]) / len(rows) * 100
        print(f"{a:>6.2f} {mt:>10.1f} {mr:>10.1f} {pl:>6.1f} {pq:>6.3f} {ln:>8.0f} {coll_pct:>9.0f}%")
    print()


if __name__ == "__main__":
    main()