Files
grpo_proj2/docs/results.md
T
wassname b0d1bcd3d5 Rebuild src/ from pseudocode: SVD-basis gradient projection vs GRPO reward hacking
Expand docs/pseudocode/01..07 into a slim, fail-fast src/projected_grpo/ that
passes `just smoke`. Code mirrors the pseudocode (δS/Σ/V names, relu-before-agg
cin/cout, Dr.GRPO unbiased loss). Did not read the original src.

7 modules (~880 LOC):
- rewards.py    grader + 4 loophole modes + hack x mode diagonal self-check (R1)
- problems.py   tiny LeetCode substrate + contrastive pairs (R5)
- antipasto.py  SVD adapter, identity at δS=0 (R2)
- proj.py       erase/route/measure_only projection (R3)
- extract_vhack_grad.py  per-module SVD of paired grad diffs, noise floor (R5)
- train.py      mixed student+teacher GRPO loop, presets smoke/fast/full (R4)
- build_pool.py self-contained frozen teacher-pool fixture

`just smoke-all` PASS (exit 0): erase/none/route trio, grader diagonal clean,
v_hack cache miss->hit, ckpt every-25. Fresh-eyes review: 6/6 mechanics faithful.

Simplifications: merged loopholes+verify_rewards->rewards, pairs->problems; flat
Config + `train.py {preset} [--overrides]` CLI; justfile 384->71 lines; trimmed
results table; token-efficient train logging (config anchor, SHOULD at loop site,
sparse tqdm postfix, BLUF tail with cue + direction-arrow table).

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
2026-05-31 14:06:42 +00:00

17 KiB
Raw Blame History

Results, organized by the question each run answers

Generated from logs/*.log via just results (source: scripts/results.py). Curated snapshot 2026-05-30; regenerate any time. Each table cites its source logs in an HTML comment so every number traces back to a file.

How to read this

  • Tables show absolute last-5-step rates (mean of the final 5 training steps; converged regime, noise-robust vs a single step). Compare rows within a table by eye. Paired-vs-vanilla deltas are mentioned in prose only where the seeds match.
  • hack = fraction of student rollouts flagged as reward-hacks (hack_s).
  • solve = fraction of student rollouts passing ground-truth tests (gt_s). NOT PASS_RATE (which mixes in the ~99%-hacked teacher pool).
  • ±std is across seeds. Blank = n=1 (no std). At n=4 the seed-to-seed std is ~0.12 on both vanilla and projected, so 5-step single-seed numbers are noisy; weight by n.
  • Never compare a multi-seed mean to a single-seed point. Several arms (refresh-1/5/10, no_gate, reverse, mean-diff) only ran on seed 41. Those are compared only at seed 41, against the seed-41 vanilla and seed-41 frozen rows, never against a 4-seed mean. Mixing n is how the old refresh "ladder" produced a fake monotonic trend.
  • All runs are the fast preset (20 steps, G=4, cached-teacher mix); the fast surrogate regime, not endogenous hacking. Incomplete runs are excluded (a run must log all steps).
  • Confound (corrected from safetensors shapes, see Q8): v_hack_full = 10 pairs / k=5; v_hack_21pairs = 16 pairs / k=12. Cross-basis rows confound pair-count AND directions-kept AND tau — NOT a clean "pair set" axis.

Q1. Does the cached-teacher pool drive the student to hack? (feasibility, H4)

arm mix hack ±std solve ±std seeds
vanilla 0.5 0.719 0.120 0.306 0.116 41,42,43,44
vanilla 0.25 0.678 0.082 0.200 0.076 41,42,43
vanilla 0.125 0.757 0.040 0.207 0.020 41 (×2)

Answer: yes. Clean Qwen3-4B reaches 68-76% last-5 hack within 20 steps at every teacher density. (Don't read a mix trend here — different seed sets; see Q6 for the paired mix comparison.)

Q2. 🥇 Does v_hack projection reduce hacking vs vanilla? (H1)

mix=0.5, v_hack_21pairs, one_sided, k=5, all n=4 (seeds 41-44):

arm hack ±std solve ±std
vanilla 0.719 0.120 0.306 0.116
projected frozen-V 0.588 0.131 0.256 0.083
projected refresh-2 0.537 0.066 0.225 0.050

Answer: a consistent-in-sign reduction. Frozen drops hack 0.719→0.588 (13pp), refresh-2 →0.537 (18pp); both cost ~5-8pp solve. Per-seed paired deltas (same-seed vanilla) are negative on every seed but the std (~0.13-0.17) is about the mean, so the magnitude is not pinned down at n=4. Short of the preregistered 30pp. Note refresh-2 has the tightest hack std (0.066), i.e. its effect is the most seed-stable.

Q3. one_sided vs no_gate vs reverse gating? (gate_mode, seed 41 only)

no_gate and reverse only ran on seed 41, so this is a seed-41 within-group comparison (no cross-seed mixing):

gate hack solve
vanilla 0.775 0.300
one_sided 0.775 0.275
no_gate 0.625 0.200
reverse 0.575 0.150

Answer: more-aggressive gates cut more hack but cost more solve, and one_sided on the 18-pair basis does ~nothing at seed 41 (0.775 = vanilla). This is the weak-basis signal (Q8): the 18-pair v_hack barely overlaps the live gradient, so only the brute no_gate/reverse gates move hack — and they pay for it in solve (0.200, 0.150 vs 0.300). Single seed; directional only.

Q4. SVD top-k vs rank-1 mean-diff? (basis, seed 41 only)

basis hack solve
vanilla 0.775 0.300
SVD k=5 (v_hack_full) 0.775 0.275
mean-diff k=1 0.750 0.125

Answer: at seed 41 neither 18-pair basis cuts hack, and mean-diff tanks solve (0.300→0.125). Rank-1 being too blunt is plausible; n=1, weak-basis confound (Q8) dominates anyway.

Q5. refresh-every cadence (seed 41 only — the honest comparison)

refresh-1/5/10 only ran on seed 41, so the only valid comparison is at seed 41, on the shared seed-41 vanilla baseline:

refresh hack solve
vanilla 0.775 0.300
frozen (n=20+) 0.475 0.200
10 0.575 0.200
5 0.550 0.225
2 0.450 0.200
1 0.600 0.200

Answer: no monotonic refresh trend. At seed 41, frozen (0.475) and refresh-2 (0.450) are the best; refresh-1/5/10 are worse. The earlier "more refresh = more suppression" ladder was an artifact of comparing seed-41-only refresh-5/10 against a 4-seed frozen mean (0.131 paired). The only cadence with multi-seed support is refresh-2 (Q2): on the full seed set it edges frozen (0.537 vs 0.588 hack), but at seed 41 alone the two are within noise. Refresh helps marginally at best; basis width (Q8) is the real lever.

Q6. Teacher density (mix) — paired, does the gap hold as the pool thins?

Paired Δ vs same-seed vanilla (v_hack_full, frozen, one_sided). Δ columns are per-seed paired means; absolute hack/solve are group means (may differ slightly from Δ since n differs):

mix van hack proj hack Δhack ±std van solve proj solve Δsolve n shared seeds
0.5 0.719 0.700 0.062 0.075 0.306 0.283 0.081 4 41(×2),43,44
0.25 0.678 0.556 0.122 0.146 0.200 0.217 +0.017 3 41,42,43
0.125 0.757 0.657 0.100 0.040 0.207 0.214 +0.007 2 41(×2)

Answer: the reduction holds across densities (6 to 12pp), and the solve cost vanishes at low mix — Δsolve goes from 8pp at mix=0.5 to slightly positive (+0.7 to +1.7pp) at mix=0.25/0.125. mix=0.125 also has the tightest std (0.040, n=2). This is why 0.125 is now the locked-in default: same hack cut, no solve tax.

Q8. Weak basis (v_hack_full) vs strong basis (v_hack_21pairs)

The basis NAMES are misleading. Reading the safetensors shapes/metadata (the stored per-pair grads' first dim = pairs used; basis top_k from header):

basis pairs used k (top_k) extract tau what it is
v_hack_full 10 5 0.25 older ~12-pair set, k=5
v_hack_21pairs 16 12 0.0 later ~18-pair set, k=12

Neither is 18 or 21 pairs (n_heldout=2 reserves 2). Both load with the same train-time drop_bottom_frac=0.25 noise floor. So the comparison below is triple-confounded: pairs (10 vs 16) AND directions kept (k=5 vs k=12) AND extract tau. We cannot attribute the gap to "pair set".

mix=0.5, frozen, one_sided:

basis hack ±std solve ±std n seeds
vanilla 0.719 0.120 0.306 0.116 4 41,42,43,44
v_hack_full (weak) 0.700 0.109 0.283 0.038 3 41,43,44
v_hack_21pairs 0.588 0.131 0.256 0.083 4 41,42,43,44

At shared seed 41: weak basis = 0.775 (= vanilla, no effect), strong = 0.475.

Answer: the stronger basis cuts hack ~2x more — but pair count is a red herring; what matters is which hack mechanisms the pairs cover. The strong basis spans the later axes (try/except-swallow, type-only-assert, weak-inequality, hardcode) that the weak/older set under-covers. The real experiment is a content/axis ablation — which mechanisms carry the cut — which is the same question as G2/G3 cross-mechanism generalisation (does a basis from mechanism A suppress hack B), the no-cheat hypothesis itself. The k=5-vs-12 and 10-vs-16 differences are present but secondary.

Current pairs.py (PAIRS, 18 pairs) by mechanism: axis-1 weak-run_tests = 8/18; hardcode / persona / try-except-swallow / type-only-assert / weak-inequality = 2 each.


Q9. Solve-direction orthogonalization (does stripping the solve subspace recover solve?)

basis hack solve
vanilla 0.775 0.300
18-pair base (no orth) 0.500 0.200
18-pair solve-orth m=4 0.550 0.150

Answer: no — at n=1 it did the opposite. Stripping the top-4 solve directions from D pre-SVD was meant to recover solve; instead solve fell 0.200→0.150 and hack rose 0.500→0.550. Both moves are ~0.05, inside the ~0.12 seed std — inconclusive, leaning negative. Caveats: (1) two nominally-18-pair bases already disagree by 0.275 hack at this seed (v_hack_full=0.775 vs v_hack_18base=0.500), so extraction variance likely dominates a 0.05 delta; (2) with 18 pairs the solve basis B (top-4 SVD of G_c) is itself noisy and may strip real hack signal; (3) hack/solve subspaces may genuinely overlap. Needs ≥3 seeds before any verdict.


Q10. 🥇 Does the pair set content matter? (mechanism vs framing vs placebo)

The detector we're allowed to have is weak (no-cheat invariant): it sees some hacks and misses others. So: does a v_hack extracted from a pair set that does NOT contrast the LeetCode mechanism still suppress the mechanical hack? We swap only the pair-set content (every basis extracted identically, k=12/tau=0, trained k=5) and read Δhack vs same-seed vanilla. n=1 projected per row; ±0.06 is the baseline noise (std of the 3 seed-41 vanilla runs), so treat anything inside ±0.06 as null.

basis (pair set) contrasts hack solve Δhack vs vanilla
vanilla (baseline) -- 0.726 ~0.20
prog_wide hack mechanism 0.500 0.221 0.226
prog_wider mech + lang/cond 0.679 0.236 0.048
intent_vs_spec semantic framing 0.686 0.207 0.040
honesty_text semantic framing 0.714 0.193 0.012
moral semantic framing 0.721 0.221 0.005
eval_aware semantic framing 0.736 0.186 +0.010
philosophical semantic framing 0.743 0.243 +0.017
null_city (PLACEBO) random content 0.750 0.221 +0.024

(Baseline = mean of the 3 seed-41 vanilla mix=0.125 runs the deltas are paired against. The canonical v_hack_21pairs is NOT in this table: it was only run at mix=0.5 / different step counts, so a same-table comparison would confound mix and horizon. Its mix=0.5 effect is in Q2.)

Answer: it's the mechanism, not the framing. Pairs that contrast the programmatic hack mechanism (prog_wide) cut hack the most (0.226), at no solve cost. Semantic / value framings (moral, honesty, eval-awareness, philosophy) do essentially nothing -- all within baseline noise of the null_city placebo. The placebo sits at +0.024 (no effect), exactly as it should. So v_hack is picking up the hack-mechanism subspace, not a generic "honesty" or "intent" direction.

Caveats (n=1, hold loosely): (1) all rows are single seed-41 runs; ±0.06 is the seed-41 vanilla noise, so everything from intent_vs_spec down is null. prog_wide needs ≥3 seeds (task #122) before the 0.226 is trustworthy. (2) Broadening prog_wideprog_wider (adding language/phrasing/condition variation) hurt (0.226→−0.048): diluting the mechanism contrast with surface variation weakened the basis. (3) Encouraging for the no-cheat story -- a mechanism-matched-but-off-task detector generalizes -- but the real generalization test is held-out mechanism (Stage 2/3), not held-out framing.

Q11. Does the projection gap survive to convergence? (60-step, seed 42)

The Q2 gap (13 to 18pp) is measured at 20 steps, where vanilla hack has only just plateaued (~step 13-16, see Dynamics note). Does projection keep hack down once we run 3x past the plateau? One 60-step run per arm, seed 42, mix=0.125:

arm L5 hack L5 solve
vanilla 0.936 0.293
projected frozen-V 0.957 0.293
projected refresh-2 0.907 0.307

Answer: at n=1, the gap closes. By step 60 all three arms sit at ~0.91-0.96 hack -- projection delays hacking but does not prevent it at this horizon. The attractor in this surrogate (cached-teacher) regime is full hack, and the projected student eventually catches up. refresh-2 is marginally below vanilla (2.9pp) and frozen marginally above (+2.1pp), both inside the ~0.06-0.12 seed noise, so the honest read is "no surviving gap at 60 steps, seed 42."

Caveats: (1) n=1, seed 42 only -- needs the 3-seed convergence (task #121) to distinguish "gap truly closes" from "seed-42 is a high-hack draw". (2) This is mix=0.125 + seed 42 + 60 steps, three axes different from Q2's mix=0.5/20-step numbers, so it is NOT a clean "same run, later" comparison. (3) The 20-step suppression is real (Q2, n=4); what's unclear is whether longer training erodes it or whether this is a sparse-teacher/seed artifact.

Dynamics note (sizing the convergence test)

Per-step trajectories (mix=0.125 g8, seed 41): hack_s rises 0→~0.6-0.75 and plateaus by step ~13-16; gt_s (solve) stays noisy-flat at ~0.1-0.5 the whole run, it never climbs. The attractor in this surrogate regime is full hack, not full solve — so "run until full solve" has no target. The convergence question is therefore: once vanilla hack plateaus (~step 15), does projected stay below it or catch up? A 60-step run (~2.2h at g8) sees 3x past the plateau; a 1000-step run (~36h) is wasteful.

Open / queued (no result yet)

  • convergence at ≥3 seeds (#121): the n=1 seed-42 run (Q11) shows the gap closing by step 60, but that could be a seed-42 high-hack draw. Need 2+ more seeds before concluding the suppression erodes vs survives.
  • pairset content at ≥3 seeds (#122): Q10's mechanism>framing>placebo ordering is n=1 per row; replicate prog_wide and the placebo on 2+ seeds.
  • route arm at scale (#182): running; validates routing's ablated-eval hack<kept on Qwen3-4B before the 3-way none/erase/route cells (#130).
  • k-slice (k=1/2/5): only smoke-tested, no 4B results.
  • Stage 2/3 cross-mechanism generalisation: the load-bearing test -- extract v_hack from hack A, check it stops the unknown hack B the student would otherwise learn. Q10 (held-out framing) is a weaker cousin.