Subtree dropout

Subtree Dropout — Per-Epoch Random Root-Child Masking

Status (2026-04-29): closed. Modest improvement at mid-SE budgets, neutral or slightly negative at high-SE budgets. The experiment's most important byproduct was exposing that AGPT was severely undertrained — see ../agpt-epoch-scaling/ for the major finding that came out of this work. As a standalone training improvement, subtree dropout is useful but small.

Hypothesis

AGPT does 65 deterministic root-child subtree firings per super-epoch (one Adam step per subtree at d=32 with no-accumulate). Each SE walks all 65 in the same order. Hypothesis: random per-epoch subset selection would (a) increase trajectory variety across SE, (b) provide a dropout-like regularization, (c) effectively give more "kinds of optimization steps" without changing AGPT's loss objective.

The framing was: AGPT might be stuck in smooth gradient valleys due to common-prefix gradient washout. Random subtree masking is a within-AGPT-design way to inject variety, unlike the SGD interleave hybrid (which cross-paradigm-breaks AGPT's optimizer state and loss).

Implementation

Added env var AGPT_SUBTREE_DROPOUT=p (default 0). Each super-epoch samples a keep mask: each of n_root_children root-children is kept with prob (1-p), dropped with prob p. Skipped subtrees contribute no training events that epoch. Different keep mask each epoch (LCG seeded).

If everything got dropped (rare at small p), the largest-mass subtree is force-kept.

Results

Recipe: rmsprop wc lr=3e-3 entropy-λ=1.0 mass-weight=log no-accumulate. Eval at seq=32 (matched).

Initial sweep (3 reps each)

The headline read at first: p=0.3 with 5 SE beats no-dropout 3 SE by 0.69 PPL. Compensating for the dropout's per-epoch reduction by adding more epochs gave a real ~6% improvement.

Follow-up: dropout helps at mid-SE, NOT at high-SE

At 5-10 SE: dropout adds nothing meaningful when compared to just training pure AGPT for the same number of epochs.

• p=0 5SE 9.64 vs p=0.3 5SE 10.13 — pure AGPT WINS
• p=0 10SE 8.24 vs p=0.3 10SE 8.62 — pure AGPT WINS
• p=0 15SE 7.21 vs p=0.3 15SE 7.69 — pure AGPT WINS

The "p=0.3 5SE beats p=0 3SE" win was just "more epochs help" plus a small variety bonus that didn't compound at higher SE. Dropout is at best a regularization for low-budget training; it's not a fundamental improvement to AGPT optimization.

Combination with joint-mass

Joint-mass and subtree-dropout do NOT compound. Combining them gives a WORSE result than either alone. The two mechanisms target different modes (joint-mass weights events; dropout removes events) and the interaction destabilizes both.

Why this is "closed"

Subtree dropout's apparent ~6% improvement at 5 SE was an artifact of comparing it to under-trained pure AGPT (3 SE). When you give pure AGPT the same total epoch count, it wins. The diagnostic value of this experiment was in exposing the undertraining — see ../agpt-epoch-scaling/ for the consequence.

Subtree dropout itself has a narrow use case: regularization for very low-SE training budgets where the model is at risk of subtree-order artifact memorization. For any serious training run (10+ SE), it's not a useful knob. Leaving the code in place behind the AGPT_SUBTREE_DROPOUT env var (default off — no behavior change) for future experimentation.

Recipe (default off)

AGPT_SUBTREE_DROPOUT=0.3 bin/agpt_train ... --epochs 5 ...

AGPT_SUBTREE_DROPOUT_SEED env var also available to control the LCG seed for reproducibility.

Reproduce

bash rnd/subtree-dropout/run_dropout_sweep.sh

Runs p=0/0.3/0.4/0.5 across 3-15 SE budgets × 3 reps. ~30 min.

See also

• ../agpt-epoch-scaling/ — the major finding that emerged from this experiment's follow-up sweep. Pure AGPT 40 SE → 5.39 PPL@32, far below anything dropout achieves.
• ../trie-attention-framing/ — the depth-routing / joint-mass line. Joint-mass + dropout don't compound, as documented above.