4.5 KiB
Interim report: workspace-loop retrofit, through wave 3
2026-07-13/14 · gemma-4-E2B-it (frozen throughout) · DGX Spark
All eval numbers: held-out test splits, greedy decode, execution/answer-verified.
Full protocol: PROTOCOL_UNIFIED.md; methods: ../WORKSPACE_LOOPING.md.
Headline finding
A 1.6M-parameter merge adapter (0.03% of the model) at the L13→L14 boundary, trained ~30 min on ~280 self-labeled items, lets the frozen model loop its workspace band (L14–30) over the prompt k times before writing code — converting plan-dependent failures into successes:
MBPP full test set (500 items), hard bucket (55 plan-only items), k=4:
| seed | hard pass@1 | overall |
|---|---|---|
| baseline (k=0) | 5.5% | 51.8% |
| seed 0 | 43.6% | 53.6% |
| seed 1 | 41.8% | 53.8% |
| seed 2 | 30.9% | 51.8% |
Mean hard-bucket effect ≈ 7× baseline; overall at-or-above baseline for all seeds (no net tax). Convergent mechanism evidence: the trained loop is a hard fixed-point iteration (cos→1.000 by k≈4), and the J-lens shows latent-concept sharpening across iterations (~8× over untrained control on the probe task).
Attribution: the loop is the ingredient (code), not the weights
Same data, same parameter count, same boundary — only the mechanism varies (MBPP 250-item subset, hard bucket):
| arm | hard pass@1 |
|---|---|
| baseline | 3.6% |
| untrained loop (α-merge only) | 17.9% |
| trained NO-loop adapter (weights control) | 17.9% |
| trained loop | 42.9–46.4% |
The trained feedforward control lands exactly on the untrained-loop number: ~18% is what perturbation + format alignment buys; the remaining ~28 points require the recurrence.
The boundary: math inverts the picture
No recurrent variant beats plain weights on GSM8K. Four-arm grid (hard bucket):
| GSM8K hard | prompt-side only | touches generation |
|---|---|---|
| feedforward weights | 11.8% (best) | 4.7% (pause control) |
| recurrence | 6.3–8.7% (prompt-loop) | 9.4% (carry, beats FF 2× in-harness) |
Two orthogonal effects: (1) perturbing free-running generation positions is costly for either mechanism; (2) recurrence beats weights only where a state must evolve (generation side), and loses where it doesn't (static prompt). Interpretation: the loop performs plan refinement. Code needs a plan (loop wins); math needs answer-time computation that one frozen band pass per step cannot supply at 2B (CoT tokens remain load-bearing). Overall GSM accuracy: no variant beats the 10.5% baseline. (Hard-bucket cells carry an outcome-selection caveat — greedy labels; sampled relabeling in progress — so the math story rests on the overall numbers.)
Secondary findings
- Unified (mixed-task) training regressed both tasks vs dedicated adapters (MBPP 46.8% vs 52.0% at pre-registered k=2; GSM no better). CE-level parity between arms did NOT predict generation parity.
- Gated deployment (logistic probe on the k=0 workspace state, STaR labels as supervision): 52.0% overall with easy items fully preserved (97.5%); gate recall 64%, precision 19% — the current ceiling; threshold/ feature work is CPU-only follow-up.
- Frozen-prompt KV-cache equivalence: looped prompt states are constant across token steps (causality), so loop once + cached generation is bit-identical and ≥3.5× faster. Latent planning is prefill-shaped (parallel, compute-dense) — its economics improve with model scale, unlike serial CoT decode.
- Cross-token carry (design C) is stable and the first recurrent variant to beat its own baseline on math overall (12.9% vs 10.2%), via latent pause positions, but does not beat the weights control; p=6 ≤ p=2 (no latent-depth scaling).
- Fixed operational traps documented in
../LESSONS.md: plan-pass truncation masquerading as "planning hurts"; stop-string matching the opening fence (pass@1=0 artifact); full-vocab logits OOM (earlyoom silent kills); KV-cache append in layer re-runs; small-pool overfitting ~step 300; val-CE checkpoint selection does not track generation accuracy.
Status of caveats
Resolved: post-hoc k (pre-registered k=2 before unified test numbers); weights-vs-loop attribution (FF arms); seed dependence (3 seeds, wave 5 adds 2 more); same-harness baselines. In progress tonight: outcome-selection on hard buckets (sampled relabeling); equal-FLOPs explicit-planning reference; depth generalization beyond trained k; lens battery at N=20; 12B replication (the scale question). Not addressed: second synthesis domain, base-model diversity beyond gemma, easy-dip elimination.