The two-circuit method on harder math, and starting from a correct base model

Weekly update · week of 2026-07-06 · self-contained for other teams · follows last week’s weight-projection readout.

✓  Verified by Shamane

1. This week, in one line

The result that matters: even from the generic Qwen2.5-1.5B-Instruct base (not a math model), our two-circuit method stays stable on the harder, more random-looking Big-Math task, as long as we keep the staleness right.

2. Where we were, and what is new

Rank-1 holds on both tasks; only the easy one is projectable consecutive updates aligned → accumulated move is rank-1 (one direction) → rank-1 but incoherent (nothing to project along) GSM8K rank-1 yes · aligned 0.86 PROJECTABLE Big-Math rank-1 yes · aligned 0.15 NOT projectable
Figure 1. Two timescales (schematic). Both tasks sit on the right: the accumulated weight change is one rank-1 direction on both (the part worth keeping). They split on the vertical axis, the alignment of consecutive updates: GSM8K 0.86 (a line can ride it) versus Big-Math 0.15 (near-orthogonal, nothing to project along). Rank-1 alone does not license projection.

3. The harder task is random, and the two-circuit method stays stable

Big-Math (generic base): a fresh anchor holds, a stale one collapses, projection does not help 0.30 0.35 0.40 0.45 0.50 0.55 0 20 40 60 training step dense (comm-eff OFF), ceiling two-circuit method, fresh anchor (β=0.00), holds two-circuit method, stale anchor (β=0.90), collapses weight projection, refreshed periodically, no help
Figure 2. Big-Math training reward, generic base (real curves, cadence 20, single seed). α is how strongly the clean anchor sign is applied, β is how fresh the anchor memory M is. The two-circuit method with a fresh anchor (green) holds just under the dense ceiling; the same method with a stale anchor (pink, β=0.90) peels down after step 55. Weight projection (blue, refreshed periodically) does not beat the plain method. The knobs, not projection, are what control stability. (The projection curve is from the companion Big-Math projection sweep, matched in scale.)

For contrast: the easy, coherent task

On the easy task the trajectory is coherent, but staleness bites the same way. At cadence 20 the anchor reference is about 20 steps behind, and applying that stale sign on every step compounds a directional bias until it clearly collapses, while refreshing periodically (or projecting) holds. Same lesson as the hard task: keep the reference fresh.

The easy coherent task (GSM8K): a stale sign applied every step still collapses 0.00 0.20 0.40 0.60 0.80 0 20 40 60 80 training step dense (no compression) two-circuit, stale sign applied every step (collapses) two-circuit, reference refreshed every 20 steps (holds) weight projection, every 20 steps (holds)
Figure 3. The easy, coherent task (GSM8K), for contrast (real curves, single seed). At cadence 20 the anchor reference is about 20 steps stale. Applying that sign on every step (red) compounds a systematic directional bias and clearly collapses near step 65 before clawing back; refreshing only when the anchor fires (teal) or adding a projection (blue) holds on the dense line. The lever is the same as on the hard task: keep the reference fresh.

The surprise: our compression-and-efficiency work is more robust on the harder, random-looking regime than on the tidy easy one, because the load-bearing part never relied on a coherent path.

4. The big shift: correct base models change how we do RL research

This is the finding most likely to shape our RL work going forward, and especially how we evaluate it: the strong, efficient RLVR results in the literature depend on starting from a domain-adapted or distilled base model, so our efficiency claims have to be validated in that setting, not on a generic model. Worth a proper conversation.
DeepScaleR + R1-Distill-1.5B (16k): training reward stays on the dense line 0.30 0.35 0.40 0.45 0.50 0.55 20 40 60 80 100 training step dense (no compression) two-circuit method, no projection two-circuit method + weight projection (cut at step 50)
Figure 4. Correct base, hard task, long reasoning (run #63, real training reward). DeepSeek-R1-Distill-Qwen-1.5B, DeepScaleR, 16k tokens. The two-circuit method (green) and the two-circuit method plus weight projection (blue) both sit on the dense line (gray); over the shared window the projection variant is fractionally ahead (mean 0.495 versus 0.493, ahead on about half the steps, i.e. within noise). The projection arm was operator-cut at step 50. Held-out AIME is 0.21 for the compressed arm versus 0.25 for dense, inside the noise band of a 30-problem benchmark, at roughly 8 times the policy entropy of dense.
Why it works on a correct base (the part worth explaining)

A domain-adapted or distilled base already has the capability the RL is meant to elicit, so RL is re-weighting an existing skill, not installing a new one. Three consequences line up with what we see:

5. A second efficiency axis: is one block enough?

Result · block-only training tracks full-parameter training (issue #64, two-seed PASS)

Setup: freeze everything except the middle block (layers 11 to 15), train only that block with vanilla GRPO for 75 steps, comm-eff off, everything else identical to the full-parameter (dense) control. Base is the generic Qwen2.5-1.5B-Instruct, which is not a math model. Two seeds, both datasets. We read this as how closely the block-only curve tracks the full-parameter curve, not as a base-relative recovery ratio (a ratio distorts things when the base-to-full gap is small, or inflated by an answer-format quirk).

All gates green; vanilla GRPO, comm-eff off; a clean two-seed PASS. It green-lights blockwise-independent training as a real second comm-eff axis for pipeline parallelism.

Block-only training lands next to full-parameter training 0.20 0.40 0.60 0.80 0.00 0.78 0.76 in-domain (easier) 0.61 0.58 harder split full-parameter GRPO middle block only (layers 11 to 15) final validation accuracy
Figure 5. Block-only versus full-parameter, final validation (run #64, seed-averaged over two seeds). Train only decoder layers 11 to 15 of Qwen2.5-1.5B-Instruct, comm-eff off, versus the full-parameter control. Block-only lands right next to full-parameter on both datasets (a gap of about 0.02 on the easier task, 0.03 on the harder one), from a generic non-math base.

6. Takeaways

7. Next steps

References

Reports and runs from this and prior weeks:

Same rank-1, different projectability Big-Math signed-EMA (α,β) sweep Big-Math training reward MOAT projector verdict Run #63: DeepScaleR + R1-Distill Run #64: middle-block-freeze GRPO

Papers and substrate:

Weekly update for the week of 2026-07-06. Curves are real training data (Big-Math and GSM8K reward from the offline sweeps and the live runs; run #63 from WandB; run #64 final validation, two seeds). The (α,β) sweep and run #63 are single-seed and, for #63, operator-truncated, so treat them as directional. Figure 1 is a schematic. Verified by Shamane.