comm-eff · look-ahead anchor · weight projection

Two adaptive projectors, two different directions in weight space

Both cold-start at fixed_linear and learn a per-block correction from retrospective error — no peek at live weights. The real difference is what each one learns, and therefore which direction in weight space it can actually correct.

mode 1 · fixed linear fixed_linear mode 2 · offset residual learned_linear mode 3 · step scale learned_step_scale

mode 1 · the baseline

Fixed linear projection

The frozen seed both adaptive modes start from — a plain linear extrapolation along the recent weight trajectory, with no feedback and nothing learned.

Projection

θ̂ = S₀ + (α·h/g) · (S₀ − S₁)

At the operating point (α = 1, h = g) this is the AsyncPP seed 2·S₀ − S₁ — coefficients (2, −1). A pure function of two stale snapshots: fully deterministic, no per-block state.

cost: 2 sources (S₀, S₁) · no θ̂ retained · 0 learned params — the cheapest mode

Its role

This is the cold start of both learned modes. β = 1 (step-scale) and r = 0 (offset) each reduce to exactly this line — so the adaptive modes only ever improve on fixed_linear, and their first fire is byte-identical to it.

Its limit

It is static: the same extrapolation for every block, forever. With no feedback it can't sense that a block over- or under-shot, so where the trajectory bends — k-collapse, Big-Math phase changes — it lands wrong and stays wrong. Closing that fixed gap is the whole job of the two adaptive modes below.

The anchor wants to sit at the live weights ● θ_true, but only has stale snapshots on the trajectory S₁ → S₀. Offset can only shove the whole matrix along the all-ones diagonal — a direction the error barely has. Step-scale slides along the trajectory itself, landing on the closest point to θ_true. Drag β and watch it close the gap.

step = S₀ − S₁ 𝟙 (all-ones / DC) e = θ_true − θ̂ β=1 · fixed_linear S₁ = θ[t−2K] S₀ = θ[t−K] θ_true (live weights) offset: r·𝟙 ≈ 0 step-scale θ̂(β)
β1.35
dist step-scale25
dist offset / β=178
mode 2 · the offset residual
learned_linear_with_fixed_linear_cold_start
Learns
a per-block additive residual r (a DC offset), init 0.
Correction
θ̂ = fixed_linear + r  (r broadcast over the whole matrix = r·𝟙)
Update rule
r += 0.1 · mean(θ_true − θ̂)  ·  clip |r| ≤ 1e-3
The catch
mean(e) is the error's component along the all-ones direction. Weight-update matrices are ≈ zero-mean, so mean(e) ≈ 0 — the signal it trains on is nearly empty, and r stays tiny.
Geometry
shifts the matrix up/down by a constant — moves θ̂ only along the DC diagonal, not toward the trajectory.
Cost
3 sources (S₀,S₁,S₂) + θ̂ = 4 copies
mode 3 · the step scale NEW
learned_step_scale_with_fixed_linear_cold_start
Learns
a per-block multiplicative scale β on the extrapolation step, init 1.
Correction
θ̂ = S₀ + β · (α·h/g) · (S₀ − S₁)
Update rule
β += 0.2 · ρ,  ρ = ⟨e, step⟩ / ‖step‖²  ·  β ∈ [0, 2]
Why it works
⟨e, step⟩ projects the error onto the step actually taken — a genuine inner product, not a mean. Signed and ≈ nonzero: >0 = "under-shot, go further", <0 = "over-shot". It answers how far to project.
Geometry
slides θ̂ along the trajectory to the point closest to the live weights. β→0 degrades to raw-stale where the path isn't linear.
Cost
2 sources (S₀,S₁) + θ̂ = 3 copies (lighter)

Side by side

Same skeleton, one decisive swap: the basis direction the per-block scalar is learned along.

 offset residual (mode 2)step scale (mode 3)
Learned quantityadditive residual rmultiplicative scale β
Init (cold start)r = 0β = 1
Basis direction𝟙 (all-ones / DC)S₀ − S₁ (trajectory)
Training signalmean(e) ≈ 0 → weak⟨e, step⟩ ≠ 0 → informative
What it can fixthe mean level of a matrixhow far along the trajectory θ̂ sits
Bound|r| ≤ 1e-3 (absolute)β ∈ [0, 2] trust region
Source snapshots3  (S₀,S₁,S₂)2  (S₀,S₁)
Full-model copies43
Ordersecond-order-readyfirst-order

What they share

The step-scale mode is a drop-in sibling — it keeps every invariant the offset mode already guarantees.

Why the direction matters

The anchor's whole job is a clean reference gradient close to the fast network.

At staleness K the raw θ[t−K] has rotated nearly orthogonal to the live weights — the k-collapse — so a gradient computed there is a poor reference. Projection moves θ̂ back toward the live weights, so G_anchor lands near the fast net and the correction feedback becomes trustworthy.

The offset mode can only correct a DC level the error essentially doesn't have. The step-scale mode corrects the one thing that actually collapses: how far along the trajectory the anchor sits — learned directly from whether the last projection under- or over-shot.