verl comm-eff research · offline weight-projection · rank-1 trajectory · task-dependence test · 30-tick Big-Math

Does the rank-1 “one straight line” story survive a harder task?
Big-Math vs GSM8K, same model & recipe

On GSM8K the RELEX picture held cleanly (see the GSM8K report): an LLM’s RLVR weight update is almost perfectly rank-1 (moves along one fixed direction), the distance along it grows linearly, and that direction is locally stable. The operator’s hunch: maybe that is a property of an easy task, and a harder task — where the model has more to learn and a messier optimization path — would look less like one clean line. We re-ran the identical analysis on Big-Math (EXP-58), same Qwen2.5-1.5B-Instruct, same GRPO, and put the two side by side.

EXP-58 Big-Math fp32 · 30 checkpoints (steps 20–800) vs EXP-57 GSM8K fp32 · 160 ticks panel = layers {0,7,13,20,27} + norm · 61 matrices same code path
Verdict — the hunch is right, with a sharpened conclusion. Rank-1 itself is not what breaks: on Big-Math one direction still captures the bulk of every tensor’s move (EVR₁ 95.6% at an 8-checkpoint window). What degrades on the harder task is the durability of that single line — how much of the trajectory stays on it and how long its direction lasts. Over the same number of checkpoints, 25.1% of Big-Math’s accumulated move lies off the line vs only 4.5% on GSM8K (≈6× more), the along-line clock is less linear (R² 0.939 vs 0.985), and the direction rotates further from where it started (|cos| 0.46 vs 0.64 over each run). So: rank-1 is universal; “one fixed line for the whole run” is task-dependent, and the harder the task the more it wants a recent, not global, line.

Read the magnitudes with one caveat. The two traces are sampled at very different rates: GSM8K checkpoints are ≈0.5 optimizer steps apart, Big-Math’s are 20 steps apart (≈40× coarser). So a “window of W checkpoints” spans far more training on Big-Math, which by itself lowers EVR₁ and raises the off-line share. We therefore compare at a fixed number of checkpoints (the unit a staleness-compensator actually sees) and normalize the drift curve to each run’s own length — but the size of every gap below mixes “harder task” with “coarser snapshots.” The direction of every effect is consistent and robust; treat the exact multipliers as indicative, not exact.
On this page: Scorecard at a glance 1 · Still rank-1? 2 · Still linear? 3 · Still one durable direction? 4 · Predicting future weights 5 · What it means

Scorecard at a glance

metric (fixed #checkpoints)GSM8KBig-Math Big-Math is…what it measures
EVR₁ — rank-1 energy share (W=8)99.9%95.6%lowerone direction's share of the move
EVR₁ — rank-1 energy share (W=16)99.7%89.9%lower
coef-linearity R² (W=16)0.9850.939lowerhow straight the along-line clock is
off-line share of Δθ (W=8)4.5%25.1%higherfraction of the move NOT on the line
direction |cos| vs run start0.6400.462lowerhow durable the single direction is

“Big-Math is…” colours green when the harder task looks better on that metric, red when worse. Every stability/durability metric is worse; raw rank-1-ness is barely touched.

1 · Are the updates still rank-1? (yes — on both)

EVR₁ — rank-1 energy share (W=16)→ higher = more rank-10.900.920.940.960.981.00λ₁ / trace(G) · 1.0 = pure line
GSM8KBig-Math
Dashed lines = medians. Both tasks pile up near 1.0: over a fixed 16-checkpoint window one direction still explains most of each tensor’s change (GSM8K 99.7%, Big-Math 89.9%). Big-Math’s mass is shifted left — a bit less rank-1 — but it is emphatically not the thing a harder task destroys. The single-direction assumption survives.

2 · Does the distance grow linearly? (less cleanly on Big-Math)

coef-linearity R² (W=16)→ higher = more linear0.400.520.640.760.881.00R² of c(t)=a·t+b, fixed window
GSM8KBig-Math
GSM8K median R² 0.985 (65.6% of tensors clear the paper’s 0.98 bar); Big-Math 0.939 (0.0% clear it). The “how-far-along” coefficient still trends linearly on Big-Math but with visibly more wobble.
coef-linearity R² by module (W=8) — GSM8K vs Big-Math→ higher = more linear0.60.70.80.91paper bar 0.98qkvogateupdownnormbiascoef R² (median)
GSM8KBig-Math
Per module: the gap is broad, not driven by one outlier block — attention and MLP projections alike are a little less linear on the harder task.

3 · Is it still one durable direction, or does it drift? (the real difference)

Direction drift: v₁ now vs the earliest window (x = fraction of the probed span)→ higher = more stable (1 = frozen direction)0.40.60.810%10%20%30%40%50%60%70%80%90%100%training progress (normalized)|cos| vs earliest v₁
GSM8KBig-Math
v₁ estimated from a sliding window vs v₁ from the earliest window (|cos|=1 → identical direction). Both start at 1.0; the steeper the fall, the more the “shared” direction rotates across training. Over its own run Big-Math falls to 0.46 vs GSM8K’s 0.64 — the harder task’s single direction is markedly less durable.
Off-line share of accumulated Δθ, by window← lower = stays on one line00.10.20.30.40.5W=8W=16prefixfraction off the v₁ line
GSM8KBig-Math
The cleanest same-#checkpoints view: what fraction of the accumulated move sits OFF the single fitted line. At W=8, Big-Math 25.1% vs GSM8K 4.5% — the line leaves far more of the harder task’s movement unexplained, and the gap widens with window size.

Why this is the crux. A rank-1 predictor assumes the future move continues along one stored direction. If a large slice of the real move is already off that line — and rising — then the stored line is a weaker guide to where the weights go next. On the harder task that off-line slice is several times larger, which is exactly why “freeze one global direction” degrades and a recent-window line is needed to keep up with the faster rotation.

4 · Does either actually predict future weights?

Projection skill by horizon — GSM8K vs Big-Math (weight_proj_ratio, W=8)↓ lower = betterabove the line = HARMFUL (worse than doing nothing)below the line = GOOD (beats holding stale weights)0.40.50.711.5212510ratio = 1 (hold-stale / do nothing)horizon h (ticks ahead of anchor)weight_proj_ratio3.0↑
GSM8K rank1_anchored[8]Big-Math rank1_anchored[8]GSM8K naive_last2Big-Math naive_last2
weight_proj_ratio: below 1 beats holding the stale weights, above 1 is worse than doing nothing. Anchored rank-1 hugs 1.0 on both (it describes position, not velocity). The sharp contrast is momentum (naive_last2): a strong short-horizon winner on GSM8K but harmful on Big-Math — largely because 20-step-apart checkpoints decorrelate consecutive deltas, so “last step ⇒ next step” no longer holds. This panel is the most cadence-sensitive; read it as “coarse checkpoints kill naive momentum,” which is itself a reason the harder/coarser regime needs the anchored line as the safe fallback.

5 · What this means for the comm-eff design

Rank-1 is a property of the optimizer/model, not the task — it replicates on Big-Math, so the trajectory-clock state (v₁ + a couple of scalars per tensor) is still well-defined and cheap. The task-dependent part is durability: the harder task’s trajectory leaves more of itself off the single line and rotates that line further, so a frozen global line fits worse. The design implication is the same as GSM8K’s and reinforced: use the rank-1 line as a cheap position / reconstruction / regularization state maintained from a recent window, never as a frozen global velocity predictor — and the harder the task, the shorter “recent” must be. Momentum-style short-horizon repair that worked on GSM8K’s fine cadence does not survive coarse checkpoints, so on harder/longer runs the anchored rank-1 line is the never-harmful fallback to lean on.

Provenance & honest limits