Two-circuit communication-efficient GRPO, Qwen2.5-1.5B, Big-Math

Training reward: does weight projection help on Big-Math?

Each curve is a training run under a different communication-efficient configuration, against a dense (uncompressed) baseline. Metric: critic/score/mean, 9-step moving average, one seed.

“Stale signed EMA every step” is the standard stale-anchor signed-EMA gradient correction, applied on every optimizer step. “Every 20 steps” means that same correction is applied only when the anchor fires (once per 20 steps). The other axis is whether a weight projection is added on top (none, linear, or adaptive).

critic/score/mean, higher is betterhover a legend entry to isolate a curve; dot = last step

Runs (last column: mean minus dense)

Configurationmeanpeakfinalvs dense
Dense (no compression)0.5040.6060.5800.000
No projection, stale signed EMA every step0.4450.5680.440-0.059
No projection, stale signed EMA every 20 steps (fires only)0.4330.5280.483-0.071
Linear weight projection, stale signed EMA every 20 steps0.4090.4890.443-0.095
Linear weight projection, stale signed EMA every step0.3780.5070.358-0.125
Adaptive projection (method 1), stale signed EMA every 20 steps0.4370.5200.447-0.066
Adaptive projection (method 1), stale signed EMA every step0.4180.5330.303-0.086
Adaptive projection (method 2), stale signed EMA every step0.3760.5100.376-0.128

Takeaways

Why: the weight trajectory differs by dataset

Weight-trajectory analysisGSM8KBig-Math
Linearity (predictive linear-fit R²)≈ 0.54≈ −1.0 (no linear structure)
Direction coherence (consecutive-update alignment)≈ 0.86≈ 0.15 (near-orthogonal)

On GSM8K the weights move coherently and roughly linearly, so an intervention could help there: applying the signed EMA only on fire steps, or adding a projection, gave a measurable lift. On Big-Math the updates are near-orthogonal with essentially no linear structure, so there is no consistent direction to project along or average, and no variant beats plain stale signed EMA every step. Big-Math simply moves more randomly.

Appendix

GSM8K: the coherent-trajectory contrast

The same configurations on GSM8K, whose weight trajectory is coherent and roughly linear. Same metric, critic/score/mean, 9-step moving average, one seed. On GSM8K the pattern reverses from Big-Math: correcting every step with the stale reference collapses without projection and fades with it, while correcting only on fires, or adding a projection, tracks the dense ceiling.

critic/score/mean, higher is betterhover a legend entry to isolate a curve; dot = last step
Configurationmeanpeakfinal
Dense (no compression)0.7990.9060.863
No projection, stale signed EMA every 20 steps (fires only)0.4850.7660.719
No projection, stale signed EMA every step0.4010.6640.372
Linear weight projection, stale signed EMA every 20 steps0.5900.8040.762
Linear weight projection, stale signed EMA every step0.5830.8540.646

Even though the optimization path is a coherent near-straight line, a reference frozen twenty steps back points where the weights used to be, and because the correction is a systematic directional bias rather than random noise, re-applying that stale direction on every step steadily pushes training the wrong way and the bias compounds until it collapses; refreshing the reference or projecting it forward realigns it with where the weights are now, so the lift returns.

Values are the training reward (critic/score/mean), not held-out accuracy; single seed, 9-step moving average. Some Big-Math runs were stopped early once the ordering was clear. Coherence and linearity figures from the offline weight-trajectory ablation.