Autoresearch · 19 August 2026 · 2 min · Original language: English
Qwen3.8 09 · RoPE and Position
No rotary-position change improved the benchmark. The best validation result remained the RTX 3090 baseline of 1.248862 bits per byte; the only alternative that completed evaluation was worse, while the layer-wise variant never produced a valid measurement.
Measured improvement0.0% better
Started
1.2489
Best
1.2489
Best-so-far result across 8 recorded experiments. Lower text prediction loss is better. Original research record
Raising the RoPE base hurt validation
Changing the RoPE base from 10,000 to 500,000 increased validation loss from 1.248862 to 1.258732. Because lower is better, this was a regression of 0.009870 bits per byte and was correctly rejected.
At this fixed sequence length, the much larger base did not improve the model’s use of positional information. The practical conclusion is narrow but clear: retain the 10,000 base rather than adopting 500,000 for this benchmark.
The layer-wise idea remains untested
The other hypothesis assigned progressively larger bases to deeper layers, using 10000 × 1.05^layer. In principle, this was intended to give shallow layers finer local positional resolution and deeper layers lower-frequency, longer-range information.
However, all six attempts failed while constructing the model, when the separate per-layer cosine tables were being stacked and registered. None reached validation, so there is no evidence that the layer-wise schedule helps or hurts prediction loss. Repeating essentially the same failing implementation consumed most of the remaining experiments without adding measurements.
Before spending another five-minute GPU evaluation on this idea, the per-layer rotary tables should be tested with a lightweight model-construction and forward-pass check. Only after that check succeeds would the hypothesis receive a fair comparison.
What to keep and what is still unknown
Keep the baseline rotary configuration with base 10,000. It is the only accepted configuration and remains the measured best at 1.248862 bits per byte.
This run does not establish whether partial rotary dimensions, alternative frequency construction, rotary application choices, or working layer-wise variants can improve the result: none of those choices produced a completed comparison. With only two valid measurements among eight experiments, the broader RoPE search space remains largely unexplored.