Prompt Quantity Affects Results in MoE Experiments
eliebakouch · x · 2026-07-18
This reply details a set of experiments regarding jspace:
- For larger MoE models, the author notes there is still some "jspace" available, but fable makes a significant approximation by only computing on very few layers.
- During their experiments, they let the model decide some experimental setups, though it didn't always consider the most reasonable controls, such as:
- Validating the impact on a smaller model first;
- Checking the impact of varying prompt quantities.
- Additionally, the model downloaded the nvfp4 checkpoint for speed, explicitly acknowledging that this might introduce errors.
- The author also mentions that due to the model's massive size, they had to use fewer prompts to estimate jspace (250 vs 1000) and shared previous results on gemma 3-27B showing how prompt count affects jspace shape.
Related event: MoE Experiments Reveal Prompt Quantity Impacts Jspace Estimation(2 posts)→
More from Research
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27
- arXiv advice from T. Dietterich: wrong primary category will almost surely get your paper on hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27
- Community-trained MOSS Sharpness Slider LoRA: 72 photos, 500 steps, +/-1 focus control — aoleg77 · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27