Fewer Prompts Alter jspace Shape
eliebakouch · x · 2026-07-18
While estimating jspace, the author mentioned using only 250 prompts this time due to the larger model size, compared to the 1000 prompts used in previous neuronpedia checkpoints.
They also shared previous experiments on Gemma 3-27B demonstrating that prompt quantity affects the shape of jspace. Key takeaways include:
- Estimation scale is constrained by model size
- The number of sampled prompts changes the resulting shape
- This phenomenon was already observed in prior Gemma 3-27B experiments
Related event: MoE Experiments Reveal Prompt Quantity Impacts Jspace Estimation(2 posts)→
More from Research
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27