Article
Protocol-Dependent Resolution in Finite-Sample Model Selection
2026-08-19
Abstract excerpt
<title>Abstract</title> <p>The maximum number of tokens a model is allowed to generate can reverse whether a set of checkpoints looks distinguishable or indistinguishable, and can hide a strong negative trend (ρ = −0.77) behind apparently flat performance. We document this through controlled LoRA experi- ments (Qwen2.5-7B, Mistral-7B) where moving from 128 to 256 tokens shifts accuracy ranges by 2.3×, flips pairw...
Topics
Open a Topic to create a Post that cites this publication.
Identifiers and source
- Literature Corpus work
- 399596ae-c491-5379-becd-85e59d3cc319
- DOI
- 10.21203/rs.3.rs-10469456/v1
