Guo and Liu designed controlled menus of single-segment layer-skipping and repetition programs for multiple-choice evaluation.
Input-blind controls mimic oracle gains for layer programs in multiple-choice evaluation
Guo and Liu show that cross-prompt headroom for layer-skipping and repetition programs can be replicated by input-blind perturbations that only match answer-change rate.
Academic
Yibei Guo · Rui Liu
Kent State University
Research Digest··1 min read
1-8B) across 4,413 multiple-choice items, the authors compare oracle gains with gains from input-blind controls at the same sites.
Why this paper
From Kent State University
In one line
Input-blind controls produce oracle headroom as large as layer programs, showing gains do not establish computation-specific benefits.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§