Skip to content

Author

Xuezhou Zhao

3 papers indexed here

We haven’t gathered this author’s papers yet. Follow them and we’ll fetch their work.

Not the right person? Other researchers publish under this name.

#protein folding Open access Sep 2026

esm2-layer-invariance: conditional layer invariance in protein language models (BIOINF-2026-2348 revision)

Complete analysis pipeline and per-dataset results for the revised version of BIOINF-2026-2348. Includes: (1) unified per-layer probing of ESM-2 650M and SaProt_650M_AF2 on all 63 ProteinGym substitution datasets under three feature arms (masked wild-type, unmasked wild-type, mutant) and four train-test splits (random, ProteinGym-official modulo and contiguous, position-grouped); (2) a random-position lookup control showing random-split probing performance is largely position memorization; (3) multi-scale pooling benchmarks with per-fold-trained attention; (4) fp32 logit lens scoring validated against the official logits pathway; (5) layer-wise cosine and linear CKA similarity analyses; (6) TOST equivalence tests, and (7) Foldseek 3Di structure tokens for 62 datasets. Also includes the reproduction record of the withdrawn fp16 lens curves. Code MIT; results CC-BY-4.0.

Yutao Guo, Zihan Zhang, Xuezhou Zhao et al. · 0 citations
#protein folding Open access Sep 2026

esm2-layer-invariance: conditional layer invariance in protein language models (BIOINF-2026-2348 revision)

Complete analysis pipeline and per-dataset results for the revised version of BIOINF-2026-2348. Includes: (1) unified per-layer probing of ESM-2 650M and SaProt_650M_AF2 on all 63 ProteinGym substitution datasets under three feature arms (masked wild-type, unmasked wild-type, mutant) and four train-test splits (random, ProteinGym-official modulo and contiguous, position-grouped); (2) a random-position lookup control showing random-split probing performance is largely position memorization; (3) multi-scale pooling benchmarks with per-fold-trained attention; (4) fp32 logit lens scoring validated against the official logits pathway; (5) layer-wise cosine and linear CKA similarity analyses; (6) TOST equivalence tests, and (7) Foldseek 3Di structure tokens for 62 datasets. Also includes the reproduction record of the withdrawn fp16 lens curves. Code MIT; results CC-BY-4.0.

Yutao Guo, Zihan Zhang, Xuezhou Zhao et al. · 0 citations
#protein folding Open access Sep 2026

esm2-layer-invariance: conditional layer invariance in protein language models (BIOINF-2026-2348 revision)

Complete analysis pipeline and per-dataset results for the revised version of BIOINF-2026-2348. Includes: (1) unified per-layer probing of ESM-2 650M and SaProt_650M_AF2 on all 63 ProteinGym substitution datasets under three feature arms (masked wild-type, unmasked wild-type, mutant) and four train-test splits (random, ProteinGym-official modulo and contiguous, position-grouped); (2) a random-position lookup control showing random-split probing performance is largely position memorization; (3) multi-scale pooling benchmarks with per-fold-trained attention; (4) fp32 logit lens scoring validated against the official logits pathway; (5) layer-wise cosine and linear CKA similarity analyses; (6) TOST equivalence tests, and (7) Foldseek 3Di structure tokens for 62 datasets. Also includes the reproduction record of the withdrawn fp16 lens curves. Code MIT; results CC-BY-4.0.

Yutao Guo, Zihan Zhang, Xuezhou Zhao et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.