It is shown that robustness to post-training quantization (PTQ) is a transferable direction in weight space, and rigorously that quantization vectors are well-defined and do not suffer from reparameterization symmetries, and provide a local geometric account of their effect.
Daniele Solombrino, Antonio Andrea Gargiulo, A. R. Minut et al.· arXiv.org· 2 citations
This paper introduces SAGE, Semantic Audio Generative Encoder: a compact variational autoencoder that shapes its latent by distilling embeddings from a pretrained audio-text model, combining high reconstruction fidelity, state-of-the-art semantic structure, and fast inference.
Francesco Brigante, Luca Cerovaz, Davide Marincione et al.· 0 citations
Studying two independent testbeds (activation oracles and emergent-misaligned models) across three model families, it is found that the answer is threefold: merging preserves an emergent capability that both parents carry, but cannot create an emergent capability that is superadditive in its parents.
We show that a quantized model that keeps its classification accuracy still changes $14$ to $46\%$ of its top-1 retrieval results, and that aggregate ranking metrics reveal only part of this damage. We tie this failure to the gap between the two highest model scores and use that gap to decide when to trust a quantized...
It is concluded that tail-shape estimation in the LLM toxicity-evaluation setups the authors examined is more fragile than the recent literature suggests, and recommends the protocol as a starting point for tail-index claims in similar setups.
Luca Zhou· arXiv.org· 1 citation
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.