Papers for

software developers for compressed ai

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Three-level optimization improves compression of large language models

Per-Matrix Optimality Is Not Enough: Three-Level Optimization for Low-Rank LLM Compression

Abstract: Per-matrix singular value decomposition (SVD) truncation is Eckart-Young optimal in the whitened Frobenius norm, but errors from independently compressed matrices compound through the block's nonlinear forward pass. Inspired in part by hierarchical variational optimization in quantum many-body methods, we introduce a three-level chain that widens optimization scope from individual matrices to Transformer blocks to the full model: whitened SVD~(L1), block-level joint optimization~(L2), and end-to-end language-modeling loss refinement~(L3), all from 256 calibration sequences, with no instruction or recovery data. On LLaMA-7B at 60% compression, the chain reduces WikiText-2 perplexity from 42.1 to 19.1 to 11.4. The block-level stage acts as a regularizer: skipping it worsens Penn Treebank (PTB) perplexity by 24 points, a gap that additional end-to-end training did not close in our experiments. Perplexity gains hold across 20-80% compression, five architectures up to 13B parameters, and both in-distribution and out-of-distribution benchmarks, though the cross-architecture rows use architecture-specific configurations and the ratio sweep was not run under one common protocol. With more calibration data, skipping the block-level stage becomes competitive, revealing an offline compute--data trade-off. We therefore claim improvements only in perplexity and compression fidelity; downstream accuracy remains well below the dense model.

Mon 14 SeptMachine LearningArtificial Intelligence
The gist
Compressing parts of large language models one matrix at a time can cause errors to add up, making the model less accurate. The authors propose optimizing in three stages: first on individual matrices, then on groups of matrices in a block, and finally on the whole model. This approach helps reduce these compounded errors and improves the model's accuracy after compression, even with limited calibration data. However, the compressed models still do not match the accuracy of the original full models.
Open 2609.15838v1