Beyond Variance: Knowledge-Aware LLM Compression via Fisher-Aligned Subspace Diagnostics (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for activation compression are gradient-blind and preserve high-variance dimensions regardless of their impact on factual knowledge preservation. |
| Approach: | They propose a knowledge-aware compression framework that models activation-gradient coupling by directly modeling subspaces. |
| Outcome: | The proposed framework preserves 6–8% more accuracy on knowledge-intensive benchmarks compared to variance-based methods at 50% rank reduction. |
Similar Papers
How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients (2026.acl-long)
Copied to clipboard
| Challenge: | Spectral properties of low/high-quality instruction and reasoning data are used to explain finetuning dynamics in large language models. |
| Approach: | They propose to analyze layer-wise gradients induced by low/high-quality instruction and reasoning data for LLM post-training. |
| Outcome: | The results show that higher-quality data are associated with lower nuclear norms and higher effective ranks. |
FLAT-LLM: Fine-grained Low-rank Activation Space Transformation for Large Language Model Compression (2026.findings-eacl)
Copied to clipboard
| Challenge: | Low-rank decomposition methods suffer from accuracy degradation and expensive calibration procedures. |
| Approach: | They propose a fast and accurate, training-free structural compression method based on fine-grained low-rank transformations in the activation space. |
| Outcome: | The proposed method outperforms pruning baselines in generalization and downstream performance while delivering inference speedups. |
FARSS: Fisher-Optimized Adaptive Low-Rank and Singular-Vector Selection for Knowledge-Preserving Fine-Tuning (2026.findings-acl)
Copied to clipboard
Renxing Chen, Ziwei Xiang, Peisong Wang, Hongjian Fang, Meng Li, Fanhu Zeng, Yanan Zhu, Peipei Yang, Xu-Yao Zhang, Jian Cheng
| Challenge: | Low-rank adaptation methods for large language models have limitations in preserving world knowledge and limiting updates to preserve world knowledge. |
| Approach: | They propose a Fisher-optimized adaptive low Rank and Singular-VectorSelection framework for knowledge-preserving fine-tuning that allows efficient and task-sensitive updates. |
| Outcome: | The proposed framework outperforms existing methods for knowledge-preserving fine-tuning. |
IMPACT: Importance-Aware Activation Space Reconstruction (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) achieve strong performance across domains but remain difficult to deploy in resource-constrained environments due to their massive size. |
| Approach: | They propose an importance-aware activation reconstruction framework that links compression to its effect on model performance. |
| Outcome: | Experiments show that IMPACT reduces model size by 55.4% while maintaining accuracy comparable to or better than state-of-the-art models. |
JW-SVD: Bridging the Cross-Modal Mismatch in Post-Training MLLM Compression (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for compression of Multimodal Large Language Models lack multimodal adaptation to preserve cross-modal synergy. |
| Approach: | They propose a framework that aligns vision and language manifolds via a Joint Covariance basis and propose Global Spectrum-Aware Truncation to dynamically transfer parameter budget to the sensitive Backbone. |
| Outcome: | Experiments on Qwen2.5-VL and Llama-3-Next confirm that JW-SVD retains both text and image capabilities. |
Beyond Fixed-Length Calibration for Post-Training Compression of LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a recent study has demonstrated that the sequence length of calibration data plays a crucial role in the effectiveness of post-training compression methods. |
| Approach: | They propose a calibration technique that applies masking along the sequence axis to normalized hidden states. |
| Outcome: | The proposed method improves perplexity and zero-shot downstream tasks performance. |
BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models have driven major progress in NLP, but memory and compute requirements hinder practical deployment. |
| Approach: | They propose a framework that preserves high accuracy while achieving 1-bit weight quantization . the orthogonal-kronecker transformation learns an orthogonale mapping via EM minimization - a new approach to quantization is proposed . |
| Outcome: | The proposed framework achieves 1-bit weight quantization with low activations with low-bit activations. |
Language Models Resist Alignment: Evidence From Data Compression (2025.acl-long)
Copied to clipboard
Jiaming Ji, Kaile Wang, Tianyi Alex Qiu, Boyuan Chen, Jiayi Zhou, Changye Li, Hantao Lou, Josef Dai, Yunhuai Liu, Yaodong Yang
| Challenge: | Large language models (LLMs) may exhibit undesirable behaviors due to the inevitable biases and harmful content present in training. |
| Approach: | They propose to investigate the elasticity of large language models by examining their performance. |
| Outcome: | The proposed model performance declines rapidly before reverting to the pre-training distribution, the authors show . the proposed model weight and code are available at pku-lm-res ist-alignment.github.io. |
Decomposed Trust: Privacy, Adversarial Robustness, Ethics, and Fairness in Low-Rank LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have driven major advances across domains, yet their massive size hinders deployment in resource-constrained settings. |
| Approach: | They propose to compress large language models to reduce computation and memory consumption while maintaining accuracy. |
| Outcome: | The proposed algorithms preserve training data privacy but weaken the protection of personally identifiable information during conversations. |
Pruning via Merging: Compressing LLMs via Manifold Alignment Based Layer Merging (2024.emnlp-main)
Copied to clipboard
Deyuan Liu, Zhanyue Qin, Hairu Wang, Zhao Yang, Zecheng Wang, Fangying Rong, Qingbin Liu, Yanchao Hao, Bo Li, Xi Chen, Cunhang Fan, Zhao Lv, Dianhui Chu, Zhiying Tu, Dianbo Sui
| Challenge: | Existing methods for parameter pruning fail to utilize the knowledge from pruned parameters. |
| Approach: | They propose a method that uses manifold learning and the Information Bottleneck measure to merge similar layers to preserve model performance. |
| Outcome: | The proposed method outperforms pruning methods on multiple datasets and LLMs with quantization and achieves substantial compression ratios. |