Dimensions overall score 9.0

GitHub Code Pulse

No public code linked for this paper yet.

• LOOKAT achieves 64× compression of the KV cache at 95.7% output fidelity when tested on GPT-2.result 95%
• LOOKAT achieves 32× compression of the KV cache at 95.0% output fidelity when tested on GPT-2.result 95%
• LOOKAT maintains rank correlation ρ > 0.95 between original and compressed attention scores.result 95%
• LOOKAT transforms attention calculation from memory-bound to compute-bound by using lookup tables.technical 90%
• LOOKAT applies product quantization and asymmetric distance computation to compress the KV-cache by decomposing key vectors into subspaces and learning codebooks.method 90%
• Current quantization methods for KV cache compression fail to reduce bandwidth because they require dequantizing keys from INT4/INT8 to FP16 before use in attention calculation.limitation 85%
• LOOKAT only compresses keys, so values still require full memory storage.limitation 90%
• The theoretical analysis confirms that rank correlation degrades as O(d_k/mK), with guarantees validated across sequence lengths up to 1024 tokens.technical 85%
• LOOKAT achieves 64× compression of the KV cache while maintaining 95.7% output fidelity on GPT-2.result 95%
• LOOKAT transforms attention calculation from memory-bound to compute-bound by using lookup tables.technical 90%
• LOOKAT maintains rank correlation ρ > 0.95 when compressing the KV cache.result 90%
• LOOKAT requires no architecture changes or training to implement.method 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card