Thought-Aware KV Cache Compaction for Reasoning via Adaptive Attention Matching

Name: Thought-Aware KV Cache Compaction for Reasoning via Adaptive Attention Matching
Uploaded: 2026-08-20T18:38:03.151517+00:00
Duration: 743 s
Description: The episode discusses "Thought-Aware KV Cache Compaction for Reasoning via Adaptive Attention Matching," a method to reduce memory usage in large language models. The hosts explain how the paper uses structure-aware compression to cut memory by up to 65% while maintaining high accuracy, making advanced reasoning models more deployable.

The episode discusses "Thought-Aware KV Cache Compaction for Reasoning via Adaptive Attention Matching," a method to reduce memory usage in large language models. The hosts explain how the paper uses structure-aware compression to cut memory by up to 65% while maintaining high accuracy, making advanced reasoning models more deployable.

Open the full episode