Random KV cache eviction matches every scored evictor and serves 32-43% more throughput.

Salesforce AI Research keeps the prompt, then evicts uniformly at random inside each attention head with no score computed at all. Across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput in vLLM. The explanation: the prompt is the fragile part, and the reasoning trace protects itself through redundancy at two levels, restating what it needs in the text and duplicating it across heads. Once the prompt is safe, a random draw retains enough copies. The entire scoring paradigm was doing almost nothing. Code at SalesforceAIResearch/Random-Attention. arXiv 2609.03430