Beacon queries cut KV memory by 40%

The BeaconKV method uses beacon queries that predict re-access to evicted context fragments, cutting peak KV cache memory by up to 40% without losing answer quality. On Qwen3-14B with a 1024-token cache limit, AIME24 accuracy rose by 31.7 percentage points. The method is training-free but requires manual tuning of the target compression ratio.

• Peak KV cache memory cut by up to 40% on four open reasoning models
• AIME24 accuracy gain on Qwen3-14B is 31.7 pp with a 1024-token cache
• Training-free method; compression ratio set via --decode-max-budget flag
• Savings are smaller on tasks without re-access to distant context