Search papers, labs, and topics across Lattice.
This paper introduces TwinKV, a training-free and attention-free method for improving key-value (KV) cache eviction in long-context inference by leveraging pairwise key redundancy. The authors challenge existing eviction methods by demonstrating that attention magnitude does not correlate with a token's causal contribution, leading to a novel approach that identifies and swaps orphaned tokens with redundant donors while maintaining the original scoring rules. TwinKV shows performance improvements across several eviction policies, particularly benefiting models like Llama-3.2-1B, highlighting its potential to optimize resource usage in long-context scenarios.
TwinKV reveals that optimizing KV cache eviction through redundancy can significantly enhance long-context inference without the need for training or attention mechanisms.
Long-context inference is bottlenecked by the memory footprint of the key-value (KV) cache, especially for small models under tight resource budgets. Existing KV cache eviction methods score tokens using the model's attention distribution or, in attention-free variants, each key's distance from a global reference point. Using a controlled leave-one-out probe, we find that attention magnitude is unrelated to a token's causal contribution to the answer (Spearman $\rho=-0.004$), challenging the premise behind dominant eviction methods. We introduce TwinKV, a training-free, attention-free redundancy signal that detects whether a token's key has a near-duplicate elsewhere in context. Rather than replacing existing policies, TwinKV acts as a composable repair pass: given a policy's fixed retained set, it identifies evicted tokens with no surviving duplicate (\emph{orphans}) and retained tokens whose information is duplicated elsewhere (\emph{redundant donors}), then swaps them while preserving the original budget and scoring rule. We compose TwinKV with four recent eviction policies across LongBench, LooGLE, RULER, and a short-context MMLU-Pro no-harm control at compression ratios ${0.3,0.5,0.7}$. On Qwen3-4B, TwinKV improves a majority of configurations for two policies, is near-even for a third, and helps only a minority for a fourth adaptive baseline already near a performance ceiling; gains across the three non-ceiling policies are smallest at the loosest ratio. On RULER with Llama-3.2-1B, however, that fourth policy improves in every evaluated cell because its Alone score leaves substantial room to improve. More broadly, Llama-3.2-1B shows a smaller average LongBench gain but a higher fraction of improved cells on LongBench and LooGLE than Qwen3-4B, plus a clean RULER win. We also identify few-shot classification exemplars as a task structure where TwinKV does not help on either model.
