SOTA: TwinKV composable KV redundancy repair for ruvLLM

Finding

TwinKV, arXiv 2608.27128, submitted 2026 08 27, challenges attention based KV eviction with a controlled leave one out result reporting Spearman rho near zero between attention magnitude and causal answer contribution. Its repair mode audits an existing retained set, identifies evicted orphans with no surviving key twin and retained redundant donors, then swaps them one for one without changing cache budget.

Evidence class: originating team report. Not yet independently reproduced in RuV.

RuV implementation

A bounded policy agnostic repair primitive is implemented on an isolated branch under crates/ruvllm/src/optimization/twinkv.rs with tests, ADR 330, and a pinned reproduction protocol.

The implementation computes similarity only against the retained set, giving repair specific O(n K d) work rather than constructing the paper implementation's full pairwise matrix.

Contradictions

The published effect is workload and policy dependent. Few shot classification exemplars regress, and one strong adaptive baseline gains only in a minority of Qwen3 4B cells while improving more consistently on Llama 3.2 1B. This must remain opt in until reproduced on RuV workloads.

Acceptance

Reproduce baseline policy versus identical policy plus repair on Qwen3 4B and Llama 3.2 1B across LongBench, RULER, short context controls, and one real RuV agent workload. Report task score, cache bytes, prefill and repair latency, decode throughput, peak memory, energy when available, variance, swaps, orphans, donors, and failures.

Promotion requires positive mean quality on the target workload, no short context regression above 0.5 absolute points, exact cache budget preservation, and repair overhead that does not erase end to end latency or memory gains.

Start with crates/ruvllm/src/optimization/twinkv.rs, its tests, ADR 330, and the pinned reproduction protocol. Run the specified comparisons on Qwen3 4B and Llama 3.2 1B across LongBench, RULER, short-context controls, and a RuV agent workload. Done means reporting the required quality, cache, latency, memory, variance, swap, orphan, donor, and failure metrics against the stated promotion criteria.