简体中文 · English

QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding¶

Core mechanism reproduction. Defining operators run locally. Private production data, large weights and dedicated serving are not reproduced; paper results are separate.

Paper information¶

Field	Details
Paper	arXiv 2608.05326
Institution	Indian Institute of Technology Roorkee
First public release	2026-08-05 (arXiv v1)
Author code	No official author code in the paper when checked on 2026-08-09.
Local adapter or method key	`qevict`
Local reproduction code	`src/auto_research/reproductions/qevict/`

Original paper summary¶

Background and main changes¶

Topic: Long-context KV cache. Binary keep/drop cannot handle attention drift: a low-priority window can later become relevant. QEvict uses full-precision, quantized-recoverable and deleted tiers, and dequantizes and promotes windows as accumulated attention changes.

flowchart LR
    N0["Accumulated attention"] --> N1
    N1["Full-precision tier"] --> N2
    N2["Quantized recoverable tier"] --> N3
    N3["Dynamic promotion and demotion"] --> N4
    N4["Delete lowest-confidence entries"]

Key figure from the paper¶

Figure 2: Overall workflow, stages and data flow. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\(s_i^{(t)}=\gamma s_i^{(t-1)}+a_i^{(t)},\quad q_i\rightarrow f_i\ \text{if }s_i^{(t)}>\tau\)

Offline and online results¶

The paper consistently beats representative eviction/quantization baselines on long-context comprehension, retrieval and reasoning. Its abstract gives no unified percentage.

Local reproduction¶

Under a fixed-slot budget, run three-tier caching, quantized recovery and promotion on a vector attention stream. Compare recall with irreversible eviction.

auto-research reproduce --paper qevict --dataset-dir data --seed 42

See stable metrics. No checkpoint is committed.

Comparison: Remove the paper-specific mechanism while matching split and budget; the experimental arm uses qevict. Relative change is in public-seed42.json; cross-paper percentages are not comparable.

Scope¶

Local results verify execution and direction, not reproduction at paper scale.
Private features, traffic and production serving are unavailable; paper online values are references only.
Evolve-compatible architectures are registered as candidates. Serving-only methods stay runnable standalone adapters, not trainable genomes.