TaskPress: Query-agnostic KV cache compression via task-guided pruning¶
Standalone
component_only;adapter: null. Real small-tensor K/V and position gathering without an LLM or end-to-end speed claim. Full experiment too costly and not reproduced (review DONE); no retry queue. Google EMNLP historical backfill, not a new publication.
Paper information¶
| Field | Details |
|---|---|
| Paper | Original arXiv v1; pinned v2 full text |
| Institution | First author Wonpyo Park: Google (pinned v2 byline) |
| First public release | 2026-08-04 07:51:37 UTC (arXiv v1) |
| Author code | The paper builds on kvpress but does not provide TaskPress-specific source or a verifiable license for it. Third-party kvpress code/license is not author method code. Paper: arXiv non-exclusive license. |
| Local adapter or method key | taskpress |
| Local reproduction code | src/auto_research/foundation_models/taskpress.py |
Paper method and evidence¶
At prefill, sum task-guide attention to the context, smooth with an odd-width zero-padded average pool, then multiply by per-token quantized value scale. Each KV head independently keeps its top-k within the same budget. Gather original K/V in chronological order, preserve original position IDs and discard guide K/V. The input is task-to-context sliced attention, never the actual downstream query. Caller supplies actual guides and GQA query-head to KV-head aggregation. Lower-index tie-breaking is a local convention. The method/resources follow v2 Section 3 Eqs. 1–3; no assertion of line-by-line identity to v1.
flowchart LR
A[Task guide attention] --> B[Sum and smooth]
B[Sum and smooth] --> C[Multiply quantization scale]
C[Multiply quantization scale] --> D[Select each head]
D[Select each head] --> E[Gather original positions and discard guide]Original paper figure¶
Original Figure 2, pinned v2 page 2; no redrawing. Crop, source and SHA-256. From the original paper; copyright remains with the authors.
Core equation¶
Guide length is \(L_t\) and sliced attention is \(A\in\mathbb R^{L_t\times L_c}\). Boundary smoothing divides by a fixed odd \(K\).
Reported results¶
The authors evaluate real models/data, with no production A/B. Llama 3.1-8B and Qwen3-8B/32B, LongBench/RULER with 16K–64K contexts, methods and compression ratios make a broad matrix. The paper reports about 25.2 minutes on one A100 40GB for guide evolution plus Gemini calls; this is not local timing or a full-campaign cost estimate.
Local reproduction¶
taskpress_scores consumes sliced attention and explicit value scales; symmetric_value_scale is a floating-point reference, not an int4 kernel. prune_taskpress really gathers K/V/positions, but no production Transformer cache integration or natural-language guide evolution is provided. CPU diagnostics use generated inputs and negative controls; CUDA requires real A100/A30, fails rather than falling back, and acceptance depends on frozen-source three-seed artifacts and a receipt. Operator agreement is not model capability, accuracy, writing quality or Evolve promotion. diagnostic_only=true, promotion_eligible=false, evolve_integrated=false. The full model and timing matrix was not launched.