{ "schema_version": 1, "builder_plan": { "scenario": "Long customer conversations exhaust memory or limit concurrent requests.", "change": "Change only the attention cache policy, after checking model and runtime support.", "baseline": "The same model with its ordinary attention cache.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Peak cache memory, long-context answer quality, first-token latency, and requests completed.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2610.17959", "title": "Virtual Head Attention", "summary": "VHA replaces standard GQA attention in a model trained from scratch, using a query premix and a low-rank I + AB^T postmix that fold into the query and output projections at inference, yielding a GQA-2-equivalent model with two KV heads. The cost is that the gains only exist if the architecture is adopted at pretraining time; it cannot be retrofitted onto existing pretrained checkpoints, and the reader has no pretraining budget. What can fail after adoption is unvalidated quality at scales or context lengths beyond the reported experiments, and the small accuracy deltas may not transfer to downstream fine-tuning or domain-specific workloads.", "rationale": "The technique is compelling for long-context serving but requires training a model from scratch, which the reader's constrained infrastructure and reliance on existing checkpoints or APIs preclude.", "technique": "Virtual Head Attention (VHA) for KV compression", "family": "cache_kv", "recommendation": "observar", "published": "2026-10-13", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.17959", "independent_implementations": 1, "total_implementations": 2, "stars": 5, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "VHA replaces standard GQA attention in a model trained from scratch, using a query premix and a low-rank I + AB^T postmix that fold into the query and output projections at inference, yielding a GQA-2-equivalent model with two KV heads. The cost is that the gains only exist if the architecture is adopted at pretraining time; it cannot be retrofitted onto existing pretrained checkpoints, and the reader has no pretraining budget. What can fail after adoption is unvalidated quality at scales or context lengths beyond the reported experiments, and the small accuracy deltas may not transfer to downstream fine-tuning or domain-specific workloads.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.17959", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Reduces KV cache by 4x vs. GQA-8 at dense scales and 2x vs. GQA-4 at MoE scale, cuts training/prefill FLOPs by 7.9% (1.7B, S=4096), and matches or exceeds GQA quality (+0.05% to +0.81% across four scales; +0.70% over MLA and +2.60% over GQA-4 on a 30B-A3B MoE 12-benchmark suite; RULER 84.4 vs. 73.0 for GQA at 8K-128K).", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.17959", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }