{ "schema_version": 1, "builder_plan": { "scenario": "Your workload might fit a smaller or differently structured model.", "change": "Check for available weights and a compatible runtime before planning a product trial.", "baseline": "Your current model on identical held-out tasks and the same resource budget.", "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": "Task success, latency, memory, context limits, and migration cost.", "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": "2609.16183", "title": "Anatomy of Associative Recall in Fixed-State Recurrences: A Matched-State Decomposition, an Interference Wall, and a Curriculum That Breaks It", "summary": "This replaces the assumption that fixed-state recurrences (linear attention, SSMs) are inherently poor at associative recall: it attributes the gap to the missing short convolution and to training interference, and fixes both via a convolution plus a distance or accuracy-gated curriculum. The cost is architectural modification to the recurrent cell and a curriculum schedule with per-seed retraining, all on synthetic benchmarks rather than validated production tasks. It can fail because results are confined to synthetic masked multi-query recall, lock-in remains a seed lottery (7/10, not 10/10), the uniform curriculum and simple ramp collapse at longer lengths, and transfer to real language-modeling workloads is unproven.", "rationale": "The interventions are cheap and small-team feasible, but the evidence comes from controlled synthetic recall tasks, so adoption in production models requires workload-specific validation before committing.", "technique": "Conv-armed recurrent cells with recall curriculum", "family": "arquitetura_eficiente", "recommendation": "testar", "published": "2026-09-14", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.16183", "independent_implementations": 0, "total_implementations": 1, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "This replaces the assumption that fixed-state recurrences (linear attention, SSMs) are inherently poor at associative recall: it attributes the gap to the missing short convolution and to training interference, and fixes both via a convolution plus a distance or accuracy-gated curriculum. The cost is architectural modification to the recurrent cell and a curriculum schedule with per-seed retraining, all on synthetic benchmarks rather than validated production tasks. It can fail because results are confined to synthetic masked multi-query recall, lock-in remains a seed lottery (7/10, not 10/10), the uniform curriculum and simple ramp collapse at longer lengths, and transfer to real language-modeling workloads is unproven.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.16183", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "A short causal convolution adds ~0.5 recall at matched state and training; a distance-based curriculum raises 32-pair distractor-haystack retrieval from 0.021 to 1.000 and raises seed lock-in from 1/10 to 7/10 (p=0.02), with accuracy-gated ramps reaching 6/6 at L=512 (p=0.001).", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.16183", "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": [ { "anchor": "S3.E1", "label": "(1)", "section": "3 Cells, Tasks, and Protocol", "role": "baseline", "latex": "\\mathbf{S}_{t}=\\mathbf{S}_{t-1}\\!\\left(\\mathbf{I}-\\beta_{t}k_{t}k_{t}^{\\top}\\right)+\\beta_{t}\\,v_{t}k_{t}^{\\top},\\quad o_{t}=\\mathbf{S}_{t}q_{t}.", "mathml": "𝐒t=𝐒t−1​(𝐈−βt​kt​kt⊤)+βt​vt​kt⊤,ot=𝐒t​qt.", "context": "DeltaNet (rank-1, no decay) applies the delta rule (Schlag et al., 2021; Yang et al., 2024b), with input-dependent write strength βt∈(0,1):", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E2", "label": "(2)", "section": "3 Cells, Tasks, and Protocol", "role": "proposed_method", "latex": "\\mathbf{S}_{t}=\\alpha_{t}\\,\\mathbf{S}_{t-1}\\!\\left(\\mathbf{I}-\\beta_{t}k_{t}k_{t}^{\\top}\\right)+\\beta_{t}\\,v_{t}k_{t}^{\\top}.", "mathml": "𝐒t=αt​𝐒t−1​(𝐈−βt​kt​kt⊤)+βt​vt​kt⊤.", "context": "Gated DeltaNet (gated_deltanet; rank-1 + decay) adds a learned scalar gate αt∈(0,1) (Yang et al., 2024a):", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E5", "label": "(5)", "section": "3 Cells, Tasks, and Protocol", "role": "baseline", "latex": "\\mathbf{H}_{t}=a_{t}\\,\\mathbf{H}_{t-1}+x_{t}\\,b_{t}^{\\top},\\qquad y_{t}=\\mathbf{H}_{t}\\,c_{t},", "mathml": "𝐇t=at​𝐇t−1+xt​bt⊤,yt=𝐇t​ct,", "context": "Mamba-2 reference (mamba2_ref) is a pure-PyTorch state-matched implementation of the Mamba-2 SSD recurrence (Dao and Gu, 2024): a per-head scalar selective diagonal scan over an explicit state 𝐇t∈ℝdinner×dstate,", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-10-11", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }