{ "schema_version": 1, "builder_plan": { "scenario": "Start with one repeated product task that has a measurable failure or cost.", "change": "Read the methods and identify one change you can isolate in that workflow.", "baseline": "Your current implementation on the same held-out tasks.", "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": "Successful outcomes, response time, human rework, and total cost per successful task.", "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": "2609.17042", "title": "Learning Options for Compositional Motor Control with Adapter Banks", "summary": "The method replaces monolithic multitask policies (a task-conditioned recurrent controller) with a shared recurrent core modulated by discrete, adapter-like low-rank perturbations that a frozen high-level policy sequences into novel movements. The cost is training a closed-loop biomechanical control stack end-to-end plus a discrete latent selection mechanism, which is a research training pipeline rather than an inference-time optimization. It can fail when the target domain is not closed-loop motor control: the discrete option structure and emergent low-rank subspaces are demonstrated only in biomechanical simulation, and the claim does not transfer to language, vision, or general production ML workloads.", "rationale": "The technique addresses compositional motor control in biomechanical simulation, which falls outside the reader's LLM-focused production scope and offers no adoptable component for their inference or training constraints.", "technique": "Adapter banks over shared recurrent core for motor control", "family": "outro", "recommendation": "nao_aplica", "published": "2026-09-15", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.17042", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces monolithic multitask policies (a task-conditioned recurrent controller) with a shared recurrent core modulated by discrete, adapter-like low-rank perturbations that a frozen high-level policy sequences into novel movements. The cost is training a closed-loop biomechanical control stack end-to-end plus a discrete latent selection mechanism, which is a research training pipeline rather than an inference-time optimization. It can fail when the target domain is not closed-loop motor control: the discrete option structure and emergent low-rank subspaces are demonstrated only in biomechanical simulation, and the claim does not transfer to language, vision, or general production ML workloads.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.17042", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Up to an order-of-magnitude lower generalization error than a task-input-conditioned multitask baseline on novel closed-loop motor sequences, with a frozen-core high-level policy.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.17042", "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.E2", "label": "(2)", "section": "3 Learning Adapters for Flexible Motor Control", "role": "proposed_method", "latex": "\\tilde{h}_{t}=\\tilde{h}_{t-1}+(1-\\sigma(t))\\Delta_{z_{0}}(h_{t})+\\sigma(t)\\Delta_{z_{1}}(h_{t})", "mathml": "h~t=h~t−1+(1−σ⁡(t))​Δz0​(ht)+σ⁡(t)​Δz1​(ht)", "context": "We impose no explicit rank constraint on the adapters. Each adapter (one per latent code) Δk:ℝN→ℝN is a residual modulator of the core network’s hidden state and any low-rank structure they develop in training emerges from the optimization itself, a property we examine quantitatively in Section 4. At each decoder timestep t, the core recurrent network produces a hidden state ht from observations and the previous state, and the adapter perturbs it residually. Let z(0),z(1) denote the codes sampled by the encoders for both segments respectively, and let σ⁡(t)=∑τ≤tbτ denote the cumulative probability the segment boundary has occured by time t. At decoder timestep t, the core network produces a hidden state ht and the active adapters perturb it residually:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "A1.Ex1", "label": "", "section": "Appendix A Supplementary Methods", "role": "proposed_method", "latex": "\\Delta_{k}(\\mathbf{h})=\\mathbf{W}_{k}^{\\mathrm{out}}\\tanh\\!\\left(\\mathbf{W}_{k}^{\\mathrm{in}}\\mathbf{h}\\right),", "mathml": "Δk​(𝐡)=𝐖kout​tanh⁡(𝐖kin​𝐡),", "context": "Each of the K=10 adapters is a two-layer nonlinear residual modulator:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S2.E1", "label": "(1)", "section": "2 Low-rank Adapters via Thalamocortical Architectures in Neuroscience", "role": "baseline", "latex": "\\tau\\dot{\\mathbf{h}}=-\\mathbf{h}+\\mathbf{J}_{0}\\mathbf{h}+\\sum_{k}\\mathbf{u}_{k}\\mathbf{v}_{k}^{\\top}\\mathbf{h}=-\\mathbf{h}+\\left(\\mathbf{J}_{0}+\\sum_{k}\\mathbf{u}_{k}\\mathbf{v}_{k}^{\\top}\\right)\\mathbf{h}.", "mathml": "τ​𝐡˙=−𝐡+𝐉0​𝐡+∑k𝐮k​𝐯k⊤​𝐡=−𝐡+(𝐉0+∑k𝐮k​𝐯k⊤)​𝐡.", "context": "Let 𝐡∈ℝN denote the cortical state and 𝐉0∈ℝN×N the recurrent connectivity in cortex. Each thalamocortical channel k consists of a corticothalamic readout 𝐯k∈ℝN and a thalamocortical projection 𝐮k∈ℝN. The cortical dynamics evolve as", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-10-17", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }