{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "A product assistant retrieves plausible documents but gives unsupported answers.",
    "change": "Change one retrieval or ranking step while keeping the document snapshot fixed.",
    "baseline": "Your current retrieval pipeline with identical questions and answer model.",
    "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": "Supported answers, retrieval misses, correct abstentions, index cost, and response latency.",
    "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": "2610.23807",
  "title": "Steering beyond Language: Improving Multilingual Training-Free Text Embeddings via Paired-Residual Subspace Removal",
  "summary": "It replaces standard last-token hidden-state embeddings from frozen LLMs with representations from which layer-wise low-rank language-specific subspaces (estimated from parallel-sentence residuals) have been removed. The cost is inference-time overhead: a one-time subspace estimation requiring multilingual parallel sentences plus a per-layer projection at embedding time, with no fine-tuning. It can fail if parallel data for the target language pair is unavailable or unrepresentative, if the low-rank assumption removes semantically relevant components, or if gains do not transfer to languages or models outside the evaluated set.",
  "rationale": "The method is training-free and runs on a single GPU with frozen models, fitting constrained infrastructure, but its dependence on parallel sentences and layer-wise subspace quality requires validation on the reader's specific language pairs and retrieval workload before production adoption.",
  "technique": "Language-subspace removal for multilingual LLM embeddings",
  "family": "recuperacao_e_rag",
  "recommendation": "testar",
  "published": "2026-10-16",
  "as_of": "2026-10-19",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.23807",
  "independent_implementations": 0,
  "total_implementations": 1,
  "stars": 122,
  "citations": null,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "It replaces standard last-token hidden-state embeddings from frozen LLMs with representations from which layer-wise low-rank language-specific subspaces (estimated from parallel-sentence residuals) have been removed. The cost is inference-time overhead: a one-time subspace estimation requiring multilingual parallel sentences plus a per-layer projection at embedding time, with no fine-tuning. It can fail if parallel data for the target language pair is unavailable or unrepresentative, if the low-rank assumption removes semantically relevant components, or if gains do not transfer to languages or models outside the evaluated set.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.23807",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "PRSR consistently improves training-free embedding baselines across 200+ monolingual MTEB/MMTEB evaluations in 18 languages and English-to-200+ language bitext mining, but no single quantified factor is reported in the abstract.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.23807",
      "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": ""
}