{"paper":{"title":"Locale-Conditioned Few-Shot Prompting Mitigates Demonstration Regurgitation in On-Device PII Substitution with Small Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Locale-conditioned rotating few-shot prompts stop small language models from echoing demonstration examples during on-device PII substitution.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Anuj Sadani, Deepak Kumar","submitted_at":"2026-05-13T13:47:11Z","abstract_excerpt":"Personally Identifiable Information (PII) redaction usually replaces detected entities with placeholder tokens such as [PERSON], destroying the downstream utility of the redacted text for retrieval and Named Entity Recognition (NER) training. We propose a fully on-device pipeline that substitutes PII with consistent, type-preserving fake values: a 1.5 B mixture-of-experts token classifier (openai/privacy-filter) detects spans, a 1-bit Bonsai-1.7B Small Language Model (SLM) proposes contextual surrogates for names, addresses, and dates, and a rule-based generator (faker) handles patterned field"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"With the fix, 482/482 unique Bonsai-1.7B calls succeed (no echoes) and produce locale-correct surrogates, although the SLM still copies from a small same-locale demonstration pool - a residual narrowness we quantify.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the observed NER performance gap is caused primarily by reduced variety in SLM outputs rather than by other unmeasured differences in the 160/40 subset or by the choice of XGLM-564M as the multilingual evaluator.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Locale-conditioned rotating few-shot prompting eliminates demonstration regurgitation in 1.7B SLMs for PII substitution while producing more natural text than rule-based methods, though downstream NER training benefits more from synthetic variety than naturalness.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Locale-conditioned rotating few-shot prompts stop small language models from echoing demonstration examples during on-device PII substitution.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1767d0d1798e97fd46bb4e9176590148e9eb44ae983ad5d85fb458dc47ca6388"},"source":{"id":"2605.13538","kind":"arxiv","version":1},"verdict":{"id":"fddab2d7-c488-4d6a-8201-d6bc53eabbb7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:00:57.673618Z","strongest_claim":"With the fix, 482/482 unique Bonsai-1.7B calls succeed (no echoes) and produce locale-correct surrogates, although the SLM still copies from a small same-locale demonstration pool - a residual narrowness we quantify.","one_line_summary":"Locale-conditioned rotating few-shot prompting eliminates demonstration regurgitation in 1.7B SLMs for PII substitution while producing more natural text than rule-based methods, though downstream NER training benefits more from synthetic variety than naturalness.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the observed NER performance gap is caused primarily by reduced variety in SLM outputs rather than by other unmeasured differences in the 160/40 subset or by the choice of XGLM-564M as the multilingual evaluator.","pith_extraction_headline":"Locale-conditioned rotating few-shot prompts stop small language models from echoing demonstration examples during on-device PII substitution."},"references":{"count":16,"sample":[{"doi":"","year":2024,"title":"Faker: a Python package that generates fake data","work_id":"b1918f04-be18-4cdc-8b1b-8098eac3c9ec","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"llama.cpp: Port of LLaMA models in C/C++","work_id":"6b334f6b-fac2-4e0c-910e-43ed4197b5c2","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"spaCy: Industrial-strength natural language processing in Python","work_id":"8878703d-2330-47d2-81ec-6f3b4c7eacb6","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Few-shot learning with multilingual language models","work_id":"32d56185-d394-41f6-a038-08c943a4650c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity","work_id":"c6ecb82b-59ca-437f-91d8-470531f9f7c8","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":16,"snapshot_sha256":"8f7fc02398e1eadab2ea9d9fadc8e255a1580afbb705951483156df5e9d4b78b","internal_anchors":3},"formal_canon":{"evidence_count":1,"snapshot_sha256":"13cea1224d6f07fff5395eb607961feb0d0eb9b58ef3be5215275e4827c88642"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}