{"paper":{"title":"Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs","arxiv_id":"2611.04782","primary_cat":"cs.AI","submitted_at":"2026-11-03T12:22:01+00:00","version_compiled":"v1"},"notice":"Every quoted sentence is the authors' text from the arXiv listing abstract. Labels and probabilities are the outputs of classifier-1.13.0 on the question bank abstract-p1-20260923, exactly as returned. No sentence in this record was written by a language model.","schema":"pith.compiled-record/v0","status":"compiled","abstract":{"source":"arXiv listing abstract v1","sentences":[{"i":0,"p":1.0,"id":"s0","own":false,"role":"background","text":"SAE-based feature-level interventions to suppress PII leakage in LLMs are primarily evaluated by log-probability reduction.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.07,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":1.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":1,"p":0.64,"id":"s1","own":true,"role":"result","text":"We show that this criterion cannot distinguish feature specific suppression from a magnitude artifact: a direction-agnostic effect of sufficiently large activation perturbations.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.87,"main_claim_p":0.3,"quantitative":false,"quantitative_p":0.1,"role_probabilities":{"gap":0.17,"method":0.0,"result":0.64,"background":0.01,"limitation":0.0,"future_work":0.0,"contribution":0.15,"interpretation":0.03}},{"i":2,"p":0.93,"id":"s2","own":true,"role":"method","text":"Across 36 (model, layer, PII type) cases, we compare selected SAE feature interventions with coherencematched random direction controls using logprob and rank-aware metrics.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.95,"main_claim_p":0.04,"quantitative":true,"quantitative_p":0.97,"role_probabilities":{"gap":0.0,"method":0.93,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.07,"interpretation":0.0}},{"i":3,"p":1.0,"id":"s3","own":true,"role":"result","text":"While 17 cases (47%) appear successful under logprob-only evaluation, only 1 case (2.8%) passes our multimetric framework.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.96,"main_claim_p":0.27,"quantitative":true,"quantitative_p":0.99,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":4,"p":0.97,"id":"s4","own":true,"role":"result","text":"The rest reveal two failure modes: random directions outperform selected features, and feature selection collapse across PII types.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.92,"main_claim_p":0.12,"quantitative":true,"quantitative_p":0.76,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.97,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.03}},{"i":5,"p":1.0,"id":"s5","own":true,"role":"interpretation","text":"These results indicate that existing evaluations cannot verify whether interventions target causal pathways of PII leakage.","hedge":{"p":0.95,"label":"hedged_possibility"},"own_p":0.61,"main_claim_p":0.63,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":1.0}},{"i":6,"p":0.87,"id":"s6","own":true,"role":"result","text":"A 9,000 run causal mediation feasibility study corroborates this: the verifier separates candidates reliably, but correlation-selected candidates rarely pass it.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.91,"main_claim_p":0.13,"quantitative":true,"quantitative_p":0.97,"role_probabilities":{"gap":0.0,"method":0.02,"result":0.87,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.1,"interpretation":0.01}},{"i":7,"p":1.0,"id":"s7","own":false,"role":"future_work","text":"We therefore propose a new research mission for NLP: causally grounded PII suppression through circuit-level localization and causal mediation analysis.","hedge":{"p":0.89,"label":"asserted"},"own_p":0.15,"main_claim_p":0.36,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":0.0,"limitation":0.0,"future_work":1.0,"contribution":0.0,"interpretation":0.0}}]},"compiled":{"main_claim":{"p":0.89,"text":"These results indicate that existing evaluations cannot verify whether interventions target causal pathways of PII leakage.","location":"abstract s5","hedge_strength":{"p":0.95,"label":"hedged possibility"}},"provenance":{"route":{"via":"TypeSafe API","model":"classifier-1.13.0","endpoint":"https://api.typesafe.ai/v1/systemone","provider":"TypeSafe"},"packing":"one request per paper","requests":1,"input_tokens":6205,"cost_usd_reported":0.0},"compiled_at":"2026-11-04","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.07,"states_limitation":0.31,"human_participants":0.03,"controlled_comparison":0.98,"randomized_assignment":0.08,"states_quantitative_result":0.99,"compares_with_prior_results":0.92,"mentions_code_or_data_availability":0.03},"evidential_basis":{"p":0.84,"value":"observation or experiment","probabilities":{"argument":0.0,"literature":0.0,"proof_or_derivation":0.0,"analysis_of_existing_data":0.0,"observation_or_experiment":0.84,"simulation_or_computation":0.16}},"contribution_type":{"p":0.97,"value":"empirical study","probabilities":{"review":0.0,"method_or_tool":0.0,"empirical_study":0.97,"simulation_study":0.0,"theoretical_result":0.0,"dataset_or_benchmark":0.0,"position_or_commentary":0.03}},"contribution_statement":{"p":0.73,"text":"Across 36 (model, layer, PII type) cases, we compare selected SAE feature interventions with coherencematched random direction controls using logprob and rank-aware metrics.","location":"abstract s2"}},"coverage":{"tier":"abstract","levels_run":["P abstract"],"not_examined":["full text","figures and tables","references and citing sentences","relations to other papers"],"sources_read":["arXiv listing abstract v1"],"levels_not_run":["S sections","T body sentences","C citations","R cross-paper relations"],"sentences_examined":8,"sentences_in_abstract":8},"measured_accuracy":{"url":"https://pith.science/accuracy","measured_on":"2026-10-02","method":"two independent model families labelled a fresh random sample blind; the figure is how often the published label matches their consensus","sentence_role_agreement":0.92,"main_claim_agreement":0.908,"typed_citation_agreement":0.933,"typed_citation_agreement_95_interval":[0.917,0.945],"review_comments_accurate":0.758,"review_comments_accurate_95_interval":[0.684,0.82],"review_caught_admitted_error_found_or_related":0.448,"review_comments_judged":{"judged_by":["gpt-5.6-luna","grok-4.7"],"judged_against":"Pith's own text copy of each paper: text extracted from the PDF, the first 90,000 characters","measured_on":"2026-10-04","accurate":113,"decided":149,"sampled_reviews":200,"split_by_lane":false,"by_lane":"https://pith.science/accuracy#by-lane"}}}