{"paper":{"id":"2609.00051.pdf","title":"From Detection to Refusal: Safer LLMs via Circuit-Guided Weight Scaling","authors":["Kuan-Lin Chu","Chung-En Sun","Tsui-Wei Weng"],"arxiv_id":"","url":"","source_kind":"pdf","ir_version":"1.0"},"generated_at":"2026-09-14T04:07:58.600569+00:00","counts":{"error":0,"warning":0,"info":0},"findings":[],"top_findings":[],"ir_summary":{"claims":23,"contributions":2,"method_components":7,"experiments":5,"baselines":5,"datasets":10,"metrics":7,"tables":26,"figures":3,"citations":29,"equations":0,"quantitative_results":2428,"has_repository":false},"ir_health":{"observations":[{"code":"numbers-are-identifier-fragments","severity":"warning","message":"41 of 2428 extracted numbers (2%) are digits inside identifiers, not measurements.","detail":"e.g. '2' in 'Safety Rate (%) ↑\\nLLaMA3\\nLLaMA2\\nMistral\\n', '3' in 'Safety Rate (%) ↑\\nLLaMA3\\nLLaMA2\\nMistral\\n', '2' in 'Safety Rate (%) ↑\\nLLaMA3\\nLLaMA2\\nMistral\\n'. Rules that compare numbers across sections can report a contradiction between an identifier and a real value."},{"code":"compile-warning","severity":"note","message":"references: the reference list carries no entry numbers; citation keys are enumeration order [1]..[N], not labels printed in the paper, so a key may not match the [N] used in the body","detail":""},{"code":"compile-warning","severity":"note","message":"semantic extraction: dropped 13 items whose quote was not found in the paper","detail":""}],"errors":0,"warnings":1,"trustworthy":true},"diagnostics":{"source":"/home/oceann/workspace/research-linter/eval/papers/2609.00051.pdf","rules_run":["RL001","RL002","RL003","RL004"],"rules_skipped":[{"rule":"RL005","reason":"not-applicable"}],"rule_errors":{},"rule_findings":{"RL001":0,"RL002":0,"RL003":0,"RL004":0},"rule_seconds":{"RL001":0.05,"RL002":241.92,"RL003":10.15,"RL004":0.04},"dropped_no_evidence":0,"demoted_low_confidence":0,"ir_health":{"observations":[{"code":"numbers-are-identifier-fragments","severity":"warning","message":"41 of 2428 extracted numbers (2%) are digits inside identifiers, not measurements.","detail":"e.g. '2' in 'Safety Rate (%) ↑\\nLLaMA3\\nLLaMA2\\nMistral\\n', '3' in 'Safety Rate (%) ↑\\nLLaMA3\\nLLaMA2\\nMistral\\n', '2' in 'Safety Rate (%) ↑\\nLLaMA3\\nLLaMA2\\nMistral\\n'. Rules that compare numbers across sections can report a contradiction between an identifier and a real value."},{"code":"compile-warning","severity":"note","message":"references: the reference list carries no entry numbers; citation keys are enumeration order [1]..[N], not labels printed in the paper, so a key may not match the [N] used in the body","detail":""},{"code":"compile-warning","severity":"note","message":"semantic extraction: dropped 13 items whose quote was not found in the paper","detail":""}],"errors":0,"warnings":1,"trustworthy":true},"notes":["RL001 funnel arithmetic: sentences_with_from=10 endpoint_pairs=4 claims_compared=1 emitted=0 [dropped: not-a-measurement=0]","RL001 funnel delta-column: structured_tables=18 with_exactly_one_delta_column=0 rows_checked=0 emitted=0","RL001 funnel cross-location: quant=2428 prose=1060 prose_with_metric=129 prose_with_dataset=195 prose_with_metric_and_dataset=49 distinct_buckets=2 emitted=0 [dropped: not-a-measurement=2 not-in-quote=143]","RL001 funnel identifier-drift: settings_checked=2 sections_stating_one_value=0 emitted=0","RL002 citation entailment: attempted 17 of 17 ranked cited claims (31 in-text markers from 29 bibliography entries); 5 cited works could not be identified, 7 had no retrievable text, 0 model quotes failed verbatim verification, 0 were topically unrelated to the retrieved text.","RL002 candidate ranking (why the top 17 were chosen): [21] [introduction, absolutist wording, title/sentence overlap near zero, no DOI/arXiv id]; [15] [introduction, title/sentence overlap near zero, no DOI/arXiv id]; [22] [introduction, title/sentence overlap near zero, no DOI/arXiv id]; [3] [experiment, names an artefact (PAIR), absolutist wording, title/sentence overlap near zero, no DOI/arXiv id]; [28] [experiment, names an artefact (CircuitBreakers), absolutist wording, title/sentence overlap near zero, no DOI/arXiv id]; [11] [introduction, no DOI/arXiv id]; [13] [experiment, names an artefact (WikiText), absolutist wording, title/sentence overlap near zero]; [8] [experiment, names an artefact (Malicious Instruction Dataset), title/sentence overlap near zero, no DOI/arXiv id]","RL002 label distribution over 5 classified citations: SUPPORTED=5","RL002 base-rate self-check: not tripped (0/5 = 0% NOT_SUPPORTED, threshold 40%).","RL002 attribution check (case B): 13 candidates name an artefact, 12 of them made the top-N and were checked against their cited source","RL002: 3 citations skipped because the sentence reports this paper's own results (the citation there is a comparison pointer, not evidence for the sentence)","RL002: 1 markers skipped because the claim is jointly attributed to several works","RL002: 1 enumerated-citation sentences skipped (they cite >=3 works, so the claim is about the set and cannot be refuted by any single cited work)","RL002: 5 cited works unresolvable (first 3: [{\"key\": \"[11]\", \"title\": \"Reinforcement learning from human feedback\", \"why\": \"no work record matched the bibliography title\"}, {\"key\": \"[8]\", \"title\": \"Catastrophic jailbreak of opensource llms via exploiting generation\", \"why\": \"no work record matched the bibliography title\"}, {\"key\": \"[20]\", \"title\": \"Concept bottleneck large language models\", \"why\": \"no work record matched the bibliography ti)","RL003: 3 queries -> 128 unique works (arXiv search 77, category listings 0, scholarly graph 51 [crossref 51], 0 graph works matched to an arXiv id); source errors during retrieval: arXiv 0, graph 0","RL003: parsed the reference list of 10/12 comparable papers (2 unreadable) — that is the PeerUsage denominator","RL003: enriched 1 of 2 candidates with their own abstract (1 unavailable, 0 over budget); 0 dropped as published after this paper","RL003: scholarly graph for this run = crossref","RL003 funnel (358 candidates): already_used_by_the_paper -7, cited_only_from_related_work -66, is_the_paper_itself -1, peer_usage_below_quorum -103, survey_or_artifact -38, title_names_no_method -141 || kept: citation_context_unknown +66 || survived-cheap-gates 2 -> with-abstract 1 -> comparable 2 -> above-threshold-0.65 0","RL003 candidate: 'Direct preference optimization' score 0.432 (T 0.00 D 1.00 M 0.33 P 0.40 R 0.71; 4/10 peers cite it) -> not reported","RL003 candidate: 'Inference-time intervention' score 0.412 (T 0.00 D 1.00 M 0.33 P 0.30 R 0.71; 3/10 peers cite it; title-only, no abstract available) -> not reported","RL003: 2 comparable candidates all scored below 0.65 — nothing to report for this paper","RL004 funnel: components=7 contribution_backed=3 candidates=3 ablations_found=6 matched_by_ablation=3 unmatched=0 experiments=5","RL004 component M1 'Harmful Detection Heads': covered (candidate)","RL004 component M2 'Safety Neurons': covered (candidate)","RL004 component M4 'Circuit-Based Safety Editing (CBSE)': covered (candidate)","RL004 component M5 'attention-head scaling': covered (filtered earlier)","RL004 component M6 'safety-neuron scaling': covered (filtered earlier)","RL004 component M7 'activation patching': covered (filtered earlier)","RL004 component M8 'targeted component ablation': covered (filtered earlier)","RL004 funnel: ranked=0 emitted=0 suppressed_by_cap=0 dropped_unverified_quote=0"],"missing_core_rules":[],"seconds":269.9,"llm":{"calls":5,"cache_hits":3,"errors":0,"prompt_chars":32748},"sources":{"arxiv":{"requests":21,"errors":3,"cache_hits":29,"available":1},"openalex":{"requests":1,"errors":0,"cache_hits":0,"available":0,"unavailable_reason":"OpenAlex free daily budget exhausted (resets midnight UTC)"},"crossref":{"requests":15,"errors":0,"cache_hits":5,"available":1},"s2":{"requests":0,"errors":0,"cache_hits":0,"available":1},"short_circuited":["crossref","openalex","s2"]},"cache":{"hits":37,"misses":40}},"compile_warnings":["references: the reference list carries no entry numbers; citation keys are enumeration order [1]..[N], not labels printed in the paper, so a key may not match the [N] used in the body","semantic extraction: dropped 13 items whose quote was not found in the paper"]}