{"paper":{"title":"Tracing the Dynamics of Refusal: Exploiting Latent Refusal Trajectories for Robust Jailbreak Detection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Refusal in language models follows a persistent upstream trajectory that remains detectable even when attacks suppress the final output.","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Che Wang, Jianbo Gao, Wei Yang Bryan Lim, Xulin Hu, Zhong Chen","submitted_at":"2026-05-02T14:56:37Z","abstract_excerpt":"Representation Engineering analyses often characterize refusal using static directions extracted from terminal or pooled representations. We ask whether this view misses how refusal is constructed across layer-token positions. Using causal tracing, we identify a \\textit{Refusal Trajectory}: a sparse upstream activation pattern that often persists even when attacks such as GCG suppress terminal refusal signals. Based on this observation, we propose SALO (Sparse Activation Localization Operator), a lightweight white-box detector that operates on raw hidden-state volumes from a selected layer win"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"SALO effectively recovers defense capabilities against forced-decoding attacks, improving detection rates from ~0% to >90% where methods relying on terminal states perform poorly.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the refusal trajectory uncovered by causal tracing is a stable, generalizable signature rather than an artifact of the specific models, prompts, or attack implementations tested.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Refusal in LLMs leaves a detectable upstream trajectory that SALO exploits to raise jailbreak detection from near zero to over 90 percent even under forced-decoding attacks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Refusal in language models follows a persistent upstream trajectory that remains detectable even when attacks suppress the final output.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"98f8241d2d7d8505d4472f6174cf11b249fde10448a9087056e960d42b906000"},"source":{"id":"2605.02958","kind":"arxiv","version":2},"verdict":{"id":"6cc86458-824d-4c8e-a196-1b6d49eba5b7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-09T14:08:46.459279Z","strongest_claim":"SALO effectively recovers defense capabilities against forced-decoding attacks, improving detection rates from ~0% to >90% where methods relying on terminal states perform poorly.","one_line_summary":"Refusal in LLMs leaves a detectable upstream trajectory that SALO exploits to raise jailbreak detection from near zero to over 90 percent even under forced-decoding attacks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the refusal trajectory uncovered by causal tracing is a stable, generalizable signature rather than an artifact of the specific models, prompts, or attack implementations tested.","pith_extraction_headline":"Refusal in language models follows a persistent upstream trajectory that remains detectable even when attacks suppress the final output."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.02958/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T17:41:08.913207Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:16:24.573853Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"75ce4b50a859a889d1999293782c18672c4f12c631159e8e7cd1f0e05517dd4d"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}