{"paper":{"title":"Hidden State Poisoning Attacks against Mamba-based Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Short input phrases can irreversibly overwrite hidden states in Mamba models, inducing amnesia on retrieval tasks that pure Transformers resist.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Alexandre Le Mercier, Chris Develder, Thomas Demeester","submitted_at":"2026-01-05T10:27:19Z","abstract_excerpt":"State space models (SSMs) like Mamba offer efficient alternatives to Transformer-based language models, with linear time complexity. Yet, their adversarial robustness remains critically unexplored. This paper studies the phenomenon whereby specific short input phrases induce a partial amnesia effect in such models, by irreversibly overwriting information in their hidden states, referred to as a Hidden State Poisoning Attack (HiSPA). Our benchmark RoBench-25 allows evaluating a model's information retrieval capabilities when subject to HiSPAs, and confirms the vulnerability of SSMs against such"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Specific short input phrases induce a partial amnesia effect in Mamba-based models by irreversibly overwriting information in their hidden states, confirmed by collapse on RoBench-25 while pure Transformers remain unaffected.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the observed performance drops are caused specifically by irreversible hidden-state overwriting rather than other mechanisms such as attention disruption or output formatting changes, and that RoBench-25 isolates this effect without confounding factors.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Short input phrases can irreversibly overwrite hidden states in Mamba models, impairing information retrieval on a new benchmark while leaving pure Transformer models unaffected.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Short input phrases can irreversibly overwrite hidden states in Mamba models, inducing amnesia on retrieval tasks that pure Transformers resist.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b4701f3450a2a9968c79a376510e18607e1a66592cf234c6d26d6ad247c438eb"},"source":{"id":"2601.01972","kind":"arxiv","version":4},"verdict":{"id":"2002e7c8-50e3-4adc-9373-0b2a053178d9","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T18:12:20.748575Z","strongest_claim":"Specific short input phrases induce a partial amnesia effect in Mamba-based models by irreversibly overwriting information in their hidden states, confirmed by collapse on RoBench-25 while pure Transformers remain unaffected.","one_line_summary":"Short input phrases can irreversibly overwrite hidden states in Mamba models, impairing information retrieval on a new benchmark while leaving pure Transformer models unaffected.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the observed performance drops are caused specifically by irreversible hidden-state overwriting rather than other mechanisms such as attention disruption or output formatting changes, and that RoBench-25 isolates this effect without confounding factors.","pith_extraction_headline":"Short input phrases can irreversibly overwrite hidden states in Mamba models, inducing amnesia on retrieval tasks that pure Transformers resist."},"references":{"count":25,"sample":[{"doi":"","year":null,"title":"arXiv preprint arXiv:2402.01771 , year=","work_id":"884dfd62-5356-4dbe-ad95-772813d81cca","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Nemotron-h: A family of accurate and efficient hybrid mamba-transformer models","work_id":"134d0eac-bc97-4065-9c60-e8fde9d20b48","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Mahabaleshwarkar, Shih- Yang Liu, Matthijs Van Keirsbilck, Min-Hung Chen, Yoshi Suhara, et al","work_id":"283d6d50-8219-44e7-94e7-eb6b8673dec7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Investigating the indirect object identification circuit in mamba.arXiv preprint arXiv:2407.14008,","work_id":"8d48c229-d2f2-4228-9412-6a21c701ea60","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"The Pile: An 800GB Dataset of Diverse Text for Language Modeling","work_id":"9b10667a-da61-4358-aceb-10578234d45d","ref_index":5,"cited_arxiv_id":"2101.00027","is_internal_anchor":true}],"resolved_work":25,"snapshot_sha256":"27dfefc519c323130591e8526b9e56914a3276ee7d8a3d1b2834400fa7016149","internal_anchors":5},"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"}