{"paper":{"title":"LPDS: Evaluating LLM Robustness Through Logic-Preserving Difficulty Scaling","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Logic-preserving difficulty scaling finds problem variations that cause language models to fail up to five times more often than random tests.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Dieuwke Hupkes, Jesse Dodge, Philipp Mondorf, Samuel J. Bell","submitted_at":"2026-05-14T20:26:59Z","abstract_excerpt":"As large language models (LLMs) are increasingly deployed to perform tasks with minimal human oversight, it is crucial that these models operate robustly. In particular, a model that can solve a given problem should not fail simply because certain entities$\\unicode{x2013}$such as names, numbers, or other contextual details$\\unicode{x2013}$have changed while the underlying problem logic remains the same. Prior work suggests that current LLMs still struggle with this form of robustness: they often succeed on some variations of a problem but fail on others. However, existing evaluations often lac"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We show that LPDS efficiently finds difficult problem variations for a model, resulting in performance drops up to 5 times larger compared to random sampling.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The framework assumes that difficulty of a logic-preserving variation can be quantified in a model-agnostic or at least transferable way that reliably predicts where failures will occur, and that the search procedure finds variations that are truly harder rather than merely different.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LPDS quantifies difficulty of logic-preserving problem variations and searches for the hardest ones, producing up to 5x larger performance drops than random sampling and better robustness gains from fine-tuning on difficult examples.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Logic-preserving difficulty scaling finds problem variations that cause language models to fail up to five times more often than random tests.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"4868c284b20cd34d47ec39df83f807d8e395912041b769ba1cd18b3fea78ef97"},"source":{"id":"2605.15393","kind":"arxiv","version":1},"verdict":{"id":"f3716853-91ee-4a80-8f50-c16cebf4021b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T16:28:43.234195Z","strongest_claim":"We show that LPDS efficiently finds difficult problem variations for a model, resulting in performance drops up to 5 times larger compared to random sampling.","one_line_summary":"LPDS quantifies difficulty of logic-preserving problem variations and searches for the hardest ones, producing up to 5x larger performance drops than random sampling and better robustness gains from fine-tuning on difficult examples.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The framework assumes that difficulty of a logic-preserving variation can be quantified in a model-agnostic or at least transferable way that reliably predicts where failures will occur, and that the search procedure finds variations that are truly harder rather than merely different.","pith_extraction_headline":"Logic-preserving difficulty scaling finds problem variations that cause language models to fail up to five times more often than random tests."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.15393/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T17:01:18.258220Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T16:35:47.406696Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T16:26:23.761732Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"citation_quote_validity","ran_at":"2026-05-19T15:50:56.443617Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T14:21:54.167704Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.721310Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"2a6e65153d0868f98a03a80e4de9ddc6b180ac3634a4fd94fe06996ca19c1e01"},"references":{"count":74,"sample":[{"doi":"10.1145/3641289","year":2024,"title":"and Yang, Qiang and Xie, Xing , number =","work_id":"baa1ad58-22ae-43a4-86d6-49b6915c12d6","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1145/3747588","year":2025,"title":"A survey on large language models for code generation","work_id":"2362b076-10d0-4e85-8b67-08ac034b2069","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2023.acl-long.307","year":2023,"title":"Evaluating Open-Domain Question Answering in the Era of Large Language Models","work_id":"4c6f7651-e3e6-412e-acdc-0cbc939ca956","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models","work_id":"0b361fed-cf2a-4b90-b61a-de88de4b8840","ref_index":4,"cited_arxiv_id":"2503.09567","is_internal_anchor":true},{"doi":"10.1038/s41586-025-09422-z","year":null,"title":"Nature645(8081), 633–638 (2025) https://doi.org/10.1038/s41586-025-09422-z","work_id":"9835b482-5032-4135-93dd-82a066677569","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":74,"snapshot_sha256":"8d521af3a732175332d3a8294f64a2f501aa840d73096aa720ca9f7b35373ea0","internal_anchors":12},"formal_canon":{"evidence_count":2,"snapshot_sha256":"b2597318cb21bd8a4d7b69ba05971b9da37bd0e8e8b9f7c0c1e3d77a0a99665b"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}