{"paper":{"title":"Robust Reasoning Benchmark","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Open-weight reasoning models lose up to 55 percent accuracy when AIME problems receive 14 simple text perturbations.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Evgenii Opryshko, Gennady Pekhimenko, Mark C. Jeffrey, Pavel Golikov","submitted_at":"2026-03-26T22:19:33Z","abstract_excerpt":"While Large Language Models (LLMs) achieve high performance on standard mathematical benchmarks, their problem-solving abilities depend on the context and textual formatting. We introduce the Robust Reasoning Benchmark (RRB), a pipeline of 13 deterministic textual perturbations applied to AIME 2024 and AIME 2025. Evaluating 8 state-of-the-art models, we find that frontier models are largely resilient, with the notable exception of Claude, which categorically refuses many transformed prompts. Open-weights reasoning models exhibit a range of failure modes under structural noise (cognitive thrash"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"open weights reasoning models suffer catastrophic collapses (up to 55% average accuracy drops across perturbations and up to 100% on some), exposing structural fragility. ... intermediate reasoning steps permanently pollute standard dense attention mechanisms.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the 14 perturbations preserve the underlying mathematical content and difficulty so that accuracy drops can be attributed specifically to reasoning or parsing failures rather than altered problem solvability.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Perturbations to math problem text cause up to 55% average accuracy drops in open-weight LLMs and sequential solving reveals context pollution in attention mechanisms.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Open-weight reasoning models lose up to 55 percent accuracy when AIME problems receive 14 simple text perturbations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"634fbc973c4addcee81e54eb1543476f09fba30caccf0dd07cd7f51220c576e4"},"source":{"id":"2604.08571","kind":"arxiv","version":2},"verdict":{"id":"945fc6ea-ed1b-4a0c-8f45-a1cf2c162f6b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T00:01:42.710966Z","strongest_claim":"open weights reasoning models suffer catastrophic collapses (up to 55% average accuracy drops across perturbations and up to 100% on some), exposing structural fragility. ... intermediate reasoning steps permanently pollute standard dense attention mechanisms.","one_line_summary":"Perturbations to math problem text cause up to 55% average accuracy drops in open-weight LLMs and sequential solving reveals context pollution in attention mechanisms.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the 14 perturbations preserve the underlying mathematical content and difficulty so that accuracy drops can be attributed specifically to reasoning or parsing failures rather than altered problem solvability.","pith_extraction_headline":"Open-weight reasoning models lose up to 55 percent accuracy when AIME problems receive 14 simple text perturbations."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.08571/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"915672ab613893c5da86adb476506c502bfc505360602597d4e3fd9a0a27ed24"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}