{"paper":{"title":"Epistemic Regret Minimization: Label-Free Causal Critique Beyond Outcome Reward","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Epistemic Regret Minimization identifies causal flaws in LLM reasoning traces without ground-truth labels","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Edward Y. Chang, Longling Geng","submitted_at":"2026-02-12T07:48:21Z","abstract_excerpt":"Large language models can answer causal questions correctly for the wrong reasons. Current RL methods reward \\emph{what} a model concludes but ignore \\emph{why}, reinforcing correlational shortcuts -- a failure we call \\emph{Reward Entrenchment}. We introduce \\emph{Epistemic Regret Minimization} (\\erm), a framework that critiques the causal \\emph{structure} of a model's reasoning trace rather than its answer. Applying established causal principles, \\erm flags unexamined confounders, correlation--intervention conflation, and unchecked back-door paths from exposed reasoning traces. The framework"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"A separation theorem proves outcome-only RL cannot distinguish correct from flawed causal models in confounded environments, and preliminary experiments show epistemic reward carries signal where outcome reward does not.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That causal flaws are reliably identifiable and correctable from reasoning traces alone without ground-truth labels or external verifiers, and that the observed corrections on CausalT5K and CLadder generalize beyond the tested models and scenarios.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Epistemic Regret Minimization identifies causal reasoning flaws in LLMs from traces alone, corrects stubborn models where outcome-only methods fail, and is supported by a separation theorem proving outcome-only RL cannot distinguish correct from flawed causal models in confounded settings.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Epistemic Regret Minimization identifies causal flaws in LLM reasoning traces without ground-truth labels","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3b31d6e4d0acc81bc9f0dd8e9cf8877ca93e9d0881f07e5a7cd81c242658ada2"},"source":{"id":"2602.11675","kind":"arxiv","version":4},"verdict":{"id":"c94b2c6a-f016-4dca-a646-1cde4ebdba1b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T05:15:57.563360Z","strongest_claim":"A separation theorem proves outcome-only RL cannot distinguish correct from flawed causal models in confounded environments, and preliminary experiments show epistemic reward carries signal where outcome reward does not.","one_line_summary":"Epistemic Regret Minimization identifies causal reasoning flaws in LLMs from traces alone, corrects stubborn models where outcome-only methods fail, and is supported by a separation theorem proving outcome-only RL cannot distinguish correct from flawed causal models in confounded settings.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That causal flaws are reliably identifiable and correctable from reasoning traces alone without ground-truth labels or external verifiers, and that the observed corrections on CausalT5K and CLadder generalize beyond the tested models and scenarios.","pith_extraction_headline":"Epistemic Regret Minimization identifies causal flaws in LLM reasoning traces without ground-truth labels"},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.11675/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":3,"snapshot_sha256":"60cc0fac08e1c6b18781b3250b295ff749ea533a854ae105f5fb76b8d331e0f0"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}