{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:7RVBZ3TJVGIMIBRUKW3VNGMQON","short_pith_number":"pith:7RVBZ3TJ","schema_version":"1.0","canonical_sha256":"fc6a1cee69a990c4063455b7569990735793d44a7c9e012b976c368aab7d7d84","source":{"kind":"arxiv","id":"2402.18344","version":2},"attestation_state":"computed","paper":{"title":"Focus on Your Question! Interpreting and Mitigating Toxic CoT Problems in Commonsense Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Chenhao Wang, Daojian Zeng, Jiachun Li, Jun Zhao, Kang Liu, Pengfei Cao, Yubo Chen, Zhuoran Jin","submitted_at":"2024-02-28T14:09:02Z","abstract_excerpt":"Large language models exhibit high-level commonsense reasoning abilities, especially with enhancement methods like Chain-of-Thought (CoT). However, we find these CoT-like methods lead to a considerable number of originally correct answers turning wrong, which we define as the Toxic CoT problem. To interpret and mitigate this problem, we first utilize attribution tracing and causal tracing methods to probe the internal working mechanism of the LLM during CoT reasoning. Through comparisons, we prove that the model exhibits information loss from the question over the shallow attention layers when"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2402.18344","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-02-28T14:09:02Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"f74a2ed32f57995f1977421d80610b0287283e290b733899da248bc041ef87f1","abstract_canon_sha256":"15d4edb4516c8456fd7331c4f5bbe21e25a6c0824d966af3b864881f161fc0c6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:19:26.899687Z","signature_b64":"gFELI4URogXskxpk41X1F1l8tzBMvs3VBOr4n8Sfj/DyVdNx+YWCnTt2XYhSBvvnhyybHSHkpIm/pbQzEQnDAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fc6a1cee69a990c4063455b7569990735793d44a7c9e012b976c368aab7d7d84","last_reissued_at":"2026-07-05T09:19:26.899222Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:19:26.899222Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Focus on Your Question! Interpreting and Mitigating Toxic CoT Problems in Commonsense Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Chenhao Wang, Daojian Zeng, Jiachun Li, Jun Zhao, Kang Liu, Pengfei Cao, Yubo Chen, Zhuoran Jin","submitted_at":"2024-02-28T14:09:02Z","abstract_excerpt":"Large language models exhibit high-level commonsense reasoning abilities, especially with enhancement methods like Chain-of-Thought (CoT). However, we find these CoT-like methods lead to a considerable number of originally correct answers turning wrong, which we define as the Toxic CoT problem. To interpret and mitigate this problem, we first utilize attribution tracing and causal tracing methods to probe the internal working mechanism of the LLM during CoT reasoning. Through comparisons, we prove that the model exhibits information loss from the question over the shallow attention layers when"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.18344","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2402.18344/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":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2402.18344","created_at":"2026-07-05T09:19:26.899277+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.18344v2","created_at":"2026-07-05T09:19:26.899277+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.18344","created_at":"2026-07-05T09:19:26.899277+00:00"},{"alias_kind":"pith_short_12","alias_value":"7RVBZ3TJVGIM","created_at":"2026-07-05T09:19:26.899277+00:00"},{"alias_kind":"pith_short_16","alias_value":"7RVBZ3TJVGIMIBRU","created_at":"2026-07-05T09:19:26.899277+00:00"},{"alias_kind":"pith_short_8","alias_value":"7RVBZ3TJ","created_at":"2026-07-05T09:19:26.899277+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2507.10085","citing_title":"Enhancing Chain-of-Thought Reasoning with Critical Representation Fine-tuning","ref_index":16,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON","json":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON.json","graph_json":"https://pith.science/api/pith-number/7RVBZ3TJVGIMIBRUKW3VNGMQON/graph.json","events_json":"https://pith.science/api/pith-number/7RVBZ3TJVGIMIBRUKW3VNGMQON/events.json","paper":"https://pith.science/paper/7RVBZ3TJ"},"agent_actions":{"view_html":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON","download_json":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON.json","view_paper":"https://pith.science/paper/7RVBZ3TJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.18344&json=true","fetch_graph":"https://pith.science/api/pith-number/7RVBZ3TJVGIMIBRUKW3VNGMQON/graph.json","fetch_events":"https://pith.science/api/pith-number/7RVBZ3TJVGIMIBRUKW3VNGMQON/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON/action/timestamp_anchor","attest_storage":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON/action/storage_attestation","attest_author":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON/action/author_attestation","sign_citation":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON/action/citation_signature","submit_replication":"https://pith.science/pith/7RVBZ3TJVGIMIBRUKW3VNGMQON/action/replication_record"}},"created_at":"2026-07-05T09:19:26.899277+00:00","updated_at":"2026-07-05T09:19:26.899277+00:00"}