{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:MMITW4YHDAD4YYBMXAM2EJMGY6","short_pith_number":"pith:MMITW4YH","schema_version":"1.0","canonical_sha256":"63113b73071807cc602cb819a22586c78bc07d370b2f6923eb0a3a7c92ad6f69","source":{"kind":"arxiv","id":"2503.04722","version":2},"attestation_state":"computed","paper":{"title":"Enough Coin Flips Can Make LLMs Act Bayesian","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Dan Klein, David M. Chan, Eric Wang, Jiaxin Ge, Ritwik Gupta, Rodolfo Corona, Trevor Darrell","submitted_at":"2025-03-06T18:59:23Z","abstract_excerpt":"Large language models (LLMs) exhibit the ability to generalize given few-shot examples in their input prompt, an emergent capability known as in-context learning (ICL). We investigate whether LLMs use ICL to perform structured reasoning in ways that are consistent with a Bayesian framework or rely on pattern matching. Using a controlled setting of biased coin flips, we find that: (1) LLMs often possess biased priors, causing initial divergence in zero-shot settings, (2) in-context evidence outweighs explicit bias instructions, (3) LLMs broadly follow Bayesian posterior updates, with deviations"},"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":"2503.04722","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-03-06T18:59:23Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"c218361e1991ddc254b45efd1e90d41811917b0655cc73a8886579897c387923","abstract_canon_sha256":"0c3093be0f8a55b9f0592d3262aee197fded73b83da17e9eee9433d0a3e4a206"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:29:07.825577Z","signature_b64":"BOYqt/GBNlCvqfM8nf39n6s7HvJzkjPzKKocoTfHm9J2G+jiC3fuhjOjyJlcaAFGsB5lRl1bwxjxYrYPKh87Bg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"63113b73071807cc602cb819a22586c78bc07d370b2f6923eb0a3a7c92ad6f69","last_reissued_at":"2026-07-05T11:29:07.825006Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:29:07.825006Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Enough Coin Flips Can Make LLMs Act Bayesian","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Dan Klein, David M. Chan, Eric Wang, Jiaxin Ge, Ritwik Gupta, Rodolfo Corona, Trevor Darrell","submitted_at":"2025-03-06T18:59:23Z","abstract_excerpt":"Large language models (LLMs) exhibit the ability to generalize given few-shot examples in their input prompt, an emergent capability known as in-context learning (ICL). We investigate whether LLMs use ICL to perform structured reasoning in ways that are consistent with a Bayesian framework or rely on pattern matching. Using a controlled setting of biased coin flips, we find that: (1) LLMs often possess biased priors, causing initial divergence in zero-shot settings, (2) in-context evidence outweighs explicit bias instructions, (3) LLMs broadly follow Bayesian posterior updates, with deviations"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.04722","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/2503.04722/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":"2503.04722","created_at":"2026-07-05T11:29:07.825075+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.04722v2","created_at":"2026-07-05T11:29:07.825075+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.04722","created_at":"2026-07-05T11:29:07.825075+00:00"},{"alias_kind":"pith_short_12","alias_value":"MMITW4YHDAD4","created_at":"2026-07-05T11:29:07.825075+00:00"},{"alias_kind":"pith_short_16","alias_value":"MMITW4YHDAD4YYBM","created_at":"2026-07-05T11:29:07.825075+00:00"},{"alias_kind":"pith_short_8","alias_value":"MMITW4YH","created_at":"2026-07-05T11:29:07.825075+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2506.07298","citing_title":"Pre-trained Large Language Models Learn Hidden Markov Models In-context","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2506.09998","citing_title":"Flipping Against All Odds: Reducing LLM Coin Flip Bias via Verbalized Rejection Sampling","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06201","citing_title":"Beyond Facts: Benchmarking Distributional Reading Comprehension in Large Language Models","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10455","citing_title":"EviCare: Enhancing Diagnosis Prediction with Deep Model-Guided Evidence for In-Context Reasoning","ref_index":11,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6","json":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6.json","graph_json":"https://pith.science/api/pith-number/MMITW4YHDAD4YYBMXAM2EJMGY6/graph.json","events_json":"https://pith.science/api/pith-number/MMITW4YHDAD4YYBMXAM2EJMGY6/events.json","paper":"https://pith.science/paper/MMITW4YH"},"agent_actions":{"view_html":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6","download_json":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6.json","view_paper":"https://pith.science/paper/MMITW4YH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.04722&json=true","fetch_graph":"https://pith.science/api/pith-number/MMITW4YHDAD4YYBMXAM2EJMGY6/graph.json","fetch_events":"https://pith.science/api/pith-number/MMITW4YHDAD4YYBMXAM2EJMGY6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6/action/storage_attestation","attest_author":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6/action/author_attestation","sign_citation":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6/action/citation_signature","submit_replication":"https://pith.science/pith/MMITW4YHDAD4YYBMXAM2EJMGY6/action/replication_record"}},"created_at":"2026-07-05T11:29:07.825075+00:00","updated_at":"2026-07-05T11:29:07.825075+00:00"}