{"about":{"non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","site":"https://codewithpapers.app","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page","syntology":{"site":"https://syntology.ai","developers":"https://syntology.ai/developers","mcp":{"server":"https://syntology.ai/mcp","transport":"streamable-http","server_card":"https://syntology.ai/.well-known/mcp/server-card.json","auth":{"type":"trial token, no account","trial_token":"https://syntology.ai/api/oauth/trial/token","method":"POST","docs":"https://syntology.ai/developers"}},"have":"https://syntology.ai/api/graph/have?x=<method, arXiv id or title> (free, answers coverage only)","paper_base":"https://syntology.ai/paper/","atlas_base":"https://app.syntology.ai/?focus="},"machine_readable":[{"url":"https://codewithpapers.app/llms.txt","what":"the machine catalog: every machine-readable file, counted"},{"url":"https://codewithpapers.app/index/manifest.json","what":"paper-to-code index by arXiv id, with Syntology's counts"},{"url":"https://codewithpapers.app/search/manifest.json","what":"site search index (titles, authors) and its files"},{"url":"https://codewithpapers.app/download","what":"bulk files: Syntology's layer, described there"},{"url":"https://codewithpapers.app/build_manifest.json","what":"the build record: inputs, counts, exclusions, probes"}]},"url":"/paper/towards-a-theoretical-understanding-of-the-2","title":"Towards a Theoretical Understanding of the 'Reversal Curse' via Training Dynamics","arxiv_id":"2405.04669","date":"2024-05-07","proceeding":null,"authors":["Hanlin Zhu","Baihe Huang","Shaolun Zhang","Michael Jordan","Jiantao Jiao","Yuandong Tian","Stuart Russell"],"abstract":"Auto-regressive large language models (LLMs) show impressive capacities to solve many complex reasoning tasks while struggling with some simple logical reasoning tasks such as inverse search: when trained on '$A \\to B$' (e.g., 'Tom is the parent of John'), LLM fails to directly conclude '$B \\gets A$' (e.g., 'John is the child of Tom') during inference even if the two sentences are semantically identical, which is known as the 'reversal curse'. In this paper, we theoretically analyze the reversal curse via the training dynamics of (stochastic) gradient descent for two auto-regressive models: (1) a bilinear model that can be viewed as a simplification of a one-layer transformer; (2) one-layer transformers under certain assumptions. Our analysis reveals that for both models, the reversal curse is a consequence of the (effective) model weights 'asymmetry', i.e., the increase of weights from a token $A$ to token $B$ during training does not necessarily cause the increase of the weights from $B$ to $A$, which is caused by the training dynamics under certain choice of loss function and the optimization space of model parameters. Moreover, our analysis can be naturally applied to other logical reasoning tasks such as chain-of-thought (COT), which provides a new perspective different from previous work that focuses on expressivity. Finally, we conduct experiments to validate our theory on multi-layer transformers under different settings. Our code is available at https://github.com/marlo-z/reversal_curse_analysis/.","url_abs":"https://arxiv.org/abs/2405.04669v2","url_pdf":"https://arxiv.org/pdf/2405.04669v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"towards-a-theoretical-understanding-of-the-2","repo_url":"https://github.com/marlo-z/reversal_curse_analysis","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"logical-reasoning","task_name":"Logical Reasoning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/2405.04669","atlas_url":"https://app.syntology.ai/?focus=2405.04669","mcp":{"get_harvested_code_for_paper":{"arxiv_id":"2405.04669"},"calls":[{"tool":"get_harvested_code_for_paper","arguments":{"arxiv_id":"2405.04669"},"returns":"the samples harvested from this paper, each with its run record; it does not verify the paper"},{"tool":"get_code","arguments_in":"samples[].code_sha256","returns":"one stored sample by its code_sha256; not shown on this site"}]},"developers":"https://syntology.ai/developers","held":true,"reason":"in_graph_layer","checked_against":null,"agent_calls":[{"tool":"get_harvested_code_for_paper","arguments":{"arxiv_id":"2405.04669"},"call":"get_harvested_code_for_paper(arxiv_id=\"2405.04669\")"},{"tool":"get_code","arguments_in":{"code_sha256":"syntology.samples[].code_sha256"},"call":"get_code(code_sha256=\"…\")"},{"tool":"get_citation_path","arguments":{"paper_1":"2405.04669"},"arguments_in":{"paper_2":"another paper's arXiv id or title"},"call":"get_citation_path(paper_1=\"2405.04669\", paper_2=\"…\")"},{"tool":"get_concepts_for_paper","arguments":{"arxiv_id":"2405.04669"},"call":"get_concepts_for_paper(arxiv_id=\"2405.04669\")"}],"by_repo":[{"repo":"marlo-z/reversal_curse_analysis","kind":"official","samples":7,"ran":7,"instrument":1,"unverified":0}],"by_repo_is":"one row per repository in the archive's code links, Syntology's graph links or the samples; samples 0 means no sample is linked to this paper, a statement about Syntology's coverage, not about the repository: with harvested_for_other_papers present, Syntology harvested the repository for that many other papers, and without it, Syntology harvested nothing from it; placed_by_identical_code counts samples whose link records no repository, placed here because identical code was harvested from this repository for another paper; ran counts executions on a synthesized input and instrument counts failures of Syntology's instrument, not of the code","read_at":"2026-10-05T00:36:11+00:00","read_at_is":"when the build read Syntology's graph, not when any sample ran","ran_record":{"url":"https://syntology.ai/api/ran/2405.04669.json","schema":"syntology.ran_record/1","counts_crosswalk":{"n_samples":"syntology.counts.lifted","n_ran":"syntology.counts.executed","n_ran_checked":"syntology.counts.checked","n_instrument":"syntology.counts.instrument_failures","n_unverified":"syntology.counts.no_recorded_run"},"note":"live record, measured at request time; may be newer than this page's graph read","badge":"https://syntology.ai/api/ran/2405.04669.svg"},"claim":"Per-sample execution status on synthesized fixtures; not a correctness claim about the paper. Samples come from repositories linked to the paper, official or community; repo_kind says which.","repos":[{"provenance":"deterministic:regex_extraction","url":"https://github.com/marlo-z/reversal_curse_analysis","reach":{"status":"ok","spdx":"MIT"}}],"summary":{"ran_draft_wrong":1,"ran":6},"n_samples":7,"n_ran":7,"n_constructed":0,"n_ran_checked":6,"n_instrument":1,"n_unverified":0,"n_instrument_is":"failures of Syntology's instrument, not of the code","every_run_is_an_instrument_failure":false,"key_notes":{"n_ran_checked":"legacy name, kept unchanged so existing readers do not break: it counts the samples that ran with no instrument failure (honoured, violated, and ran with no contract checked); it does not mean a contract was checked, and the pages print it as 'K with no instrument failure', not 'K checked'","n_constructed":"a sub-count of the samples that ran, never subtracted from them and never a failure: an executed sample whose run returned an instance of its own class (fixture_out_type equals the entry name): the run built an object and did not compute a result (Syntology's RAN record, counts.constructed)"},"by_repo_kind":{"official":{"samples":7,"ran":7,"n_ran_checked":6,"n_instrument":1,"repositories":1}},"repo_kind_vocabulary":{"official":"The archive marks this repository official for the paper","named_in_paper":"The archive records that the paper mentions this repository; it is not marked official","listed":"In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper","found_in_text":"Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted","community":"Not in the archive's code links for this paper; a community repository Syntology harvested"},"n_pointer_only_for_licence":0,"inline_ok_is":"inline_ok is true only when both licence gates allow Syntology to serve this code's text: this paper's link (inline_ok_link, from the licence of this paper's copy) and the stored copy of the same code that Syntology serves, which may have been first harvested from another repository. withheld_clause is Syntology's own wording of which gate keeps the text a pointer and why, and is null when inline_ok is true. This site shows no code text either way; n_pointer_only_for_licence counts samples with inline_ok false (both gates), n_pointer_only_link_allows those of them whose inline_ok_link is true","n_pointer_only_link_allows":0,"fixture_fields_is":{"signature_and_fixture_out_type":"the arguments Syntology's fixture called the sample with, each as name:type/rank/element kind (? where none was recorded), and the type of what the fixture's run returned, exactly as the harness recorded them; the arguments and the returned type were recorded by two separate runs of the same fixture. No output value is shown, and a returned type says nothing about whether the output was right","level_recorded_at":"when Syntology's graph wrote the sample's verification level, not when the code executed; absent where the level was written before the date was kept, and never filled in with the build date","present_on":"samples that ran only; a key is absent, never empty, where the graph records nothing"},"n_ran_with_level_date":7,"n_ran_without_level_date":0,"n_ran_signature_not_shown":0,"levels_written_between":{"first":"2026-09-24","last":"2026-09-24","is":"the first and last UTC dates on which Syntology's graph wrote a verification level for a sample on this page that ran, not when the code executed"},"samples":[{"code_sha256":"2a84924c2adf7a2c","code_sha256_prefix":"2a84924c2adf7a2c","entry":"create_forward_with_rotary_pe","repo":"marlo-z/reversal_curse_analysis","repo_kind":"official","path":"src/model/Add_Rotary_PE.py","file_url":"https://github.com/marlo-z/reversal_curse_analysis/blob/HEAD/src/model/Add_Rotary_PE.py","link_basis":"first_harvest_node","language":"python","status":"ran_draft_wrong","verification_level":1,"contract_check":"OUTPUT_MISDECLARED","metamorphic_tier":null,"behaviour_fingerprint":false,"signature":"rotary_embed:RotaryEmbedding/0/?","fixture_out_type":"function","level_recorded_at":"2026-09-24T15:16:18.137819+00:00","code_page":null,"census":null,"licence":"MIT","inline_ok":true,"inline_ok_link":true,"withheld_clause":null,"mcp_get_code":{"code_sha256":"2a84924c2adf7a2c"}},{"code_sha256":"2beefe019a13b6f1","code_sha256_prefix":"2beefe019a13b6f1","entry":"generate_chain_dataset","repo":"marlo-z/reversal_curse_analysis","repo_kind":"official","path":"src/datasets/ChainDataset.py","file_url":"https://github.com/marlo-z/reversal_curse_analysis/blob/HEAD/src/datasets/ChainDataset.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"signature":"vocab_size:int/0/int, word_size:int/0/int, device:str/0/str, split_ratio:float/0/float, SEED:int/0/int","fixture_out_type":"dict","level_recorded_at":"2026-09-24T15:16:18.137819+00:00","code_page":null,"census":null,"licence":"MIT","inline_ok":true,"inline_ok_link":true,"withheld_clause":null,"mcp_get_code":{"code_sha256":"2beefe019a13b6f1"}},{"code_sha256":"efa17b00486faf3f","code_sha256_prefix":"efa17b00486faf3f","entry":"generate_chain_dataset","repo":"marlo-z/reversal_curse_analysis","repo_kind":"official","path":"src/datasets/ChainDataset_logits.py","file_url":"https://github.com/marlo-z/reversal_curse_analysis/blob/HEAD/src/datasets/ChainDataset_logits.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"signature":"vocab_size:int/0/int, word_size:int/0/int, device:device/0/?, split_ratio:float/0/float, SEED:int/0/int","fixture_out_type":"dict","level_recorded_at":"2026-09-24T15:16:18.137819+00:00","code_page":null,"census":null,"licence":"MIT","inline_ok":true,"inline_ok_link":true,"withheld_clause":null,"mcp_get_code":{"code_sha256":"efa17b00486faf3f"}},{"code_sha256":"e13e1fce2d0b0ecc","code_sha256_prefix":"e13e1fce2d0b0ecc","entry":"generate_chain_dataset_v2","repo":"marlo-z/reversal_curse_analysis","repo_kind":"official","path":"src/datasets/ChainDataset_related_tokens.py","file_url":"https://github.com/marlo-z/reversal_curse_analysis/blob/HEAD/src/datasets/ChainDataset_related_tokens.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"signature":"vocab_size:int/0/int, alphabet_size:int/0/int, num_train:int/0/int, num_val:int/0/int, device:str/0/str, SEED:int/0/int","fixture_out_type":"dict","level_recorded_at":"2026-09-24T15:16:18.137819+00:00","code_page":null,"census":null,"licence":"MIT","inline_ok":true,"inline_ok_link":true,"withheld_clause":null,"mcp_get_code":{"code_sha256":"e13e1fce2d0b0ecc"}},{"code_sha256":"2a41a86260e2a69f","code_sha256_prefix":"2a41a86260e2a69f","entry":"generate_reverse_dataset","repo":"marlo-z/reversal_curse_analysis","repo_kind":"official","path":"src/datasets/ReverseDataset.py","file_url":"https://github.com/marlo-z/reversal_curse_analysis/blob/HEAD/src/datasets/ReverseDataset.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"signature":"vocab_size:int/0/int, word_size:int/0/int, device:str/0/str, split_ratio:float/0/float, SEED:int/0/int","fixture_out_type":"dict","level_recorded_at":"2026-09-24T15:16:18.137819+00:00","code_page":null,"census":null,"licence":"MIT","inline_ok":true,"inline_ok_link":true,"withheld_clause":null,"mcp_get_code":{"code_sha256":"2a41a86260e2a69f"}},{"code_sha256":"9bfc6b75de3254a1","code_sha256_prefix":"9bfc6b75de3254a1","entry":"generate_reverse_dataset","repo":"marlo-z/reversal_curse_analysis","repo_kind":"official","path":"src/datasets/ReverseDataset_logits.py","file_url":"https://github.com/marlo-z/reversal_curse_analysis/blob/HEAD/src/datasets/ReverseDataset_logits.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"signature":"vocab_size:int/0/int, word_size:int/0/int, device:device/0/?, split_ratio:float/0/float, SEED:int/0/int","fixture_out_type":"dict","level_recorded_at":"2026-09-24T15:16:18.137819+00:00","code_page":null,"census":null,"licence":"MIT","inline_ok":true,"inline_ok_link":true,"withheld_clause":null,"mcp_get_code":{"code_sha256":"9bfc6b75de3254a1"}},{"code_sha256":"22223853110ce7ce","code_sha256_prefix":"22223853110ce7ce","entry":"generate_reverse_dataset_ICL","repo":"marlo-z/reversal_curse_analysis","repo_kind":"official","path":"src/datasets/ReverseDataset_ICL.py","file_url":"https://github.com/marlo-z/reversal_curse_analysis/blob/HEAD/src/datasets/ReverseDataset_ICL.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"signature":"vocab_size:int/0/int, num_x_tokens:int/0/int, num_y_tokens:int/0/int, device:str/0/str, split_ratio:float/0/float, SEED:int/0/int","fixture_out_type":"dict","level_recorded_at":"2026-09-24T15:16:18.137819+00:00","code_page":null,"census":null,"licence":"MIT","inline_ok":true,"inline_ok_link":true,"withheld_clause":null,"mcp_get_code":{"code_sha256":"22223853110ce7ce"}}]},"code_links_note":null,"arxiv_metadata":null,"syntology_extracted_results":null,"record_sha256":"6deabc870afef75ab58575c1353aab3d0866e60525f047cfca4f5c8680f0905f","record_changed_at":"2026-10-05","record_changed_at_basis":"changed"}