{"id":"https://openalex.org/W7162563104","doi":"https://doi.org/10.48550/arxiv.2605.26343","title":"MechRL: Reinforcement Learning Agents Perform Circuit Discovery for Mechanistic Interpretability","display_name":"MechRL: Reinforcement Learning Agents Perform Circuit Discovery for Mechanistic Interpretability","publication_year":2026,"publication_date":"2026-05-25","ids":{"openalex":"https://openalex.org/W7162563104","doi":"https://doi.org/10.48550/arxiv.2605.26343"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.26343","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26343","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.26343","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5126692194","display_name":"Barsat Khadka","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Khadka, Barsat","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":1,"corresponding_author_ids":["https://openalex.org/A5126692194"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.6862000226974487,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.6862000226974487,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.06419999897480011,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.03350000083446503,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/interpretability","display_name":"Interpretability","score":0.8658999800682068},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.6658999919891357},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.4526999890804291},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.38609999418258667},{"id":"https://openalex.org/keywords/oracle","display_name":"Oracle","score":0.36059999465942383},{"id":"https://openalex.org/keywords/bridging","display_name":"Bridging (networking)","score":0.34040001034736633},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.3255999982357025},{"id":"https://openalex.org/keywords/identification","display_name":"Identification (biology)","score":0.30219998955726624}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.8658999800682068},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7275000214576721},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.6658999919891357},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6319000124931335},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.49079999327659607},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.4526999890804291},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.38609999418258667},{"id":"https://openalex.org/C55166926","wikidata":"https://www.wikidata.org/wiki/Q2892946","display_name":"Oracle","level":2,"score":0.36059999465942383},{"id":"https://openalex.org/C174348530","wikidata":"https://www.wikidata.org/wiki/Q188635","display_name":"Bridging (networking)","level":2,"score":0.34040001034736633},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.3255999982357025},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.30219998955726624},{"id":"https://openalex.org/C70437156","wikidata":"https://www.wikidata.org/wiki/Q7228652","display_name":"Pooling","level":2,"score":0.290800005197525},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.29019999504089355},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.27880001068115234},{"id":"https://openalex.org/C94124525","wikidata":"https://www.wikidata.org/wiki/Q912550","display_name":"Categorization","level":2,"score":0.27810001373291016},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.27489998936653137},{"id":"https://openalex.org/C199190896","wikidata":"https://www.wikidata.org/wiki/Q3509276","display_name":"Learning classifier system","level":3,"score":0.2581000030040741},{"id":"https://openalex.org/C2780598303","wikidata":"https://www.wikidata.org/wiki/Q65921492","display_name":"Flexibility (engineering)","level":2,"score":0.25679999589920044},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.25609999895095825},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.2529999911785126}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.26343","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26343","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"article"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.26343","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26343","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"article"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/16","display_name":"Peace, Justice and strong institutions","score":0.4091843068599701}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Mechanistic":[0],"interpretability":[1],"has":[2],"identified":[3],"small":[4,49],"sets":[5],"of":[6,47,120,156,186],"attention":[7,45],"heads":[8,46,114,119],"that":[9,64,170],"implement":[10],"specific":[11],"behaviours":[12],"in":[13,91],"transformer":[14],"language":[15],"models,":[16],"but":[17],"recovering":[18],"these":[19],"circuits":[20],"typically":[21],"requires":[22],"a":[23,36,51,58,61,92,106,177],"bespoke":[24],"analytical":[25],"pipeline":[26],"for":[27,181],"each":[28,55],"new":[29],"task.":[30,79],"We":[31],"recast":[32],"circuit":[33],"discovery":[34],"as":[35,50,130,140],"reinforcement-learning":[37],"problem.":[38],"An":[39],"agent":[40],"operates":[41],"over":[42,173],"the":[43,66,77,97,117,125,136,146,149,157,183],"144":[44],"GPT-2":[48],"discrete":[52],"action":[53,56],"space;":[54],"triggers":[57],"zero-ablation":[59],"and":[60,89,104],"contrastive":[62],"reward":[63],"subtracts":[65],"ablation's":[67],"damage":[68,75],"to":[69,76,190],"general":[70],"next-token":[71],"prediction":[72],"from":[73],"its":[74],"target":[78],"A":[80],"single":[81],"PPO":[82],"policy,":[83],"trained":[84],"on":[85,100,105,123],"two":[86],"tasks":[87,103],"(induction":[88],"IOI)":[90],"vectorised":[93],"multi-task":[94],"environment,":[95],"attains":[96],"per-episode":[98],"oracle":[99,158],"both":[101],"training":[102],"held-out":[107,150],"third":[108],"task":[109,162],"(docstring":[110],"completion).":[111],"Its":[112],"preferred":[113],"coincide":[115],"with":[116,160],"canonical":[118],"established":[121],"literature":[122],"precisely":[124],"axes":[126],"those":[127],"papers":[128],"identify":[129,139],"causally":[131],"non-redundant":[132],"under":[133],"single-head":[134,184],"ablation;":[135],"categories":[137],"they":[138],"redundant":[141],"are":[142],"correctly":[143],"de-prioritised":[144],"by":[145],"agent.":[147],"On":[148],"task,":[151],"best-of-five":[152],"planning":[153],"recovers":[154],"96\\%":[155],"ceiling":[159],"no":[161],"signal":[163],"supplied":[164],"at":[165],"evaluation.":[166],"These":[167],"results":[168],"indicate":[169],"reinforcement":[171],"learning":[172],"causal":[174],"interventions":[175],"is":[176],"viable,":[178],"transferable":[179],"substrate":[180],"identifying":[182],"bottlenecks":[185],"mechanistic":[187],"circuits,":[188],"complementary":[189],"existing":[191],"path-patching":[192],"approaches.":[193]},"counts_by_year":[],"updated_date":"2026-05-28T06:19:16.900634","created_date":"2026-05-28T00:00:00"}