{"id":"https://openalex.org/W7155960108","doi":"https://doi.org/10.48550/arxiv.2604.22082","title":"Removing Sandbagging in LLMs by Training with Weak Supervision","display_name":"Removing Sandbagging in LLMs by Training with Weak Supervision","publication_year":2026,"publication_date":"2026-04-23","ids":{"openalex":"https://openalex.org/W7155960108","doi":"https://doi.org/10.48550/arxiv.2604.22082"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.22082","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.22082","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2604.22082","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5119841385","display_name":"Emil Ryd","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ryd, Emil","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5012032714","display_name":"Henning Bartsch","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bartsch, Henning","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5025141623","display_name":"Julian Stastny","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Stastny, Julian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5056010406","display_name":"Joe Benton","orcid":"https://orcid.org/0000-0002-2103-6112"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Benton, Joe","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5134042704","display_name":"Vivek Hebbar","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hebbar, Vivek","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"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/T13197","display_name":"Spreadsheets and End-User Computing","score":0.08900000154972076,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"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/T13197","display_name":"Spreadsheets and End-User Computing","score":0.08900000154972076,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.08609999716281891,"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/T10260","display_name":"Software Engineering Research","score":0.07530000060796738,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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/exploit","display_name":"Exploit","score":0.7870000004768372},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.7152000069618225},{"id":"https://openalex.org/keywords/work","display_name":"Work (physics)","score":0.4629000127315521},{"id":"https://openalex.org/keywords/hacker","display_name":"Hacker","score":0.4277999997138977},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.37869998812675476},{"id":"https://openalex.org/keywords/coding","display_name":"Coding (social sciences)","score":0.3431999981403351}],"concepts":[{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.7870000004768372},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.7152000069618225},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4957999885082245},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.4629000127315521},{"id":"https://openalex.org/C86844869","wikidata":"https://www.wikidata.org/wiki/Q2798820","display_name":"Hacker","level":2,"score":0.4277999997138977},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.39890000224113464},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.37869998812675476},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.35499998927116394},{"id":"https://openalex.org/C179518139","wikidata":"https://www.wikidata.org/wiki/Q5140297","display_name":"Coding (social sciences)","level":2,"score":0.3431999981403351},{"id":"https://openalex.org/C56739046","wikidata":"https://www.wikidata.org/wiki/Q192060","display_name":"Knowledge management","level":1,"score":0.33000001311302185},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.30959999561309814},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.27639999985694885},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.274399995803833},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.27140000462532043},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.2712000012397766},{"id":"https://openalex.org/C67203356","wikidata":"https://www.wikidata.org/wiki/Q1321905","display_name":"Reinforcement","level":2,"score":0.26570001244544983}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.22082","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.22082","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2604.22082","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.22082","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"As":[0],"AI":[1],"systems":[2],"begin":[3],"to":[4,67,113,128,140,179],"automate":[5],"complex":[6],"tasks,":[7],"supervision":[8,87],"increasingly":[9],"relies":[10,156],"on":[11,72,104,157],"weaker":[12,149],"models":[13,92,164],"or":[14],"limited":[15],"human":[16],"oversight":[17],"that":[18,39,83,187],"cannot":[19],"fully":[20,115],"verify":[21],"output":[22],"quality.":[23],"A":[24],"model":[25,64],"more":[26],"capable":[27],"than":[28,132,150],"its":[29,46],"supervisors":[30],"could":[31],"exploit":[32],"this":[33,62,155],"gap":[34],"through":[35],"sandbagging,":[36,194],"producing":[37],"work":[38,55],"appears":[40],"acceptable":[41],"but":[42],"falls":[43],"short":[44],"of":[45,199],"true":[47],"abilities.":[48],"Can":[49],"training":[50,84,158,168,176,188,201],"elicit":[51,90,116,141],"a":[52,190],"model's":[53],"best":[54],"even":[56],"without":[57,123,137],"reliable":[58],"verification?":[59],"We":[60,81],"study":[61],"using":[63],"organisms":[65],"trained":[66],"sandbag,":[68],"testing":[69],"elicitation":[70],"techniques":[71],"problem-solving":[73],"math,":[74],"graduate-level":[75],"science,":[76],"and":[77,97,135,169],"competitive":[78],"coding":[79],"tasks.":[80],"find":[82],"with":[85],"weak":[86,105],"can":[88,165,172],"reliably":[89,121],"sandbagging":[91,109],"when":[93,144,163],"supervised":[94],"fine-tuning":[95],"(SFT)":[96],"reinforcement":[98],"learning":[99],"(RL)":[100],"are":[101],"combined:":[102],"SFT":[103,124,136],"demonstrations":[106],"breaks":[107],"the":[108,145,151,197],"behavior,":[110],"enabling":[111],"RL":[112,138],"then":[114],"performance.":[117],"Neither":[118],"method":[119],"succeeds":[120],"alone-RL":[122],"almost":[125],"always":[126],"leads":[127],"reward":[129],"hacking":[130],"rather":[131],"genuine":[133],"improvement,":[134],"fails":[139],"full":[142],"performance":[143],"supervisor":[146],"is":[147,189],"much":[148],"untrusted":[152],"model.":[153],"Critically,":[154],"being":[159],"indistinguishable":[160,202],"from":[161,203],"deployment;":[162],"distinguish":[166],"between":[167],"deployment,":[170],"they":[171],"perform":[173],"well":[174],"during":[175],"while":[177,195],"continuing":[178],"sandbag":[180],"afterward.":[181],"Our":[182],"results":[183],"provide":[184],"initial":[185],"evidence":[186],"viable":[191],"mitigation":[192],"against":[193],"highlighting":[196],"importance":[198],"making":[200],"deployment.":[204]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-04-28T00:00:00"}