{"id":"https://openalex.org/W7214740700","doi":"https://doi.org/10.48550/arxiv.2609.31796","title":"Same Probe, Different Numbers: Are Activation Probes Robust to Inference-Time Numerical Non-Determinism?","display_name":"Same Probe, Different Numbers: Are Activation Probes Robust to Inference-Time Numerical Non-Determinism?","publication_year":2026,"publication_date":"2026-09-25","ids":{"openalex":"https://openalex.org/W7214740700","doi":"https://doi.org/10.48550/arxiv.2609.31796"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.31796","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.31796","pdf_url":"https://arxiv.org/pdf/2609.31796","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.31796","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5039279679","display_name":"Alizishaan Khatri","orcid":"https://orcid.org/0000-0001-5531-3635","observed_orcids":["https://orcid.org/0000-0001-5531-3635"]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Khatri, Alizishaan","raw_affiliation_strings":["Wrynx Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Wrynx Inc","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5039279679"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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":0.9365000128746033,"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":0.9365000128746033,"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":0.017999999225139618,"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":0.006500000134110451,"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/batch-size","display_name":"batch size","score":0.6299999952316284},{"id":"https://openalex.org/keywords/floating-point-precision","display_name":"floating-point precision","score":0.8889999985694885},{"id":"https://openalex.org/keywords/bfloat16","display_name":"bfloat16","score":0.8379999995231628},{"id":"https://openalex.org/keywords/cohen-s-kappa","display_name":"Cohen's kappa","score":0.5989999771118164},{"id":"https://openalex.org/keywords/next-token-prediction","display_name":"next-token prediction","score":0.45399999618530273}],"concepts":[{"id":"https://openalex.org/C135598885","wikidata":"https://www.wikidata.org/wiki/Q1366302","display_name":"Row","level":2,"score":0.8205999732017517},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6966000199317932},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.613099992275238},{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.5202999711036682},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.4632999897003174},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.44780001044273376},{"id":"https://openalex.org/C4679612","wikidata":"https://www.wikidata.org/wiki/Q866298","display_name":"Aggregate (composite)","level":2,"score":0.43799999356269836},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.4268999993801117},{"id":"https://openalex.org/C104140500","wikidata":"https://www.wikidata.org/wiki/Q2088159","display_name":"Row and column spaces","level":3,"score":0.3553999960422516},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.34850001335144043},{"id":"https://openalex.org/C2776591724","wikidata":"https://www.wikidata.org/wiki/Q5459651","display_name":"Flip","level":3,"score":0.3440000116825104},{"id":"https://openalex.org/C127162648","wikidata":"https://www.wikidata.org/wiki/Q16858953","display_name":"Channel (broadcasting)","level":2,"score":0.3418000042438507},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.3237000107765198},{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.3052999973297119},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.2980000078678131},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.28600001335144043},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.2851000130176544},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.27880001068115234},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.26499998569488525},{"id":"https://openalex.org/C2776502983","wikidata":"https://www.wikidata.org/wiki/Q690182","display_name":"Contrast (vision)","level":2,"score":0.26080000400543213},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.25949999690055847},{"id":"https://openalex.org/C2778724333","wikidata":"https://www.wikidata.org/wiki/Q14401","display_name":"Kappa","level":2,"score":0.2515999972820282}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.31796","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.31796","pdf_url":"https://arxiv.org/pdf/2609.31796","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.31796","doi":"https://doi.org/10.48550/arxiv.2609.31796","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.31796","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,"listed_in":[],"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"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.31796","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.31796","pdf_url":"https://arxiv.org/pdf/2609.31796","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214740700.pdf","grobid_xml":"https://content.openalex.org/works/W7214740700.grobid-xml"},"referenced_works_count":30,"referenced_works":["https://openalex.org/W2036166268","https://openalex.org/W2163455955","https://openalex.org/W2529194139","https://openalex.org/W2942810103","https://openalex.org/W2947629474","https://openalex.org/W2953307569","https://openalex.org/W2970971581","https://openalex.org/W2972296365","https://openalex.org/W2980282514","https://openalex.org/W2981852735","https://openalex.org/W3122184684","https://openalex.org/W3126445456","https://openalex.org/W3166109569","https://openalex.org/W3201174429","https://openalex.org/W4385574241","https://openalex.org/W4387561538","https://openalex.org/W4388514540","https://openalex.org/W4401306886","https://openalex.org/W4402385990","https://openalex.org/W4403928636","https://openalex.org/W4407231841","https://openalex.org/W4411120389","https://openalex.org/W4414939355","https://openalex.org/W4417356736","https://openalex.org/W6947991961","https://openalex.org/W7124717434","https://openalex.org/W7133198810","https://openalex.org/W7133267101","https://openalex.org/W7167676507","https://openalex.org/W7202165717"],"related_works":["https://openalex.org/W7202165717","https://openalex.org/W7164942170","https://openalex.org/W4417356736","https://openalex.org/W7213414758","https://openalex.org/W4417168559","https://openalex.org/W7166801871","https://openalex.org/W7167580951","https://openalex.org/W4407764230","https://openalex.org/W7152969042","https://openalex.org/W7214216316"],"abstract_inverted_index":{"Activation":[0],"probes":[1,82],"are":[2,36,49,98,216],"increasingly":[3],"used":[4,20],"to":[5,121,163],"monitor":[6],"LLMs":[7],"in":[8,172,183],"deployment.":[9],"A":[10,193],"probe":[11,54],"is":[12,103,187],"typically":[13],"trained":[14,56],"under":[15,21,144],"one":[16,84],"inference":[17],"configuration,":[18,85],"then":[19],"whatever":[22],"batch":[23,69,149],"size":[24,150],"and":[25,39,66,73,75,78,91,138,206,223,288],"numerical":[26],"precision":[27],"the":[28,44,51,53,104,124,148,159,185,188,191,198,238,256,262,290],"serving":[29,291],"stack":[30],"uses.":[31],"Because":[32],"common":[33],"GPU":[34],"kernels":[35],"not":[37,50,190],"batch-invariant":[38],"floating-point":[40],"formats":[41],"round":[42],"differently,":[43],"activations":[45,232],"seen":[46],"at":[47,227],"deployment":[48],"ones":[52],"was":[55],"on.":[57],"We":[58],"measure":[59],"what":[60],"that":[61],"costs":[62],"for":[63,107,202],"Llama-3.1-8B,":[64],"Qwen3-8B":[65],"Gemma-3-4B":[67],"across":[68,135],"sizes":[70],"4,":[71],"8":[72],"16":[74],"float32,":[76],"bfloat16":[77,194,242],"float16,":[79],"training":[80],"768":[81],"on":[83,88,209],"evaluating":[86],"each":[87],"every":[89],"other,":[90],"comparing":[92],"verdicts":[93,114,142,155],"example":[94],"by":[95,116,129,212,226,247],"example.":[96],"Probes":[97,259],"stable,":[99],"but":[100,165],"aggregate":[101,280],"accuracy":[102,126],"wrong":[105],"instrument":[106],"showing":[108],"it:":[109],"it":[110],"understates":[111],"how":[112],"many":[113],"change":[115,196,244],"a":[117,241,248],"factor":[118],"of":[119,141,153,175,204,252,271],"two":[120],"nine.":[122],"At":[123],"prompt,":[125],"never":[127],"moves":[128,225],"more":[130],"than":[131,279],"0.47":[132],"percentage":[133],"points":[134],"1,392":[136],"transfers":[137],"only":[139,147,173],"0.076%":[140],"change;":[143],"float32":[145],"with":[146],"varied,":[151],"none":[152],"201,960":[154],"change.":[156],"During":[157],"decoding":[158],"flip":[160,171,182],"rate":[161],"rises":[162],"2.8%,":[164],"rows":[166,178,205],"whose":[167,179],"realised":[168],"tokens":[169,180,211],"matched":[170],"0.12-0.15%":[174],"cases,":[176],"while":[177],"diverged":[181],"12.9%:":[184],"cause":[186],"text,":[189],"arithmetic.":[192],"batch-size":[195,243],"flips":[197],"first":[199],"generated":[200],"token":[201,213],"2.1%":[203],"leaves":[207],"25%":[208],"different":[210],"20.":[214],"Flips":[215],"symmetric,":[217],"Cohen's":[218],"kappa":[219],"stays":[220],"above":[221],"0.94,":[222],"AUROC":[224],"most":[228],"0.05":[229],"points.":[230],"Underneath,":[231],"move":[233],"about":[234],"as":[235,237],"much":[236],"format's":[239],"rounding:":[240],"perturbs":[245],"them":[246],"median":[249],"relative":[250],"L2":[251],"1e-2,":[253],"roughly":[254],"8x":[255],"float16":[257],"figure.":[258],"absorb":[260],"this;":[261],"model's":[263],"own":[264],"next-token":[265],"argmax":[266],"does":[267],"not.":[268],"Robustness":[269],"evaluations":[270],"activation":[272],"monitors":[273],"should":[274],"report":[275],"per-example":[276],"agreement":[277],"rather":[278],"accuracy,":[281],"separate":[282],"representational":[283],"noise":[284],"from":[285],"input":[286],"change,":[287],"state":[289],"configuration.":[292]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-30T00:00:00"}