Yashwant, Sai, Gondimalla, Manoj Kumar, Sharma, Arjun, Shubham · Nov 5, 2026
Authors compare PII redaction designs for LLM pipelines
Markdown · JSON · Original source
Sai Yashwant, Manoj Kumar Gondimalla, Arjun Sharma and Shubham compare four ways to redact personally identifiable information (PII) before text is sent to a (LLM). Their ProfileAssist test used 100 synthetic resumes, with 900 PII values and 600 business terms that needed to survive redaction. The authors report that regex ran fastest, at a 0.20 ms median, but caught only 88.9% of PII; they say it missed names not known in advance. Presidio, a hybrid using Presidio for names and regex for structured identifiers, and a fast hybrid using spaCy’s named-entity component each caught 100% of PII and retained all business terms in this test. The authors recommend the fast hybrid as a default for LLM pipelines handling personal data, with regex as a safety net rather than the only control. Their measured median for those three designs were 37.59 ms, 30.23 ms and 13.48 ms, respectively; they compare the slowest with a 9,120 ms median LLM call and report a 0.41% addition. The results make the tradeoff concrete in this workflow: the fastest regex option missed some PII, while the more complete designs added little latency relative to the measured LLM call. Whether the comparison holds for real resumes and other LLM workflows remains untested in the reported sample.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-09T03:09:51.395Z. First observed: 2026-11-06T05:03:26.048Z. Retrieved: 2026-11-09T03:08:55.430Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported evaluation uses 100 synthetic resumes in one sample workflow, ProfileAssist; it does not report results on real resumes or other workflows. The 100% PII detection and business-term retention figures apply to this test, not necessarily to other data or deployments.
Open question
How would the four designs perform on real resumes rather than the 100 synthetic resumes used in this test?
Source-stated
The authors evaluated four redaction designs in a ProfileAssist workflow using 100 synthetic resumes, including 900 PII values and 600 business terms to retain.
using 100 synthetic resumes (523 characters on average, with 900 PII values and 600 business terms that must survive)
Source-stated
Regex had a 0.20 ms median latency and caught 88.9% of PII; the authors attribute missed names to the method's inability to find names not known in advance.
Regex ran at a 0.20 ms median but caught only 88.9% of PII, because it cannot find names that are not known in advance
Source-stated
Presidio and both hybrid designs caught all PII and retained all business terms in the test, with median latencies of 37.59 ms, 30.23 ms and 13.48 ms.
Presidio, hybrid and fast hybrid each caught 100% of PII and kept 100% of business terms, at median latencies of 37.59 ms, 30.23 ms and 13.48 ms.
Source-stated
The authors recommend the fast hybrid as a default for LLM pipelines handling personal data, while keeping regex as a safety net rather than the only control.
We therefore recommend the fast hybrid as a default for LLM pipelines that handle personal data, with regex kept as a safety net rather than as the only control.
Source-stated
The authors say the slowest of the three designs that caught all PII added 0.41% relative to the measured median LLM call.
Measured against a 9,120 ms median LLM call, even the slowest option adds only 0.41% to the request.
Source and contributor context
Yashwant, Sai, Gondimalla, Manoj Kumar, Sharma, Arjun, Shubham · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Latency
The delay between sending a request and receiving the result.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 998fdf35-e604-4973-8f23-6076949c2dc3.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T03:09:51.395Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 2eb6190348723e947275ff99024792295f69854399202bc8121818e1cfba9193. Source text hash: c6f88166167c068324a49c5d4f8232ad4e9239dfc3c5a8ac067188ef65f7529a.
Revision and source-check history
Version 1 · 2026-11-09T03:09:51.395Z
Authors compare PII redaction designs for LLM pipelines
Sai Yashwant, Manoj Kumar Gondimalla, Arjun Sharma and Shubham compare four ways to redact personally identifiable information (PII) before text is sent to a (LLM). Their ProfileAssist test used 100 synthetic resumes, with 900 PII values and 600 business terms that needed to survive redaction. The authors report that regex ran fastest, at a 0.20 ms median, but caught only 88.9% of PII; they say it missed names not known in advance. Presidio, a hybrid using Presidio for names and regex for structured identifiers, and a fast hybrid using spaCy’s named-entity component each caught 100% of PII and retained all business terms in this test. The authors recommend the fast hybrid as a default for LLM pipelines handling personal data, with regex as a safety net rather than the only control. Their measured median for those three designs were 37.59 ms, 30.23 ms and 13.48 ms, respectively; they compare the slowest with a 9,120 ms median LLM call and report a 0.41% addition. The results make the tradeoff concrete in this workflow: the fastest regex option missed some PII, while the more complete designs added little latency relative to the measured LLM call. Whether the comparison holds for real resumes and other LLM workflows remains untested in the reported sample.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-09T03:09:51.395Z. First observed: 2026-11-06T05:03:26.048Z. Retrieved: 2026-11-09T03:08:55.430Z.
Limits
The reported evaluation uses 100 synthetic resumes in one sample workflow, ProfileAssist; it does not report results on real resumes or other workflows. The 100% PII detection and business-term retention figures apply to this test, not necessarily to other data or deployments.
Open question
How would the four designs perform on real resumes rather than the 100 synthetic resumes used in this test?
Source-stated
The authors evaluated four redaction designs in a ProfileAssist workflow using 100 synthetic resumes, including 900 PII values and 600 business terms to retain.
using 100 synthetic resumes (523 characters on average, with 900 PII values and 600 business terms that must survive)
Source-stated
Regex had a 0.20 ms median latency and caught 88.9% of PII; the authors attribute missed names to the method's inability to find names not known in advance.
Regex ran at a 0.20 ms median but caught only 88.9% of PII, because it cannot find names that are not known in advance
Source-stated
Presidio and both hybrid designs caught all PII and retained all business terms in the test, with median latencies of 37.59 ms, 30.23 ms and 13.48 ms.
Presidio, hybrid and fast hybrid each caught 100% of PII and kept 100% of business terms, at median latencies of 37.59 ms, 30.23 ms and 13.48 ms.
Source-stated
The authors recommend the fast hybrid as a default for LLM pipelines handling personal data, while keeping regex as a safety net rather than the only control.
We therefore recommend the fast hybrid as a default for LLM pipelines that handle personal data, with regex kept as a safety net rather than as the only control.
Source-stated
The authors say the slowest of the three designs that caught all PII added 0.41% relative to the measured median LLM call.
Measured against a 9,120 ms median LLM call, even the slowest option adds only 0.41% to the request.
Source and contributor context
Yashwant, Sai, Gondimalla, Manoj Kumar, Sharma, Arjun, Shubham · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Latency
The delay between sending a request and receiving the result.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 998fdf35-e604-4973-8f23-6076949c2dc3.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T03:09:51.395Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 2eb6190348723e947275ff99024792295f69854399202bc8121818e1cfba9193. Source text hash: c6f88166167c068324a49c5d4f8232ad4e9239dfc3c5a8ac067188ef65f7529a.
passed · 2026-11-09T03:09:51.395Z · anthropic/claude-opus-5.5
Publication: 998fdf35-e604-4973-8f23-6076949c2dc3
attribution — pass: The draft names the four listed authors and attributes the results and recommendation to them with phrasing such as 'the authors report,' 'they say' and 'the authors recommend.' Source claims are not presented as established facts.
Authors:Sai Yashwant, Manoj Kumar Gondimalla, Arjun Sharma, Shubham
summary — pass: The draft faithfully conveys the four designs, the test setup, the numbers and the recommendation, and it expands PII and LLM. Regex, Presidio, spaCy and named-entity component are not explained in plain language; term-library tooltips should cover them, and Microsoft Presidio could use one. The source's stated motive is omitted, but this does not prevent understanding: prompts copy PII into provider systems, logs and traces, and teams lack cost data. A one-clause addition would strengthen the stakes. The title and question are neutral and follow from the evidence.
Once a prompt is sent, this data is copied into provider systems, application logs and observability traces.
date — pass: The originalPublished date, 2026-11-05, matches the v1 submission date on arXiv. The later feed date is not used.
[Submitted on 5 Nov 2026]
support — pass: All figures match the abstract: 0.20 ms, 88.9%, 100%, 37.59/30.23/13.48 ms, 9,120 ms and 0.41%. The interpretive sentence about the tradeoff is limited to 'this workflow.' The limitation about synthetic data reflects the source and is not overstated. 'The slowest' is consistent because Presidio is the slowest of all four designs.
Measured against a 9,120 ms median LLM call, even the slowest option adds only 0.41% to the request.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way to access the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceThe delay between sending a request and receiving the result.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceThe delay between sending a request and receiving the result.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition reference