Pakhomov, Egor, Koul, Anurag, Liu, Hongjun, Pang, Bo, Joty, Shafiq, Nijkamp, Erik · Nov 10, 2026
Cleaning an Agent's History Hurts When the World Still Holds the Error
Markdown · JSON · Original source
The paper’s authors, Egor Pakhomov, Anurag Koul, Hongjun Liu, Bo Pang, Shafiq Joty and Erik Nijkamp, argue that correcting an agent’s history can help or hurt depending on whether it describes has also changed. They test this by freezing an agent mid-task and letting a receiver model continue after edits to the history or world. On AgentCE-Bench, adding the correct value as a truthful note beside the preserved error improves recovery, but presenting it as the agent’s completed action lowers recovery below the unedited history. On AppWorld, a receipt reporting success lowers whole-task success when the database effect is absent, and raises it when the effect is present. The authors’ central point is that a history is not just a record to clean: its claims can guide the next agent differently depending on whether the world matches them. The results concern these tested tasks and receivers; the abstract does not establish how broadly the effects apply.
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 10, 2026 (2026-11-10).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T15:35:34.638Z. First observed: 2026-11-13T07:40:01.739Z. Retrieved: 2026-11-13T15:34:47.676Z.
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 abstract reports results on AgentCE-Bench and AppWorld; it does not establish whether the effects extend to other tasks or settings. The AppWorld comparison is described as pooled over two receivers, and the abstract does not establish how results vary across a broader range of receiver models.
Open question
How would the reported effects vary across tasks and receiver models beyond those tested?
Source-stated
The authors argue that the effect of editing an agent’s history depends on whether the described world state matches the actual world state.
A history edit's effect thus depends on the world it describes
Source-stated
On AgentCE-Bench, a truthful note giving the correct value beside the preserved error raises recovery compared with a matched neutral note, while recording the correct value as the agent’s completed action lowers recovery below the unedited history.
The note raises recovery, the share of remaining targets the receiver gets right, from 51% under a matched neutral note to 66%.
Source-stated
On AppWorld, a receipt reporting success reduces whole-task success when the database effect is absent, but increases it when the effect is present.
A success receipt lowers whole-task success from 45% to 28% while the effect is absent and raises it from 38% to 49% once present.
Source-stated
The authors’ experiments indicate that the correct value’s presence in the history alone does not predict whether the receiver performs better.
so the value's presence alone does not predict the effect.
Source and contributor context
Pakhomov, Egor, Koul, Anurag, Liu, Hongjun, Pang, Bo, Joty, Shafiq, Nijkamp, Erik · 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.
The World
The World is a public radio news program publishing international reporting on politics, society, labor and technology.
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: edc95a1e-3450-4862-878c-7be7bf7e39c2.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T15:35:34.638Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4f6a7221aac848519e81909ef7296fd1e5be1057ef5582ca5e6cd25e294dee22. Source text hash: ec91b31423e7b2ef9dfe713ce4d8b2cba036ea75a122b0867e21400e5461e8ee.
Revision and source-check history
Version 1 · 2026-11-13T15:35:34.638Z
Cleaning an Agent's History Hurts When the World Still Holds the Error
The paper’s authors, Egor Pakhomov, Anurag Koul, Hongjun Liu, Bo Pang, Shafiq Joty and Erik Nijkamp, argue that correcting an agent’s history can help or hurt depending on whether it describes has also changed. They test this by freezing an agent mid-task and letting a receiver model continue after edits to the history or world. On AgentCE-Bench, adding the correct value as a truthful note beside the preserved error improves recovery, but presenting it as the agent’s completed action lowers recovery below the unedited history. On AppWorld, a receipt reporting success lowers whole-task success when the database effect is absent, and raises it when the effect is present. The authors’ central point is that a history is not just a record to clean: its claims can guide the next agent differently depending on whether the world matches them. The results concern these tested tasks and receivers; the abstract does not establish how broadly the effects apply.
Original source
cs.AI · Article
Source details
Original source date: Nov 10, 2026 (2026-11-10).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T15:35:34.638Z. First observed: 2026-11-13T07:40:01.739Z. Retrieved: 2026-11-13T15:34:47.676Z.
Limits
The abstract reports results on AgentCE-Bench and AppWorld; it does not establish whether the effects extend to other tasks or settings. The AppWorld comparison is described as pooled over two receivers, and the abstract does not establish how results vary across a broader range of receiver models.
Open question
How would the reported effects vary across tasks and receiver models beyond those tested?
Source-stated
The authors argue that the effect of editing an agent’s history depends on whether the described world state matches the actual world state.
A history edit's effect thus depends on the world it describes
Source-stated
On AgentCE-Bench, a truthful note giving the correct value beside the preserved error raises recovery compared with a matched neutral note, while recording the correct value as the agent’s completed action lowers recovery below the unedited history.
The note raises recovery, the share of remaining targets the receiver gets right, from 51% under a matched neutral note to 66%.
Source-stated
On AppWorld, a receipt reporting success reduces whole-task success when the database effect is absent, but increases it when the effect is present.
A success receipt lowers whole-task success from 45% to 28% while the effect is absent and raises it from 38% to 49% once present.
Source-stated
The authors’ experiments indicate that the correct value’s presence in the history alone does not predict whether the receiver performs better.
so the value's presence alone does not predict the effect.
Source and contributor context
Pakhomov, Egor, Koul, Anurag, Liu, Hongjun, Pang, Bo, Joty, Shafiq, Nijkamp, Erik · 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.
The World
The World is a public radio news program publishing international reporting on politics, society, labor and technology.
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: edc95a1e-3450-4862-878c-7be7bf7e39c2.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T15:35:34.638Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4f6a7221aac848519e81909ef7296fd1e5be1057ef5582ca5e6cd25e294dee22. Source text hash: ec91b31423e7b2ef9dfe713ce4d8b2cba036ea75a122b0867e21400e5461e8ee.
uncertain · 2026-11-13T15:35:34.638Z · anthropic/claude-opus-5.5
Publication: edc95a1e-3450-4862-878c-7be7bf7e39c2
attribution — pass: The six listed authors match the arXiv record, and the findings are framed as the authors' experiments and argument rather than as established fact.
Authors:Egor Pakhomov, Anurag Koul, Hongjun Liu, Bo Pang, Shafiq Joty, Erik Nijkamp
summary — uncertain: The reading is accurate but hard for non-specialists to follow. Central terms go unexplained: agent, receiver model, recovery (which the source defines as the share of remaining targets the receiver gets right), AgentCE-Bench and AppWorld. Smallest repair: add plain-language glosses or tooltips. The summary also drops the paper's strongest conclusion, that a cleaned history can leave an agent worse off than making no edit; restoring that clause would preserve the authors' force. The lead mentions only the world condition, although the abstract also stresses how the history states the correction. Adding 'and how the history states it' would fix this.
a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all
date — pass: The original publication date of 2026-11-10 matches the v1 submission date. The feed date of Nov 13 is later and is correctly not used.
[Submitted on 10 Nov 2026]
support — uncertain: The claims and numbers match the abstract. However, the second limitation says the AppWorld comparison is pooled over two receivers and that results across a broader range of receivers are not established. That omits the abstract's report that the difference was positive in six of eight receivers at 97.5% and seven at 95%. Smallest repair: mention the eight-receiver result, then note that breadth beyond those receivers remains untested. The question about other tasks and receivers is otherwise a fair open issue.
a difference positive in six of eight receivers at 97.5% and seven at 95%
media — pass: The selected media is the verified arXiv abstract page for this exact paper, giving direct access to 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.
The World is a public radio news program publishing international reporting on politics, society, labor and technology.
Explore The WorldarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe World is a public radio news program publishing international reporting on politics, society, labor and technology.
Explore The WorldarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv