Skip to content
AI News HubLIVE
Original source5 min read

Cleaning an Agent's History Hurts When the World Still Holds the Error

Summary

A new arXiv paper shows through controlled experiments on AgentCE-Bench and AppWorld that editing an agent's history only helps when the real-world state changes along with it; making a history look clean while the error persists can leave a receiver agent worse off than no edit at all.

SourcearXiv AIAuthor: Egor Pakhomov, Anurag Koul, Hongjun Liu, Bo Pang, Shafiq Joty, Erik Nijkamp
Cleaning an Agent's History Hurts When the World Still Holds the Error
Report an error

The correction channel is not available yet. You can copy the article reference below for later.

Correction instructions
Read article

[Submitted on 10 Nov 2026]

Title:Cleaning an Agent's History Hurts When the World Still Holds the Error

View a PDF of the paper titled Cleaning an Agent's History Hurts When the World Still Holds the Error, by Egor Pakhomov and 5 other authors

View PDF HTML (experimental)

Abstract:Tool-using language-model agents accumulate histories of their own mistakes, and a system that compacts, repairs or hands off such a history decides what it says about those mistakes while their effects stay in the world. Prior work cleans the history at a fixed world, where showing a corrected step and saying it is done move together. We ask whether the same correct content helps or hurts depending on how the history states it and what the world holds. We freeze an agent mid-task with its exact world state, edit the history or the world, and let a receiver model continue. On AgentCE-Bench, after the agent commits a wrong value, we hold the world fixed and add the correct value by two natural edits: a truthful note beside the preserved error or the agent's own completed action. The note raises recovery, the share of remaining targets the receiver gets right, from 51% under a matched neutral note to 66%. The completed action lowers it below the unedited history, from 50% to 44%, so the value's presence alone does not predict the effect. On AppWorld, after a database write, we vary whether its receipt reports success or failure and whether its 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. The two effects differ by 27 points pooled over two receivers, a difference positive in six of eight receivers at 97.5% and seven at 95%. A history edit's effect thus depends on the world it describes, and a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all.

Comments: 23 pages, 6 figures; code and data in ancillary files

Subjects:

Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)

Cite as: arXiv:2611.14654 [cs.AI]

(or arXiv:2611.14654v1 [cs.AI] for this version)

https://doi.org/10.48550/arXiv.2611.14654

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Anurag Koul [view email] [v1] Tue, 10 Nov 2026 19:06:36 UTC (8,443 KB)

Full-text links:

Access Paper:

View a PDF of the paper titled Cleaning an Agent's History Hurts When the World Still Holds the Error, by Egor Pakhomov and 5 other authors

View PDF

HTML (experimental)

TeX Source

view license

Ancillary-file links:

Ancillary files (details):

MANIFEST.tsv

README.txt

packets/e29-primary-surface-packets.tar.xz

packets/e36-appworld-arm-packets.tar.xz

positioning/step-3-positioning-passages.tar.xz

readouts/e36-units-RESULT.tar.xz

readouts/e39-units-RESULT.tar.xz

readouts/endpoint-rescore.tar.xz

sources/earlier-tree/analysis/calibration_band.py

sources/earlier-tree/analysis/tables/channel-contrasts.tsv

sources/earlier-tree/analysis/tables/equivalence.tsv

sources/earlier-tree/analysis/tables/pooled-primary.tsv

sources/earlier-tree/harness/runs/E1-gate/gate-results.json

sources/earlier-tree/runs/E27-contemporaneous/analysis/contrasts.tsv

sources/earlier-tree/runs/E29-truthful-repair/ANALYSIS-CONTRACT-RECORD.json

sources/earlier-tree/runs/E29-truthful-repair/ANALYSIS-CONTRACT.md

sources/earlier-tree/runs/E29-truthful-repair/CONFIRMATION-DESIGN-RECORD.json

sources/earlier-tree/runs/E29-truthful-repair/CONFIRMATION-MANIFEST.json

sources/earlier-tree/runs/E29-truthful-repair/DESIGN.md

sources/earlier-tree/runs/E29-truthful-repair/analyze_confirmation.py

sources/earlier-tree/runs/E29-truthful-repair/confirmation-analysis/contrasts.tsv

sources/earlier-tree/runs/E29-truthful-repair/confirmation_analysis.py

sources/earlier-tree/runs/E29-truthful-repair/secondary-analysis/secondary-contrasts.tsv

sources/earlier-tree/runs/E29-truthful-repair/verification/released-report.json

sources/earlier-tree/runs/E31-global-progress/analysis/global-progress.tsv

sources/earlier-tree/runs/E31-global-progress/analysis/slot-decomposition.tsv

sources/earlier-tree/runs/E32-error-flag/analysis/contrasts.tsv

sources/earlier-tree/runs/E33-appworld-diagnosis/analysis/contrasts.tsv

sources/earlier-tree/runs/E34-history-world-interaction/analysis/contrasts.tsv

sources/earlier-tree/runs/E34-history-world-interaction/closed-frame-precheck/VERIFIED.json

sources/earlier-tree/runs/E9-ceiling-probe/READOUT.txt

sources/this-tree/experiments/E36-direct-treatment-scale/collection/FRAME-VERIFIED.json

sources/this-tree/experiments/E36-direct-treatment-scale/collection/PLAN-SEALED.json

sources/this-tree/experiments/E36-direct-treatment-scale/collection/PLAN.json

sources/this-tree/experiments/E36-direct-treatment-scale/dependencies/agentce-harness/config.py

sources/this-tree/experiments/E36-direct-treatment-scale/dependencies/appworld-package/appworld-0.1.3.post1.dist-info/METADATA

sources/this-tree/experiments/E36-direct-treatment-scale/dependencies/appworld-package/pip-freeze.txt

sources/this-tree/experiments/E36-direct-treatment-scale/design/E36-CONFIRMATION-ANALYSIS-CONTRACT.json

sources/this-tree/experiments/E36-direct-treatment-scale/design/E36-CONFIRMATION-DESIGN.md

sources/this-tree/experiments/E36-direct-treatment-scale/design/E36-PROCESS-DIAGNOSTICS.md

sources/this-tree/experiments/E36-direct-treatment-scale/implementation/analysis/analyze_e36_confirmation.py

sources/this-tree/experiments/E36-direct-treatment-scale/implementation/analysis/check_e36_confirmation_frame.py

sources/this-tree/experiments/E36-direct-treatment-scale/implementation/analysis/freeze_e36_confirmation_frame.py

sources/this-tree/experiments/E36-direct-treatment-scale/implementation/analysis/policy_estimand_bounds.py

sources/this-tree/experiments/E36-direct-treatment-scale/implementation/analysis/reproduce_e36_confirmation.py

sources/this-tree/experiments/E36-direct-treatment-scale/implementation/analysis/verify_e36_confirmation.py

sources/this-tree/experiments/E36-direct-treatment-scale/implementation/collection/workflow_diagnostics.py

sources/this-tree/experiments/E36-direct-treatment-scale/planning-evidence/e36-statistical-pilot/analysis/READOUT.json

sources/this-tree/experiments/E38-channel-isolation/RECOMPUTED.json

sources/this-tree/experiments/E38-channel-isolation/check_pooled_claims.py

sources/this-tree/experiments/E38-channel-isolation/integrity_check.py

sources/this-tree/experiments/E38-channel-isolation/levels_readout.py

sources/this-tree/experiments/E38-channel-isolation/recompute_contrasts.py

sources/this-tree/experiments/E38-channel-isolation/validate_inheritance.py

sources/this-tree/experiments/E39-receiver-expansion/collection/gemini-2026-09-28/COMPLETE.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gemini-2026-09-28/PLAN.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt-rungs-2026-09-28/COMPLETE.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt-rungs-2026-09-28/PLAN.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt4o-2026-09-30-partial/COMPLETE.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt4o-2026-09-30-partial/PLAN.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt4o-2026-10-03/COMPLETE.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt4o-2026-10-03/PLAN.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt4omini-2026-09-29/COMPLETE.json

sources/this-tree/experiments/E39-receiver-expansion/collection/gpt4omini-2026-09-29/PLAN.json

sources/this-tree/experiments/E39-receiver-expansion/collection/grok-2026-09-28/COMPLETE.json

sources/this-tree/experiments/E39-receiver-expansion/collection/grok-2026-09-28/PLAN.json

sources/this-tree/experiments/E39-receiver-expansion/collection/pilot-2026-09-28/COMPLETE.json

sources/this-tree/experiments/E39-receiver-expansion/collection/pilot-2026-09-28/PLAN.json

sources/this-tree/experiments/E39-receiver-expansion/endpoint-rescore/SUMMARY-gpt-5-mini.json

sources/this-tree/experiments/E39-receiver-expansion/endpoint-rescore/SUMMARY-gpt-5.json

sources/this-tree/experiments/E39-receiver-expansion/endpoint-rescore/rescore-gemini.jsonl

sources/this-tree/experiments/E39-receiver-expansion/endpoint-rescore/rescore-gpt.jsonl

sources/this-tree/experiments/E39-receiver-expansion/endpoint-rescore/rescore-gpt4o.jsonl

sources/this-tree/experiments/E39-receiver-expansion/endpoint-rescore/rescore-gpt4omini.jsonl

sources/this-tree/experiments/E39-receiver-expansion/endpoint-rescore/rescore-grok.jsonl

sources/this-tree/experiments/E39-receiver-expansion/implementation/build_contrasts_tsv.py

sources/this-tree/experiments/E39-receiver-expansion/implementation/e39_common.py

sources/this-tree/experiments/E39-receiver-expansion/implementation/estimate_e39.py

sources/this-tree/experiments/E39-receiver-expansion/readout/ALL-FIVE-E39.json

sources/this-tree/experiments/E39-receiver-expansion/readout/ALL-SIX-E39.json

sources/this-tree/experiments/E39-receiver-expansion/readout/E36-PUBLISHED-65FAM.json

sources/this-tree/experiments/E39-receiver-expansion/readout/ESTIMATOR-VALIDATION.json

sources/this-tree/experiments/E39-receiver-expansion/readout/GEMINI-FULL-65FAM.json

sources/this-tree/experiments/E39-receiver-expansion/readout/GPT-RUNGS-FULL65.json

sources/this-tree/experiments/E39-receiver-expansion/readout/GPT4O-FULL65.json

sources/this-tree/experiments/E39-receiver-expansion/readout/GPT4O-SENS-SINGLE-INVOCATION-BLOCKS.json

sources/this-tree/experiments/E39-receiver-expansion/readout/GPT4OMINI-FULL65.json

sources/this-tree/experiments/E39-receiver-expansion/readout/GROK.json

sources/this-tree/experiments/E39-receiver-expansion/readout/contrasts-all-eight-receivers-2026-10-03.tsv

sources/this-tree/experiments/E39-receiver-expansion/readout/contrasts-all-seven-receivers-2026-09-29.tsv

sources/this-tree/experiments/analyses-2026-10-07/R1-e29-four-arm-means/RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R1-e29-four-arm-means/r1_four_arm_means.py

sources/this-tree/experiments/analyses-2026-10-07/R2-table1-graded/RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R2-table1-graded/r2_table1_graded.py

sources/this-tree/experiments/analyses-2026-10-07/R3-e34-interaction/RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R3-e34-interaction/r3_e34_interaction.py

sources/this-tree/experiments/analyses-2026-10-07/R4-e39-975/RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R4-e39-975/r4_e39_975.py

sources/this-tree/experiments/analyses-2026-10-07/R5-step-budget/RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R5-step-budget/r5_step_budget.py

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/R6a-REFUSALS.json

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/R6a-RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/R6b-RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/R6c-RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/r6a_refusal_path.py

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/r6a_refusals_observed.py

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/r6b_independent_estimator.py

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/replay/diagnose-orderitem.json

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/replay/diagnose.jsonl

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/replay/diagnose_disagreements.py

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/replay/diagnose_orderitem.py

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/replay/reexecute_driver.py

sources/this-tree/experiments/analyses-2026-10-07/R6-e36-provenance/replay/replay.jsonl

sources/this-tree/experiments/analyses-2026-10-07/R7-e35-preregistered-frame/RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R7-e35-preregistered-frame/r7_e35_preregistered_frame.py

sources/this-tree/experiments/analyses-2026-10-07/R8-counts/R8a-RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R8-counts/R8b-RESULTS.json

sources/this-tree/experiments/analyses-2026-10-07/R8-counts/r8a_attempt_counts.py

sources/this-tree/experiments/analyses-2026-10-07/R8-counts/r8b_natural_disagreement.py

sources/this-tree/experiments/budget-exhaustion/RESULTS.json

sources/this-tree/experiments/e29-readout/slot-split/RESULTS.json

sources/this-tree/experiments/e35-readout/COMPARISON.json

sources/this-tree/experiments/e35-readout/PLAN.json

sources/this-tree/experiments/e35-readout/RECOMPUTED-READOUT.json

sources/this-tree/experiments/e35-readout/recompute_e35.py

sources/this-tree/experiments/e35-readout/step-attribution-v1/STEP-ATTRIBUTION-READOUT.json

sources/this-tree/experiments/e36-readout/READOUT.json

sources/this-tree/experiments/e36-readout/endpoint-rescore/SUMMARY.json

sources/this-tree/experiments/e36-readout/process-measures/RESULTS.json

sources/this-tree/experiments/e36-readout/process-measures/process_measures.py

sources/this-tree/experiments/e36-readout/process-measures/tables/contrasts-design-receivers.tsv

sources/this-tree/experiments/e36-readout/process-measures/tables/e36-per-unit.tsv

sources/this-tree/experiments/e36-readout/process-measures/tables/focal-operations.tsv

sources/this-tree/experiments/e36-readout/process-measures/tables/per-cell.tsv

sources/this-tree/experiments/e36-readout/process-measures/tables/per-unit.tsv

sources/this-tree/manuscript-v2/evidence/E36-RECEIVERS-SHARED63-2026-10-02.json

sources/this-tree/manuscript-v2/evidence/release/NUMBERS.tsv

sources/this-tree/prior-art/step-3-positioning/params/our-dimensions.md

sources/this-tree/prior-art/step-3-positioning/params/params.json

(134 additional files not shown)

Additional Features

Audio Summary

Current browse context:

cs.AI

new | recent | 2026-11

Change to browse by:

cs cs.CL cs.LG

References & Citations

SciX

Google Scholar

Semantic Scholar

Loading...

Data provided by:

Bibliographic Tools

Bibliographic and Citation Tools

Bibliographic Explorer Toggle

Bibliographic Explorer (What is the Explorer?)

Connected Papers Toggle

Connected Papers (What is Connected Papers?)

Litmaps Toggle

Litmaps (What is Litmaps?)

scite.ai Toggle

scite Smart Citations (What are Smart Citations?)

Code, Data, Media

Code, Data and Media Associated with this Article

alphaXiv Toggle

alphaXiv (What is alphaXiv?)

Links to Code Toggle

CatalyzeX Code Finder for Papers (What is CatalyzeX?)

DagsHub Toggle

DagsHub (What is DagsHub?)

GotitPub Toggle

Gotit.pub (What is GotitPub?)

Huggingface Toggle

Hugging Face (What is Huggingface?)

ScienceCast Toggle

ScienceCast (What is ScienceCast?)

Demos

Demos

Replicate Toggle

Replicate (What is Replicate?)

Spaces Toggle

Hugging Face Spaces (What is Spaces?)

Spaces Toggle

TXYZ.AI (What is TXYZ.AI?)

Related Papers

Recommenders and Search Tools

Link to Influence Flower

Influence Flower (What are Influence Flowers?)

Core recommender toggle

CORE Recommender (What is CORE?)

Author

Venue

Institution

Topic

About arXivLabs

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?)

Key points and analysis

Article intelligence

EngineersAdvanced

Key points

  • On AgentCE-Bench, a truthful note beside the preserved error lifted recovery from 51% to 66%, while writing the fix as a completed action dropped it from 50% to 44%.
  • On AppWorld, a success receipt cut whole-task success from 45% to 28% when the effect was absent, but raised it from 38% to 49% once the effect was present.
  • The two effects differed by 27 points pooled across receivers, positive in six of eight receivers at 97.5% and seven at 95%.
  • A history edit's value depends on the world it describes; a history that looks clean while the error remains in the world can hurt the agent.

Highlights and analysis are generated automatically and may contain errors. Check the original source.