Woo, Sunghyeun, Conde, Javier, Reviriego, Pedro · Nov 8, 2026
Same Error, Different Recovery: How Delivery Changes LLM Agent Behavior
Markdown · JSON · Original source
Sunghyeun Woo, Javier Conde and Pedro Reviriego argue that the way an error is delivered in a conversation can change how an agent recovers, even when the error feedback is the same. Across two studies involving 13 open models, they varied where feedback appeared, what guidance was provided, how much history remained, and how turns were structured. They report that starting a new user turn while retaining the failed exchange avoided a severe stopping failure seen in Llama models when that exchange was removed. This approach was non-inferior to history rewriting in pooled synthetic comparisons, but was not best for every model. In ToolMaze, an opt-in recovery layer raised success on explicit errors; the reported gains differed by model size. Loud/Quiet, a set of 24 multi-turn tasks with visible failures and silent corruption, showed no measurable gain. Giving the client the exact fault kind performed better still. The authors say these findings make delivery a distinct factor in recovery, not an engine-exclusive solution or a universal retry policy. The abstract reports results on these evaluations; it does not establish that the same effects occur in other settings.
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 8, 2026 (2026-11-08).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T06:02:22.582Z. First observed: 2026-11-10T08:39:49.307Z. Retrieved: 2026-11-12T05:48:09.427Z.
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 source evidence is the abstract, not the full paper; details of the study methods and results are therefore limited. The abstract reports that results vary by model and task, including no measurable gain on Loud/Quiet; it does not establish effects beyond the evaluations it describes.
Open question
How do the reported effects of delivery vary across models and tasks beyond the evaluations described in the abstract?
Source-stated
The authors say that identical error feedback can prompt different actions depending on where it appears in the conversation.
the same feedback can lead to different actions depending on where it appears in the conversation.
Source-stated
Across their studies, retaining the failed exchange in a new user turn avoided a severe stopping failure observed in Llama models when the exchange was removed.
A new user turn that retains the failed exchange avoids the severe stopping failure observed in Llama models when that exchange is removed.
Source-stated
The new-turn approach was non-inferior to history rewriting in pooled synthetic comparisons, but was not best for every model.
It is non-inferior to history rewriting in pooled synthetic comparisons, but is not best for every model.
Source-stated
The recovery layer raised success on explicit errors in ToolMaze, while Loud/Quiet showed no measurable gain.
On ToolMaze, it raises success on explicit errors by 14.2 points with 7-8B models and 10.5 points with 32-70B models; Loud/Quiet, 24 multi-turn tasks with visible failures and silent corruption, scored by final-state agreement, shows no measurable gain.
Source-stated
The authors characterize delivery as a distinct recovery factor, not a universal retry policy or an engine-exclusive solution.
The findings support delivery as a distinct factor in recovery, not an engine-exclusive solution or a universal retry policy.
Source and contributor context
Woo, Sunghyeun, Conde, Javier, Reviriego, Pedro · 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.
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: be7dc5b5-9b00-4dee-843c-ec283c72a4ab.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T06:02:22.582Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b37b65fe580bb1f167b2cbf395e09cf6ab31358afb1be91b1b23f46f28a6da4f. Source text hash: e73425b40bad119cf00c8bc165ce42ff7f9939cfca0020b697d952fa8092b16b.
Revision and source-check history
Version 1 · 2026-11-12T06:02:22.582Z
Same Error, Different Recovery: How Delivery Changes LLM Agent Behavior
Sunghyeun Woo, Javier Conde and Pedro Reviriego argue that the way an error is delivered in a conversation can change how an agent recovers, even when the error feedback is the same. Across two studies involving 13 open models, they varied where feedback appeared, what guidance was provided, how much history remained, and how turns were structured. They report that starting a new user turn while retaining the failed exchange avoided a severe stopping failure seen in Llama models when that exchange was removed. This approach was non-inferior to history rewriting in pooled synthetic comparisons, but was not best for every model. In ToolMaze, an opt-in recovery layer raised success on explicit errors; the reported gains differed by model size. Loud/Quiet, a set of 24 multi-turn tasks with visible failures and silent corruption, showed no measurable gain. Giving the client the exact fault kind performed better still. The authors say these findings make delivery a distinct factor in recovery, not an engine-exclusive solution or a universal retry policy. The abstract reports results on these evaluations; it does not establish that the same effects occur in other settings.
Original source
cs.AI · Article
Source details
Original source date: Nov 8, 2026 (2026-11-08).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T06:02:22.582Z. First observed: 2026-11-10T08:39:49.307Z. Retrieved: 2026-11-12T05:48:09.427Z.
Limits
The source evidence is the abstract, not the full paper; details of the study methods and results are therefore limited. The abstract reports that results vary by model and task, including no measurable gain on Loud/Quiet; it does not establish effects beyond the evaluations it describes.
Open question
How do the reported effects of delivery vary across models and tasks beyond the evaluations described in the abstract?
Source-stated
The authors say that identical error feedback can prompt different actions depending on where it appears in the conversation.
the same feedback can lead to different actions depending on where it appears in the conversation.
Source-stated
Across their studies, retaining the failed exchange in a new user turn avoided a severe stopping failure observed in Llama models when the exchange was removed.
A new user turn that retains the failed exchange avoids the severe stopping failure observed in Llama models when that exchange is removed.
Source-stated
The new-turn approach was non-inferior to history rewriting in pooled synthetic comparisons, but was not best for every model.
It is non-inferior to history rewriting in pooled synthetic comparisons, but is not best for every model.
Source-stated
The recovery layer raised success on explicit errors in ToolMaze, while Loud/Quiet showed no measurable gain.
On ToolMaze, it raises success on explicit errors by 14.2 points with 7-8B models and 10.5 points with 32-70B models; Loud/Quiet, 24 multi-turn tasks with visible failures and silent corruption, scored by final-state agreement, shows no measurable gain.
Source-stated
The authors characterize delivery as a distinct recovery factor, not a universal retry policy or an engine-exclusive solution.
The findings support delivery as a distinct factor in recovery, not an engine-exclusive solution or a universal retry policy.
Source and contributor context
Woo, Sunghyeun, Conde, Javier, Reviriego, Pedro · 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.
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: be7dc5b5-9b00-4dee-843c-ec283c72a4ab.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T06:02:22.582Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b37b65fe580bb1f167b2cbf395e09cf6ab31358afb1be91b1b23f46f28a6da4f. Source text hash: e73425b40bad119cf00c8bc165ce42ff7f9939cfca0020b697d952fa8092b16b.
uncertain · 2026-11-12T06:02:22.582Z · anthropic/claude-opus-5.5
Publication: be7dc5b5-9b00-4dee-843c-ec283c72a4ab
attribution — pass: The reading correctly attributes the work and its claims to Sunghyeun Woo, Javier Conde and Pedro Reviriego, and it keeps the findings framed as the authors' reported results.
Authors:Sunghyeun Woo, Javier Conde, Pedro Reviriego
summary — uncertain: The summary is faithful, but a general reader may not follow it. It omits the abstract's framing: a tool-using model must choose whether to retry a failed call, correct it, or stop and tell the user. Without that, 'recovery' and 'stopping failure' are unclear. Central terms are also unexplained: history rewriting, non-inferior, pooled synthetic comparisons, client, and engine-exclusive. Smallest repair: open with the retry/correct/stop framing, and add term-library tooltips for history rewriting, non-inferior and LLM agent.
A tool-using language model must decide whether to retry a failed call, correct it, or stop and tell the user.
date — pass: The originalPublished date of 2026-11-08 matches the v1 submission date in the arXiv record. The feed date of 2026-11-10 is later and is not the date displayed.
[Submitted on 8 Nov 2026]
support — pass: Each claim matches an exact excerpt from the abstract. Mixed results are preserved, including that the approach was not best for every model and showed no measurable gain on Loud/Quiet. The limitation that effects beyond these evaluations are not established is appropriately cautious, and no benefits are overstated.
Loud/Quiet, 24 multi-turn tasks with visible failures and silent corruption, scored by final-state agreement, shows no measurable gain.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way to reach 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 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 referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv