Jiang, Yuhua, Xu, Chang, Gao, Feifei, Zhao, Li · Oct 23, 2026
HarnessMM: Recursive Self-Improvement for Harness Token Efficiency
Markdown · JSON · Original source
Yuhua Jiang, Chang Xu, Feifei Gao and Li Zhao present HarnessMM, a research project that aims to reduce repeated reasoning in agent harnesses: software that coordinates model reasoning, tool execution and feedback. Their approach uses to refine prompts, context settings and adapter rules with training feedback, while validation selects configurations. At inference , a controller reuses decisions when observed feedback meets specified guards; the actor’s weights stay fixed, and the actor resumes when further reasoning is needed. The authors report results across coding, tool-based and visual-agent settings. On HumanEval, they report preserving Codex success while reducing tokens per task; on DeepSeek Harness, they report both improved success and fewer tokens. An MBPP ablation also reports lower cost than two comparison approaches. These results illustrate the report’s proposed tradeoff: avoid redundant reasoning callbacks without replacing the actor’s ability to reason when reuse conditions fail. The abstract summarizes empirical results but does not provide the task-level evidence or evaluation details needed to assess how broadly the results 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: Oct 23, 2026 (2026-10-23).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T20:32:30.401Z. First observed: 2026-10-26T06:01:14.319Z. Retrieved: 2026-11-12T18:20:12.673Z.
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 evidence is the abstract, which summarizes results but does not include task-level evaluation details. The abstract says optimization overhead and execution are documented where measured, but does not give those measurements here. The reported results are tied to the named and settings; the abstract does not establish performance in other settings.
Open question
How do the reported results vary across tasks and evaluation conditions within the coding, tool-based and visual-agent settings?
Source-stated
The authors describe HarnessMM as reducing repeated reasoning by reusing decisions when observed feedback satisfies guards, while keeping actor weights fixed and resuming reasoning when needed.
At inference time, a controller reuses resolved decisions while keeping actor weights fixed. Executable Decision Graphs encode actions, arguments, dependencies, and guards.
Source-stated
The authors report that on HumanEval, HarnessMM preserves Codex success while reducing tokens per task from 66.72k to 47.02k.
On HumanEval, HarnessMM preserves Codex success at 90.82\% while reducing tokens per task from 66.72k to 47.02k (29.54\% saving).
Source-stated
For DeepSeek Harness, the authors report higher success and fewer tokens with HarnessMM than the comparison figures given in the abstract.
For DeepSeek Harness, success improves from 71.43\% to 85.71\% while tokens decrease from 96.28k to 11.62k (87.93\% saving).
Source-stated
The authors report that an MBPP ablation found lower cost than both single-use graphs and retained callbacks.
An MBPP ablation shows 26.91\% lower cost than single-use graphs and 19.52\% lower cost than retained callbacks.
Escape Meridian interpretation
The report covers coding, tool-based and visual-agent settings, but its abstract does not give task-level evaluation details for assessing how broadly the findings apply.
This technical report consolidates its shared framework, implementations, and empirical studies across coding, tool-based, and visual-agent settings.
Source and contributor context
Jiang, Yuhua, Xu, Chang, Gao, Feifei, Zhao, Li · 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.
Recursive self-improvement
An AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 7d6a3549-28a3-4431-8a19-651918be8ec8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T20:32:30.401Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: be9533160645fa7ef15f8e798936cdc25ca6afe8659755686bfd47fb2f17f49d. Source text hash: 4c07b99b152eb88da223e0d873c699966594192fc7b67b061e6e096c83d5cca5.
Revision and source-check history
Version 1 · 2026-11-12T20:32:30.401Z
HarnessMM: Recursive Self-Improvement for Harness Token Efficiency
Yuhua Jiang, Chang Xu, Feifei Gao and Li Zhao present HarnessMM, a research project that aims to reduce repeated reasoning in agent harnesses: software that coordinates model reasoning, tool execution and feedback. Their approach uses to refine prompts, context settings and adapter rules with training feedback, while validation selects configurations. At inference , a controller reuses decisions when observed feedback meets specified guards; the actor’s weights stay fixed, and the actor resumes when further reasoning is needed. The authors report results across coding, tool-based and visual-agent settings. On HumanEval, they report preserving Codex success while reducing tokens per task; on DeepSeek Harness, they report both improved success and fewer tokens. An MBPP ablation also reports lower cost than two comparison approaches. These results illustrate the report’s proposed tradeoff: avoid redundant reasoning callbacks without replacing the actor’s ability to reason when reuse conditions fail. The abstract summarizes empirical results but does not provide the task-level evidence or evaluation details needed to assess how broadly the results apply.
Original source
cs.AI · Article
Source details
Original source date: Oct 23, 2026 (2026-10-23).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T20:32:30.401Z. First observed: 2026-10-26T06:01:14.319Z. Retrieved: 2026-11-12T18:20:12.673Z.
Limits
The evidence is the abstract, which summarizes results but does not include task-level evaluation details. The abstract says optimization overhead and execution are documented where measured, but does not give those measurements here. The reported results are tied to the named and settings; the abstract does not establish performance in other settings.
Open question
How do the reported results vary across tasks and evaluation conditions within the coding, tool-based and visual-agent settings?
Source-stated
The authors describe HarnessMM as reducing repeated reasoning by reusing decisions when observed feedback satisfies guards, while keeping actor weights fixed and resuming reasoning when needed.
At inference time, a controller reuses resolved decisions while keeping actor weights fixed. Executable Decision Graphs encode actions, arguments, dependencies, and guards.
Source-stated
The authors report that on HumanEval, HarnessMM preserves Codex success while reducing tokens per task from 66.72k to 47.02k.
On HumanEval, HarnessMM preserves Codex success at 90.82\% while reducing tokens per task from 66.72k to 47.02k (29.54\% saving).
Source-stated
For DeepSeek Harness, the authors report higher success and fewer tokens with HarnessMM than the comparison figures given in the abstract.
For DeepSeek Harness, success improves from 71.43\% to 85.71\% while tokens decrease from 96.28k to 11.62k (87.93\% saving).
Source-stated
The authors report that an MBPP ablation found lower cost than both single-use graphs and retained callbacks.
An MBPP ablation shows 26.91\% lower cost than single-use graphs and 19.52\% lower cost than retained callbacks.
Escape Meridian interpretation
The report covers coding, tool-based and visual-agent settings, but its abstract does not give task-level evaluation details for assessing how broadly the findings apply.
This technical report consolidates its shared framework, implementations, and empirical studies across coding, tool-based, and visual-agent settings.
Source and contributor context
Jiang, Yuhua, Xu, Chang, Gao, Feifei, Zhao, Li · 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.
Recursive self-improvement
An AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 7d6a3549-28a3-4431-8a19-651918be8ec8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T20:32:30.401Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: be9533160645fa7ef15f8e798936cdc25ca6afe8659755686bfd47fb2f17f49d. Source text hash: 4c07b99b152eb88da223e0d873c699966594192fc7b67b061e6e096c83d5cca5.
passed · 2026-11-12T20:32:30.401Z · anthropic/claude-opus-5.5
Publication: 7d6a3549-28a3-4431-8a19-651918be8ec8
attribution — pass: The reading names the four listed authors and attributes the method description and every result to them with phrasing such as "they report." It does not present the benchmark outcomes as independently verified.
Authors:Yuhua Jiang, Chang Xu, Feifei Gao, Li Zhao
summary — pass: The reading leads with the actors and the goal of cutting repeated reasoning. It defines an agent harness in plain language and explains what recursive self-improvement refines here: prompts, context settings and adapter rules, not model weights. It describes the guarded reuse and fallback mechanism and keeps scope limits. Benchmark names such as HumanEval, MBPP and Codex are left to tooltips but are not essential to the main point. The question is open rather than leading. The summary omits the headline percentages, but the claims carry them.
Agent harnesses often invoke reasoning after each tool result, even when earlier calls have already determined subsequent actions.
date — pass: originalPublished is 2026-10-23, which matches the v1 submission date. It does not use the later v2 revision date (6 Nov) or the feed date (9 Nov).
[Submitted on 23 Oct 2026 (v1), last revised 6 Nov 2026 (this version, v2)]
support — pass: Each claim quotes matching abstract text, and the figures are copied accurately. The single interpretive claim is labeled as interpretation and stays within the evidence by noting only that the abstract lacks task-level details. The limitations correctly note that overhead and execution-time measurements are not shown. No benefits, costs or forecasts are invented.
The report also summarizes the HarnessMM-V visual-agent study and documents optimization overhead and execution time where measured.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, 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.
An AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceAn AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference