HARNESS RSITRACEABLE EVIDENCE DATASET + RESEARCH ATLAS
Same model.
A different way to work.
A curated dataset of reported harness changes and outcomes. Trace what changed, what helped, what failed, and what the evidence cannot establish. Explore the comparisons or download a versioned snapshot.
Preserve a multiline edit as a file payload instead of dispatching its contents to the shell.
Inspect the reported change ↗THE RESEARCH QUESTION
What changed in the harness? What did that change accomplish? Under which conditions might it be useful again?
An agent’s harness is the prompts, tools, memory and execution logic around its model. Our survey connects how the improvement workflow works, whether the improvement procedure itself changes, and what reusable mechanisms emerge across tasks.
A FEW CHANGES, UP CLOSE
Selected results from different studies and settings. Each card states what was measured; these are not head-to-head comparisons.
01 / INSIDE THE RESEARCH
A harness change persists beyond one answer. Follow a candidate from proposal to execution, evidence and reuse.
THE RECURSIVE QUESTION
A fixed improvement procedure searches for better task-agent harnesses.
A revised agent, improvement skill or search policy participates in generating later revisions.
Task gains alone do not establish better subsequent optimization or general RSI. Our survey examines reported configurations and their evidence separately. Related RSI framework ↗
02 / CROSS-DOMAIN DISCOVERIES
Choose a question, then compare two domains.
See the shared function, the actual edits and the conditions for reuse.
Follow each comparison to the source-linked observation, then explore the underlying benchmark records in the atlas below.
03 / THE EVOLUTION ATLAS
Four task domains. Concrete before-and-after changes.
Representative evidence and its limits, together.
Selected synthesis from the survey’s domain evidence. Each card compares reported reusable changes; it is not one shared implementation or one experiment.
THE DATA / REPORTED, NOT REPRODUCED
A versioned evidence dataset, not a leaderboard.
One row is a source within a task subcategory—not an independent intervention.
4/95 records screened for structured signals · 6 scoped observations · 91 not yet reviewed
The 95-record survey snapshot is unchanged. Outcome/retention labels are a selected pilot screened from existing curated text, not a new independent audit of every source. Unreviewed ≠ no failures; usage ≠ benefit; retained ≠ effective. No dataset-specific reuse license has been granted.
Dataset could not be loaded. The versioned download links remain available.
One record may contain multiple outcomes under different comparisons. Each signal has a scope, attribution type, retention status, exact supporting record excerpt, and inherited source locator. Empty benchmark applicability does not mean every benchmark. The remaining records and reserved control/selection fields are explicitly unassessed. This is not a representative sample of attempted edits; do not calculate a global success or failure rate. The separate living bibliography below includes newer papers that have not entered this evidence snapshot.
04 / THE EVIDENCE NOTEBOOK
Explore what a harness can change, what was actually observed, and what the comparison supports. Every record keeps its original setting and source.
CURRENT VIEW
Open a record for the starting harness, exact change, observed result, retention status and primary-source location. Missing details stay visible.
Benchmark evidence could not be loaded. Download the data or refresh this page.
05 / LIVING COLLECTION
Publication trends and actual collection updates.
Explore cumulative growth or monthly activity.
| Month | Papers | Cumulative |
|---|
Counts cover the survey collection, including related methods, surveys, and benchmarks.
Publication months use first arXiv submission where available, otherwise bibliographic publication dates.
Recursive reuse: Harness², RSIFlow and Harmony Harness distinguish task-harness changes, retained improvement guidance and supplied search machinery.
Failures and generalization: failure-feedback directedness, GSO and Harness Evolution Hits a Ceiling separate target repair, regressions, held-out effects and trained-model carryover.
Memory and definitions: MemSkill, ERSkill, SESA and Self-Evolving Agent Harness distinguish stored content, reusable procedures, trained policies and conceptual scope.
Read the source-linked comparison supplement ↗
This targeted synthesis covers already-collected literature through October 20 and is now incorporated in the revised GitHub reading PDF. The October 19 PDF remains archived; submitted manuscripts and the dataset are unchanged. The frozen evidence still covers October 1: 95 records, dataset v0.1.0. Unreported failures are not recovered by this update.
Loading collection history…
Collection statistics could not be loaded. Please refresh this page.
06 / BUILT TO KEEP EVOLVING
The paper captures a research snapshot.
This atlas is where the evidence keeps growing.
Suggest a study ↗ · Correct an evidence record ↗ · Read the curation guide ↗. Community suggestions and source-backed corrections are welcome.
Update details could not be loaded. Please refresh this page.
07 / HOW TO READ THE ATLAS
Useful comparisons need more than a headline score. Here is what the records do—and do not—tell you.
A reusable change to the prompts, interfaces, tools, context handling, skills or control logic around a target model. A one-off answer, generated image or task patch is not automatically a harness change. The records distinguish supplied mechanisms from modifications discovered during evolution.
The living bibliography includes surveys, benchmarks and adjacent methods. The frozen domain evidence is a narrower set: 95 category–source records from 62 research papers and one benchmark resource. A paper can contribute a record to more than one task category.
Read the evidence label. A full-system gain does not isolate one edit. An ablation tests the components actually removed; a paired case remains a case. A rejected or unused proposal is not presented as a final retained capability.
Use it as a design hypothesis. Compare the failure trigger, available observations and corrective actions. An analogy across domains is not an established transfer result; unchanged-harness transfer is identified separately when reported.
New studies are screened, concrete revisions are traced to primary sources, and comparisons are revised when the evidence changes. Dated release notes distinguish website changes, evidence updates and bibliography imports. The paper’s October 1, 2026 evidence snapshot remains identifiable.
08 / PROJECT
Read the evidence, reuse the data,
or suggest a paper or correction.
From Harness Optimization to Recursive Self-Improvement: A Survey of Language Model Agents. Revised author preprint, 20 October 2026; 74 pages, independent non-ACL layout; not peer-reviewed. The October 19 reading copy remains archived in GitHub; submitted manuscripts are unchanged.
Read the survey (PDF)Curated sources, classifications and publication-date provenance.
Download BibTeXSource code and supporting research material.
GitHub Public repository09 / THE TEAM
The survey and its living research companion.
The domain navigation could not be loaded. Please refresh this page.