File size: 4,130 Bytes
dc95327 f726fd6 dc95327 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f726fd6 52c4957 f8e4370 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 | ---
license: apache-2.0
language:
- en
tags:
- agents
- benchmark
configs:
- config_name: tasks
default: true
data_files:
- split: test
path: preview/tasks.jsonl
---
<div align="center">
<h1>ReScopeBench</h1>
<h3>Towards Robust Long-Lived Agents</h3>
<p><i>Can agents keep experience useful when their environment changes?</i></p>
[Homepage](https://rescopebench.github.io/) |
[Code](https://github.com/JoeYing1019/ReScopeBench) |
[Quick Start](https://github.com/JoeYing1019/ReScopeBench#quick-start) |
[Main Results](https://github.com/JoeYing1019/ReScopeBench#main-results)
<p>
<b>18</b> stateful environments ·
<b>352</b> tasks ·
<b>26</b> drift events
</p>
</div>
## Overview
**ReScopeBench evaluates the lifecycle of agent experience:** acquiring local
knowledge, reusing it for new goals, and selectively revising it when the
environment changes. Its task streams share structural and rule knowledge,
while unannounced drift events make only part of that knowledge obsolete.
- **Structural experience:** how resources are located, connected, and accessed.
- **Rule experience:** the local conditions and consequences that govern actions.
- **Evolving tasks:** streams of 16, 24, or 32 tasks with persistent business
state and one, two, or three drift events.
This dataset contains the executable **benchmark cases**. The
[GitHub repository](https://github.com/JoeYing1019/ReScopeBench) provides the agent
harness, model API adapters, metric computation, and report generation.
## Dataset Contents
- **`cases/`** contains environment implementations, initial states, task streams,
drift definitions, and evaluator assets. Each environment has one `driver.py`
and one `formal-binding.json`.
- **`preview/tasks.jsonl`** exposes the 352 public task prompts for browsing.
The `case_path` field identifies the corresponding environment directory.
- **`cases-manifest.json`** records the file inventory and SHA-256 checksums used
by the installer.
The preview is a view of the task streams, **not an independent-question
evaluation set**. Evaluating a row in isolation does not reproduce the benchmark.
## Run the Benchmark
Follow the [installation guide](https://github.com/JoeYing1019/ReScopeBench#installation)
to set up Python 3.10, Docker, and the evaluation repository. From its root,
download the matching cases:
```bash
python scripts/download_cases.py
```
The installer selects the immutable dataset revision pinned in the code
repository's `config/cases.json` and verifies the downloaded files. Use that
matched code/data pair rather than combining arbitrary revisions.
For every model and environment, run **both complete settings**:
- **w exp:** retain naturally produced memory, skills, and permitted history.
- **w/o exp:** clear explicit experience between tasks.
The two streams start from the same initial environment and then maintain their
own business state. Clearing experience does not reset that state. See the
[quick start](https://github.com/JoeYing1019/ReScopeBench#quick-start) for model
configuration, batched tool calls, and execution commands.
Reference knowledge and grading assets belong to the evaluator, not to learner
inputs. These cases use the ReScopeBench harness; they are not native Harbor
task bundles. Preserve applicable third-party notices and licenses.
## Evaluate Results
The [metric pipeline](https://github.com/JoeYing1019/ReScopeBench#metrics) reports
task performance, structural and rule experience coverage, immediate updating,
and recovery timeliness, with CSV and HTML outputs. Experience scores measure
coverage of selected reference knowledge and accept equivalent expressions;
task graders also allow other valid strategies.
The [main results](https://github.com/JoeYing1019/ReScopeBench#main-results)
compare 20 models across all 18 environments under both settings.
## Acknowledgements
This research was financially supported by the Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), under Grant No. GML-KF-26-06.
|