File size: 2,634 Bytes
01c21fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
task_categories:
- other
language:
- en
tags:
- agents
- llm-agents
- tool-use
- appworld
- api-discovery
pretty_name: CEG  AppWorld Hidden-API Discovery Benchmark
---

# CEG — AppWorld Hidden-API Discovery Benchmark

Benchmark **specification** for the *hidden-API discovery* setting used in the CEG project
(**Observation-Verified Partial-Skill Memory for Training-Free LLM Agents**).

A set of advanced-feature APIs is **removed from the agent's documentation** (their existence,
signature, and docs are redacted) while they remain **executable**, forcing a *frozen* agent to
(i) **discover** each capability from its own observations and (ii) **use it reliably** to complete
multi-requirement [AppWorld](https://appworld.dev) tasks.

> This repo ships the **selection spec**, not raw task content. The actual AppWorld tasks come from
> the `appworld` Python package (`pip install appworld`); these files tell you *which tasks* to run
> and *which APIs to hide*.

## Contents

```
appworld_hidden_api/
  config_13api/        # main paper set — 48 tasks, 13 hidden APIs
    hidden_apis.json   #   {"hard": [ 13 fully-qualified API names to redact ]}
    task_ids.json      #   [ 48 AppWorld task ids ]  (16 templates × 3 variations)
    task_meta.json     #   task_id -> {"split": ..., "gold_hidden": [APIs the gold solution calls]}
  config_30api/        # larger variant — 75 tasks, 30 hidden APIs (same schema)
```

## How the sets were built (reward-free, no hardcoding)

1. **Pool** AppWorld tasks across `train`+`dev` splits.
2. Keep a task only if its **ground-truth solution genuinely calls one of the hidden APIs**
   (recorded per task in `task_meta.json["gold_hidden"]`) — so discovery is *required*, not optional.
3. The hidden APIs are **non-basic "new-feature" actions** (e.g. `spotify.follow_artist`,
   `spotify.add_to_queue`, `venmo.remind_payment_request`, `file_system.compress_directory`);
   each app's authentication, reads (`show_/list_/search_`), and primary create/send/play actions
   stay **visible**.

## Usage

```python
import json
hidden = json.load(open("appworld_hidden_api/config_13api/hidden_apis.json"))["hard"]
task_ids = json.load(open("appworld_hidden_api/config_13api/task_ids.json"))
# Run each task_id in AppWorld while redacting `hidden` from the API catalog
# (show_api_descriptions / show_api_doc / search_api_docs) but keeping them CALLABLE.
```

The 13-API config is the headline benchmark; official **Task Goal Completion (TGC)** is the metric.

## Citation

CEG — *Observation-Verified Partial-Skill Memory for Training-Free LLM Agents* (in progress).