File size: 6,691 Bytes
31d8fc0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29c59b5
31d8fc0
29c59b5
31d8fc0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29c59b5
 
31d8fc0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
---
dataset_info:
  features:
    - name: task_id
      dtype: string
    - name: title
      dtype: string
    - name: category
      dtype: string
    - name: difficulty
      dtype: string
    - name: has_generator
      dtype: bool
    - name: ablation_scores
      dtype: string
    - name: tni
      dtype: float32
    - name: classification
      dtype: string
  splits:
    - name: train
      num_examples: 153
    - name: test
      num_examples: 120
license: mit
task_categories:
  - text-generation
tags:
  - multi-agent
  - benchmark
  - software-engineering
  - teamwork
  - evaluation
  - llm-evaluation
pretty_name: TeamBench
---

# TeamBench: A Multi-Agent Teamwork Benchmark

## Overview

TeamBench is a rigorous benchmark for evaluating whether LLM-based agent *teams* outperform single oracle agents on realistic software engineering tasks. Each task is executed under five ablation conditions (oracle, restricted, full team, team without planning, team without verification), enabling fine-grained measurement of when and why teamwork helps. The core metric is the **Teamwork Necessity Index (TNI)**, which quantifies how much a task requires coordinated multi-agent effort beyond what a single capable agent can achieve alone.

TeamBench uses OS-enforced role separation — Planner, Executor, and Verifier agents operate in isolated sandboxes with distinct tool allow-lists — ensuring that role boundaries are structurally enforced rather than prompt-based. Tasks span 18 categories including security, data engineering, adversarial specification traps, distributed systems, and long-horizon planning, with 153 tasks totalling 459 parameterized instances (seeds 0–2).

---

## Task Categories

| Category | Tasks | Description |
|---|---|---|
| Software Eng. | 22 | Hidden specs, backward compatibility, refactoring |
| Security | 17 | Vulnerability patching, cryptographic correctness, audit triage |
| Operations | 16 | Incident root-cause, container debugging, monitoring |
| Data Engineering | 12 | Schema drift, ETL repair, query optimization |
| Testing | 11 | Spec-to-tests, mutation resistance, property-based testing |
| Incident Response | 11 | Cascade failure, memory leak, rollback planning |
| Information Retrieval | 8 | Evidence QA, misinformation traps, multi-source retrieval |
| Policy | 8 | Access control, data retention, license compliance |
| Distributed Systems | 7 | Race conditions, Raft consensus, idempotency |
| Adversarial | 7 | Spec conflicts, false bug reports, security theater |
| Code Review | 6 | API review, style enforcement, test coverage |
| Multi-language | 6 | Go concurrency, JavaScript XSS, polyglot debugging |
| Long-Horizon | 6 | Multi-step migrations, staged deployments, audit trails |
| Pipeline | 6 | ETL fix, API gateway, message queues |
| Cross-System Integration | 5 | API contract mismatches, schema evolution, auth federation |
| Specification | 3 | Feature implementation from RFC, config schema design |
| Integration | 1 | Pipeline repair, API versioning |
| Negotiation | 1 | Trade-off configuration under competing constraints |

**Difficulty breakdown:** 104 hard, 26 medium, 16 expert, 7 easy (78% hard or expert).

---

## Ablation Conditions

Each task is evaluated under five conditions:

| Condition | Description |
|---|---|
| `oracle` | Single powerful agent with full tool access and no role restrictions |
| `restricted` | Single agent with executor-only tool access (no planning/verification tools) |
| `team` | Full three-role team: Planner → Executor → Verifier |
| `team_no_plan` | Two-role team: Executor → Verifier (planning phase skipped) |
| `team_no_verify` | Two-role team: Planner → Executor (verification phase skipped) |

Scores are in [0, 1] and represent the fraction of grader checks passed.

---

## TNI Metric

The **Teamwork Necessity Index (TNI)** measures how much a task *requires* teamwork:

```
TNI = team_uplift / necessity_gap
    = (team - oracle) / (1 - restricted)
```

- `TNI > 0`: team outperforms oracle relative to the task's difficulty ceiling
- `TNI = 1.0`: team achieves the maximum possible improvement over oracle
- `TNI > 1.0`: team substantially exceeds oracle (rare; indicates strong synergy)
- `TNI < 0`: team underperforms oracle (teamwork overhead hurts)

**Classification thresholds:**
- `HIGH-TNI`: TNI ≥ 0.5 and team > oracle
- `TEAM-HELPS`: team > oracle but TNI < 0.5
- `NEUTRAL`: |team - oracle| ≤ 0.05
- `TEAM-HURTS`: team < oracle

Of the 153 tasks: TEAM-HELPS 53, NEUTRAL 51, TEAM-HURTS 28, HIGH-TNI 16.

---

## Usage

```python
import json

with open("teambench_dataset.json") as f:
    tasks = json.load(f)

# Filter to tasks where teamwork helps
team_helps = [t for t in tasks if t.get("classification") in ("TEAM-HELPS", "HIGH-TNI")]
print(f"Tasks where team > oracle: {len(team_helps)}")

# Get hard tasks with ablation scores
hard_with_scores = [
    t for t in tasks
    if t["difficulty"] in ("hard", "expert") and "ablation_scores" in t
]
print(f"Hard/expert tasks with ablation data: {len(hard_with_scores)}")

# Compute average team uplift
uplifts = [
    t["ablation_scores"]["team"] - t["ablation_scores"]["oracle"]
    for t in tasks
    if "ablation_scores" in t
]
print(f"Mean team uplift: {sum(uplifts)/len(uplifts):.3f}")
```

### Loading via HuggingFace datasets

```python
from datasets import load_dataset

ds = load_dataset("ybkim95/teambench", split="train")  # all 153 tasks
# ds = load_dataset("ybkim95/teambench", split="test")  # 120 hard/expert tasks only
print(ds[0])
```

---

## Benchmark Results

Cross-model evaluation across Gemini and OpenAI model families shows that:

- Team outperforms oracle on **43.9%** of tasks (68/155 with full ablation data)
- Average TNI: **0.744** across tasks with a measurable necessity gap
- Weaker models benefit *more* from teamwork (larger relative uplift)
- `team_no_verify` is often the strongest condition — verifier overhead can hurt on average
- The **Expertise-Asymmetry (EA)** variant (5 tasks) shows TNI > 1.0 with capable models, meaning specialized role knowledge pushes teams beyond what any single oracle achieves

Full results and leaderboard: [GitHub](https://github.com/ybkim95/TeamBench)

---

## Citation

```bibtex
@dataset{kim2026teambench,
  author    = {Yubin Kim},
  title     = {TeamBench: A Multi-Agent Teamwork Benchmark for LLM Evaluation},
  year      = {2026},
  url       = {https://huggingface.co/datasets/ybkim95/teambench},
  note      = {153 tasks across 18 categories with OS-enforced role separation}
}
```

---

## License

MIT License. See [LICENSE](https://github.com/ybkim95/TeamBench/blob/main/LICENSE) for details.