File size: 4,789 Bytes
b18f94f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75b499d
 
b18f94f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75b499d
b18f94f
 
 
75b499d
b18f94f
 
 
 
75b499d
 
b18f94f
75b499d
b18f94f
 
 
75b499d
 
 
 
 
 
 
 
 
 
b18f94f
 
 
 
 
 
 
350bfab
b18f94f
 
 
 
 
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
---
license: apache-2.0
pretty_name: "QuorumBench: Asynchronous Multi-Agent LLM Consensus & Straggler Herding Benchmark"
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - llm-agents
  - multi-agent-systems
  - consensus
  - tail-latency
  - semantic-entropy
  - vllm
size_categories:
  - n<1K
models:
  - Qwen/Qwen2.5-3B-Instruct
  - Qwen/Qwen2.5-7B-Instruct
  - Qwen/Qwen2.5-Coder-14B-Instruct
  - HuggingFaceTB/SmolLM2-1.7B-Instruct
  - ibm-granite/granite-3.1-8b-instruct
configs:
  - config_name: policy_decisions
    default: true
    data_files:
      - split: train
        path: data/policy_decisions/train.jsonl
  - config_name: tasks
    data_files:
      - split: train
        path: data/tasks/train.jsonl
  - config_name: agent_observations
    data_files:
      - split: train
        path: data/agent_observations/train.jsonl
  - config_name: gpu_contention
    data_files:
      - split: train
        path: data/gpu_contention/train.jsonl
---

# QuorumBench v1.0

Benchmark tasks and Apple M5 Pro (`48 GB` unified memory, `MPS` FP16) telemetry for **[QuorumGuard](https://github.com/batistado/quorumguard)**: asynchronous multi-agent consensus across `Qwen/Qwen2.5-3B-Instruct`, `Qwen/Qwen2.5-7B-Instruct`, `Qwen/Qwen2.5-Coder-14B-Instruct`, `HuggingFaceTB/SmolLM2-1.7B-Instruct`, and `ibm-granite/granite-3.1-8b-instruct` with isolated Python sandbox verification on both `QuorumBench` and `CRUXEval-50`.

* **Interactive Explorer:** [`batistado/QuorumGuard-Explorer`](https://huggingface.co/spaces/batistado/QuorumGuard-Explorer)
* **Scheduler & Controller Package:** [`batistado/QuorumGuard`](https://huggingface.co/batistado/QuorumGuard)
* **Compiled Paper PDF (10 pages):** [`paper/main.pdf`](paper/main.pdf)
* **GitHub Repository:** [`github.com/batistado/quorumguard`](https://github.com/batistado/quorumguard)

## Dataset Configurations

1. `policy_decisions` (`400` rows): Stage-1, Stage-2, and end-to-end latency/accuracy outcomes across 5 core consensus policies (`Fast_k_of_N`, `CVaR_Static_Cutoff`, `Sync_N_of_N`, `Genealogy_Sync_HPR`, `QuorumGuard`) on both the 50-task 3-tier suite and 30-task 2-tier pilot (`raw/revision_ablations_and_sweeps.json` adds `Fast_3_of_N_Maj`, `TW-Fast`, `Entropy-Cascade`, 7-point CVaR sweep, 10-seed variances, 2,000-bootstrap 95% CIs, hardware-only latencies, and 4-axis/`psi_i` sensitivity).
2. `tasks` (`50` rows): Two-stage systems, numerical precision, SQL, and algorithmic verification workflows (`20` routine tasks and `30` boundary traps) with shared upstream triage notes (`raw/cruxeval50_crossfamily.json` adds external `CRUXEval-50` & `SmolLM2-1.7B + Granite-3.1-8B` runs).
3. `agent_observations` (`560` rows): Per-agent MPS wall-clock times, completion timestamps, pre-commitment 4-option Shannon entropies, top-2 logit margins, provenance roots, and Python sandbox execution receipts.
4. `gpu_contention` (`12` rows): Trace-driven 8-token KV-cached decode quantum contention on `Qwen2.5-7B-Instruct` across loads `rho in {0.58, 0.72, 0.85, 0.93}` (`raw/live_mps_quantum_server.json` adds the live 64-request concurrent multi-`DynamicCache` MPS server benchmark).

## Summary Results (50 Two-Stage Workflows, 3B + 7B + 14B-Coder)

| Policy | S1 / S2 Acc | Joint [95% CI] | 10-Seed (Trap) | Mean / P95 [95% CI] | HW-Only P95 |
| :--- | :---: | :---: | :---: | :---: | :---: |
| `Fast_k_of_N (k=3)` | 46.0% / 52.0% | 28.0% [16, 42] | 27.8+-0.6% (6.7%) | 1.18s / 1.49s [1.4, 1.6] | 0.32s |
| `Fast_3_of_N_Maj` | 46.0% / 52.0% | 28.0% [16, 42] | 28.0+-0.9% (6.7%) | 1.34s / 1.58s [1.5, 1.6] | 0.34s |
| `CVaR_Cutoff (B=2.8s)` | 46.0% / 52.0% | 28.0% [16, 42] | 26.2+-1.1% (6.7%) | 5.60s / 5.60s [5.6, 5.6] | 5.60s |
| `TW-Fast (Static Tier)` | 62.0% / 68.0% | 42.0% [28, 56] | 44.2+-3.0% (23.3%) | 9.70s / 27.07s [19.7, 35.0] | 16.05s |
| `Entropy-Cascade` | 68.0% / 84.0% | 58.0% [44, 72] | 57.4+-2.3% (43.3%) | 25.12s / 45.06s [40.5, 47.9] | 28.81s |
| `Sync_N_of_N` | 66.0% / 68.0% | 42.0% [28, 56] | 42.6+-1.9% (23.3%) | 35.64s / 46.83s [43.4, 53.6] | 30.19s |
| `Genealogy_Sync_HPR` | 84.0% / 86.0% | 72.0% [58, 84] | 72.0+-1.6% (63.3%) | 36.34s / 47.39s [43.7, 53.6] | 30.19s |
| **`QuorumGuard`** | **84.0% / 86.0%** | **72.0% [58, 84]** | **72.0+-1.6% (63.3%)** | **10.18s / 15.09s [13.8, 16.1]** | **13.24s** |

![Pareto Frontier](figures/fig3_pareto_latency_accuracy.png)

## Citation

```bibtex
@misc{syed2026quorumbench,
  author       = {Syed, Mohammed Kamran and Rentala, Anudeep and Rahul, Somesh and Pai, Raghuram Malpe and Dara, Krishna Suman},
  title        = {{QuorumBench}: Asynchronous Multi-Agent {LLM} Consensus and Straggler Herding Benchmark},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/batistado/QuorumBench}}
}
```