batistado's picture
Publish CoherenceBench-Agent v1.0 multi-model & multi-agent telemetry
2c7461b verified
|
Raw History Blame Contribute Delete
4.77 kB
metadata
license: apache-2.0
pretty_name: >-
  CoherenceBench-Agent: Multi-Turn Mutable-Workspace KV-Cache Coherence
  Benchmark
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - kv-cache
  - cache-coherence
  - llm-agents
  - vllm
  - lmcache
  - rope
  - surgical-invalidation
size_categories:
  - n<10K
models:
  - Qwen/Qwen2.5-3B-Instruct
  - Qwen/Qwen2.5-7B-Instruct
  - Qwen/Qwen2.5-Coder-14B-Instruct
configs:
  - config_name: trajectories
    default: true
    data_files:
      - split: train
        path: data/trajectories/train.jsonl
  - config_name: tasks
    data_files:
      - split: train
        path: data/tasks/train.jsonl
  - config_name: block_taint_dag
    data_files:
      - split: train
        path: data/block_taint_dag/train.jsonl
  - config_name: tau_pareto
    data_files:
      - split: train
        path: data/tau_pareto/train.jsonl
  - config_name: multi_agent_leases
    data_files:
      - split: train
        path: data/multi_agent_leases/train.jsonl
  - config_name: long_horizon_scaling
    data_files:
      - split: train
        path: data/long_horizon_scaling/train.jsonl

CoherenceBench-Agent v1.0

Official Benchmark Suite & Multi-Model Hardware Telemetry for MESI-KV
Interactive Explorer Space: https://huggingface.co/spaces/batistado/MESI-KV-Explorer | Hub System Repo: https://huggingface.co/batistado/MESI-KV | GitHub: https://github.com/batistado/mesikv

CoherenceBench-Agent evaluates cross-layer KV-cache and semantic tool-cache coherence when external workspaces mutate mid-trajectory across 24 full-factorial 15–16-turn agent scenarios (plus 8.5K and 16.5K long-horizon extensions) spanning Python AST modules, Microservice Config manifests, and Live SQLite operational tables with row-version triggers, covering 1-hop and 2-hop transitive chains, Latent Symbol-Bridge vs. Explicit CoT Scratchpads, and Same-Length (Delta = 0) vs. Variable-Length (Delta != 0) RoPE mutations.

Architecture Overview

Verified Headline Summary (tau_kv = 0.12, Order-Statistic Calibrated)

Model Leyline Oracle Match PatchKV Oracle Match MESI-KV Oracle Match MESI-KV Downstream KV Saved (Tok / Wall)
Qwen/Qwen2.5-Coder-14B-Instruct (N=24) 0.00% 0.00% 100.00% 73.95% (3.41x tok / 1.94x wall)
Qwen/Qwen2.5-3B-Instruct (N=24) 16.67% 16.67% 100.00% 89.14% (6.86x tok / 1.72x wall)
Qwen/Qwen2.5-7B-Instruct (N=24) 16.67% 16.67% 91.67% 84.45% (5.23x tok / 1.66x wall)

Load Any Dataset Subset via datasets

from datasets import load_dataset

# 1. Per-scenario x per-method accuracy, KV savings, Frobenius drift, and KL divergence
trajectories = load_dataset("batistado/CoherenceBench-Agent", "trajectories", split="train")

# 2. 24 full-factorial mutable-workspace tasks + ground-truth pre/post-mutation answers
tasks = load_dataset("batistado/CoherenceBench-Agent", "tasks", split="train")

# 3. Turn-by-turn direct & transitive attention-taint DAG profiles
taint = load_dataset("batistado/CoherenceBench-Agent", "block_taint_dag", split="train")

# 4. Accuracy-vs-KV-savings Pareto curves (Order-Stat calibrated vs. Raw Max)
pareto = load_dataset("batistado/CoherenceBench-Agent", "tau_pareto", split="train")

# 5. Concurrent multi-agent (P=1,2,4,8) lease safety, collateral isolation & us latency
leases = load_dataset("batistado/CoherenceBench-Agent", "multi_agent_leases", split="train")

Preprint & Artifacts

Citation

@article{syed2026mesikv,
  title   = {{MESI-KV}: Cross-Layer Directory Coherence and Attention-Taint-Bounded Invalidation Across Mutable Workspaces, Semantic Tool Caches, and Disaggregated {KV} Pools},
  author  = {Syed, Mohammed Kamran and Rentala, Anudeep and Rahul, Somesh and Pai, Raghuram Malpe and Dara, Krishna Suman},
  journal = {IEEE Transactions on Parallel and Distributed Systems (TPDS)},
  year    = {2026},
  url     = {https://github.com/batistado/mesikv}
}