Loopserv Eval

loopserv-eval

LoopServe: An Adaptive Dual-phase LLM Inference Acceleration System for Multi-Turn Dialogues — Li et al. (2025) (arXiv:2507.13681, 2025)

What this evaluates

Evaluates an adaptive dual-phase LLM inference acceleration system for multi-turn dialogues, probing its ability to maintain generation accuracy and reduce computational overhead across varying query positions in long-context conversations. It specifically tests whether the system can generalize beyond positional heuristics used by static KV cache compression methods.

Datasets

• MFQA-en — total ?; splits: test (-1)
• 2WikiMQA — total ?; splits: test (-1)
• Musique — total ?; splits: test (-1)
• HotpotQA — total ?; splits: test (-1)
• NrtvQA — total ?; splits: test (-1)
• Qasper — total ?; splits: test (-1)
• MultiNews — total ?; splits: test (-1)
• GovReport — total ?; splits: test (-1)
• QMSum — total ?; splits: test (-1)
• TREC — total ?; splits: test (-1)
• SAMSUM — total ?; splits: test (-1)

Metrics

• Accuracy (primary) — range: [0, 1]
  • Exact match ratio: fraction of predictions that exactly match the ground truth label or answer. Standard for QA and Few-shot classification tasks.
• F1 — range: [0, 1]
  • Harmonic mean of precision and recall, typically computed as macro or micro average depending on the dataset. Used for QA tasks where partial matches or entity overlap matter.
• Rouge-L — range: [0, 1]
  • Recall-Oriented Understudy for Gisting Evaluation (Longest Common Subsequence). Measures the longest matching subsequence between generated summary and reference summary, normalized by reference length.

Input / output format

Input: Multi-turn dialogue context containing multiple rounds with diverse query positions (beginning, middle, or end) and dependencies. Covers Question Answering, Summarization, and Few-shot Learning tasks.

Output: Model-generated text response or class label corresponding to the target query in the dialogue turn.

Scoring recipe

def evaluate(predictions, golds, task_type):
    if task_type == 'QA':
        return sum(1 for p, g in zip(predictions, golds) if p == g) / len(golds)
    elif task_type == 'Summarization':
        return rouge_l_score(predictions, golds)
    elif task_type == 'FewShot':
        return sum(1 for p, g in zip(predictions, golds) if p == g) / len(golds)
    return 0.0

Common pitfalls

• Assuming KV cache baselines generalize well across all query positions; they often degrade significantly when queries are at the beginning or middle due to reliance on positional heuristics.
• Setting the token budget B or sparsity threshold α too aggressively without checking task-specific sensitivity, which can cause disproportionate accuracy drops without meaningful efficiency gains.

Evidence (verbatim from paper)

For each dataset, we compare LoopServe with six state-of-the-art KV cache acceleration baselines and two base LLMs, using F1, Rouge-L, or Accuracy as appropriate. As shown in Table 1, LoopServe achieves the best or comparable results across most datasets and query positions.

Citation

@misc{li2025loopserv,
  title={LoopServe: An Adaptive Dual-phase LLM Inference Acceleration System for Multi-Turn Dialogues},
  author={Li et al. (2025)},
  year={2025},
  note={arXiv:2507.13681}
}

• arXiv: 2507.13681