qhjqhj00 omni-dpo-eval
Omni-DPO: A Dual-Perspective Paradigm for Dynamic Preference Learning of LLMs — Peng et al. (2025) (arXiv:2506.10054, 2025)
What this evaluates
Evaluates the instruction-following and mathematical reasoning capabilities of LLMs fine-tuned with a dual-perspective preference optimization method. It measures conversational quality, adherence to instructions, and problem-solving accuracy across diverse open-ended and quantitative benchmarks.
Datasets
- AlpacaEval 2.0 — total ?; splits: test (-1)
- Arena-Hard v0.1 — total ?; splits: test (-1)
- IFEval — total ?; splits: test (-1)
- SedarEval — total ?; splits: test (-1)
- GSM8K — total ?; splits: test (-1)
- MATH 500 — total ?; splits: test (-1)
- AIME 2024 — total ?; splits: test (-1)
- AMC 2023 — total ?; splits: test (-1)
Metrics
LC(%)(primary) — range: percent- Percentage of wins against a reference model, adjusted for response length to mitigate length bias.
WR(%)— range: percent- Percentage of times the model's response is preferred over a baseline response by an LLM judge.
Accuracy (Acc.)— range: percent- Percentage of correctly solved problems, reported as Strict, Loose, or Overall depending on the benchmark.
Input / output format
Input: Open-ended instruction prompts or mathematical problems. For math tasks, zero-shot chain-of-thought prompting is applied.
Output: Natural language responses for instruction tasks; step-by-step reasoning followed by a final answer for math tasks.
Scoring recipe
def score(predictions, golds, metric_type):
if metric_type in ['LC', 'WR']:
# Pairwise LLM-as-a-judge comparison against baseline
wins = count_wins(predictions, baselines)
return (wins / len(predictions)) * 100
elif metric_type == 'Acc':
correct = sum(1 for pred, gold in zip(predictions, golds) if exact_match(pred, gold))
return (correct / len(predictions)) * 100
Common pitfalls
- Win rates are heavily biased by response length; always report Length-Controlled Win Rate (LC) for fair comparison.
- Different benchmarks use different LLM judges or evaluation scripts (e.g., AlpacaEval vs. Arena-Hard), so results are not directly comparable across benchmarks.
- Math benchmarks require consistent zero-shot CoT prompting and greedy decoding; varying decoding strategies will change accuracy scores.
Evidence (verbatim from paper)
We primarily evaluate our method on four widely adopted open-ended instruction-following benchmarks: AlpacaEval 2.0, Arena-Hard v0.1, IFEval, and SedarEval. ... Table 1: Main result of textual understanding. WR denotes the Win Rate, LC denotes the Length-Controlled win rate, and Acc. denotes the Accuracy.
Citation
@misc{peng2025omnidpo,
title={Omni-DPO: A Dual-Perspective Paradigm for Dynamic Preference Learning of LLMs},
author={Peng et al. (2025)},
year={2025},
note={arXiv:2506.10054}
}
- arXiv: 2506.10054