Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators

By Yann Dubois, Bal'azs Galambosi, Percy Liang, Tatsunori B. Hashimoto. 2024

873 citations recorded by the source corpus.

LLM-based auto-annotators have become a key component of the LLM development process due to their cost-effectiveness and scalability compared to human-based evaluation. However, these auto-annotators can introduce biases that are hard to remove. Even simple, known confounders such as preference for longer outputs remain in existing automated evaluation metrics. We propose a simple regression analysis approach for controlling biases in auto-evaluations. As a real case study, we focus on reducing the length bias of AlpacaEval, a fast and affordable benchmark for instruction-tuned LLMs that uses LLMs to estimate response quality. Despite being highly correlated with human preferences, AlpacaEval is known to favor models that generate longer outputs. We introduce a length-controlled AlpacaEval that aims to answer the counterfactual question:"What would the preference be if the model's and baseline's output had the same length?"To achieve this, we first fit a generalized linear model to predict the biased auto-annotator's preferences based on the mediators we want to control for (length difference) and other relevant features. We then obtain length-controlled preferences by predicting preferences while conditioning the GLM with a zero difference in lengths. Length-controlling not only improves the robustness of the metric to manipulations in model verbosity, but we also find that it increases the Spearman correlation with LMSYS Chatbot Arena from 0.94 to 0.98.

Subfields: Artificial Intelligence, Large Language Models, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Large Language Models

In Evaluation & Benchmarks

Citations in the corpus

From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline (2024, methodology): “1’s separability and alignment with humans against other widely used benchmarks, MT-Bench [51] and AlpacaEval 2.0 Length Controlled [13], we obtain 95% confidence interval of model performances via applying 100 rounds of bootstrapping on judgment results for each benchmark.”
LiveBench: A Challenging, Contamination-Free LLM Benchmark (2024, background): “Furthermore, LLMs often favor their own answers over other LLMs, and LLMs favor more verbose answers [18, 35, 36].”
SimPO: Simple Preference Optimization with a Reference-Free Reward (2024, positive): “The performance advantage is consistent across various training setups and extensive chat-based evaluations, including AlpacaEval 2 [51, 28] and the challenging Arena-Hard [50] benchmark.”
Mixture-of-Agents Enhances Large Language Model Capabilities (2024, background): “Figure 1 showcases the LC win rate on the AlpacaEval 2.0 benchmark (Dubois et al., 2024) for 6 popular LLMs. 1 Our code can be found in: https://github.com/togethercomputer/moa .”
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model (2024, methodology): “Notably, DeepSeek-V2 Chat (RL) achieves 38.9 length-controlled win rate on AlpacaEval 2.0 (Dubois et al., 2024), 8.97 overall score on MT-Bench (Zheng et al., 2023), and 7.91 overall score on AlignBench (Liu et al., 2023).”
From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge (2024, positive): “Following them, Wu et al. (2024) introduce meta-rewarding that judges the LLMs’ judge and uses the feedback to refine their judging skills.”
DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence (2024, neutral): “On the other hand, DeepSeek-V2 Chat demonstrates slightly better results in benchmarks such as MT-bench (Zheng et al., 2023), AlpacaEval 2.0 (Dubois et al., 2024), and AlignBench (Liu et al., 2023c).”
TÜLU 3: Pushing Frontiers in Open Language Model Post-Training (2024, positive): “AlpacaEval 2 [Dubois et al., 2024] contains a set of prompts sourced that reflect real human usages of LMs, and compares model outputs to GPT-4 turbo responses with an additional length control to avoid longer answers being unfairly favoured.”
Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing (2024, background): “The second metric is the length-controlled win rate (LC) [18], a debiased version of WR.”
WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild (2024, methodology): “Previous studies have shown that LLM judges tend to prefer longer responses [6].”
S-SPPO: Semantic-Calibrated Self-Play Preference Optimization (2026, unknown): “To measure instruction-following proficiency in realistic user interaction scenarios, we report win rates and length-controlled win rates on AlpacaEval 2.0 (Dubois et al., 2024).”
Better Instruction-Following Through Minimum Bayes Risk (2024, unknown): “Recent works find that “LLM-as-a-Judge” frameworks (Zheng et al., 2023; Li et al., 2023; Dubois et al., 2024a) demonstrate strong correlation with human evaluations and significantly outperform lexical (Lin, 2004b; Papineni et al., 2002) and embedding-based methods (Zhang et al.; Yuan et al., 2021;…”