TL;DR
Evaluating large language models (LLMs) during training is computationally expensive, especially for generation tasks. This paper proposes reformulating these costly generation tasks into cheaper Natural Language Understanding (NLU) formats (multiple-choice or log-likelihood) to enable more frequent and efficient monitoring of LLM capabilities. The authors demonstrate a strong correlation between the original and reformulated tasks, achieving significant speedups (over 35x) without sacrificing the ability to assess model performance.
What Problem Does This Paper Address?
Training large language models (LLMs) involves iterative evaluation to track their development. While Natural Language Understanding (NLU) tasks, which involve selecting from predefined options, are computationally cheap to evaluate, Natural Language Generation (NLG) tasks, which require generating text token-by-token, are significantly more time- and compute-intensive. This costliness often leads to NLG capabilities, such as reasoning and code generation, being neglected during the crucial pre-training and model selection phases. The paper aims to bridge this gap by providing a more efficient way to evaluate these critical NLG capabilities.
Methods
The core of this work is the reformulation of NLG tasks into computationally cheaper NLU alternatives. Specifically, they explore two main approaches:
- Multiple Choice (MC) tasks: The model is presented with a question and a set of answer options (one correct, multiple incorrect) and must select the most probable answer. To create these incorrect options (distractors), they employ two strategies:
- Random distractors: Answers are randomly sampled from other questions within the same benchmark.
- Smart distractors: Distractors are generated using LLMs to match the semantic context of the question, making them more plausible.
- Log-likelihood (LL) tasks: The model calculates the probability (log-likelihood) of the correct answer being generated, without needing incorrect options.
They also investigate "cross-benchmark pairing," where existing NLU benchmarks are used to evaluate capabilities that were originally assessed by NLG benchmarks in different domains.
Datasets
The paper evaluates their approach on several established benchmarks across different domains:
- Mathematical Reasoning: GSM8K
- Code Generation: HumanEvalPack
- Factual Knowledge: TriviaQA
- Reading Comprehension: SQuAD 2.0
For cross-benchmark pairings, they use:
- Factual Knowledge: MMLU
- Reading Comprehension: BoolQ
Key Insights and Ideas
- NLU as a Proxy for NLG: The fundamental idea that NLU tasks can serve as effective proxies for monitoring NLG capabilities during training is strongly supported by their findings. This is a significant shift in how LLM evaluation can be approached.
- Strong Correlation Across Formats: The paper consistently finds a high correlation between the original NLG tasks and their NLU reformulations (both MC and LL), across various models and capabilities. This suggests that the relative performance trends observed in NLU tasks accurately reflect those in NLG tasks.
- MC vs. LL Trade-offs: While both MC and LL reformulations offer significant speedups, LL is slightly more efficient as it doesn't require generating distractors. However, MC tasks, especially with smart distractors, can sometimes offer a more nuanced evaluation.
- Random vs. Smart Distractors: Interestingly, random distractors often perform on par with smart distractors, especially for larger models. This suggests that simpler methods for generating distractors can be sufficient, reducing the overhead of LLM-based generation.
- Predicting NLG Performance: The study shows that it's possible to predict NLG performance using a few NLG evaluations combined with NLU performance scores, further reducing the need for frequent, expensive NLG evaluations.
Key Results
- Significant Speedup: The proposed NLU reformulations achieve an average reduction in evaluation time of over 35x, with some tasks seeing reductions from nearly 2 hours to under a minute.
- High Correlation: Pearson correlation coefficients between NLG and NLU variants are consistently high (often above 0.8), indicating that the NLU tasks effectively capture the same performance trends as NLG tasks. Spearman correlation also shows strong agreement in model rankings.
- MC vs. MCrnd Performance: MC and MCrnd (random distractors) perform similarly, with MCrnd showing a slight advantage for smaller models, possibly because less capable models are more easily misled by implausible random answers.
- LL Performance: LL formulation also shows strong correlations, though slightly lower on average than MC for some tasks. It offers the advantage of not requiring distractor generation.
- Cross-Benchmark Pairing Success: Using MMLU as an MC pairing for TriviaQA yielded strong correlations, even outperforming their custom MC version of TriviaQA.
Limitations and Future Work
- Model Size: The study primarily focused on small to medium-sized models (up to 7B parameters). The authors acknowledge that larger models might behave differently, and their results may not directly translate. Future work could explore these larger models.
- Safety and Ethics: The current work did not include safety-related domains or ethical evaluation benchmarks, as the models used were considered too small to meaningfully handle such complex tasks. Future work could explore task reformulation for these domains, especially with instruction-tuned models.
- Distractor Difficulty: For some tasks, like MC versions of GSM8K, the performance drop when removing reasoning steps suggests that more difficult distractor options might be needed for MC tasks to be truly challenging.
- Generalizability: While the study covered four diverse domains, exploring a wider range of capabilities and benchmarks would further validate the approach.
Summary updated