Pack of LLMs: Model Fusion at Test-Time via Perplexity Optimization

Summary

This paper provides an alternative method for combining the responses from multiple LLMs. Weights were introduced to be used to re-weight the output logits. To determine the criterion for obtaining the optimal weights, the authors first discovered that the perplexity on the given prompts can be already served as a good approach to fuse the outputs of models. Next, the authors propose an greedy optimization-based algorithm that effectively fuses the outputs from different models. Experimental results show that the proposed algorithm can merge the outputs from models better than baseline methods such as ensemble and using only the model with lowest perplexity.

Rating

7

Confidence

3

Ethics flag

1

Reasons to accept

- The motivation is clearly described and the topic is interesting. - The empirical performance seems to be strong. - The techniques introduced in this paper are technically sound.

Reasons to reject

- I think the authors may also want to compare with always picking the responses from the largest model, because selecting the model with the lowest perplexity does not mean that model can generalize well. Selecting always the largest model may yield better performance. - I would also like to see how this model compared with model merging. I understand it is not always possible to merge models, especially when the number of parameters and the structures are diverse across models. However, I think the authors can at least find several publicly available models fine-tuned from llama-7b on different datasets and see if the proposed algorithm can outperform model merging.

Summary

This paper presents a test-time LLM fusion method to leverage knowledge from the user-specified LLMs during inference. Specifically, the authors optimize the perplexity over the input prompt to perform model fusion. The experimental results demonstrate the effectiveness of the proposed method.

Rating

5

Confidence

4

Ethics flag

1

Reasons to accept

This work presents a simple but effective test-time LLM fusing method and provides valuable insights. The paper is well-written and clear. The authors conducted experiments on multiple datasets, illustrating that this method generalizes well to different domains.

Reasons to reject

In the main experimental results reported in Table 1, the proposed method outperforms the baseline models. However, it is important to note that the experiments involving some test-time fusion methods, such as cBTM and DExperts, are not fully reported. This lack of complete data undermines the reliability of these results. This issue is similar to the one observed in Table 2, where the comparison with learning-based fusion models becomes less convincing due to incomplete experimental data.

Questions to authors

- Is it possible to complete the results of the test-time baseline experiment? - It is advised to provide some case studies to show that the proposed ppl-based fusion method is superior to other methods.

Summary

This paper proposes PackLLMs, which utilizes the findings that perplexity of input tokens is a good indicator of LLM's performance in domains. The authors propose PackLLM-simp that utilize the heuristics above and PackLLM-Opt, which approximates an optimization objective with a greedy search algorithm. Experiments show improvement over baselines such as cBTM, DExperts. The approach is further modular and easy to use compared to other fusion methods that requires training LLMs.

Rating

7

Confidence

3

Ethics flag

1

Reasons to accept

- The proposed approach is simple and effective, and is broadly applicable to any combinations of LLMs with either same or different architectures. - The choice of baselines are recent and comprehensive. The results also show clear improvements of LLM. - The analysis of coefficients obtained (\lambda), and relationships between perplexity and prompt length in Section 6 are quite helpful. - The writing is quite clear.

Reasons to reject

- The authors should include discussions about the computational overhead of the approach. How much overhead is introduced by the grid search part, given the high dimensionality of logits (sequence length * vocab size for a single instance)? Is this overhead significant compared to a forward pass of the LM? I also encourage the authors to add a limitations sections to discuss potential limitations of the approach.

Questions to authors

I think the paper is quite solid and has addressed my initial questions in their later sections. A minor comment is that the second paragraph of the related works is repetitive given the second paragraph in introduction.

Summary

This paper introduces a new method, PackLLM, to fuse LLMs at test-time as weighted ensemble, where the importance weights for seed LLMs are determined by minimizing the perplexity of the input prompt. There are two variants of PackLLM, one that directly uses normalized perplexity scores of individual LLMs as weights, and the other that performs an efficient greedy algorithm to search the optimal scores that minimize the prompt perplexity of the ensemble. The paper conducts extensive experiments to validate the performance of PackLLM. On pretraining perplexity, both variants of PackLLM outperform existing test-time model ensemble baselines like cBTM and DExperts, especially when the number of seed LLMs increases. On downstream tasks, PackLLM also outperforms simple baselines like uniform ensemble. While it trails other learning-based fusion methods like FuseLLM and FoE with the same set of seed LLMs, the learning-free nature allows PackLLM to achieve better performance by swapping in stronger seed LLMs. The paper also includes some analysis of the weight values, the impact of prompt length, and the time complexity of the two variants.

Rating

7

Confidence

3

Ethics flag

1

Reasons to accept

- Proposes a simple yet effective test-time fusion method for LLMs - Comprehensive empirical study by comparing PackLLM with different baselines across various settings - Well-written, clearly articulated motivation

Reasons to reject

I don’t have major concerns. I do have some questions listed below.

Questions to authors

- I’m curious to know the impact of aligning different tokenizers. Have you done any analysis comparing the performance when fusing similar-sized models with the same vs. different vocabularies? - In experimental results, I’d like to also see a comparison with the best seed LLM performance on each task, especially for Table 4, where most of the baseline results are not available. - While not an exhaustive search, the greedy search algorithm still seems a very expensive operation that needs to be done for every test-time input. Can we situate the time complexity analysis in Table 3 by comparing the additional time cost from greedy search with the actual inference time of LLMs? This would be important to help understand the real-world applicability of the proposed method.