Large Language Models are Demonstration Pre-Selectors for Themselves

We thank Reviewer mKZf for recognizing our novelty, acknowledging its efficiency and performance improvement, appreciating our comprehensive evaluations on multiple LLMs and tasks, and noting its compatibility with existing demonstration selection strategies.

Q1. More comparison beyound basic selection methods (Random, Similarity, and Diversity)

R1. Thank you for your issue. We have compared FEEDER with advanced ICL techniques, specifically including the uncertainty-based, the clustering-based, and the recent latent variable-based demonstration selection method proposed in [1]. This method is explicitly denoted as "Latent" in our paper (Section 5.1, line 317-319, Section A5.1, line 1186-1188 left).

Corresponding experimental results are reported in Appendix Tables A2, A4, and A6, clearly demonstrating FEEDER’s performance relative to this advanced baseline.

To further address your suggestion, we will highlight this comparison more explicitly in revision and include additional similarity-based methods.

[1] Large Language Models Are Latent Variable Models: Explaining and Finding Good Demonstrations for In-Context Learning. NeurIPS 2023

Q2. While FEEDER shows performance improvements in most cases, the reported gains are not particularly significant.

We have added a significance test for the main results on the GSM8K dataset using the Llama-3 8B model in the ICL setting, as shown below. Results for other datasets and LLMs will be included in our revision.

Table R2. * indicates p < 0.05 in significance tests compared to using TRAIN.

Q3. Larger models (>8B)

R3. We add Qwen 2.5 32B model as our base model on GSM8K and GPQA. Please refer to the R1 in response of Reviewer oHZY. Results for the other benchmarks and LLMs will be included in the revision.

Q4. Whether these improvements justify the additional computational complexity.

Thanks for your question. We have included the runtime of running FEEDER in Figure A6. Running FEEDER only requires pre-running the inference over half of the number of training data points (33.45s for GPT-neo 1.3B on COLA dataset).

For the fine-tuning, it costs 867.93s for training GPT-neo 1.3B on COLA dataset for one-epoch on the entire data points, whereas training GPT-neo 1.3B on COLA dataset on our FEEDER data costs 707.79s. As training is significantly more costly than inference, performing pre-selection is always cost-effective if around 10% can be filtered out.

Furthermore, we also would like to emphaize that our FEEDER operates at the pre-selection stage which allows for pre-computation over the entire training dataset, and serves various test data.

Q5. The concepts of sufficiency and necessity are somewhat abstract, and the paper could benefit from more intuitive, visual examples to help readers understand their impact on demonstration selection

R5. We have provided detailed descriptions of sufficiency and necessity, along with examples, in Appendix A2. Additionally, we also presented a case study in A11.2 to illustrate the impact of sufficiency and necessity on LLM performance.

Q6. Does FEEDER maintain its effectiveness and selection quality degrade as dataset sizes increase?

R6. Thanks for your question. To examine the impact of dataset size, we evaluate the effect of randomly reducing the SUBJ dataset by 20%, 40%, 60%, and 80% using the Similarity demonstration selector. Results are summarized below.

Table R3.

The results above demonstrate that our FEEDER is effective across different dataset scales.

Please let us know if you have further questions -- thank you so much!

We thank Reviewer S3Tb for recognizing our novelty, solid theoretical contributions and good writing. Below, we respond to each of your questions in detail.

Q1. The paper is not contextualized well within the active learning literature.

R1. We clarify the relationship between our FEEDER method and active learning research line as follows:

1. Similarity: Both FEEDER and active learning methods focus on selecting representative and informative samples.
2. Difference: Active learning typically selects samples for annotation during training, while FEEDER selects a representative subset from already labeled data to reduce computation.
3. Our novelty: FEEDER proposes sufficiency and necessity metrics tailored specifically to LLMs, efficiently identifying critical demonstrations and removing redundancy.

We listed some related active learning works discussed some of them in Appendix A1.3, we will highlight it and add more discussion in revision.

Q2. Performance of the approach is only benchmarked on text classification tasks.

R2.

1. As mentioned in our experiment section, in addition to 6 text classification tasks, we have included the reasoning task GSM8K and the semantic parsing task SMCALFlow, with corresponding results shown in Tables 2 and A6.
2. Furthermore, we add the science QA task GPQA in our revision. Partial results with the Qwen2.5 32 model are available in Table R1 below. These results further demonstrates FEEDER's ability to enhance LLM performance in math and science tasks.

Table R1. Performance comparisons on GSM8K and GPQA datasets are conducted in the ICL setting. * indicates p < 0.05 in significance tests compared to using TRAIN.

Q3. No runtime numbers. How do runtimes compare to fine-tuning on the whole dataset?

R3.

Thanks for your question. We have included the runtime of running FEEDER in Figure A6. Running FEEDER only requires pre-running the inference over half of number of training data points (33.45s for GPT-neo 1.3B on COLA dataset).

For the fine-tuning, it costs 867.93s for training GPT-neo 1.3B on COLA dataset for one-epoch on the entire data points, whereas training GPT-neo 1.3B on COLA dataset on our FEEDER data costs 707.79s. As training is significantly more costly than inference, performing pre-selection is always cost-effective if around 10% can be filtered out. Furthermore, we also would like to emphaize that our FEEDER operates at the pre-selection stage which allows for pre-computation over the entire training dataset, and serves various test data.

Q4. Do the authors know why COLA+LAG+2 performance is worse with an increasing number of rounds?

R4. We apologize for mixing the results of different demonstration retrievers when drawing this subfigure in Figure 3. The corrected results for COLA+LAG+2 should be 0.389 (for #Round = 0), 0.412 (for #Round = 1), 0.438 (for #Round = 5), and 0.430 (for #Round = 10), which follow the same trend as other settings.

Please let us know if we have properly addressed your questions and we are more than happy to discuss more!

We thank Reviewer s2sv for recognizing our extensive experiments across multiple benchmarks and LLMs, showing consistent improvements of our framework.

Below, we respond to each of your questions in detail.

Q1.1 The differences discussion between this submission and related work [1]. Some overlap in the selection of datasets and LLMs for the experiments.

R1.1 Thanks for pointing it out. Our approach differs significantly from theirs in both theoretical analysis and experimental implementation.

1. Theoretically, their work investigates the causal relationship between the input X and the output Y, whereas ours analyzes the causal relationship between the input X and the selected demonstration C.
2. Empricially, their method operates at the demonstration selection stage, while our approach introduces a pre-selection stage before demonstration selection. This allows us to pre-compute a core subset of the training dataset as the demonstration selection pool.
3. Furthermore, their approach focuses on the demonstration selection stage and requires fine-tuning the embeddings of latent variables, which limits its applicability to smaller LLMs. In contrast, our method, which operates at the pre-selection stage, does not require fine-tuning LLMs.

Q1.2 No direct comparison and more detailed discussion between the two methods (FEEDER and [1]) regarding their effectiveness for ICL.

R1.2 We directly compared FEEDER with [1] in our experiments, we have included [1] as one of the baseline demonstration selectors, note that we denoted method [1] as "Latent" (Section 5.1 line 317-319, Section A5.1. line 1186-1188 left). We provided the corresponding results in Appendix Tables A1&A2, A3&A4, 2&A6, main findings of effectiveness are listed below:

1. Comparing FEEDER (ours) + SUBJ (Similarity) against TRAIN + SUBJ (Latent [1]), we observed that applying a similarity-based demonstration selector on our FEEDER-generated core subset outperforms their advanced demonstration selector applied to the entire dataset. One possible explanation is that their method does not explicitly consider the relationships among candidate demonstrations.
2. Comparing FEEDER (ours) + SUBJ (Latent [1]) with FEEDER [1] + SUBJ (Similarity), we find that our method also improves LLM performance. However, the improvements brought by our method and theirs would have some overlap, as both approaches condition the selected demonstrations on the LLMs being used.

Q2. Clear writing sugguestion: In Table 1, COLA results on Gemma-2, the 1-shot results under Similarity and Diversity should highlight the D_train values in bold.

R2. Thanks for your suggestion. We will address these issues in our revision.

Reference:

[1] Large Language Models Are Latent Variable Models: Explaining and Finding Good Demonstrations for In-Context Learning. NeurIPS 2023

Please let us know if you have further questions -- thank you so much!

We would like to thank Reviewer oHZY for recognizing our method is efficient and practical for real-world deployment, as well as we proposing an interpretable way to explain the demonstration.

Below, we respond to each of your questions in detail.

Q1. Should consider larger LLM.

R1. We add Qwen 2.5 32B model as our base model. The corresponding results on GSM8K and GPQA are listed below. Results for the other benchmarks and LLMs will be included in the revision.

Table R1. Performance comparisons on GSM8K and GPQA datasets are conducted in the ICL setting. * indicates p < 0.05 in significance tests compared to using TRAIN.

These results demonstrate FEEDER can be extended to LLMs larger than 8B.

Q2. Paper writing issue: lot of blank spaces between sections.

R2. Thanks for your suggestion. We will address these issues in our revision.

Q3. Lack of Evaluation of Tasks. Authors should consider ICL methods on variety tasks, now only consider 6 text classification tasks.

R3. As mentioned in our experiment section, in addition to 6 text classification tasks, we have included the reasoning task GSM8K and the semantic parsing task SMCALFlow, with corresponding results shown in Tables 2 and A6 in the paper.

Furthermore, we add the science QA task GPQA in our revision. Partial results with the Qwen2.5 32B model are available in Table R1 above, which further demonstrates FEEDER's ability to enhance LLM performance in math and science tasks.

We are eager to hear your feedback. We’d deeply appreciate it if you could let us know whether your concerns have been addressed.