Calibrate to Discriminate:Improve In-Context Learning with Label-Free Comparative Inference

We thank reviewer for carefully reading through our paper. We have addressed all the feedbacks as much as we can but there are some confusions that may lead to the misinterpretation of this paper's major contributions. It looks like review is mainly focusing on the 4.2 (post-hoc section) and our results (e.g. Table 2) but we'd like to point out the most highlighted parts are 4.1 and corresponding results table 1 and figure 1,2. Please check our comments below and the revised paper.

For Weaknesses:

1. Key Related Works Missing: The statement (“While such methods can decrease ECE, it does not help with improving model classification performance,”) is only for the temperature scaling method (as claimed by the original paper, that it does not change classification accuracy). Almost all other post-hoc calibration methods will have impacts on both accuracy and ECE. We emphasized the temperature scaling here, because the approach is very popular and easy to use (E.g. many LLM APIs provide temperature changes). To avoid the confusion, we made this statement clear in our revision. Now it’s: “While temperature scaling can decrease ECE, it does not help with improving model classification performance”
2. Limited baselines: This is where we believe the major confusion happened. We acknowledge that these papers are relevant to discuss but we don’t think it’s necessary to compare them in experiments as they are all post-hoc calibration only (not prompt level approach). We think there are some priorities that we want to restate. Our contributions are mainly three fold: i). identify indiscriminate miscalibration (Figure 1a) in LLMs ii). a prompt level label free method (main results in Table 1) iii). a post-hoc calibration method based on logits (main results in Table 2) i)&ii) with Figure 1 and Table 1 are the more novel&important findings. iii). is inspired/built on top of existing post-hoc calibration methods as we claimed on line 366-367.. We are including iii). as a section here mainly because ii). naturally brought a LF post-hoc method that is new to the field. For the 4 papers that reviewer mentioned, they are all post-hoc calibration, not prompt level based method. Outputs of our method ii), in theory, can be further calibrated with any of these methods or ours (iii) for post-hoc calibration. However, only ICL itself is a prompt based method (comparable with ii). When combining them, the results are simply expected to be better (similar to combining ICL+ours, ours+agg). We now added these papers and make it clear at the end of section 4.2.
3. Limited evaluation: There are some capacity constraints. We listed this in the section 7.
4. Unfair comparison: We think we can agree that the cost rank is: 3X + 3Y>3X>0. We are not calling 2 input samples 2-shot because there’s no input-answer information provided and shouldn’t be referred to as few-shot. Few shot prompting is mostly referred to when answers are provided (e.g. https://arxiv.org/pdf/2301.00234). And more importantly, if you compare LF-10-shot, with regular 10-shot, then yes, additional unlabeled samples are the extra cost. But compared with labels, these are much easier to access. For example, you can use other training samples. On the other hand, LF-ICL 0-shot also outperforms or on-par with regular ICL 3-shot (e.g. F1 for Table 1 L2-2B, L3-8B), and LF-ICL 3-shot also outperforms or on-par regular 10-shot (e.g. F1 for Table 1 L3-8B, L3-70B) as well. For ECE, LF-ICL generally outperforms regular few-shot (with more demonstrations).
5. We addressed all the suboptimal writing in the revised paper.

For Questions:

1. Figure 1 uses simulated data for illustrating the indiscriminate calibration and regular miscalibration with different probabilities patterns (similar probabilities across different samples in 1(a) and different probabilities across different samples in 1 (b)). Experiment setting uses the real world datasets with actual LLM inference.
2. For With-labels version: Can you clarify this question? For LF-0-shot, if adding labels for 2 additional input samples, that’s gonna be like LF-2-shot.

1. Key Related Works Missing and 2. Limited baselines

I have gone through the paper thoroughly and I do acknowledge the main contribution in Table 1 and Figures 1, 2, that includes identifying indiscriminate miscalibration, a label free method, and the pos-hoc method using logits. However, I am still not convinced that the authors did not use the mentioned papers as baselines. While these papers are post-hoc methods, they even compare with the raw prompt based ICL method. If authors claim that ‘When combining them, the results are simply expected to be better ‘, I expect to see these results.

4.Unfair comparison.

By including two additional input samples in the prompt, the zero-shot setup effectively becomes 2-shot, and the three-shot setup becomes five-shot. While it’s true that these additional samples are provided without labels, they still contribute additional context and information. I believe it is crucial to ensure an apples-to-apples comparison by using fairer baselines that incorporate the same amount of information, regardless of whether the information includes labels or not. This would make it clear whether the improvements come from the method itself or simply from the extra information.

Can you clarify this question? For LF-0-shot, if adding labels for 2 additional input samples, that’s gonna be like LF-2-shot.

Yes, I am referring to LF-2-shot. It would be nice to include such ablations to see the impact of adding label information.

I appreciate that the authors have improved the writing per my feedback, so I’ve raised the Presentation score to 3. However, considering the aforementioned reservations, I am not entirely convinced and will respectfully maintain my rating.

We would like to thank the reviewer for their detailed evaluation of our manuscript and for providing valuable insights. We now added a new section for addressing the major concern of lacking in-depth analysis. Please check below for comments and our revised paper.

For Weaknesses:

1. To better provide better explanations. We added a new section in the results (section 5,4). We used the LF-0 shot VS regular 0-shot as an ablation study. Our hypothesis is that by providing comparison samples, it enforces models to contrast and compare. With multiple variations and differences of samples exposed, the model's confidence is ‘regularized’ and calibrated. To see this, we plot the bootstrapped (for balancing the correct and incorrect samples) histogram of probability of correct prediction and incorrect prediction . And we can see that with our method, the differences are more outstanding. As evidence that the model indeed is comparing and contrasting different samples, we provide a breakdown across different datasets. We found that for the financial dataset, such contrast is more obvious. This is likely because the financial task is an ‘ordinal’ dataset where it has ‘positive’ and ‘negative’ labels where ‘contrast’ and ‘comparison’ are supposed to be more effective in practice (in terms of showing the difference of probabilities, not accuracy). And we do observe that the probabilities are more outstanding there.
2. This is also brought up by another reviewer. We will make this clear in the paper. We listed ensemble (agg) methods; it's a natural method introduced without comparative inference. But that should not be the highlight of the paper.

For Questions: 1.Yes, LF-0-shot is our method alone. LF-3-shot is our method combined with few-shot prompting. You can also check appendix 1 for a real example to see what we used for prompt and how we combined the method. 2. We presented DKL in the figure 3 with MACROCE (as these are new proposed and a relatively new metric in the field focused on the special patterns of miscalibration). ECE is mostly used in literature as a general metric measuring calibration so we listed it as a main metric in the table along with accuracy.

We appreciate reviewer's constructive feedback. We think there's a major confusion that lead to reviewer think our method has more cost which we clarified here and we also address some comments in the new revision, please see our comments below and revised paper.

For Weaknesses:

1. We wanted to evaluate the model on multiple model sizes and the Llama family is the best choice to our knowledge and it can provide a good comparison across the model sizes to draw conclusions (as less bias is introduced by different model families). Different models and more datasets are definitely worth trying, but we are limited by some capacity constraints. We acknowledge this is one of the limitations in section 7.
2. This is a good point. We don’t have strong justification that bias can be averaged out (e.g. to 0) but this is just an assumption. The theoretical analysis corresponding to this would be figure 5 where you can see the metrics move with more aggregation examples. Though it plateaus at certain points which requires further studies. We list this as future work in the discussion section.
3. This is an important point and we appreciate the reviewer for pointing this out. Please check 4.1.1, we always list samples to infer as the first example. You can cut the output after the first word as that’s the only answer you will need which is what we used for experiments as well. And if LLM is not generating more tokens/answers (for comparison samples) and there will be no multiple forward passes and no additional cost. We made this clear in our paper now (end of section 4.1.1).

For Questions:

1. We added a new section in the results (section 5,4) as an ablation section for understanding the behavior of our method and showed that Financial dataset is an ‘ordinal’ task that can be more sensitive/effective to the comparison approach (Please see the revision or the respond below for the second reviewer). Also, figure 5 is the evidence for the aggregation part. We are inspired by some heuristic examples. For instance, considering a sentiment question, the outcome can be positive or negative. If you compare your sample with a very positive example, then LLM might be biased to give a negative answer and vice versa. But if you average them, then it’s averaged out. This would suggest that in aggregation, we need to draw samples from a balanced distribution for comparisons. In our studies, we simply random sampled from the data source we have.
2. This is a very interesting topic and we believe for each dataset, LLM, there may have very different conclusions. There should be a better strategy but that may also indicate additional cost. We listed this as future work (See section 7).
3. Copied from another response: This is an important point and we appreciate the reviewer for pointing this out. Please check 4.1.1, we always list samples to infer as the first example to avoid bias. You can cut the output when it generates the answer for your input. And if LLM is not generating more tokens (for comparison samples) and there will be no multiple forward passes and no additional cost. 4.similar to 2.
4. Are you asking for the next token/word prediction type of tasks? We believe this is interesting but not sure until someone really tries it. Challenges would come from: 1) need a proper prompt for comparison, can potentially do something like ‘comparing the following words as the next word’, etc. 2) Need to have a good pool of candidates for comparison. An heuristic is to first run inference with original text and prompt (without comparison) and then choose a few from the ones that have higher probabilities as possible ones for comparison. (Though low probability ones can also be useful but will require some handcraft work for selection). This is not very straightforward compared with well-defined supervised classification tasks though.