Thank you for taking the time to thoughtfully review our work. Below, we provide some further context and try to clarify the questions identified in your review.

Theoretical Support of Pareto-Optimality

We are not able to make theoretical claims that AttriBoT is Pareto-optimal in general because this finding involves comparison to methods that work quite differently than LOO attribution (e.g. attention-based attribution).

Closed Source Language Models

The only assumption made by our methods (and LOO attribution in general) is that we can compute the conditional likelihood of a response given a context. Thus, our method can be applied to any model that provides likelihoods as part of its API. Whether AttriBoT could be applied to a closed-source LLM therefore depends on whether the LLM's API provides likelihood values.

Non-Transformer Architectures

In principle, the hierarchical, proxy modeling, and pruning methods within AttriBoT could be applied to non-Transformer language models. KV caching would only extend to other architectures if, as in causal self-attention, it is valid to reuse past KV states.

Evaluation Outside of Open-Book QA

We focus on open-book QA in our work because it is realistic and representative of many typical LLM use-cases (i.e. any query that involves conditioning on specific spans within a larger context) while being straightforward to evaluate. Other LLM use-cases, such as abstractive summarization, most or many context spans might be informative, so measuring the LOO error might not provide useful or actionable insights. If you have specific benchmarks/settings that you would like us to include, please let us know and we will do our best to add them.

Thank you for your response to the issues raised.

Thank you for taking the time to thoroughly read our work. Below, we discuss some of the concerns brought up in your review.

Simple Methods

Regarding KV caching, while KV caching is a general trick for speeding up inference, we note that standard KV caching (where the entire context is cached and re-used as-is) is not applicable to LOO attribution since the context changes for each pass of leave-one-out. While we agree that adapting KV caching to LOO attribution is not a fundamentally new method, we argue that it does involve some novelty. Furthermore, we would like to note that KV caching actually cannot be applied out of the box to any attribution method that performs repeated inference. For instance, ContextCite gains almost no improvement in performance from KV caching, since for every forward pass a randomly sampled set of sources are removed from the context resulting in a low probability of prefix-sharing across forward passes.

More broadly, we agree that the methods proposed in this work are quite straightforward. However, seeing as how these methods actually work quite well for efficiently approximating LOO attributions, we argue that this should not be considered a weakness. Especially in newer research areas, like context attribution, it is important to start with simple methods to understand areas of improvement that actually require complex solutions.

Furthermore, we make explicit the assumptions of each method (e.g., sum of k leave-one-out attributions is approximately equal to leave-k-out attribution, proxy model attributions approximate target model attributions when both models are attributing the same response, etc.) and provide evidence that these assumptions hold in realistic settings.

Experiments on Multiple Models and additional findings

We agree with this point. Our main experiments used the Llama model family and we also reported results on the HotPotQA dataset using the Qwen model family to show that these results generalize across models. Since receiving your feedback we have replicated all of the experiments (QASPER, SQuAD, and HotPotQA) using the Qwen model family (Figure 6) and see nearly identical results, demonstrating the generalizability of our method. We also included new results considering mismatched proxy models including Mistral 7B (Figures 7-10) and find that, while still effective, using a mismatched proxy model can degrade performance somewhat.

Thank you for taking the time to read our paper and for providing this thoughtful response. Below, we provide some further context and discuss the weaknesses identified in your review.

Importance of the LOO Error and its approximation

The LOO error measures the change in a model’s prediction when a single piece of data (e.g., one training example, one context source, or other notions of a single datum) is removed from the model. This provides a simple and interpretable score measuring how important that piece of data is to a model’s prediction or behavior. In fact, approximating the LOO error is the objective of many methods that aim to attribute back to training data (e.g., including LOGRA, but also datapoint attribution methods like influence functions and TRAK) or to context sources (e.g., ContextCite).

In training data attribution works, such as the LOGRA paper highlighted in the review, computing the exact LOO error for a sample requires re-training the model with that sample excluded from the training set, which makes computing the exact LOO error for every training example infeasible. In fact, the basis of many training data attribution methods, like LOGRA, is to approximate the LOO error since it is not possible to compute directly. This is discussed in Section 2: Scalability Bottlenecks in Influence Functions from the LOGRA paper.

Since the exact LOO error is infeasible to compute in training data attribution, this also makes it prohibitively expensive to evaluate training data attribution methods in terms of how well they approximate the LOO error. As a result, most training data attribution methods resort to evaluating with proxy metrics like task accuracy as a function of the number of high-attribution training examples removed.

For context attribution, the picture is a bit different. In this setting, it is expensive, but feasible, to compute exact LOO attributions since each attribution score requires a single forward pass rather than a full training run. Thus, in our work, we are actually able to measure how well our approximate LOO attribution methods align with the ground truth LOO attributions. To improve the clarity of the paper, we will be adding parts of this explanation to our related work.

Efficient LOO Attribution and Large Model Compression

The goals of efficient LOO attribution and model compression do seem somewhat similar – in both settings we aim to emulate some behavior of large models efficiently. However, model compression focuses on retaining performance on downstream tasks while we primarily care about retaining the LOO attribution behavior of models. As such, there may be some similarity between the hierarchical, pruning, and proxy modeling approaches we propose and methods used for model compression, but we develop novel variants of these ideas and specifically test that the underlying assumptions of each of these methods hold true in the LOO attribution setting. Based on your feedback, we would be happy to add some references to the relevant literature on model compression, and if you have any suggestions for relevant papers, we would be happy to add them.

Recovering Attributions of More Than the Top Sources

Most applications of context attribution (via LOO error or otherwise) primarily aim to identify highly contributive sources because the insight that many spans contributed a small amount is not particularly actionable. We note that AttriBoT will absolutely provide attribution scores for all sources in the context, but we focus on evaluating the highly contributive sources to ensure a reliable and realistic evaluation. If there is a different metric you would be interested in seeing on HotpotQA or some other dataset, we would be happy to compute and report it.

Organization of Experimental Results Section

Due to space constraints, we had to defer some figures to the appendix while referring to them in the main text’s experimental results. We will try to make this section more clear, and if our submission is accepted, we plan to use the extra page to move these Appendix figures back into the main text as you've suggested.

Thanks for the authors' detailed responses. My main concerns have been addressed. Therefore, I am happy to maintain my positive rating.

Thank you for taking the time to thoughtfully review our work. Below, we provide some further context and discuss the weaknesses identified in your review.

Evaluation Other Than Open-Book QA

We focus on open-book QA in our work because it is realistic and representative of many typical LLM use-cases (i.e. any query that involves conditioning on specific spans within a larger context) while being straightforward to evaluate. Other LLM use-cases, such as abstractive summarization, most or many context spans might be informative, so measuring the LOO error might not provide useful or actionable insights. If you have specific benchmarks/settings that you would like us to include, please let us know and we will do our best to add them.

Hyperparameter Selection

We agree that with a set of methods like AttriBoT, that can be mixed and matched to achieve different efficiency vs. accuracy tradeoffs, selecting an algorithm combination and associated hyperparameters can be difficult. To mitigate this we do two things:

1. We amend Figure 2 to show which algorithm combinations are represented by different points on the AttriBoT Pareto front. This gives readers better insight into which combinations of methods tend to provide higher accuracy and lower efficiency vs. lower accuracy and higher efficiency.
2. We have added a “Practitioner’s Guide to AttriBoT” as Appendix D, in which we summarize the relative efficiency of different algorithm combinations, provide insight into hyperparameter (e.g., hierarchical thresholds) selection based on our experiments, and advise users on best practices for data preprocessing before performing context attribution.

Thank you for taking the time to thoughtfully review our work. Below, we provide some further context and discuss the questions and weaknesses identified in your review.

KV Caching

We agree with the reviewer, that KV caching comes at little implementation cost. However, we do note that our use of KV caching is notable and novel in this context for the following reasons:

1. Standard KV caching (where the entire context is cached and re-used as-is) is not applicable to LOO attribution since the context changes for each pass of leave-one-out. AttriBoT's novel application of KV caching involves identifying the longest possible pre-computed prefix that can be used based on the current context and past contexts, with additional complications when processing a context hierarchically.
2. KV caching does not provide a speedup by default for all context attribution methods that involve multiple forward passes. An example of this is ContextCite, which performs each forward pass on a context with a randomly selected set of sources removed. Since the source removal is probabilistic, contexts on average have very short prefixes in common resulting in virtually zero performance gain from KV caching.

Proxy Modeling for Complex Tasks

This is an excellent point that certain complex tasks are only possible for large models, thus potentially breaking the proxy modeling assumption. This is seen in our proxy modeling results (Figure 3) where larger proxy models yield better approximate LOO attributions than smaller proxy models. One important caveat to note is that for proxy modeling to work, it is not necessary that the proxy model be able to perform the task on its own. This is because when we approximate the LOO attributions for a large target model using a small proxy model, we use the response generated by the target model and evaluate this response’s likelihood (given different contexts) using the proxy model. Thus, the proxy does not need to produce the same response as the target model, but rather must emulate the conditional likelihood of a response given a context. We do note that we include results on HotpotQA, which involves reasoning over multiple "hops" to answer questions. We have added corresponding clarifications to the paper.

Duplicate Context Information

It is a valid point that duplicate contextual information can cause problems for the methods we propose. If the same piece of information appears multiple multiple times in the context, then simply removing one instance of that information would likely not change the model’s response likelihood. However, this is not unique to our methods, but rather to any attribution method that aims to approximate the LOO error. For instance, many popular training data attribution methods, such as influence functions, also suffer from this issue since they approximate LOO error for training data. In our newly added ‘Practitioner’s Guide to AttriBoT’ in Appendix D, we suggest users to deduplicate input before performing LOO attribution.

An alternative approach for situations with duplicate context information would be to compute attributions based on notions of value other than the LOO error – such as the Shapley value. In general, however, Shapley values are much more expensive to compute than the LOO error, as they require performing an exponential (in the number of context sources) number of forward passes compared to a linear number of forward passes needed for the LOO error.

Thank you for taking the time to engage with our paper. Regarding your questions:

1. We have amended Figure 2 in our submission to identify which points on the Pareto front correspond to different methods. Across most experiments, the rough ordering of methods on the Pareto front from most efficient to least efficient is as follows: Hierarchical + Proxy + KV Caching, Proxy + KV Caching, Hierarchical + KV Caching, Pruning + KV Caching, and KV Caching.

2. Our original experiments all used the Llama model family except for one experiment on HotPotQA that we replicated using the Qwen 2.5 model family. The intention of this Qwen experiment was to show that these methods generalize across model families. Since receiving this feedback, we have gone one step further and replicated experiments on each of the datasets with Qwen 2.5 (Figure 6), finding nearly identical results. Additionally, we have also included an experiment where we use Llama 70B as the target model and compare Llama 8B, Qwen 7B, and Mistral 7B as proxy models (Figures 7-10). We find that using a proxy model from the same model family as the target model performs best, providing evidence that a similar training distribution between the target and proxy model is important for accurate proxy modeling.

Thank you for taking the time to review our work. We agree that context attribution is a topic of crucial importance for understanding how LLMs “think” as well as allowing us to easily verify and validate their context-dependent claims. Below, we would like to discuss the three weaknesses you identified in your review:

Analysis of Proxy Model Fidelity

We agree that the AttriBoT methods that rely on proxy modeling degrade as the similarity between the proxy model and the target model decreases. This is seen in our results where larger proxy models achieve higher context attribution fidelity than smaller proxy models (Figure 3). Spurred by your suggestions, we have expanded on these results by investigating how differences in training distribution between the target and proxy model affect context attribution. In particular, we evaluate the use of Llama 8B, Qwen 7B, and Mistral 7B as a proxy for Llama 70B and find the LOO attribution approximation error is higher with proxy models from a different model family than the target model (mAP of 0.81 for Llama 8B vs. 0.75 Mistral 7B and 0.71 for Qwen 7B). This gives evidence that matching the training distribution between the target and proxy model is important, and we have added these results to our paper (Figures 7-10, appendix). However, given that most LLM model families include models of various sizes (e.g. the Llama 3 collection now contains 5 models from 1B to 405B parameters), we think practitioners will likely be able to choose a proxy model from the same family.

Generalization to other model classes

We focus on decoder-only transformers in this work for two reasons. First, nearly all LLMs are decoder-only transformers, and second, many decoder-only transformers have been trained at a variety of model sizes allowing for methods that leverage multiple models from the same family. To maximize the impact of our work, we chose to focus on this model class. However, if the reviewer has a specific suggestion of other language model architectures that are state-of-the-art and come in multiple sizes, we would certainly be interested in evaluating them too.

Incorporating Context Attribution into the Human-Machine Interaction Process

This would be a useful and interesting application of our research on efficient LOO context attribution. Our specific goal in this paper was to lay the algorithmic foundation for an application like the one you've proposed. However, we will be open sourcing a library (see supplementary files for a recent version) with an easy-to-use API for performing LOO attribution with the AttriBoT set of methods. We hope that this will empower practitioners to enhance the usability of LLMs deployed in their products.

Baselines

We believe we have covered all relevant baselines (5 including standard LOO) for context attribution. If the reviewer has any specific suggestions on additional baselines for context attribution, we would be happy to test them against our existing baselines and AttriBoT.