Improving Context-Aware Preference Modeling for Language Models

While finetuning language models from pairwise preferences has proven remarkably effective, the underspecified nature of natural language presents critical challenges. Direct preference feedback is uninterpretable, difficult to provide where multidimensional criteria may apply, and often inconsistent, either because it is based on incomplete instructions or provided by diverse principals. To address these challenges, we consider the two-step preference modeling procedure that first resolves the under-specification by selecting a context, and then evaluates preference with respect to the chosen context. We decompose reward modeling error according to these two steps, which suggests that supervising context in addition to context-specific preference may be a viable approach to aligning models with diverse human preferences. For this to work, the ability of models to evaluate context-specific preference is critical. To this end, we contribute context-conditioned preference datasets and accompanying experiments that investigate the ability of language models to evaluate context-specific preference. We use our datasets to (1) show that existing preference models benefit from, but fail to fully consider, added context, (2) finetune a context-aware reward model with context-specific performance exceeding that of GPT-4 and Llama 3 70B on tested datasets, and (3) investigate the value of context-aware preference modeling.

Paper

References (86)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer rZtt6/10 · confidence 4/52024-07-09

Summary

The paper focuses on fine-tuning LLMs to improve their ability to handle context-aware preference modeling. The authors address the challenge of the underspecified and ambiguous nature of natural language preferences by introducing a two-step modeling process. This includes selecting a context and evaluating preferences within that context. The approach is backed by the introduction of new datasets, named RPR, which are designed to test the effectiveness of context-specific preference modeling. The study provides extensive experimental evidence showing how context-aware models can surpass traditional methods in handling ambiguous and context-specific scenarios.

Strengths

1. The paper tackles a crucial issue in the realm of language modeling by enhancing the LLM's ability to understand and process user preferences in context-dependent scenarios. This is particularly important as LLMs are increasingly used in diverse real-world applications. 2. The introduction of the RPR datasets is a notable contribution, as these datasets specifically aim to disentangle context-specific preferences from general preferences, offering a valuable resource for further research. 3. The paper provides thorough experimental results that not only demonstrate the effectiveness of the proposed method but also explore various aspects of context-aware preference modeling, showing improvements over existing models like GPT-4 and Llama 3.

Weaknesses

1. The paper could benefit from testing the proposed method across a broader range of general benchmarks, such as MMLU and AGI-Eval, to assess how fine-tuning for context-aware preferences might affect the LLM's original capabilities or general applicability. 2. It remains unclear how applicable the proposed method is to other LMs, especially those of different sizes or from different series. Addressing this would help validate the robustness and versatility of the method.

Questions

Please see my concerns in weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

4

Limitations

n/a

Reviewer t1sE5/10 · confidence 4/52024-07-11

Summary

This paper divides preference modeling into two steps: first estimating the user's intent then evaluate the generated text within the context of this intent. The paper makes the Reasonable Preference Reversal Datasets, which encompass criteria and scenarios for preference data. Experiments find that models can achieve higher performance when providing intent contexts during evaluation.

Strengths

- The approach of first estimating user intent before evaluation may be promising. - This work constructs the first open-source context-conditioned preference datasets, which may be useful for future work.

Weaknesses

1. Lack of citations and comparisons to related work. For example, Li et al. [1] generate the criteria for evaluation and then get the final answer according to the criteria, which is similar to this work. 2. In Table 2, the author aims to demonstrate that fine-tuning a context-aware reward model significantly enhances context-specific performance. However, the observed improvement in Table 2 could be attributed to the model being trained on data from the same distribution as the test set. Even for new datasets that are not context-specific, training the model on such datasets will likely improve its performance on its test sets. 3. The lack of experiments illustrating that two-step preference modeling achieves better performance compared to traditional reward modeling on general preference datasets is a concern. In Table 4, the model is prompted by context generated by GPT-4, leading to unfair comparisons with similar models. To demonstrate that two-step preference modeling is superior, it is essential to show its advantages for models of equal ability. [1] Li J, Sun S, Yuan W, et al. Generative judge for evaluating alignment[J]. arXiv preprint arXiv:2310.05470, 2023.

Questions

Why is two-step preference modeling necessary? The user's intent is inherently contained in the query, and traditional preference modeling implicitly evaluates whether the response aligns with the user's intent. Therefore, why should the estimated intent be treated independently? Additionally, ambiguity exists in estimating intent, as different individuals may have inconsistent interpretations of intent for the same query.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Please see Weakness part.

Reviewer t1sE2024-08-11

Response to Authors

Thank you for the authors' detailed response that solves some of my questions through additional experiments and explanations. However, my core question still exists, which is ``whether two-stage modeling is better than traditional reward modeling''. I understand that using experiments to illustrate this can be difficult, but it should be an important part. In real conversation and preference annotation scenarios, the intent of a query is ambiguous and has multiple possibilities. The two-stage approach may lead to better modeling results through decomposition, but it may also result in error accumulation. Moreover, as Weakness1 pointed out, previous work has already explored developing evaluation criteria before conducting evaluations. This results in a limited contribution of this paper. I think the author's research direction is promising, but there are still many points that can be improved. So, I keep my rating.

Reviewer mLs14/10 · confidence 4/52024-07-14

Summary

This paper points out that the preference label can be reversed by inserting additional context into the prompt. Based on this observation, the authors build a paired preference dataset. The authors also try to provide some theoretical analysis.

Strengths

* This paper studies a specific and interesting problem in preference data. * The proposed data augmentation method may benefit preference optimization of LLMs and inspire further research.

Weaknesses

My major concern lies in the experiments. The current experiments do not demonstrate the general advantage of using a paired dataset with reversed preference labels. On widely used preference datasets, the model trained on the proposed dataset does not outperform the baselines. It only performs better on the test set, which is built in the same way as the training set. Note that the preference labels on this test set are not verified by humans and could be unreliable.

Questions

See weaknesses.

Rating

4

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

This paper does not have a limitation section.

Reviewer j5YU4/10 · confidence 3/52024-07-14

Summary

The motivation behind this paper is to address the critical challenges of finetuning language models (LMs) from pairwise preferences due to the underspecified nature of natural language. Direct preference feedback is often uninterpretable, inconsistent, and difficult to provide, especially when multidimensional criteria are involved. These issues arise from incomplete instructions or the diverse backgrounds of the individuals providing the feedback. To tackle these challenges, the authors propose a two-step preference modeling approach: first, selecting a context to resolve under-specification, and second, evaluating preference with respect to the chosen context. There are mainly 4 contributions the authors have made: **Decomposition of Reward Modeling Error**: The paper introduces a method to decompose reward modeling error into two components: context inference error and context-specific reward modeling error. This decomposition supports the idea that supervising both context and context-specific preference could align models more effectively with diverse human preferences. **Context-Conditioned Preference Datasets**: The authors contribute several novel datasets designed to investigate the ability of LMs to evaluate context-specific preferences. These datasets, termed "preference reversal" datasets, isolate context-specific capabilities by disentangling them from general preferences. **Context-Aware Reward Model**: The paper demonstrates the development and finetuning of a context-aware reward model, showing that this model achieves performance comparable to or exceeding that of state-of-the-art models like GPT-4 and Llama 3 70B. The context-aware model also shows improved performance in context-specific tasks. **Experiments and Benchmarking**: The authors conduct experiments to benchmark the context-specific performance of various models, highlighting that current models benefit from additional context but often fail to fully utilize it. Finetuning with the preference reversal datasets significantly enhances the models' context-specific performance.

Strengths

This paper presents a novel approach to preference modeling in LMs by integrating context-specific evaluation. The key contributions include the introduction of context-conditioned preference datasets, the decomposition of reward modeling error, the development of a context-aware reward model, and comprehensive experiments to validate the approach.

Weaknesses

## Under-supported claims In section 3.3, the author included various claims that I believe to be interesting but only so when they are better-substantiated. **Assumption about Cardinality** The hypothesis that the cardinality of the space of contexts is smaller than that of the space of possible completions given a prompt is stated without empirical or theoretical justification. This assumption is critical to the argument about data efficiency in context annotation versus preference annotation. Without evidence or a rationale to support this assumption, the argument remains speculative. ## Details of the dataset curation The reviewer appreciate the author's efforts in curating the dataset. Section 4 briefly mentions that the dataset was generated using GPT-4 Turbo with a series of prompts designed to maximize validity. However, it lacks specific details about the prompts, the criteria for selection, and the methodology for ensuring the validity of the samples. This ambiguity makes it difficult to assess the soundness of the curation process. ## Clarity While the paper presents interesting ideas, there are some areas in writing and presentation that could benefit from improvement: - The introduction and related work sections are detailed, but they could be more concise to better highlight the key contributions and their significance. - The figures and tables, although informative, could be optimized to more effectively convey the main points. Some elements might appear redundant or unclear. - Enhancing the narrative flow would help in making the logical progression of the arguments clearer and more accessible to the reader.

Questions

## Annotation and intent inference I particularly do not understand the assumption related to Bernoulli distribution: Section 3.1 states that an annotator implicitly infers intent i from prompt x and samples a preference from a Bernoulli distribution. This process of intent inference by annotators is critical yet highly under-explained. There is no empirical evidence or studies cited that demonstrate annotators' ability to accurately infer intents from prompts. Additionally, the variability in annotators' interpretations and the potential biases they introduce are not addressed. ## Intent distribution The paper asserts that both users and annotators may possess or infer a distribution of intents and that annotation for most preference queries involves a distribution rather than a specific intent. This is a substantial claim that needs empirical support. ## Data curation process I would appreciate it if the authors can provide further details on data curation. In particular, how the quality of the synthetic data is maintained. ## Human validation What is the scale of this validation, and how are the human validators selected? Is the sample size sufficient to generalize the results to the entire dataset?

Rating

4

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

Yes.

Authorsrebuttal2024-08-12

Thank for taking the time to review our work and provide suggestions. You mention that there is not enough information about the data filtering process, and that the information surrounding dataset generation is "vague". In order for us to address these concerns, we would appreciate if the reviewer could be a bit more specific on this point. But for the sampling parameters used for GPT-4 (e.g. temperature), which we will specify, we believe the information in Appendix B is sufficiently detailed to reproduce the datasets, and are uncertain about what additional information should be included. With respect to the other points, we will expand upon our note on limitations at L317-320 in the submitted draft (which does acknowledge limitations in quality/validity/diversity due to the data being synthesized, albeit briefly), and will include the information on human annotations that we have provided to Reviewers mLs1 and j5YU below.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC