Reward models are widely used as proxies for human preferences when aligning or evaluating LLMs. However, reward models are black boxes, and it is often unclear what, exactly, they are actually rewarding. In this paper we develop Rewrite-based Attribute Treatment Estimator (RATE) as an effective method for measuring the sensitivity of a reward model to high-level attributes of responses, such as sentiment, helpfulness, or complexity. Importantly, RATE measures the causal effect of an attribute on the reward. RATE uses LLMs to rewrite responses to produce imperfect counterfactuals examples that can be used to measure causal effects. A key challenge is that these rewrites are imperfect in a manner that can induce substantial bias in the estimated sensitivity of the reward model to the attribute. The core idea of RATE is to adjust for this imperfect-rewrite effect by rewriting twice. We establish the validity of the RATE procedure and show empirically that it is an effective estimator.

This paper concerns the problem of aligning samples from large language models to human preferences using *best-of-$n$* sampling, where we draw $n$ samples, rank them, and return the best one. We consider two fundamental problems. First: what is the relationship between best-of-$n$ and other (RLHF-type) approaches to aligning LLMs? In particular, when should one be preferred to the other? We show that the best-of-$n$ sampling distribution is essentially equivalent to the policy learned by RLHF if we apply a particular monotone transformation to the reward function. Moreover, we show that this transformation yields the best possible trade-off between win-rate against the base model vs KL distance from the base model. Then, best-of-$n$ is a Pareto-optimal win-rate vs KL solution. The second problem we consider is how to fine-tune a model to mimic the best-of-$n$ sampling distribution, to avoid drawing $n$ samples for each inference. We derive *BonBon Alignment* as a method for achieving this. Experiments show that BonBon alignment yields a model that achieves high win rates while minimally affecting off-target aspects of the generations.

Cristina Garbacea, Mengtian Guo, Samuel Carton, Qiaozhu Mei. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Cristina Garbacea, Samuel Carton, Shiyan Yan, Qiaozhu Mei. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

In order to efficiently transmit and store speech signals, speech codecs create a minimally redundant representation of the input signal which is then decoded at the receiver with the best possible perceptual quality. In this work we demonstrate that a neural network architecture based on VQ-VAE with a WaveNet decoder can be used to perform very low bit-rate speech coding with high reconstruction quality. A prosody-transparent and speaker-independent model trained on the LibriSpeech corpus coding audio at 1.6 kbps exhibits perceptual quality which is around halfway between the MELP codec at 2.4 kbps and AMR-WB codec at 23.05 kbps. In addition, when training on high-quality recorded speech with the test speaker included in the training set, a model coding speech at 1.6 kbps produces output of similar perceptual quality to that generated by AMR-WB at 23.05 kbps.

We address the task of detecting the reputation polarity of social media updates, that is, deciding whether the content of an update has positive or negative implications for the reputation of a given entity. Typical approaches to this task include sentiment lexicons and linguistic features. However, they fall short in the social media domain because of its unedited and noisy nature, and, more importantly, because reputation polarity is not only encoded in sentiment-bearing words but it is also embedded in other word usage. To this end, automatic methods for extracting discriminative features for reputation polarity detection can play a role. We propose a data-driven, supervised approach for extracting textual features, which we use to train a reputation polarity classifier. Experiments on the RepLab 2013 collection show that our model outperforms the state-of-the-art method based on sentiment analysis by 20% accuracy.