The accurate evaluation of differential treatment in language models to specific groups is critical to ensuring a positive and safe user experience. An ideal evaluation should have the properties of being robust, extendable to new groups or attributes, and being able to capture biases that appear in typical usage (rather than just extreme, rare cases). Relatedly, bias evaluation should surface not only egregious biases but also ones that are subtle and commonplace, such as a likelihood for talking about appearances with regard to women. We present FairPair, an evaluation framework for assessing differential treatment that occurs during ordinary usage. FairPair operates through counterfactual pairs, but crucially, the paired continuations are grounded in the same demographic group, which ensures equivalent comparison. Additionally, unlike prior work, our method factors in the inherent variability that comes from the generation process itself by measuring the sampling variability. We present an evaluation of several commonly used generative models and a qualitative analysis that indicates a preference for discussing family and hobbies with regard to women.
Reinforcement Learning from Human Feedback (\textbf{RLHF}) has emerged as a dominant approach for aligning LLM outputs with human preferences. Inspired by the success of RLHF, we study the performance of multiple algorithms that learn from feedback (Expert Iteration, Proximal Policy Optimization (\textbf{PPO}), Return-Conditioned RL) on improving LLM reasoning capabilities. We investigate both sparse and dense rewards provided to the LLM both heuristically and via a learned reward model. We additionally start from multiple model sizes and initializations both with and without supervised fine-tuning (\textbf{SFT}) data. Overall, we find all algorithms perform comparably, with Expert Iteration performing best in most cases. Surprisingly, we find the sample complexity of Expert Iteration is similar to that of PPO, requiring at most on the order of $10^6$ samples to converge from a pretrained checkpoint. We investigate why this is the case, concluding that during RL training models fail to explore significantly beyond solutions already produced by SFT models. Additionally, we discuss a trade off between maj@1 and pass@96 metric performance during SFT training and how conversely RL training improves both simultaneously. We then conclude by discussing the implications of our findings for RLHF and the future role of RL in LLM fine-tuning.
Dense retrievers compress source documents into (possibly lossy) vector representations, yet there is little analysis of what information is lost versus preserved, and how it affects downstream tasks. We conduct the first analysis of the information captured by dense retrievers compared to the language models they are based on (e.g., BERT versus Contriever). We use 25 MultiBert checkpoints as randomized initialisations to train MultiContrievers, a set of 25 contriever models. We test whether specific pieces of information -- such as gender and occupation -- can be extracted from contriever vectors of wikipedia-like documents. We measure this extractability via information theoretic probing. We then examine the relationship of extractability to performance and gender bias, as well as the sensitivity of these results to many random initialisations and data shuffles. We find that (1) contriever models have significantly increased extractability, but extractability usually correlates poorly with benchmark performance 2) gender bias is present, but is not caused by the contriever representations 3) there is high sensitivity to both random initialisation and to data shuffle, suggesting that future retrieval research should test across a wider spread of both.
Teaching language models to use tools is an important milestone towards building general assistants, but remains an open problem. While there has been significant progress on learning to use specific tools via fine-tuning, language models still struggle with learning how to robustly use new tools from only a few demonstrations. In this work we introduce a self-verification method which distinguishes between close candidates by self-asking contrastive questions during (1) tool selection; and (2) parameter generation. We construct synthetic, high-quality, self-generated data for this goal using Llama-2 70B, which we intend to release publicly. Extensive experiments on 4 tasks from the ToolBench benchmark, consisting of 17 unseen tools, demonstrate an average improvement of 22% over few-shot baselines, even in scenarios where the distinctions between candidate tools are finely nuanced.
State-of-the-art language models can exhibit impressive reasoning refinement capabilities on math, science or coding tasks. However, recent work demonstrates that even the best models struggle to identify \textit{when and where to refine} without access to external feedback. Outcome-based Reward Models (\textbf{ORMs}), trained to predict correctness of the final answer indicating when to refine, offer one convenient solution for deciding when to refine. Process Based Reward Models (\textbf{PRMs}), trained to predict correctness of intermediate steps, can then be used to indicate where to refine. But they are expensive to train, requiring extensive human annotations. In this paper, we propose Stepwise ORMs (\textbf{SORMs}) which are trained, only on synthetic data, to approximate the expected future reward of the optimal policy or $V^{\star}$. More specifically, SORMs are trained to predict the correctness of the final answer when sampling the current policy many times (rather than only once as in the case of ORMs). Our experiments show that SORMs can more accurately detect incorrect reasoning steps compared to ORMs, thus improving downstream accuracy when doing refinements. We then train \textit{global} refinement models, which take only the question and a draft solution as input and predict a corrected solution, and \textit{local} refinement models which also take as input a critique indicating the location of the first reasoning error. We generate training data for both models synthetically by reusing data used to train the SORM. We find combining global and local refinements, using the ORM as a reranker, significantly outperforms either one individually, as well as a best of three sample baseline. With this strategy we can improve the accuracy of a LLaMA-2 13B model (already fine-tuned with RL) on GSM8K from 53\% to 65\% when greedily sampled.
To achieve faithful reasoning that aligns with human expectations, large language models (LLMs) need to ground their reasoning to real-world knowledge (e.g., web facts, math and physical rules). Tools help LLMs access this external knowledge, but there remains challenges for fine-tuning LLM agents (e.g., Toolformer) to invoke tools in multi-step reasoning problems, where inter-connected tool calls require holistic and efficient tool usage planning. In this work, we propose a new method for LLMs to better leverage tools in multi-step reasoning. Our method, Chain-of-Abstraction (CoA), trains LLMs to first decode reasoning chains with abstract placeholders, and then call domain tools to reify each reasoning chain by filling in specific knowledge. This planning with abstract chains enables LLMs to learn more general reasoning strategies, which are robust to shifts of domain knowledge (e.g., math results) relevant to different reasoning questions. It also allows LLMs to perform decoding and calling of external tools in parallel, which avoids the inference delay caused by waiting for tool responses. In mathematical reasoning and Wiki QA domains, we show that our method consistently outperforms previous chain-of-thought and tool-augmented baselines on both in-distribution and out-of-distribution test sets, with an average ~6% absolute QA accuracy improvement. LLM agents trained with our method also show more efficient tool use, with inference speed being on average ~1.4x faster than baseline tool-augmented LLMs.
As generative large language models (LLMs) grow more performant and prevalent, we must develop comprehensive enough tools to measure and improve their fairness. Different prompt-based datasets can be used to measure social bias across multiple text domains and demographic axes, meaning that testing LLMs on more datasets can potentially help us characterize their biases more fully, and better ensure equal and equitable treatment of marginalized demographic groups. In this work, our focus is two-fold: (1) Benchmarking: a comparison of 6 different prompt-based bias and toxicity metrics across 12 demographic axes and 5 families of generative LLMs. Out of those 6 metrics, AdvPromptSet and HolisticBiasR are novel datasets proposed in the paper. The comparison of those benchmarks gives us insights about the bias and toxicity of the compared models. Therefore, we explore the frequency of demographic terms in common LLM pre-training corpora and how this may relate to model biases. (2) Mitigation: we conduct a comprehensive study of how well 3 bias/toxicity mitigation techniques perform across our suite of measurements. ROBBIE aims to provide insights for practitioners while deploying a model, emphasizing the need to not only measure potential harms, but also understand how they arise by characterizing the data, mitigate harms once found, and balance any trade-offs. We open-source our analysis code in hopes of encouraging broader measurements of bias in future LLMs.
As increasingly sophisticated language models emerge, their trustworthiness becomes a pivotal issue, especially in tasks such as summarization and question-answering. Ensuring their responses are contextually grounded and faithful is challenging due to the linguistic diversity and the myriad of possible answers. In this paper, we introduce a novel approach to evaluate faithfulness of machine-generated text by computing the longest noncontinuous substring of the claim that is supported by the context, which we refer to as the Longest Supported Subsequence (LSS). Using a new human-annotated dataset, we finetune a model to generate LSS. We introduce a new method of evaluation and demonstrate that these metrics correlate better with human ratings when LSS is employed, as opposed to when it is not. Our proposed metric demonstrates an 18% enhancement over the prevailing state-of-the-art metric for faithfulness on our dataset. Our metric consistently outperforms other metrics on a summarization dataset across six different models. Finally, we compare several popular Large Language Models (LLMs) for faithfulness using this metric. We release the human-annotated dataset built for predicting LSS and our fine-tuned model for evaluating faithfulness.
As large language models improve, there is increasing interest in techniques that leverage these models' capabilities to refine their own outputs. In this work, we introduce Shepherd, a language model specifically tuned to critique responses and suggest refinements, extending beyond the capabilities of an untuned model to identify diverse errors and provide suggestions to remedy them. At the core of our approach is a high quality feedback dataset, which we curate from community feedback and human annotations. Even though Shepherd is small (7B parameters), its critiques are either equivalent or preferred to those from established models including ChatGPT. Using GPT-4 for evaluation, Shepherd reaches an average win-rate of 53-87% compared to competitive alternatives. In human evaluation, Shepherd strictly outperforms other models and on average closely ties with ChatGPT.
Lifelogs are descriptions of experiences that a person had during their life. Lifelogs are created by fusing data from the multitude of digital services, such as online photos, maps, shopping and content streaming services. Question answering over lifelogs can offer personal assistants a critical resource when they try to provide advice in context. However, obtaining answers to questions over lifelogs is beyond the current state of the art of question answering techniques for a variety of reasons, the most pronounced of which is that lifelogs combine free text with some degree of structure such as temporal and geographical information. We create and publicly release TimelineQA1, a benchmark for accelerating progress on querying lifelogs. TimelineQA generates lifelogs of imaginary people. The episodes in the lifelog range from major life episodes such as high school graduation to those that occur on a daily basis such as going for a run. We describe a set of experiments on TimelineQA with several state-of-the-art QA models. Our experiments reveal that for atomic queries, an extractive QA system significantly out-performs a state-of-the-art retrieval-augmented QA system. For multi-hop queries involving aggregates, we show that the best result is obtained with a state-of-the-art table QA technique, assuming the ground truth set of episodes for deriving the answer is available.