Abstract:In this paper, we study format biases in reinforcement learning from human feedback (RLHF). We observe that many widely-used preference models, including human evaluators, GPT-4, and top-ranking models on the RewardBench benchmark, exhibit strong biases towards specific format patterns, such as lists, links, bold text, and emojis. Furthermore, large language models (LLMs) can exploit these biases to achieve higher rankings on popular benchmarks like AlpacaEval and LMSYS Chatbot Arena. One notable example of this is verbosity bias, where current preference models favor longer responses that appear more comprehensive, even when their quality is equal to or lower than shorter, competing responses. However, format biases beyond verbosity remain largely underexplored in the literature. In this work, we extend the study of biases in preference learning beyond the commonly recognized length bias, offering a comprehensive analysis of a wider range of format biases. Additionally, we show that with a small amount of biased data (less than 1%), we can inject significant bias into the reward model. Moreover, these format biases can also be easily exploited by downstream alignment algorithms, such as best-of-n sampling and online iterative DPO, as it is usually easier to manipulate the format than to improve the quality of responses. Our findings emphasize the need to disentangle format and content both for designing alignment algorithms and evaluating models.
Abstract:Reinforcement learning with human feedback~(RLHF) is critical for aligning Large Language Models (LLMs) with human preference. Compared to the widely studied offline version of RLHF, \emph{e.g.} direct preference optimization (DPO), recent works have shown that the online variants achieve even better alignment. However, online alignment requires on-the-fly generation of new training data, which is costly, hard to parallelize, and suffers from varying quality and utility. In this paper, we propose a more efficient data exploration strategy for online preference tuning (OPTune), which does not rely on human-curated or pre-collected teacher responses but dynamically samples informative responses for on-policy preference alignment. During data generation, OPTune only selects prompts whose (re)generated responses can potentially provide more informative and higher-quality training signals than the existing responses. In the training objective, OPTune reweights each generated response (pair) by its utility in improving the alignment so that learning can be focused on the most helpful samples. Throughout our evaluations, OPTune'd LLMs maintain the instruction-following benefits provided by standard preference tuning whilst enjoying 1.27-1.56x faster training speed due to the efficient data exploration strategy.
Abstract:Instruction Fine-Tuning (IFT) significantly enhances the zero-shot capabilities of pretrained Large Language Models (LLMs). While coding data is known to boost reasoning abilities during LLM pretraining, its role in activating internal reasoning capacities during IFT remains understudied. This paper investigates a key question: How does coding data impact LLMs' reasoning capacities during the IFT stage? To explore this, we thoroughly examine the impact of coding data across different coding data proportions, model families, sizes, and reasoning domains, from various perspectives. Specifically, we create three IFT datasets with increasing coding data proportions, fine-tune six LLM backbones across different families and scales on these datasets, evaluate the tuned models' performance across twelve tasks in three reasoning domains, and analyze the outcomes from three broad-to-granular perspectives: overall, domain-level, and task-specific. Our holistic analysis provides valuable insights in each perspective. First, coding data tuning enhances the overall reasoning capabilities of LLMs across different model families and scales. Moreover, the effect of coding data varies among different domains but shows consistent trends across model families and scales within each domain. Additionally, coding data generally yields comparable task-specific benefits across different model families, with the optimal coding data proportions in IFT datasets being task-specific.
Abstract:Existing 3D mesh shape evaluation metrics mainly focus on the overall shape but are usually less sensitive to local details. This makes them inconsistent with human evaluation, as human perception cares about both overall and detailed shape. In this paper, we propose an analytic metric named Spectrum Area Under the Curve Difference (SAUCD) that demonstrates better consistency with human evaluation. To compare the difference between two shapes, we first transform the 3D mesh to the spectrum domain using the discrete Laplace-Beltrami operator and Fourier transform. Then, we calculate the Area Under the Curve (AUC) difference between the two spectrums, so that each frequency band that captures either the overall or detailed shape is equitably considered. Taking human sensitivity across frequency bands into account, we further extend our metric by learning suitable weights for each frequency band which better aligns with human perception. To measure the performance of SAUCD, we build a 3D mesh evaluation dataset called Shape Grading, along with manual annotations from more than 800 subjects. By measuring the correlation between our metric and human evaluation, we demonstrate that SAUCD is well aligned with human evaluation, and outperforms previous 3D mesh metrics.
Abstract:Data selection in instruction tuning emerges as a pivotal process for acquiring high-quality data and training instruction-following large language models (LLMs), but it is still a new and unexplored research area for vision-language models (VLMs). Existing data selection approaches on LLMs either rely on single unreliable scores, or use downstream tasks for selection, which is time-consuming and can lead to potential over-fitting on the chosen evaluation datasets. To address this challenge, we introduce a novel dataset selection method, Self-Filter, that utilizes the VLM itself as a filter. This approach is inspired by the observation that VLMs benefit from training with the most challenging instructions. Self-Filter operates in two stages. In the first stage, we devise a scoring network to evaluate the difficulty of training instructions, which is co-trained with the VLM. In the second stage, we use the trained score net to measure the difficulty of each instruction, select the most challenging samples, and penalize similar samples to encourage diversity. Comprehensive experiments on LLaVA and MiniGPT-4 show that Self-Filter can reach better results compared to full data settings with merely about 15% samples, and can achieve superior performance against competitive baselines.
Abstract:Making LLMs speak for different, especially minority groups of people, and generate statements supporting their diverse or even controversial perspectives is critical to creating an inclusive environment. However, existing LLMs lack sufficient controllability to the stance of their generated content, which often contains inconsistent, neutral, or biased statements. In this paper, we improve the controllability of LLMs in generating statements supporting an argument the user defined in the prompt. We find that multi-round debates between two LLMs with opposite stances generate higher-quality and more salient statements for each, which are important training data to improve the controllability of LLMs. Motivated by this, we develop a novel debate & tuning ("DEBATunE") pipeline finetuning LLMs to generate the statements obtained via debate. To examine DEBATunE, we curate the largest dataset of debate topics so far, which covers 710 controversial topics and corresponding arguments for each topic. Evaluations by the GPT-4 judge with a novel controversy controllability metric show that LLMs' capability of expressing diverse perspectives is significantly improved by DEBATunE. Moreover, such controllability can be generalized to unseen topics, generating high-quality statements supporting controversial arguments. Our codes, models, and data will be released at https://github.com/tianyi-lab/DEBATunE.
Abstract:Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM's reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacher-student collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs. Our codes, models, and data will be released at https://github.com/tianyi-lab/Reflection_Tuning.
Abstract:In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators to achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by varying training hyperparameters. Based on this evaluation, we conduct large-scale studies, where the results shed insights into the efficacy of hyperparameters and tricks used in RL on mitigating length bias. We further propose to improve the reward model by jointly training two linear heads on shared feature representations to predict the rewards, one trained to correlate with length, and the other trained to decorrelate with length and therefore focus more on the actual content. We then discard the length head in RL to prevent reward hacking on length. Experiments demonstrate that our approach almost eliminates the reward correlation with length, and improves the obtained policy by a significant margin.
Abstract:This technical report delves into the application of GPT-4 Vision (GPT-4V) in the nuanced realm of COVID-19 image classification, leveraging the transformative potential of in-context learning to enhance diagnostic processes.
Abstract:Large language models (LLMs), after being aligned with vision models and integrated into vision-language models (VLMs), can bring impressive improvement in image reasoning tasks. This was shown by the recently released GPT-4V(ison), LLaVA-1.5, etc. However, the strong language prior in these SOTA LVLMs can be a double-edged sword: they may ignore the image context and solely rely on the (even contradictory) language prior for reasoning. In contrast, the vision modules in VLMs are weaker than LLMs and may result in misleading visual representations, which are then translated to confident mistakes by LLMs. To study these two types of VLM mistakes, i.e., language hallucination and visual illusion, we curated HallusionBench, an image-context reasoning benchmark that is still challenging to even GPT-4V and LLaVA-1.5. We provide a detailed analysis of examples in HallusionBench, which sheds novel insights on the illusion or hallucination of VLMs and how to improve them in the future. The benchmark and codebase will be released at https://github.com/tianyi-lab/HallusionBench.