Prejudice and Caprice: A Statistical Framework for Measuring Social Discrimination in Large Language Models

The growing integration of large language models (LLMs) into social operations amplifies their impact on decisions in crucial areas such as economics, law, education, and healthcare, raising public concerns about these models' discrimination-related safety and reliability. However, prior discrimination measuring frameworks solely assess the average discriminatory behavior of LLMs, often proving inadequate due to the overlook of an additional discrimination-leading factor, i.e., the LLMs' prediction variation across diverse contexts. In this work, we present the Prejudice-Caprice Framework (PCF) that comprehensively measures discrimination in LLMs by considering both their consistently biased preference and preference variation across diverse contexts. Specifically, we mathematically dissect the aggregated contextualized discrimination risk of LLMs into prejudice risk, originating from LLMs' persistent prejudice, and caprice risk, stemming from their generation inconsistency. In addition, we utilize a data-mining approach to gather preference-detecting probes from sentence skeletons, devoid of attribute indications, to approximate LLMs' applied contexts. While initially intended for assessing discrimination in LLMs, our proposed PCF facilitates the comprehensive and flexible measurement of any inductive biases, including knowledge alongside prejudice, across various modality models. We apply our discrimination-measuring framework to 12 common LLMs, yielding intriguing findings: i) modern LLMs demonstrate significant pro-male stereotypes, ii) LLMs' exhibited discrimination correlates with several social and economic factors, iii) prejudice risk dominates the overall discrimination risk and follows a normal distribution, and iv) caprice risk contributes minimally to the overall risk but follows a fat-tailed distribution, suggesting that it is wild risk requiring enhanced surveillance.

HandDiffuse: Generative Controllers for Two-Hand Interactions via Diffusion Models

Existing hands datasets are largely short-range and the interaction is weak due to the self-occlusion and self-similarity of hands, which can not yet fit the need for interacting hands motion generation. To rescue the data scarcity, we propose HandDiffuse12.5M, a novel dataset that consists of temporal sequences with strong two-hand interactions. HandDiffuse12.5M has the largest scale and richest interactions among the existing two-hand datasets. We further present a strong baseline method HandDiffuse for the controllable motion generation of interacting hands using various controllers. Specifically, we apply the diffusion model as the backbone and design two motion representations for different controllers. To reduce artifacts, we also propose Interaction Loss which explicitly quantifies the dynamic interaction process. Our HandDiffuse enables various applications with vivid two-hand interactions, i.e., motion in-betweening and trajectory control. Experiments show that our method outperforms the state-of-the-art techniques in motion generation and can also contribute to data augmentation for other datasets. Our dataset, corresponding codes, and pre-trained models will be disseminated to the community for future research towards two-hand interaction modeling.

Tuna: Instruction Tuning using Feedback from Large Language Models

Instruction tuning of open-source large language models (LLMs) like LLaMA, using direct outputs from more powerful LLMs such as Instruct-GPT and GPT-4, has proven to be a cost-effective way to align model behaviors with human preferences. However, the instruction-tuned model has only seen one response per instruction, lacking the knowledge of potentially better responses. In this paper, we propose finetuning an instruction-tuned LLM using our novel \textit{probabilistic ranking} and \textit{contextual ranking} approaches to increase the likelihood of generating better responses. Probabilistic ranking enables the instruction-tuned model to inherit the relative rankings of high-quality and low-quality responses from the teacher LLM. On the other hand, learning with contextual ranking allows the model to refine its own response distribution using the contextual understanding ability of stronger LLMs. Furthermore, we apply probabilistic ranking and contextual ranking sequentially to the instruction-tuned LLM. The resulting model, which we call \textbf{Tuna}, consistently improves the performance on Super Natural Instructions (119 test tasks), LMentry (25 test tasks), Vicuna QA, and can even obtain better results than several strong reinforcement learning baselines. Our code and data are available at \url{ https://github.com/microsoft/LMOps}.

Momentum Calibration for Text Generation

The input and output of most text generation tasks can be transformed to two sequences of tokens and they can be modeled using sequence-to-sequence learning modeling tools such as Transformers. These models are usually trained by maximizing the likelihood the output text sequence and assumes the input sequence and all gold preceding tokens are given during training, while during inference the model suffers from the exposure bias problem (i.e., it only has access to its previously predicted tokens rather gold tokens during beam search). In this paper, we propose MoCa ({\bf Mo}mentum {\bf Ca}libration) for text generation. MoCa is an online method that dynamically generates slowly evolving (but consistent) samples using a momentum moving average generator with beam search and MoCa learns to align its model scores of these samples with their actual qualities. Experiments on four text generation datasets (i.e., CNN/DailyMail, XSum, SAMSum and Gigaword) show MoCa consistently improves strong pre-trained transformers using vanilla fine-tuning and we achieve the state-of-the-art results on CNN/DailyMail and SAMSum datasets.

Does Debiasing Inevitably Degrade the Model Performance

Gender bias in language models has attracted sufficient attention because it threatens social justice. However, most of the current debiasing methods degraded the model's performance on other tasks while the degradation mechanism is still mysterious. We propose a theoretical framework explaining the three candidate mechanisms of the language model's gender bias. We use our theoretical framework to explain why the current debiasing methods cause performance degradation. We also discover a pathway through which debiasing will not degrade the model performance. We further develop a causality-detection fine-tuning approach to correct gender bias. The numerical experiment demonstrates that our method is able to lead to double dividends: partially mitigating gender bias while avoiding performance degradation.