Xudong Shen

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi et al. · allen-ai, amazon-science

How well can NLP models generalize to a variety of unseen tasks when provided with task instructions? To address this question, we first introduce Super-NaturalInstructions, a benchmark of 1,616 diverse NLP tasks and their expert-written instructions. Our collection covers 76 distinct task types, including but not limited to classification, extraction, infilling, sequence tagging, text rewriting, and text composition. This large and diverse collection of tasks enables rigorous benchmarking of cross-task generalization under instructions -- training models to follow instructions on a subset of tasks and evaluating them on the remaining unseen ones. Furthermore, we build Tk-Instruct, a transformer model trained to follow a variety of in-context instructions (plain language task definitions or k-shot examples). Our experiments show that Tk-Instruct outperforms existing instruction-following models such as InstructGPT by over 9% on our benchmark despite being an order of magnitude smaller. We further analyze generalization as a function of various scaling parameters, such as the number of observed tasks, the number of instances per task, and model sizes. We hope our dataset and model facilitate future progress towards more general-purpose NLP models.

Ian R. McKenzie, Alexander Lyzhov, Michael Pieler et al. · stanford, utoronto

Work on scaling laws has found that large language models (LMs) show predictable improvements to overall loss with increased scale (model size, training data, and compute). Here, we present evidence for the claim that LMs may show inverse scaling, or worse task performance with increased scale, e.g., due to flaws in the training objective and data. We present empirical evidence of inverse scaling on 11 datasets collected by running a public contest, the Inverse Scaling Prize, with a substantial prize pool. Through analysis of the datasets, along with other examples found in the literature, we identify four potential causes of inverse scaling: (i) preference to repeat memorized sequences over following in-context instructions, (ii) imitation of undesirable patterns in the training data, (iii) tasks containing an easy distractor task which LMs could focus on, rather than the harder real task, and (iv) correct but misleading few-shot demonstrations of the task. We release the winning datasets at https://inversescaling.com/data to allow for further investigation of inverse scaling. Our tasks have helped drive the discovery of U-shaped and inverted-U scaling trends, where an initial trend reverses, suggesting that scaling trends are less reliable at predicting the behavior of larger-scale models than previously understood. Overall, our results suggest that there are tasks for which increased model scale alone may not lead to progress, and that more careful thought needs to go into the data and objectives for training language models.

Xudong Shen, Chao Du, Tianyu Pang et al. · tsinghua

The rapid adoption of text-to-image diffusion models in society underscores an urgent need to address their biases. Without interventions, these biases could propagate a skewed worldview and restrict opportunities for minority groups. In this work, we frame fairness as a distributional alignment problem. Our solution consists of two main technical contributions: (1) a distributional alignment loss that steers specific characteristics of the generated images towards a user-defined target distribution, and (2) adjusted direct finetuning of diffusion model's sampling process (adjusted DFT), which leverages an adjusted gradient to directly optimize losses defined on the generated images. Empirically, our method markedly reduces gender, racial, and their intersectional biases for occupational prompts. Gender bias is significantly reduced even when finetuning just five soft tokens. Crucially, our method supports diverse perspectives of fairness beyond absolute equality, which is demonstrated by controlling age to a $75\%$ young and $25\%$ old distribution while simultaneously debiasing gender and race. Finally, our method is scalable: it can debias multiple concepts at once by simply including these prompts in the finetuning data. We share code and various fair diffusion model adaptors at https://sail-sg.github.io/finetune-fair-diffusion/.

Xudong Shen, Hannah Brown, Jiashu Tao et al.

There is increasing attention being given to how to regulate AI systems. As governing bodies grapple with what values to encapsulate into regulation, we consider the technical half of the question: To what extent can AI experts vet an AI system for adherence to regulatory requirements? We investigate this question through the lens of two public sector procurement checklists, identifying what we can do now, what should be possible with technical innovation, and what requirements need a more interdisciplinary approach.

Ayse Gizem Yasar, Andrew Chong, Evan Dong et al.

As AI technology advances rapidly, concerns over the risks of bigness in digital markets are also growing. The EU's Digital Markets Act (DMA) aims to address these risks. Still, the current framework may not adequately cover generative AI systems that could become gateways for AI-based services. This paper argues for integrating certain AI software as core platform services and classifying certain developers as gatekeepers under the DMA. We also propose an assessment of gatekeeper obligations to ensure they cover generative AI services. As the EU considers generative AI-specific rules and possible DMA amendments, this paper provides insights towards diversity and openness in generative AI services.

Shihao Zou, Xianying Huang, Xudong Shen

Emotion Recognition in Conversation (ERC) plays an important role in driving the development of human-machine interaction. Emotions can exist in multiple modalities, and multimodal ERC mainly faces two problems: (1) the noise problem in the cross-modal information fusion process, and (2) the prediction problem of less sample emotion labels that are semantically similar but different categories. To address these issues and fully utilize the features of each modality, we adopted the following strategies: first, deep emotion cues extraction was performed on modalities with strong representation ability, and feature filters were designed as multimodal prompt information for modalities with weak representation ability. Then, we designed a Multimodal Prompt Transformer (MPT) to perform cross-modal information fusion. MPT embeds multimodal fusion information into each attention layer of the Transformer, allowing prompt information to participate in encoding textual features and being fused with multi-level textual information to obtain better multimodal fusion features. Finally, we used the Hybrid Contrastive Learning (HCL) strategy to optimize the model's ability to handle labels with few samples. This strategy uses unsupervised contrastive learning to improve the representation ability of multimodal fusion and supervised contrastive learning to mine the information of labels with few samples. Experimental results show that our proposed model outperforms state-of-the-art models in ERC on two benchmark datasets.

Keyuan Cheng, Xudong Shen, Yihao Yang et al.

Large language models (LLMs) have shown remarkable capabilities across various software engineering tasks; however, their effectiveness in code migration, adapting code to run in different environments, remains insufficiently studied. In this work, we introduce CODEMENV: Code Migration Across Environment, a new benchmark specifically designed to assess LLMs' abilities in code migration scenarios. CODEMENV consists of 922 examples spanning 19 Python and Java packages, and covers three core tasks: (1) identifying functions incompatible with specific versions, (2) detecting changes in function definitions, and (3) adapting code to target environments. Experimental evaluation with seven LLMs on CODEMENV yields an average pass@1 rate of 26.50%, with GPT-4O achieving the highest score at 43.84%. Key findings include: (i) LLMs tend to be more proficient with newer function versions, which aids in migrating legacy code, and (ii) LLMs sometimes exhibit logical inconsistencies by identifying function changes irrelevant to the intended migration environment. The datasets are available at https://github.com/xdshen-ai/Benchmark-of-Code-Migration.

Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann et al.

Data augmentation is an important component in the robustness evaluation of models in natural language processing (NLP) and in enhancing the diversity of the data they are trained on. In this paper, we present NL-Augmenter, a new participatory Python-based natural language augmentation framework which supports the creation of both transformations (modifications to the data) and filters (data splits according to specific features). We describe the framework and an initial set of 117 transformations and 23 filters for a variety of natural language tasks. We demonstrate the efficacy of NL-Augmenter by using several of its transformations to analyze the robustness of popular natural language models. The infrastructure, datacards and robustness analysis results are available publicly on the NL-Augmenter repository (https://github.com/GEM-benchmark/NL-Augmenter).

Kai Wang, Zhene Zou, Minghao Zhao et al.

Reinforcement learning based recommender systems (RL-based RS) aim at learning a good policy from a batch of collected data, by casting recommendations to multi-step decision-making tasks. However, current RL-based RS research commonly has a large reality gap. In this paper, we introduce the first open-source real-world dataset, RL4RS, hoping to replace the artificial datasets and semi-simulated RS datasets previous studies used due to the resource limitation of the RL-based RS domain. Unlike academic RL research, RL-based RS suffers from the difficulties of being well-validated before deployment. We attempt to propose a new systematic evaluation framework, including evaluation of environment simulation, evaluation on environments, counterfactual policy evaluation, and evaluation on environments built from test set. In summary, the RL4RS (Reinforcement Learning for Recommender Systems), a new resource with special concerns on the reality gaps, contains two real-world datasets, data understanding tools, tuned simulation environments, related advanced RL baselines, batch RL baselines, and counterfactual policy evaluation algorithms. The RL4RS suite can be found at https://github.com/fuxiAIlab/RL4RS. In addition to the RL-based recommender systems, we expect the resource to contribute to research in applied reinforcement learning.

Bertie Vidgen, Adarsh Agrawal, Ahmed M. Ahmed et al. · deepmind, oxford

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

Paul Röttger, Giuseppe Attanasio, Felix Friedrich et al.

Vision-language models (VLMs), which process image and text inputs, are increasingly integrated into chat assistants and other consumer AI applications. Without proper safeguards, however, VLMs may give harmful advice (e.g. how to self-harm) or encourage unsafe behaviours (e.g. to consume drugs). Despite these clear hazards, little work so far has evaluated VLM safety and the novel risks created by multimodal inputs. To address this gap, we introduce MSTS, a Multimodal Safety Test Suite for VLMs. MSTS comprises 400 test prompts across 40 fine-grained hazard categories. Each test prompt consists of a text and an image that only in combination reveal their full unsafe meaning. With MSTS, we find clear safety issues in several open VLMs. We also find some VLMs to be safe by accident, meaning that they are safe because they fail to understand even simple test prompts. We translate MSTS into ten languages, showing non-English prompts to increase the rate of unsafe model responses. We also show models to be safer when tested with text only rather than multimodal prompts. Finally, we explore the automation of VLM safety assessments, finding even the best safety classifiers to be lacking.

Li Yuan, Yi Cai, Xudong Shen et al.

Multimodal Information Extraction (MIE) has gained attention for extracting structured information from multimedia sources. Traditional methods tackle MIE tasks separately, missing opportunities to share knowledge across tasks. Recent approaches unify these tasks into a generation problem using instruction-based T5 models with visual adaptors, optimized through full-parameter fine-tuning. However, this method is computationally intensive, and multi-task fine-tuning often faces gradient conflicts, limiting performance. To address these challenges, we propose collaborative multi-LoRA experts with achievement-based multi-task loss (C-LoRAE) for MIE tasks. C-LoRAE extends the low-rank adaptation (LoRA) method by incorporating a universal expert to learn shared multimodal knowledge from cross-MIE tasks and task-specific experts to learn specialized instructional task features. This configuration enhances the model's generalization ability across multiple tasks while maintaining the independence of various instruction tasks and mitigating gradient conflicts. Additionally, we propose an achievement-based multi-task loss to balance training progress across tasks, addressing the imbalance caused by varying numbers of training samples in MIE tasks. Experimental results on seven benchmark datasets across three key MIE tasks demonstrate that C-LoRAE achieves superior overall performance compared to traditional fine-tuning methods and LoRA methods while utilizing a comparable number of training parameters to LoRA.

Ziwei Xu, Xudong Shen, Yongkang Wong et al.

We propose the Motion Capsule Autoencoder (MCAE), which addresses a key challenge in the unsupervised learning of motion representations: transformation invariance. MCAE models motion in a two-level hierarchy. In the lower level, a spatio-temporal motion signal is divided into short, local, and semantic-agnostic snippets. In the higher level, the snippets are aggregated to form full-length semantic-aware segments. For both levels, we represent motion with a set of learned transformation invariant templates and the corresponding geometric transformations by using capsule autoencoders of a novel design. This leads to a robust and efficient encoding of viewpoint changes. MCAE is evaluated on a novel Trajectory20 motion dataset and various real-world skeleton-based human action datasets. Notably, it achieves better results than baselines on Trajectory20 with considerably fewer parameters and state-of-the-art performance on the unsupervised skeleton-based action recognition task.

Xudong Shen, Yongkang Wong, Mohan Kankanhalli

Motivated by scenarios where data is used for diverse prediction tasks, we study whether fair representation can be used to guarantee fairness for unknown tasks and for multiple fairness notions simultaneously. We consider seven group fairness notions that cover the concepts of independence, separation, and calibration. Against the backdrop of the fairness impossibility results, we explore approximate fairness. We prove that, although fair representation might not guarantee fairness for all prediction tasks, it does guarantee fairness for an important subset of tasks -- the tasks for which the representation is discriminative. Specifically, all seven group fairness notions are linearly controlled by fairness and discriminativeness of the representation. When an incompatibility exists between different fairness notions, fair and discriminative representation hits the sweet spot that approximately satisfies all notions. Motivated by our theoretical findings, we propose to learn both fair and discriminative representations using pretext loss which self-supervises learning, and Maximum Mean Discrepancy as a fair regularizer. Experiments on tabular, image, and face datasets show that using the learned representation, downstream predictions that we are unaware of when learning the representation indeed become fairer for seven group fairness notions, and the fairness guarantees computed from our theoretical results are all valid.