Language model agents (LMA) recently emerged as a promising paradigm on muti-step decision making tasks, often outperforming humans and other reinforcement learning agents. Despite the promise, their performance on real-world applications that often involve combinations of tasks is still underexplored. In this work, we introduce a new benchmark, called CompWoB -- 50 new compositional web automation tasks reflecting more realistic assumptions. We show that while existing prompted LMAs (gpt-3.5-turbo or gpt-4) achieve 94.0% average success rate on base tasks, their performance degrades to 24.9% success rate on compositional tasks. On the other hand, transferred LMAs (finetuned only on base tasks) show less generalization gap, dropping from 85.4% to 54.8%. By balancing data distribution across tasks, we train a new model, HTML-T5++, that surpasses human-level performance (95.2%) on MiniWoB, and achieves the best zero-shot performance on CompWoB (61.5%). While these highlight the promise of small-scale finetuned and transferred models for task compositionality, their performance further degrades under different instruction compositions changing combinational order. In contrast to the recent remarkable success of LMA, our benchmark and detailed analysis emphasize the necessity of building LMAs that are robust and generalizable to task compositionality for real-world deployment.

Do machines and humans process language in similar ways? Recent research has hinted in the affirmative, finding that brain signals can be effectively predicted using the internal representations of language models (LMs). Although such results are thought to reflect shared computational principles between LMs and human brains, there are also clear differences in how LMs and humans represent and use language. In this work, we systematically explore the divergences between human and machine language processing by examining the differences between LM representations and human brain responses to language as measured by Magnetoencephalography (MEG) across two datasets in which subjects read and listened to narrative stories. Using a data-driven approach, we identify two domains that are not captured well by LMs: social/emotional intelligence and physical commonsense. We then validate these domains with human behavioral experiments and show that fine-tuning LMs on these domains can improve their alignment with human brain responses.

Recently, unified speech-text models, such as SpeechGPT, VioLA, and AudioPaLM, have achieved remarkable performance on various speech tasks. These models discretize speech signals into tokens (speech discretization) and use a shared vocabulary for both text and speech tokens. Then they train a single decoder-only Transformer on a mixture of speech tasks. However, these models rely on the Loss Masking strategy for the ASR task, which ignores the dependency among speech tokens. In this paper, we propose to model speech tokens in an autoregressive way, similar to text. We find that applying the conventional cross-entropy loss on input speech tokens does not consistently improve the ASR performance over the Loss Masking approach. To address this issue, we propose a novel approach denoted Smoothed Label Distillation (SLD), which applies a KL divergence loss with smoothed labels on speech tokens. Our experiments show that SLD effectively models speech tokens and outperforms Loss Masking for decoder-only Transformers in ASR tasks with different speech discretization methods. The source code can be found here: https://github.com/alibaba-damo-academy/SpokenNLP/tree/main/sld

In this paper, we unveil that Language Models (LMs) can acquire new capabilities by assimilating parameters from homologous models without retraining or GPUs. We first introduce DARE to set most delta parameters (i.e., the disparity between fine-tuned and pre-trained parameters) to zeros without affecting the abilities of Supervised Fine-Tuning (SFT) LMs, which randomly Drops delta parameters with a ratio p And REscales the remaining ones by 1/(1 - p) to approximate the original embeddings. Then, we use DARE as a versatile plug-and-play technique to sparsify delta parameters of multiple SFT homologous models for mitigating parameter interference and merge them into a single model by parameter fusing. We experiment with encoder- and decoder-based LMs, showing that: (1) SFT delta parameter value ranges are typically small (within 0.005) with extreme redundancy, and DARE can effortlessly eliminate 90% or even 99% of them. (2) DARE can merge multiple task-specific LMs into one LM with diverse capabilities. For instance, the amalgamation of WizardLM and WizardMath significantly enhances the GSM8K zero-shot accuracy of WizardLM from 2.2 to 66.3, retaining the instruction-following proficiency while surpassing WizardMath's 64.2 performance. Our merged LM also ranks first among models with 7 billion parameters on the Open LLM Leaderboard.

We open-source a state-of-the-art 4B-parameter generative model series for Vietnamese, which includes the base pre-trained monolingual model PhoGPT-4B and its chat variant, PhoGPT-4B-Chat. The base model, PhoGPT-4B, with exactly 3.7B parameters, is pre-trained from scratch on a Vietnamese corpus of 102B tokens, with an 8192 context length, employing a vocabulary of 20480 token types. The chat variant, PhoGPT-4B-Chat, is the modeling output obtained by fine-tuning PhoGPT-4B on a dataset of 70K instructional prompts and their responses, along with an additional 290K conversations. We demonstrate its strong performance compared to previous closed-source and open-source 7B-parameter models. Our PhoGPT models are available at: https://github.com/VinAIResearch/PhoGPT

Recent progress in vision language foundation models has shown their ability to understand multimodal data and resolve complicated vision language tasks, including robotics manipulation. We seek a straightforward way of making use of existing vision-language models (VLMs) with simple fine-tuning on robotics data. To this end, we derive a simple and novel vision-language manipulation framework, dubbed RoboFlamingo, built upon the open-source VLMs, OpenFlamingo. Unlike prior works, RoboFlamingo utilizes pre-trained VLMs for single-step vision-language comprehension, models sequential history information with an explicit policy head, and is slightly fine-tuned by imitation learning only on language-conditioned manipulation datasets. Such a decomposition provides RoboFlamingo the flexibility for open-loop control and deployment on low-performance platforms. By exceeding the state-of-the-art performance with a large margin on the tested benchmark, we show RoboFlamingo can be an effective and competitive alternative to adapt VLMs to robot control. Our extensive experimental results also reveal several interesting conclusions regarding the behavior of different pre-trained VLMs on manipulation tasks. We believe RoboFlamingo has the potential to be a cost-effective and easy-to-use solution for robotics manipulation, empowering everyone with the ability to fine-tune their own robotics policy.

Motivated by recent advances in large language models for Natural Language Processing (NLP), we design a time-series foundation model for forecasting whose out-of-the-box zero-shot performance on a variety of public datasets comes close to the accuracy of state-of-the-art supervised forecasting models for each individual dataset. Our model is based on pretraining a patched-decoder style attention model on a large time-series corpus, and can work well across different forecasting history lengths, prediction lengths and temporal granularities.

The rapid development of large language models (LLMs) has not only provided numerous opportunities but also presented significant challenges. This becomes particularly evident when LLMs inadvertently generate harmful or toxic content, either unintentionally or because of intentional inducement. Conversely, this study proposes a novel alignment technique based on mistake analysis, which deliberately exposes LLMs to erroneous content to learn the reasons for mistakes and how to avoid them. In this case, mistakes are repurposed into valuable data for alignment, effectively helping to avoid the production of erroneous responses. Without external models or human annotations, our method leverages a model's intrinsic ability to discern undesirable mistakes and improves the safety of its generated responses. Experimental results reveal that our method outperforms existing alignment approaches in enhancing model safety while maintaining the overall utility.

Today, using Large-scale generative Language Models (LLMs) it is possible to simulate free responses to interview questions like those traditionally analyzed using qualitative research methods. Qualitative methodology encompasses a broad family of techniques involving manual analysis of open-ended interviews or conversations conducted freely in natural language. Here we consider whether artificial "silicon participants" generated by LLMs may be productively studied using qualitative methods aiming to produce insights that could generalize to real human populations. The key concept in our analysis is algorithmic fidelity, a term introduced by Argyle et al. (2023) capturing the degree to which LLM-generated outputs mirror human sub-populations' beliefs and attitudes. By definition, high algorithmic fidelity suggests latent beliefs elicited from LLMs may generalize to real humans, whereas low algorithmic fidelity renders such research invalid. Here we used an LLM to generate interviews with silicon participants matching specific demographic characteristics one-for-one with a set of human participants. Using framework-based qualitative analysis, we showed the key themes obtained from both human and silicon participants were strikingly similar. However, when we analyzed the structure and tone of the interviews we found even more striking differences. We also found evidence of the hyper-accuracy distortion described by Aher et al. (2023). We conclude that the LLM we tested (GPT-3.5) does not have sufficient algorithmic fidelity to expect research on it to generalize to human populations. However, the rapid pace of LLM research makes it plausible this could change in the future. Thus we stress the need to establish epistemic norms now around how to assess validity of LLM-based qualitative research, especially concerning the need to ensure representation of heterogeneous lived experiences.

The automatic generation of visualizations is an old task that, through the years, has shown more and more interest from the research and practitioner communities. Recently, large language models (LLM) have become an interesting option for supporting generative tasks related to visualization, demonstrating initial promising results. At the same time, several pitfalls, like the multiple ways of instructing an LLM to generate the desired result, the different perspectives leading the generation (code-based, image-based, grammar-based), and the presence of hallucinations even for the visualization generation task, make their usage less affordable than expected. Following similar initiatives for benchmarking LLMs, this paper copes with the problem of modeling the evaluation of a generated visualization through an LLM. We propose a theoretical evaluation stack, EvaLLM, that decomposes the evaluation effort in its atomic components, characterizes their nature, and provides an overview of how to implement and interpret them. We also designed and implemented an evaluation platform that provides a benchmarking resource for the visualization generation task. The platform supports automatic and manual scoring conducted by multiple assessors to support a fine-grained and semantic evaluation based on the EvaLLM stack. Two case studies on GPT3.5-turbo with Code Interpreter and Llama2-70-b models show the benefits of EvaLLM and illustrate interesting results on the current state-of-the-art LLM-generated visualizations.