We present Lil-Bevo, our submission to the BabyLM Challenge. We pretrained our masked language models with three ingredients: an initial pretraining with music data, training on shorter sequences before training on longer ones, and masking specific tokens to target some of the BLiMP subtasks. Overall, our baseline models performed above chance, but far below the performance levels of larger LLMs trained on more data. We found that training on short sequences performed better than training on longer sequences.Pretraining on music may help performance marginally, but, if so, the effect seems small. Our targeted Masked Language Modeling augmentation did not seem to improve model performance in general, but did seem to help on some of the specific BLiMP tasks that we were targeting (e.g., Negative Polarity Items). Training performant LLMs on small amounts of data is a difficult but potentially informative task. While some of our techniques showed some promise, more work is needed to explore whether they can improve performance more than the modest gains here. Our code is available at https://github.com/venkatasg/Lil-Bevo and out models at https://huggingface.co/collections/venkatasg/babylm-653591cdb66f4bf68922873a

With LLMs shifting their role from statistical modeling of language to serving as general-purpose AI agents, how should LLM evaluations change? Arguably, a key ability of an AI agent is to flexibly combine, as needed, the basic skills it has learned. The capability to combine skills plays an important role in (human) pedagogy and also in a paper on emergence phenomena (Arora & Goyal, 2023). This work introduces Skill-Mix, a new evaluation to measure ability to combine skills. Using a list of $N$ skills the evaluator repeatedly picks random subsets of $k$ skills and asks the LLM to produce text combining that subset of skills. Since the number of subsets grows like $N^k$, for even modest $k$ this evaluation will, with high probability, require the LLM to produce text significantly different from any text in the training set. The paper develops a methodology for (a) designing and administering such an evaluation, and (b) automatic grading (plus spot-checking by humans) of the results using GPT-4 as well as the open LLaMA-2 70B model. Administering a version of to popular chatbots gave results that, while generally in line with prior expectations, contained surprises. Sizeable differences exist among model capabilities that are not captured by their ranking on popular LLM leaderboards ("cramming for the leaderboard"). Furthermore, simple probability calculations indicate that GPT-4's reasonable performance on $k=5$ is suggestive of going beyond "stochastic parrot" behavior (Bender et al., 2021), i.e., it combines skills in ways that it had not seen during training. We sketch how the methodology can lead to a Skill-Mix based eco-system of open evaluations for AI capabilities of future models.

The ability to learn and refine behavior after deployment has become ever more important for robots as we design them to operate in unstructured environments like households. In this work, we design a new learning system based on large language model (LLM), OLAF, that allows everyday users to teach a robot using verbal corrections when the robot makes mistakes, e.g., by saying "Stop what you're doing. You should move closer to the cup." A key feature of OLAF is its ability to update the robot's visuomotor neural policy based on the verbal feedback to avoid repeating mistakes in the future. This is in contrast to existing LLM-based robotic systems, which only follow verbal commands or corrections but not learn from them. We demonstrate the efficacy of our design in experiments where a user teaches a robot to perform long-horizon manipulation tasks both in simulation and on physical hardware, achieving on average 20.0% improvement in policy success rate. Videos and more results are at https://ut-austin-rpl.github.io/olaf/

Large language models (LLMs) have been widely used as agents to complete different tasks, such as personal assistance or event planning. While most work has focused on cooperation and collaboration between agents, little work explores competition, another important mechanism that fosters the development of society and economy. In this paper, we seek to examine the competition behaviors in LLM-based agents. We first propose a general framework to study the competition between agents. Then, we implement a practical competitive environment using GPT-4 to simulate a virtual town with two types of agents, including restaurant agents and customer agents. Specifically, restaurant agents compete with each other to attract more customers, where the competition fosters them to transform, such as cultivating new operating strategies. The results of our experiments reveal several interesting findings ranging from social learning to Matthew Effect, which aligns well with existing sociological and economic theories. We believe that competition between agents deserves further investigation to help us understand society better. The code will be released soon.

Future wireless communication networks are in a position to move beyond data-centric, device-oriented connectivity and offer intelligent, immersive experiences based on task-oriented connections, especially in the context of the thriving development of pre-trained foundation models (PFM) and the evolving vision of 6G native artificial intelligence (AI). Therefore, redefining modes of collaboration between devices and servers and constructing native intelligence libraries become critically important in 6G. In this paper, we analyze the challenges of achieving 6G native AI from the perspectives of data, intelligence, and networks. Then, we propose a 6G native AI framework based on foundation models, provide a customization approach for intent-aware PFM, present a construction of a task-oriented AI toolkit, and outline a novel cloud-edge-end collaboration paradigm. As a practical use case, we apply this framework for orchestration, achieving the maximum sum rate within a wireless communication system, and presenting preliminary evaluation results. Finally, we outline research directions for achieving native AI in 6G.

We propose a comprehensive study of one-stage elicitation techniques for querying a large pre-trained generative transformer (GPT-3.5-turbo) in the rhetorical role prediction task of legal cases. This task is known as requiring textual context to be addressed. Our study explores strategies such as zero-few shots, task specification with definitions and clarification of annotation ambiguities, textual context and reasoning with general prompts and specific questions. We show that the number of examples, the definition of labels, the presentation of the (labelled) textual context and specific questions about this context have a positive influence on the performance of the model. Given non-equivalent test set configurations, we observed that prompting with a few labelled examples from direct context can lead the model to a better performance than a supervised fined-tuned multi-class classifier based on the BERT encoder (weighted F1 score of = 72%). But there is still a gap to reach the performance of the best systems = 86%) in the LegalEval 2023 task which, on the other hand, require dedicated resources, architectures and training.

Can a machine understand the meanings of natural language? Recent developments in the generative large language models (LLMs) of artificial intelligence have led to the belief that traditional philosophical assumptions about machine understanding of language need to be revised. This article critically evaluates the prevailing tendency to regard machine language performance as mere syntactic manipulation and the simulation of understanding, which is only partial and very shallow, without sufficient referential grounding in the world. The aim is to highlight the conditions crucial to attributing natural language understanding to state-of-the-art LLMs, where it can be legitimately argued that LLMs not only use syntax but also semantics, their understanding not being simulated but duplicated; and determine how they ground the meanings of linguistic expressions.

Despite remarkable advances that large language models have achieved in chatbots, maintaining a non-toxic user-AI interactive environment has become increasingly critical nowadays. However, previous efforts in toxicity detection have been mostly based on benchmarks derived from social media content, leaving the unique challenges inherent to real-world user-AI interactions insufficiently explored. In this work, we introduce ToxicChat, a novel benchmark based on real user queries from an open-source chatbot. This benchmark contains the rich, nuanced phenomena that can be tricky for current toxicity detection models to identify, revealing a significant domain difference compared to social media content. Our systematic evaluation of models trained on existing toxicity datasets has shown their shortcomings when applied to this unique domain of ToxicChat. Our work illuminates the potentially overlooked challenges of toxicity detection in real-world user-AI conversations. In the future, ToxicChat can be a valuable resource to drive further advancements toward building a safe and healthy environment for user-AI interactions.

Simulation is an invaluable tool for developing and evaluating controllers for self-driving cars. Current simulation frameworks are driven by highly-specialist domain specific languages, and so a natural language interface would greatly enhance usability. But there is often a gap, consisting of tacit assumptions the user is making, between a concise English utterance and the executable code that captures the user's intent. In this paper we describe a system that addresses this issue by supporting an extended multimodal interaction: the user can follow up prior instructions with refinements or revisions, in reaction to the simulations that have been generated from their utterances so far. We use Large Language Models (LLMs) to map the user's English utterances in this interaction into domain-specific code, and so we explore the extent to which LLMs capture the context sensitivity that's necessary for computing the speaker's intended message in discourse.

Building upon the considerable advances in Large Language Models (LLMs), we are now equipped to address more sophisticated tasks demanding a nuanced understanding of cross-cultural contexts. A key example is recipe adaptation, which goes beyond simple translation to include a grasp of ingredients, culinary techniques, and dietary preferences specific to a given culture. We introduce a new task involving the translation and cultural adaptation of recipes between Chinese and English-speaking cuisines. To support this investigation, we present CulturalRecipes, a unique dataset comprised of automatically paired recipes written in Mandarin Chinese and English. This dataset is further enriched with a human-written and curated test set. In this intricate task of cross-cultural recipe adaptation, we evaluate the performance of various methods, including GPT-4 and other LLMs, traditional machine translation, and information retrieval techniques. Our comprehensive analysis includes both automatic and human evaluation metrics. While GPT-4 exhibits impressive abilities in adapting Chinese recipes into English, it still lags behind human expertise when translating English recipes into Chinese. This underscores the multifaceted nature of cultural adaptations. We anticipate that these insights will significantly contribute to future research on culturally-aware language models and their practical application in culturally diverse contexts.