The enormous scale of state-of-the-art foundation models has limited their accessibility to scientists, because customized experiments at large model sizes require costly hardware and complex engineering that is impractical for most researchers. To alleviate these problems, we introduce NNsight, an open-source Python package with a simple, flexible API that can express interventions on any PyTorch model by building computation graphs. We also introduce NDIF, a collaborative research platform providing researchers access to foundation-scale LLMs via the NNsight API. Code, documentation, and tutorials are available at https://www.nnsight.net.

In recent years, Large Language Models (LLM) have demonstrated impressive capabilities in the field of natural language processing (NLP). This paper explores the application of LLMs in negotiation transcript analysis by the Vanderbilt AI Negotiation Lab. Starting in September 2022, we applied multiple strategies using LLMs from zero shot learning to fine tuning models to in-context learning). The final strategy we developed is explained, along with how to access and use the model. This study provides a sense of both the opportunities and roadblocks for the implementation of LLMs in real life applications and offers a model for how LLMs can be applied to coding in other fields.

Extractive question answering over clinical text is a crucial need to help deal with the deluge of clinical text generated in hospitals. While encoder models (e.g., BERT) have been popular for this reading comprehension task, recently encoder-decoder models (e.g., T5) are on the rise. There is also the emergence of preference optimization techniques to align decoder-only LLMs with human preferences. In this paper, we combine encoder-decoder models with the direct preference optimization (DPO) method to improve over prior state of the art for the RadQA radiology question answering task by 12-15 F1 points. To the best of our knowledge, this effort is the first to show that DPO method also works for reading comprehension via novel heuristics to generate preference data without human inputs.

API call generation is the cornerstone of large language models' tool-using ability that provides access to the larger world. However, existing supervised and in-context learning approaches suffer from high training costs, poor data efficiency, and generated API calls that can be unfaithful to the API documentation and the user's request. To address these limitations, we propose an output-side optimization approach called FANTASE. Two of the unique contributions of FANTASE are its State-Tracked Constrained Decoding (SCD) and Reranking components. SCD dynamically incorporates appropriate API constraints in the form of Token Search Trie for efficient and guaranteed generation faithfulness with respect to the API documentation. The Reranking component efficiently brings in the supervised signal by leveraging a lightweight model as the discriminator to rerank the beam-searched candidate generations of the large language model. We demonstrate the superior performance of FANTASE in API call generation accuracy, inference efficiency, and context efficiency with DSTC8 and API Bank datasets.

This paper investigates the presence of political bias in emotion inference models used for sentiment analysis (SA) in social science research. Machine learning models often reflect biases in their training data, impacting the validity of their outcomes. While previous research has highlighted gender and race biases, our study focuses on political bias - an underexplored yet pervasive issue that can skew the interpretation of text data across a wide array of studies. We conducted a bias audit on a Polish sentiment analysis model developed in our lab. By analyzing valence predictions for names and sentences involving Polish politicians, we uncovered systematic differences influenced by political affiliations. Our findings indicate that annotations by human raters propagate political biases into the model's predictions. To mitigate this, we pruned the training dataset of texts mentioning these politicians and observed a reduction in bias, though not its complete elimination. Given the significant implications of political bias in SA, our study emphasizes caution in employing these models for social science research. We recommend a critical examination of SA results and propose using lexicon-based systems as a more ideologically neutral alternative. This paper underscores the necessity for ongoing scrutiny and methodological adjustments to ensure the reliability and impartiality of the use of machine learning in academic and applied contexts.

For effective interactions with the open world, robots should understand how interactions with known and novel objects help them towards their goal. A key aspect of this understanding lies in detecting an object's affordances, which represent the potential effects that can be achieved by manipulating the object in various ways. Our approach leverages a dialogue of large language models (LLMs) and vision-language models (VLMs) to achieve open-world affordance detection. Given open-vocabulary descriptions of intended actions and effects, the useful objects in the environment are found. By grounding our system in the physical world, we account for the robot's embodiment and the intrinsic properties of the objects it encounters. In our experiments, we have shown that our method produces tailored outputs based on different embodiments or intended effects. The method was able to select a useful object from a set of distractors. Finetuning the VLM for physical properties improved overall performance. These results underline the importance of grounding the affordance search in the physical world, by taking into account robot embodiment and the physical properties of objects.

Constraint Programming (CP) and Machine Learning (ML) face challenges in text generation due to CP's struggle with implementing "meaning'' and ML's difficulty with structural constraints. This paper proposes a solution by combining both approaches and embedding a Large Language Model (LLM) in CP. The LLM handles word generation and meaning, while CP manages structural constraints. This approach builds on GenCP, an improved version of On-the-fly Constraint Programming Search (OTFS) using LLM-generated domains. Compared to Beam Search (BS), a standard NLP method, this combined approach (GenCP with LLM) is faster and produces better results, ensuring all constraints are satisfied. This fusion of CP and ML presents new possibilities for enhancing text generation under constraints.

Foundation models and vision-language pre-training have notably advanced Vision Language Models (VLMs), enabling multimodal processing of visual and linguistic data. However, their performance has been typically assessed on general scene understanding - recognizing objects, attributes, and actions - rather than cultural comprehension. This study introduces CulturalVQA, a visual question-answering benchmark aimed at assessing VLM's geo-diverse cultural understanding. We curate a collection of 2,378 image-question pairs with 1-5 answers per question representing cultures from 11 countries across 5 continents. The questions probe understanding of various facets of culture such as clothing, food, drinks, rituals, and traditions. Benchmarking VLMs on CulturalVQA, including GPT-4V and Gemini, reveals disparity in their level of cultural understanding across regions, with strong cultural understanding capabilities for North America while significantly lower performance for Africa. We observe disparity in their performance across cultural facets too, with clothing, rituals, and traditions seeing higher performances than food and drink. These disparities help us identify areas where VLMs lack cultural understanding and demonstrate the potential of CulturalVQA as a comprehensive evaluation set for gauging VLM progress in understanding diverse cultures.

Direct Preference Optimization (DPO) has become a widely used training method for the instruction fine-tuning of large language models (LLMs). In this work, we explore an under-investigated aspect of DPO - its dependency on the reference model or policy. Such reference policies, typically instantiated as the model to be further fine-tuned, are important since they can impose an upper limit on DPO's effectiveness. Therefore, we address three related research questions in this work. First, we explore the optimal strength of the KL-divergence constraint in DPO, which penalizes deviations from the reference policy, and find that DPO is sensitive to this strength. Next, we examine the necessity of reference policies for instruction fine-tuning by providing both theoretical and empirical comparisons between DPO and related learning objectives, demonstrating DPO's superiority. Additionally, we investigate whether DPO benefits from stronger reference policies, finding that a stronger reference policy can lead to improved performance, but only when it is similar to the model being fine-tuned. Our findings highlight the confounding role of reference policies in DPO and offer insights for best practices, while also identifying open research questions for future studies.

Commercial large language models (LLMs), like OpenAI's GPT-4 powering ChatGPT and Anthropic's Claude 3 Opus, have dominated natural language processing (NLP) benchmarks across different domains. New competing Open-Source alternatives like Mixtral 8x7B or Llama 3 have emerged and seem to be closing the gap while often offering higher throughput and being less costly to use. Open-Source LLMs can also be self-hosted, which makes them interesting for enterprise and clinical use cases where sensitive data should not be processed by third parties. We participated in the 12th BioASQ challenge, which is a retrieval augmented generation (RAG) setting, and explored the performance of current GPT models Claude 3 Opus, GPT-3.5-turbo and Mixtral 8x7b with in-context learning (zero-shot, few-shot) and QLoRa fine-tuning. We also explored how additional relevant knowledge from Wikipedia added to the context-window of the LLM might improve their performance. Mixtral 8x7b was competitive in the 10-shot setting, both with and without fine-tuning, but failed to produce usable results in the zero-shot setting. QLoRa fine-tuning and Wikipedia context did not lead to measurable performance gains. Our results indicate that the performance gap between commercial and open-source models in RAG setups exists mainly in the zero-shot setting and can be closed by simply collecting few-shot examples for domain-specific use cases. The code needed to rerun these experiments is available through GitHub.