Shashank Gupta

Aman Madaan, Niket Tandon, Prakhar Gupta et al. · allen-ai, cmu

Like humans, large language models (LLMs) do not always generate the best output on their first try. Motivated by how humans refine their written text, we introduce Self-Refine, an approach for improving initial outputs from LLMs through iterative feedback and refinement. The main idea is to generate an initial output using an LLMs; then, the same LLMs provides feedback for its output and uses it to refine itself, iteratively. Self-Refine does not require any supervised training data, additional training, or reinforcement learning, and instead uses a single LLM as the generator, refiner, and feedback provider. We evaluate Self-Refine across 7 diverse tasks, ranging from dialog response generation to mathematical reasoning, using state-of-the-art (GPT-3.5, ChatGPT, and GPT-4) LLMs. Across all evaluated tasks, outputs generated with Self-Refine are preferred by humans and automatic metrics over those generated with the same LLM using conventional one-step generation, improving by ~20% absolute on average in task performance. Our work demonstrates that even state-of-the-art LLMs like GPT-4 can be further improved at test time using our simple, standalone approach.

Ruibo Liu, Guoqing Zheng, Shashank Gupta et al. · deepmind

Pre-trained language models (LMs) have been shown to memorize a substantial amount of knowledge from the pre-training corpora; however, they are still limited in recalling factually correct knowledge given a certain context. Hence, they tend to suffer from counterfactual or hallucinatory generation when used in knowledge-intensive natural language generation (NLG) tasks. Recent remedies to this problem focus on modifying either the pre-training or task fine-tuning objectives to incorporate knowledge, which normally require additional costly training or architecture modification of LMs for practical applications. We present Knowledge Infused Decoding (KID) -- a novel decoding algorithm for generative LMs, which dynamically infuses external knowledge into each step of the LM decoding. Specifically, we maintain a local knowledge memory based on the current context, interacting with a dynamically created external knowledge trie, and continuously update the local memory as a knowledge-aware constraint to guide decoding via reinforcement learning. On six diverse knowledge-intensive NLG tasks, task-agnostic LMs (e.g., GPT-2 and BART) armed with KID outperform many task-optimized state-of-the-art models, and show particularly strong performance in few-shot scenarios over seven related knowledge-infusion techniques. Human evaluation confirms KID's ability to generate more relevant and factual language for the input context when compared with multiple baselines. Finally, KID also alleviates exposure bias and provides stable generation quality when generating longer sequences. Code for KID is available at https://github.com/microsoft/KID.

Shashank Gupta, Vaishnavi Shrivastava, Ameet Deshpande et al.

Recent works have showcased the ability of LLMs to embody diverse personas in their responses, exemplified by prompts like 'You are Yoda. Explain the Theory of Relativity.' While this ability allows personalization of LLMs and enables human behavior simulation, its effect on LLMs' capabilities remains unclear. To fill this gap, we present the first extensive study of the unintended side-effects of persona assignment on the ability of LLMs to perform basic reasoning tasks. Our study covers 24 reasoning datasets, 4 LLMs, and 19 diverse personas (e.g. an Asian person) spanning 5 socio-demographic groups. Our experiments unveil that LLMs harbor deep rooted bias against various socio-demographics underneath a veneer of fairness. While they overtly reject stereotypes when explicitly asked ('Are Black people less skilled at mathematics?'), they manifest stereotypical and erroneous presumptions when asked to answer questions while adopting a persona. These can be observed as abstentions in responses, e.g., 'As a Black person, I can't answer this question as it requires math knowledge', and generally result in a substantial performance drop. Our experiments with ChatGPT-3.5 show that this bias is ubiquitous - 80% of our personas demonstrate bias; it is significant - some datasets show performance drops of 70%+; and can be especially harmful for certain groups - some personas suffer statistically significant drops on 80%+ of the datasets. Overall, all 4 LLMs exhibit this bias to varying extents, with GPT-4-Turbo showing the least but still a problematic amount of bias (evident in 42% of the personas). Further analysis shows that these persona-induced errors can be hard-to-discern and hard-to-avoid. Our findings serve as a cautionary tale that the practice of assigning personas to LLMs - a trend on the rise - can surface their deep-rooted biases and have unforeseeable and detrimental side-effects.

Shashank Gupta, Subhabrata Mukherjee, Krishan Subudhi et al.

Traditional multi-task learning (MTL) methods use dense networks that use the same set of shared weights across several different tasks. This often creates interference where two or more tasks compete to pull model parameters in different directions. In this work, we study whether sparsely activated Mixture-of-Experts (MoE) improve multi-task learning by specializing some weights for learning shared representations and using the others for learning task-specific information. To this end, we devise task-aware gating functions to route examples from different tasks to specialized experts which share subsets of network weights conditioned on the task. This results in a sparsely activated multi-task model with a large number of parameters, but with the same computational cost as that of a dense model. We demonstrate such sparse networks to improve multi-task learning along three key dimensions: (i) transfer to low-resource tasks from related tasks in the training mixture; (ii) sample-efficient generalization to tasks not seen during training by making use of task-aware routing from seen related tasks; (iii) robustness to the addition of unrelated tasks by avoiding catastrophic forgetting of existing tasks.

Shashank Gupta, Xuguang Ai, Ramakanth Kavuluru

End-to-end relation extraction (E2ERE) is an important and realistic application of natural language processing (NLP) in biomedicine. In this paper, we aim to compare three prevailing paradigms for E2ERE using a complex dataset focused on rare diseases involving discontinuous and nested entities. We use the RareDis information extraction dataset to evaluate three competing approaches (for E2ERE): NER $\rightarrow$ RE pipelines, joint sequence to sequence models, and generative pre-trained transformer (GPT) models. We use comparable state-of-the-art models and best practices for each of these approaches and conduct error analyses to assess their failure modes. Our findings reveal that pipeline models are still the best, while sequence-to-sequence models are not far behind; GPT models with eight times as many parameters are worse than even sequence-to-sequence models and lose to pipeline models by over 10 F1 points. Partial matches and discontinuous entities caused many NER errors contributing to lower overall E2E performances. We also verify these findings on a second E2ERE dataset for chemical-protein interactions. Although generative LM-based methods are more suitable for zero-shot settings, when training data is available, our results show that it is better to work with more conventional models trained and tailored for E2ERE. More innovative methods are needed to marry the best of the both worlds from smaller encoder-decoder pipeline models and the larger GPT models to improve E2ERE. As of now, we see that well designed pipeline models offer substantial performance gains at a lower cost and carbon footprint for E2ERE. Our contribution is also the first to conduct E2ERE for the RareDis dataset.

Aviv Brokman, Xuguang Ai, Yuhang Jiang et al.

Objective: Zero-shot methodology promises to cut down on costs of dataset annotation and domain expertise needed to make use of NLP. Generative large language models trained to align with human goals have achieved high zero-shot performance across a wide variety of tasks. As of yet, it is unclear how well these models perform on biomedical relation extraction (RE). To address this knowledge gap, we explore patterns in the performance of OpenAI LLMs across a diverse sampling of RE tasks. Methods: We use OpenAI GPT-4-turbo and OpenAI's reasoning models o1 and GPT-OSS to conduct end-to-end RE experiments on seven datasets. We use the JSON generation capabilities of GPT models to generate structured output in two ways: (1) by defining an explicit schema describing the structure of relations, and (2) using a setting that infers the structure from the prompt language. Results: Our work is the first to study and compare the performance of the GPT-4, o1 and GPT-OSS for the end-to-end zero-shot biomedical RE task across a broad array of datasets. We found the zero-shot performances to be proximal to that of fine-tuned methods. The limitations of this approach are that it performs poorly on instances containing many relations and errs on the boundaries of textual mentions. Conclusion: LLMs exhibit promising zero-shot capabilities in complex biomedical RE tasks, offering competitive performance with reduced dataset curation costs and NLP modeling needs but with increased perpetual compute costs. Addressing the limitations we identify could further boost reliability. The code, data, and prompts for all our experiments are publicly available for additional benchmarking by the community: https://github.com/bionlproc/ZeroShotRE

Shashank Gupta

Question answering is a task that answers factoid questions using a large collection of documents. It aims to provide precise answers in response to the user's questions in natural language. Question answering relies on efficient passage retrieval to select candidate contexts, where traditional sparse vector space models, such as TF-IDF or BM25, are the de facto method. On the web, there is no single article that could provide all the possible answers available on the internet to the question of the problem asked by the user. The existing Dense Passage Retrieval model has been trained on Wikipedia dump from Dec. 20, 2018, as the source documents for answering questions. Question answering (QA) has made big strides with several open-domain and machine comprehension systems built using large-scale annotated datasets. However, in the clinical domain, this problem remains relatively unexplored. According to multiple surveys, Biomedical Questions cannot be answered correctly from Wikipedia Articles. In this work, we work on the existing DPR framework for the biomedical domain and retrieve answers from the Pubmed articles which is a reliable source to answer medical questions. When evaluated on a BioASQ QA dataset, our fine-tuned dense retriever results in a 0.81 F1 score.

Pilhwa Lee, Shashank Gupta

Minimum attention applies the least action principle in the changes of control concerning state and time, first proposed by Brockett. The involved regularization is highly relevant in emulating biological control, such as motor learning. We apply minimum attention in reinforcement learning (RL) as part of the rewards and investigate its connection to meta-learning and stabilization. Specifically, model-based meta-learning with minimum attention is explored in high-dimensional nonlinear dynamics. Ensemble-based model learning and gradient-based meta-policy learning are alternately performed. Empirically, we show that the minimum attention does show outperforming competence in comparison to the state-of-the-art algorithms in model-free and model-based RL, i.e., fast adaptation in few shots and variance reduction from the perturbations of the model and environment. Furthermore, the minimum attention demonstrates the improvement in energy efficiency.

Vaishnavi Shrivastava, Radhika Gaonkar, Shashank Gupta et al.

Fine-tuning pre-trained language models improves the quality of commercial reply suggestion systems, but at the cost of unsustainable training times. Popular training time reduction approaches are resource intensive, thus we explore low-cost model compression techniques like Layer Dropping and Layer Freezing. We demonstrate the efficacy of these techniques in large-data scenarios, enabling the training time reduction for a commercial email reply suggestion system by 42%, without affecting the model relevance or user engagement. We further study the robustness of these techniques to pre-trained model and dataset size ablation, and share several insights and recommendations for commercial applications.

Shashank Gupta, Antonio Robles-Kelly, Mohamed Reda Bouadjenek

Combining symbolic human knowledge with neural networks provides a rule-based ante-hoc explanation of the output. In this paper, we propose feature extracting functions for integrating human knowledge abstracted as logic rules into the predictive behavior of a neural network. These functions are embodied as programming functions, which represent the applicable domain knowledge as a set of logical instructions and provide a modified distribution of independent features on input data. Unlike other existing neural logic approaches, the programmatic nature of these functions implies that they do not require any kind of special mathematical encoding, which makes our method very general and flexible in nature. We illustrate the performance of our approach for sentiment classification and compare our results to those obtained using two baselines.

Anirudh Anil Ojha, Rohin Garg, Shashank Gupta et al.

This paper describes our efforts in tackling Task 5 of SemEval-2020. The task involved detecting a class of textual expressions known as counterfactuals and separating them into their constituent elements. Counterfactual statements describe events that have not or could not have occurred and the possible implications of such events. While counterfactual reasoning is natural for humans, understanding these expressions is difficult for artificial agents due to a variety of linguistic subtleties. Our final submitted approaches were an ensemble of various fine-tuned transformer-based and CNN-based models for the first subtask and a transformer model with dependency tree information for the second subtask. We ranked 4-th and 9-th in the overall leaderboard. We also explored various other approaches that involved the use of classical methods, other neural architectures and the incorporation of different linguistic features.