Akshat Gupta

Akshat Gupta, Xiaoyang Song, Gopala Anumanchipalli

As large language models (LLM) evolve in their capabilities, various recent studies have tried to quantify their behavior using psychological tools created to study human behavior. One such example is the measurement of "personality" of LLMs using self-assessment personality tests developed to measure human personality. Yet almost none of these works verify the applicability of these tests on LLMs. In this paper, we analyze the reliability of LLM personality scores obtained from self-assessment personality tests using two simple experiments. We first introduce the property of prompt sensitivity, where three semantically equivalent prompts representing three intuitive ways of administering self-assessment tests on LLMs are used to measure the personality of the same LLM. We find that all three prompts lead to very different personality scores, a difference that is statistically significant for all traits in a large majority of scenarios. We then introduce the property of option-order symmetry for personality measurement of LLMs. Since most of the self-assessment tests exist in the form of multiple choice question (MCQ) questions, we argue that the scores should also be robust to not just the prompt template but also the order in which the options are presented. This test unsurprisingly reveals that the self-assessment test scores are not robust to the order of the options. These simple tests, done on ChatGPT and three Llama2 models of different sizes, show that self-assessment personality tests created for humans are unreliable measures of personality in LLMs.

Mert Cemri, Shubham Agrawal, Akshat Gupta et al.

The paradigm of automated program generation is shifting from one-shot generation to inference-time search, where Large Language Models (LLMs) function as semantic mutation operators within evolutionary loops. While effective, these systems are currently governed by static schedules that fail to account for the non-stationary dynamics of the search process. This rigidity results in substantial computational waste, as resources are indiscriminately allocated to stagnating populations while promising frontiers remain under-exploited. We introduce AdaEvolve, a framework that reformulates LLM-driven evolution as a hierarchical adaptive optimization problem. AdaEvolve uses an "accumulated improvement signal" to unify decisions across three levels: Local Adaptation, which dynamically modulates the exploration intensity within a population of solution candidates; Global Adaptation, which routes the global resource budget via bandit-based scheduling across different solution candidate populations; and Meta-Guidance which generates novel solution tactics based on the previously generated solutions and their corresponding improvements when the progress stalls. We demonstrate that AdaEvolve consistently outperforms the open-sourced baselines across 185 different open-ended optimization problems including combinatorial, systems optimization and algorithm design problems.

Fazal Mittu, Yihuan Bu, Akshat Gupta et al.

While the language modeling objective has been shown to be deeply connected with compression, it is surprising that modern LLMs are not employed in practical text compression systems. In this paper, we provide an in-depth analysis of neural network and transformer-based compression techniques to answer this question. We compare traditional text compression systems with neural network and LLM-based text compression methods. Although LLM-based systems significantly outperform conventional compression methods, they are highly impractical. Specifically, LLMZip, a recent text compression system using Llama3-8B requires 9.5 days to compress just 10 MB of text, although with huge improvements in compression ratios. To overcome this, we present FineZip - a novel LLM-based text compression system that combines ideas of online memorization and dynamic context to reduce the compression time immensely. FineZip can compress the above corpus in approximately 4 hours compared to 9.5 days, a 54 times improvement over LLMZip and comparable performance. FineZip outperforms traditional algorithmic compression methods with a large margin, improving compression ratios by approximately 50\%. With this work, we take the first step towards making lossless text compression with LLMs a reality. While FineZip presents a significant step in that direction, LLMs are still not a viable solution for large-scale text compression. We hope our work paves the way for future research and innovation to solve this problem.

Akshat Gupta

Since the introduction of ChatGPT and GPT-4, these models have been tested across a large number of tasks. Their adeptness across domains is evident, but their aptitude in playing games, and specifically their aptitude in the realm of poker has remained unexplored. Poker is a game that requires decision making under uncertainty and incomplete information. In this paper, we put ChatGPT and GPT-4 through the poker test and evaluate their poker skills. Our findings reveal that while both models display an advanced understanding of poker, encompassing concepts like the valuation of starting hands, playing positions and other intricacies of game theory optimal (GTO) poker, both ChatGPT and GPT-4 are NOT game theory optimal poker players. Profitable strategies in poker are evaluated in expectations over large samples. Through a series of experiments, we first discover the characteristics of optimal prompts and model parameters for playing poker with these models. Our observations then unveil the distinct playing personas of the two models. We first conclude that GPT-4 is a more advanced poker player than ChatGPT. This exploration then sheds light on the divergent poker tactics of the two models: ChatGPT's conservativeness juxtaposed against GPT-4's aggression. In poker vernacular, when tasked to play GTO poker, ChatGPT plays like a nit, which means that it has a propensity to only engage with premium hands and folds a majority of hands. When subjected to the same directive, GPT-4 plays like a maniac, showcasing a loose and aggressive style of play. Both strategies, although relatively advanced, are not game theory optimal.

Akshat Gupta, Atahan Ozdemir, Gopala Anumanchipalli

This paper presents a novel geometric interpretation of LayerNorm and explores how LayerNorm influences the norm and orientation of hidden vectors in the representation space. With these geometric insights, we prepare the foundation for comparing LayerNorm with RMSNorm. We show that the definition of LayerNorm is innately linked to the uniform vector, defined as $\boldsymbol{1} = [1, 1, 1, 1, \cdots, 1]^T \in \mathbb{R}^d$. We then show that the standardization step in LayerNorm can be understood in three simple steps: (i) remove the component of a vector along the uniform vector, (ii) normalize the remaining vector, and (iii) scale the resultant vector by $\sqrt{d}$, where $d$ is the dimensionality of the representation space. We also provide additional insights into how LayerNorm operates at inference time. Finally, we compare the hidden representations of LayerNorm-based LLMs with models trained using RMSNorm and show that all LLMs naturally operate orthogonal to the uniform vector at inference time, that is, on average they do not have a component along the uniform vector during inference. This presents the first mechanistic evidence that removing the component along the uniform vector in LayerNorm is a redundant step. These results advocate for using RMSNorm over LayerNorm which is also more computationally efficient.

Akshat Gupta

With their increasing size, large language models (LLMs) are becoming increasingly good at language understanding tasks. But even with high performance on specific downstream task, LLMs fail at simple linguistic tests for negation or quantifier understanding. Previous work on quantifier understanding in LLMs show inverse scaling in understanding few-type quantifiers. In this paper, we question the claims of of previous work and show that it is a result of inappropriate testing methodology. We also present alternate methods to measure quantifier comprehension in LLMs and show that LLMs are able to better understand the difference between the meaning of few-type and most-type quantifiers as their size increases, although they are not particularly good at it. We also observe inverse scaling for most-type quantifier understanding, which is contrary to human psycho-linguistic experiments and previous work, where the model's understanding of most-type quantifier gets worse as the model size increases. We do this evaluation on models ranging from 125M-175B parameters, which suggests that LLMs do not do as well as expected with quantifiers. We also discuss the possible reasons for this and the relevance of quantifier understanding in evaluating language understanding in LLMs.

Akshat Gupta

Spoken dialog systems are slowly becoming and integral part of the human experience due to their various advantages over textual interfaces. Spoken language understanding (SLU) systems are fundamental building blocks of spoken dialog systems. But creating SLU systems for low resourced languages is still a challenge. In a large number of low resourced language, we don't have access to enough data to build automatic speech recognition (ASR) technologies, which are fundamental to any SLU system. Also, ASR based SLU systems do not generalize to unwritten languages. In this paper, we present a series of experiments to explore extremely low-resourced settings where we perform intent classification with systems trained on as low as one data-point per intent and with only one speaker in the dataset. We also work in a low-resourced setting where we do not use language specific ASR systems to transcribe input speech, which compounds the challenge of building SLU systems to simulate a true low-resourced setting. We test our system on Belgian Dutch (Flemish) and English and find that using phonetic transcriptions to make intent classification systems in such low-resourced setting performs significantly better than using speech features. Specifically, when using a phonetic transcription based system over a feature based system, we see average improvements of 12.37% and 13.08% for binary and four-class classification problems respectively, when averaged over 49 different experimental settings.

Immanuel Abdi, Akshat Gupta, Micah Mok et al.

One of the biggest missing capabilities in current AI systems is the ability to learn continuously after deployment. Implementing such continually learning systems have several challenges, one of which is the large memory requirement of gradient-based algorithms that are used to train state-of-the-art LLMs. Evolutionary Strategies (ES) have recently re-emerged as a gradient-free alternative to traditional learning algorithms and have shown encouraging performance on specific tasks in LLMs. In this paper, we perform a comprehensive analysis of ES and specifically evaluate its forgetting curves when training for an increasing number of update steps. We first find that ES is able to reach performance numbers close to GRPO for math and reasoning tasks with a comparable compute budget. However, and most importantly for continual learning, the performance gains in ES is accompanied by significant forgetting of prior abilities, limiting its applicability for training models online. We also explore the reason behind this behavior and show that the updates made using ES are much less sparse and have orders of magnitude larger $\ell_2$ norm compared to corresponding GRPO updates, explaining the contrasting forgetting curves between the two algorithms. With this study, we aim to highlight the issue of forgetting in gradient-free algorithms like ES and hope to inspire future work to mitigate these issues.

Alec Candidato, Akshat Gupta, Xiaomo Liu et al.

This paper presents our submission for the SMM4H 2022-Shared Task on the classification of self-reported intimate partner violence on Twitter (in English). The goal of this task was to accurately determine if the contents of a given tweet demonstrated someone reporting their own experience with intimate partner violence. The submitted system is an ensemble of five RoBERTa models each weighted by their respective F1-scores on the validation data-set. This system performed 13% better than the baseline and was the best performing system overall for this shared task.

Zihan Wang, Qi Meng, HaiFeng Lan et al.

Speech emotion recognition (SER) classifies audio into emotion categories such as Happy, Angry, Fear, Disgust and Neutral. While Speech Emotion Recognition (SER) is a common application for popular languages, it continues to be a problem for low-resourced languages, i.e., languages with no pretrained speech-to-text recognition models. This paper firstly proposes a language-specific model that extract emotional information from multiple pre-trained speech models, and then designs a multi-domain model that simultaneously performs SER for various languages. Our multidomain model employs a multi-gating mechanism to generate unique weighted feature combination for each language, and also searches for specific neural network structure for each language through a neural architecture search module. In addition, we introduce a contrastive auxiliary loss to build more separable representations for audio data. Our experiments show that our model raises the state-of-the-art accuracy by 3% for German and 14.3% for French.

Antonio Lopardo, Avyukth Harish, Catherine Arnett et al.

Weight tying, i.e. sharing parameters between input and output embedding matrices, is common practice in language model design, yet its impact on the learned embedding space remains poorly understood. In this paper, we show that tied embedding matrices align more closely with output (unembedding) matrices than with input embeddings of comparable untied models, indicating that the shared matrix is shaped primarily for output prediction rather than input representation. This unembedding bias arises because output gradients dominate early in training. Using tuned lens analysis, we show this negatively affects early-layer computations, which contribute less effectively to the residual stream. Scaling input gradients during training reduces this bias, providing causal evidence for the role of gradient imbalance. This is mechanistic evidence that weight tying optimizes the embedding matrix for output prediction, compromising its role in input representation. These results help explain why weight tying can harm performance at scale and have implications for training smaller LLMs, where the embedding matrix contributes substantially to total parameter count.

Akshat Gupta, Xiaomo Liu, Sameena Shah

Domain adaptation is an important and widely studied problem in natural language processing. A large body of literature tries to solve this problem by adapting models trained on the source domain to the target domain. In this paper, we instead solve this problem from a dataset perspective. We modify the source domain dataset with simple lexical transformations to reduce the domain shift between the source dataset distribution and the target dataset distribution. We find that models trained on the transformed source domain dataset performs significantly better than zero-shot models. Using our proposed transformations to convert standard English to tweets, we reach an unsupervised part-of-speech (POS) tagging accuracy of 92.14% (from 81.54% zero shot accuracy), which is only slightly below the supervised performance of 94.45%. We also use our proposed transformations to synthetically generate tweets and augment the Twitter dataset to achieve state-of-the-art performance for POS tagging.

Sirui Xu, Dongting Li, Yucheng Zhang et al.

While large-scale human motion capture datasets have advanced human motion generation, modeling and generating dynamic 3D human-object interactions (HOIs) remain challenging due to dataset limitations. Existing datasets often lack extensive, high-quality motion and annotation and exhibit artifacts such as contact penetration, floating, and incorrect hand motions. To address these issues, we introduce InterAct, a large-scale 3D HOI benchmark featuring dataset and methodological advancements. First, we consolidate and standardize 21.81 hours of HOI data from diverse sources, enriching it with detailed textual annotations. Second, we propose a unified optimization framework to enhance data quality by reducing artifacts and correcting hand motions. Leveraging the principle of contact invariance, we maintain human-object relationships while introducing motion variations, expanding the dataset to 30.70 hours. Third, we define six benchmarking tasks and develop a unified HOI generative modeling perspective, achieving state-of-the-art performance. Extensive experiments validate the utility of our dataset as a foundational resource for advancing 3D human-object interaction generation. To support continued research in this area, the dataset is publicly available at https://github.com/wzyabcas/InterAct, and will be actively maintained.

Anant Singh, Akshat Gupta

Recent advancements in transformer-based speech representation models have greatly transformed speech processing. However, there has been limited research conducted on evaluating these models for speech emotion recognition (SER) across multiple languages and examining their internal representations. This article addresses these gaps by presenting a comprehensive benchmark for SER with eight speech representation models and six different languages. We conducted probing experiments to gain insights into inner workings of these models for SER. We find that using features from a single optimal layer of a speech model reduces the error rate by 32\% on average across seven datasets when compared to systems where features from all layers of speech models are used. We also achieve state-of-the-art results for German and Persian languages. Our probing results indicate that the middle layers of speech models capture the most important emotional information for speech emotion recognition.

Akshat Gupta, Anurag Rao, Gopala Anumanchipalli

Editing knowledge in large language models is an attractive capability to have which allows us to correct incorrectly learnt facts during pre-training, as well as update the model with an ever-growing list of new facts. While existing model editing techniques have shown promise, they are usually evaluated using metrics for reliability, specificity and generalization over one or few edits. We argue that for model editing to have practical utility, we must be able to make multiple edits to the same model. With this in mind, we evaluate the current model editing methods at scale, focusing on two state of the art methods: ROME and MEMIT. We find that as the model is edited sequentially with multiple facts, it continually forgets previously edited facts and the ability to perform downstream tasks. This forgetting happens in two phases -- an initial gradual but progressive forgetting phase followed by abrupt or catastrophic forgetting phase. Both gradual and catastrophic forgetting limit the usefulness of model editing methods at scale -- the former making model editing less effective as multiple edits are made to the model while the latter caps the scalability of such model editing methods. Our analysis also highlights other key limitations of ROME and MEMIT at scale. With our work, we push for the development and evaluation of model editing methods keeping scalability in mind.

Akshat Gupta, Dev Sajnani, Gopala Anumanchipalli

ROME and MEMIT are largely believed to be two different model editing algorithms, with the major difference between them being the ability to perform batched edits. In this paper, we unify these two algorithms under a single conceptual umbrella, optimizing for the same goal, which we call the preservation-memorization objective. ROME uses an equality constraint to optimize this objective to perform one edit at a time, whereas MEMIT employs a more flexible least-square constraint that allows for batched edits. We generalize ROME and enable batched editing with equality constraint in the form of EMMET - an Equality-constrained Mass Model Editing algorithm for Transformers, a new batched memory-editing algorithm. EMMET can perform batched-edits up to a batch-size of 10,000, with very similar performance to MEMIT across multiple dimensions. With the introduction of EMMET, we truly unify ROME and MEMIT and show that both algorithms are equivalent in terms of their optimization objective, their abilities (singular and batched editing), their model editing performance and their limitations.

Akshat Gupta, Sidharth Baskaran, Gopala Anumanchipalli

Recent work using Rank-One Model Editing (ROME), a popular model editing method, has shown that there are certain facts that the algorithm is unable to edit without breaking the model. Such edits have previously been called disabling edits. These disabling edits cause immediate model collapse and limits the use of ROME for sequential editing. In this paper, we show that disabling edits are an artifact of irregularities in the implementation of ROME. With this paper, we provide a more stable implementation ROME, which we call r-ROME and show that model collapse is no longer observed when making large scale sequential edits with r-ROME, while further improving generalization and locality of model editing compared to the original implementation of ROME. We also provide a detailed mathematical explanation of the reason behind disabling edits.

Akshat Gupta, Laxman Singh Tomar, Ridhima Garg

Cyber attacks deceive machines into believing something that does not exist in the first place. However, there are some to which even humans fall prey. One such famous attack that attackers have used over the years to exploit the vulnerability of vision is known to be a Homoglyph attack. It employs a primary yet effective mechanism to create illegitimate domains that are hard to differentiate from legit ones. Moreover, as the difference is pretty indistinguishable for a user to notice, they cannot stop themselves from clicking on these homoglyph domain names. In many cases, that results in either information theft or malware attack on their systems. Existing approaches use simple, string-based comparison techniques applied in primary language-based tasks. Although they are impactful to some extent, they usually fail because they are not robust to different types of homoglyphs and are computationally not feasible because of their time requirement proportional to the string length. Similarly, neural network-based approaches are employed to determine real domain strings from fake ones. Nevertheless, the problem with both methods is that they require paired sequences of real and fake domain strings to work with, which is often not the case in the real world, as the attacker only sends the illegitimate or homoglyph domain to the vulnerable user. Therefore, existing approaches are not suitable for practical scenarios in the real world. In our work, we created GlyphNet, an image dataset that contains 4M domains, both real and homoglyphs. Additionally, we introduce a baseline method for a homoglyph attack detection system using an attention-based convolutional Neural Network. We show that our model can reach state-of-the-art accuracy in detecting homoglyph attacks with a 0.93 AUC on our dataset.

Xiaoyang Song, Yuta Adachi, Jessie Feng et al.

As Large Language Models (LLMs) are integrated with human daily applications rapidly, many societal and ethical concerns are raised regarding the behavior of LLMs. One of the ways to comprehend LLMs' behavior is to analyze their personalities. Many recent studies quantify LLMs' personalities using self-assessment tests that are created for humans. Yet many critiques question the applicability and reliability of these self-assessment tests when applied to LLMs. In this paper, we investigate LLM personalities using an alternate personality measurement method, which we refer to as the external evaluation method, where instead of prompting LLMs with multiple-choice questions in the Likert scale, we evaluate LLMs' personalities by analyzing their responses toward open-ended situational questions using an external machine learning model. We first fine-tuned a Llama2-7B model as the MBTI personality predictor that outperforms the state-of-the-art models as the tool to analyze LLMs' responses. Then, we prompt the LLMs with situational questions and ask them to generate Twitter posts and comments, respectively, in order to assess their personalities when playing two different roles. Using the external personality evaluation method, we identify that the obtained personality types for LLMs are significantly different when generating posts versus comments, whereas humans show a consistent personality profile in these two different situations. This shows that LLMs can exhibit different personalities based on different scenarios, thus highlighting a fundamental difference between personality in LLMs and humans. With our work, we call for a re-evaluation of personality definition and measurement in LLMs.

Junsang Yoon, Akshat Gupta, Gopala Anumanchipalli

This study presents a targeted model editing analysis focused on the latest large language model, Llama-3. We explore the efficacy of popular model editing techniques - ROME, MEMIT, and EMMET, which are designed for precise layer interventions. We identify the most effective layers for targeted edits through an evaluation that encompasses up to 4096 edits across three distinct strategies: sequential editing, batch editing, and a hybrid approach we call as sequential-batch editing. Our findings indicate that increasing edit batch-sizes may degrade model performance more significantly than using smaller edit batches sequentially for equal number of edits. With this, we argue that sequential model editing is an important component for scaling model editing methods and future research should focus on methods that combine both batched and sequential editing. This observation suggests a potential limitation in current model editing methods which push towards bigger edit batch sizes, and we hope it paves way for future investigations into optimizing batch sizes and model editing performance.

Richard Zhuang, Akshat Gupta, Richard Yang et al.

We introduce PokerBench - a benchmark for evaluating the poker-playing abilities of large language models (LLMs). As LLMs excel in traditional NLP tasks, their application to complex, strategic games like poker poses a new challenge. Poker, an incomplete information game, demands a multitude of skills such as mathematics, reasoning, planning, strategy, and a deep understanding of game theory and human psychology. This makes Poker the ideal next frontier for large language models. PokerBench consists of a comprehensive compilation of 11,000 most important scenarios, split between pre-flop and post-flop play, developed in collaboration with trained poker players. We evaluate prominent models including GPT-4, ChatGPT 3.5, and various Llama and Gemma series models, finding that all state-of-the-art LLMs underperform in playing optimal poker. However, after fine-tuning, these models show marked improvements. We validate PokerBench by having models with different scores compete with each other, demonstrating that higher scores on PokerBench lead to higher win rates in actual poker games. Through gameplay between our fine-tuned model and GPT-4, we also identify limitations of simple supervised fine-tuning for learning optimal playing strategy, suggesting the need for more advanced methodologies for effectively training language models to excel in games. PokerBench thus presents a unique benchmark for a quick and reliable evaluation of the poker-playing ability of LLMs as well as a comprehensive benchmark to study the progress of LLMs in complex game-playing scenarios.

Saptarshi Dasgupta, Akshat Gupta, Shreshth Tuli et al.

Manipulating unseen objects is challenging without a 3D representation, as objects generally have occluded surfaces. This requires physical interaction with objects to build their internal representations. This paper presents an approach that enables a robot to rapidly learn the complete 3D model of a given object for manipulation in unfamiliar orientations. We use an ensemble of partially constructed NeRF models to quantify model uncertainty to determine the next action (a visual or re-orientation action) by optimizing informativeness and feasibility. Further, our approach determines when and how to grasp and re-orient an object given its partial NeRF model and re-estimates the object pose to rectify misalignments introduced during the interaction. Experiments with a simulated Franka Emika Robot Manipulator operating in a tabletop environment with benchmark objects demonstrate an improvement of (i) 14% in visual reconstruction quality (PSNR), (ii) 20% in the geometric/depth reconstruction of the object surface (F-score) and (iii) 71% in the task success rate of manipulating objects a-priori unseen orientations/stable configurations in the scene; over current methods. The project page can be found here: https://actnerf.github.io.

Alex Li, Ilyas Bankole-Hameed, Ranadeep Singh et al.

In hope of expanding training data, researchers often want to merge two or more datasets that are created using different labeling schemes. This paper considers two datasets that label part-of-speech (POS) tags under different tagging schemes and leverage the supervised labels of one dataset to help generate labels for the other dataset. This paper further discusses the theoretical difficulties of this approach and proposes a novel supervised architecture employing Transformers to tackle the problem of consolidating two completely disjoint datasets. The results diverge from initial expectations and discourage exploration into the use of disjoint labels to consolidate datasets with different labels.

Akshat Gupta, Phudish Prateepamornkul, Maochuan Lu et al.

Knowledge editing is a promising way to improve factuality in large language models, but recent studies have shown significant model degradation during sequential editing. In this paper, we formalize the popular locate-then-edit methods as a two-step fine-tuning process, allowing us to precisely identify the root cause of this degradation. We show that model degradation occurs due to (1) over-optimization of internal activations and (2) continuous norm-growth of edited matrices. To mitigate these issues, we introduce two regularization techniques: (1) Most-Probable Early Stopping (MPES) and (2) explicit Frobenius norm-constraint. We demonstrate that applying these simple yet effective regularization techniques at key points in the editing process can substantially mitigate model degradation. Combining these regularization methods enables scaling locate-then-edit methods to 10,000 edits while reducing editing time by 42-61%. These results show that targeted regularization is essential for lifelong knowledge editing.

Akshat Gupta, Jay Yeung, Gopala Anumanchipalli et al.

Growing evidence suggests that large language models do not use their depth uniformly, yet we still lack a fine-grained understanding of their layer-wise prediction dynamics. In this paper, we trace the intermediate representations of several open-weight models during inference and reveal a structured and nuanced use of depth. Specifically, we propose a "Guess-then-Refine" framework that explains how LLMs internally structure their computations to make predictions. We first show that the top-ranked predictions in early LLM layers are composed primarily of high-frequency tokens, which act as statistical guesses proposed by the model early on due to the lack of appropriate contextual information. As contextual information develops deeper into the model, these initial guesses get refined into contextually appropriate tokens. Even high-frequency token predictions from early layers get refined >70% of the time, indicating that correct token prediction is not "one-and-done". We then go beyond frequency-based prediction to examine the dynamic usage of layer depth across three case studies. (i) Part-of-speech analysis shows that function words are, on average, the earliest to be predicted correctly. (ii) Fact recall task analysis shows that, in a multi-token answer, the first token requires more computational depth than the rest. (iii) Multiple-choice task analysis shows that the model identifies the format of the response within the first half of the layers, but finalizes its response only toward the end. Together, our results provide a detailed view of depth usage in LLMs, shedding light on the layer-by-layer computations that underlie successful predictions and providing insights for future works to improve computational efficiency in transformer-based models.

Zixun Chen, Petr Babkin, Akshat Gupta et al.

Dialogue is one of the landmark abilities of large language models (LLMs). Despite its ubiquity, few studies actually distinguish specific ingredients underpinning dialogue behavior emerging during post-training. We employ a comprehensive suite of model-based metrics, each targeting a distinct fine-grained aspect of dialogue, motivated by linguistic theory. We evaluate how the performance of pre-trained Pythia models changes with respect to each of those dimensions, depending on model size and as a result of supervised fine-tuning on conversational datasets. We observe only a mild impact of raw model size on most metrics, whereas fine-tuning quickly saturates the scores for all but the smallest models tested. Somewhat contrary to our expectations, many metrics show very similar trends, especially if they are all rooted in the same evaluator model, which raises the question of their reliability in measuring a specific dimension. To that end, we conduct additional analyses of score distributions, metric correlations, and term frequencies in generated responses to help explain our observations.

Akshat Gupta, Maochuan Lu, Thomas Hartvigsen et al.

Knowledge editing methods like MEMIT are able to make data and compute efficient updates of factual knowledge by using a single sentence to update facts and their consequences. However, what is often overlooked is a "precomputation step", which requires a one-time but significant computational cost. The authors of MEMIT originally precompute approximately 44 million hidden vectors per edited layer, which requires a forward pass over 44 million tokens. For GPT-J (6B), this precomputation step takes 36 hours on a single GPU, while it takes approximately 40 hours for Llama2-7B. Additionally, this precomputation time grows with model size. In this paper, we show that this excessive computational cost is unnecessary. Knowledge editing using MEMIT and related methods, such as ROME and EMMET, can be performed by pre-computing a very small portion of the 44 million hidden vectors. We first present the theoretical minimum number of hidden vector precomputation required for solutions of these editing methods to exist. We then empirically show that knowledge editing using these methods can be done by pre-computing significantly fewer hidden vectors. Specifically, we show that the precomputation step can be done with less than 0.3% of the originally stipulated number of hidden vectors. This saves a significant amount of precomputation time and allows users to begin editing new models within a few minutes.

Svetlana Churina, Akshat Gupta, Insyirah Mujtahid et al.

Code-mixing involves the seamless integration of linguistic elements from multiple languages within a single discourse, reflecting natural multilingual communication patterns. Despite its prominence in informal interactions such as social media, chat messages and instant-messaging exchanges, there has been a lack of publicly available corpora that are author-labeled and suitable for modeling human conversations and relationships. This study introduces the first labeled and general-purpose corpus for understanding code-mixing in context while maintaining rigorous privacy and ethical standards. Our live project will continuously gather, verify, and integrate code-mixed messages into a structured dataset released in JSON format, accompanied by detailed metadata and linguistic statistics. To date, it includes over 355,641 messages spanning various code-mixing patterns, with a primary focus on English, Mandarin, and other languages. We expect the Codemix Corpus to serve as a foundational dataset for research in computational linguistics, sociolinguistics, and NLP applications.

Akshat Gupta, Christine Fang, Atahan Ozdemir et al.

This study investigates the impact of localized updates to large language models (LLMs), specifically in the context of knowledge editing - a task aimed at incorporating or modifying specific facts without altering broader model capabilities. We first show that across different post-training interventions like continuous pre-training, full fine-tuning and LORA-based fine-tuning, the Frobenius norm of the updated matrices always increases. This increasing norm is especially detrimental for localized knowledge editing, where only a subset of matrices are updated in a model . We reveal a consistent phenomenon across various editing techniques, including fine-tuning, hypernetwork-based approaches, and locate-and-edit methods: the norm of the updated matrix invariably increases with successive updates. Such growth disrupts model balance, particularly when isolated matrices are updated while the rest of the model remains static, leading to potential instability and degradation of downstream performance. Upon deeper investigations of the intermediate activation vectors, we find that the norm of internal activations decreases and is accompanied by shifts in the subspaces occupied by these activations, which shows that these activation vectors now occupy completely different regions in the representation space compared to the unedited model. With our paper, we highlight the technical challenges with continuous and localized sequential knowledge editing and their implications for maintaining model stability and utility.

Xiaoyang Song, Akshat Gupta, Kiyan Mohebbizadeh et al.

Have Large Language Models (LLMs) developed a personality? The short answer is a resounding "We Don't Know!". In this paper, we show that we do not yet have the right tools to measure personality in language models. Personality is an important characteristic that influences behavior. As LLMs emulate human-like intelligence and performance in various tasks, a natural question to ask is whether these models have developed a personality. Previous works have evaluated machine personality through self-assessment personality tests, which are a set of multiple-choice questions created to evaluate personality in humans. A fundamental assumption here is that human personality tests can accurately measure personality in machines. In this paper, we investigate the emergence of personality in five LLMs of different sizes ranging from 1.5B to 30B. We propose the Option-Order Symmetry property as a necessary condition for the reliability of these self-assessment tests. Under this condition, the answer to self-assessment questions is invariant to the order in which the options are presented. We find that many LLMs personality test responses do not preserve option-order symmetry. We take a deeper look at LLMs test responses where option-order symmetry is preserved to find that in these cases, LLMs do not take into account the situational statement being tested and produce the exact same answer irrespective of the situation being tested. We also identify the existence of inherent biases in these LLMs which is the root cause of the aforementioned phenomenon and makes self-assessment tests unreliable. These observations indicate that self-assessment tests are not the correct tools to measure personality in LLMs. Through this paper, we hope to draw attention to the shortcomings of current literature in measuring personality in LLMs and call for developing tools for machine personality measurement.

Simerjot Kaur, Charese Smiley, Akshat Gupta et al.

A number of datasets for Relation Extraction (RE) have been created to aide downstream tasks such as information retrieval, semantic search, question answering and textual entailment. However, these datasets fail to capture financial-domain specific challenges since most of these datasets are compiled using general knowledge sources such as Wikipedia, web-based text and news articles, hindering real-life progress and adoption within the financial world. To address this limitation, we propose REFinD, the first large-scale annotated dataset of relations, with $\sim$29K instances and 22 relations amongst 8 types of entity pairs, generated entirely over financial documents. We also provide an empirical evaluation with various state-of-the-art models as benchmarks for the RE task and highlight the challenges posed by our dataset. We observed that various state-of-the-art deep learning models struggle with numeric inference, relational and directional ambiguity.

Hemant Yadav, Akshat Gupta, Sai Krishna Rallabandi et al.

Building Spoken Language Understanding (SLU) systems that do not rely on language specific Automatic Speech Recognition (ASR) is an important yet less explored problem in language processing. In this paper, we present a comparative study aimed at employing a pre-trained acoustic model to perform SLU in low resource scenarios. Specifically, we use three different embeddings extracted using Allosaurus, a pre-trained universal phone decoder: (1) Phone (2) Panphone, and (3) Allo embeddings. These embeddings are then used in identifying the spoken intent. We perform experiments across three different languages: English, Sinhala, and Tamil each with different data sizes to simulate high, medium, and low resource scenarios. Our system improves on the state-of-the-art (SOTA) intent classification accuracy by approximately 2.11% for Sinhala and 7.00% for Tamil and achieves competitive results on English. Furthermore, we present a quantitative analysis of how the performance scales with the number of training examples used per intent.

Zhiyuan Guo, Yuexin Li, Guo Chen et al.

Spoken dialogue systems such as Siri and Alexa provide great convenience to people's everyday life. However, current spoken language understanding (SLU) pipelines largely depend on automatic speech recognition (ASR) modules, which require a large amount of language-specific training data. In this paper, we propose a Transformer-based SLU system that works directly on phones. This acoustic-based SLU system consists of only two blocks and does not require the presence of ASR module. The first block is a universal phone recognition system, and the second block is a Transformer-based language model for phones. We verify the effectiveness of the system on an intent classification dataset in Mandarin Chinese.

Akshat Gupta, Olivia Deng, Akruti Kushwaha et al.

Intent Recognition and Slot Identification are crucial components in spoken language understanding (SLU) systems. In this paper, we present a novel approach towards both these tasks in the context of low resourced and unwritten languages. We present an acoustic based SLU system that converts speech to its phonetic transcription using a universal phone recognition system. We build a word-free natural language understanding module that does intent recognition and slot identification from these phonetic transcription. Our proposed SLU system performs competitively for resource rich scenarios and significantly outperforms existing approaches as the amount of available data reduces. We observe more than 10% improvement for intent classification in Tamil and more than 5% improvement for intent classification in Sinhala. We also present a novel approach towards unsupervised slot identification using normalized attention scores. This approach can be used for unsupervised slot labelling, data augmentation and to generate data for a new slot in a one-shot way with only one speech recording

Akshat Gupta, Sargam Menghani, Sai Krishna Rallabandi et al.

Sentiment analysis is an important task in understanding social media content like customer reviews, Twitter and Facebook feeds etc. In multilingual communities around the world, a large amount of social media text is characterized by the presence of Code-Switching. Thus, it has become important to build models that can handle code-switched data. However, annotated code-switched data is scarce and there is a need for unsupervised models and algorithms. We propose a general framework called Unsupervised Self-Training and show its applications for the specific use case of sentiment analysis of code-switched data. We use the power of pre-trained BERT models for initialization and fine-tune them in an unsupervised manner, only using pseudo labels produced by zero-shot transfer. We test our algorithm on multiple code-switched languages and provide a detailed analysis of the learning dynamics of the algorithm with the aim of answering the question - `Does our unsupervised model understand the Code-Switched languages or does it just learn its representations?'. Our unsupervised models compete well with their supervised counterparts, with their performance reaching within 1-7\% (weighted F1 scores) when compared to supervised models trained for a two class problem.

Akshat Gupta, Sai Krishna Rallabandi, Alan Black

Using task-specific pre-training and leveraging cross-lingual transfer are two of the most popular ways to handle code-switched data. In this paper, we aim to compare the effects of both for the task of sentiment analysis. We work with two Dravidian Code-Switched languages - Tamil-Engish and Malayalam-English and four different BERT based models. We compare the effects of task-specific pre-training and cross-lingual transfer and find that task-specific pre-training results in superior zero-shot and supervised performance when compared to performance achieved by leveraging cross-lingual transfer from multilingual BERT models.

Sai Muralidhar Jayanthi, Akshat Gupta

In this paper we present our submission for the EACL 2021-Shared Task on Offensive Language Identification in Dravidian languages. Our final system is an ensemble of mBERT and XLM-RoBERTa models which leverage task-adaptive pre-training of multilingual BERT models with a masked language modeling objective. Our system was ranked 1st for Kannada, 2nd for Malayalam and 3rd for Tamil.

Akshat Gupta, Xinjian Li, Sai Krishna Rallabandi et al.

With recent advancements in language technologies, humans are now speaking to devices. Increasing the reach of spoken language technologies requires building systems in local languages. A major bottleneck here are the underlying data-intensive parts that make up such systems, including automatic speech recognition (ASR) systems that require large amounts of labelled data. With the aim of aiding development of spoken dialog systems in low resourced languages, we propose a novel acoustics based intent recognition system that uses discovered phonetic units for intent classification. The system is made up of two blocks - the first block is a universal phone recognition system that generates a transcript of discovered phonetic units for the input audio, and the second block performs intent classification from the generated phonetic transcripts. We propose a CNN+LSTM based architecture and present results for two languages families - Indic languages and Romance languages, for two different intent recognition tasks. We also perform multilingual training of our intent classifier and show improved cross-lingual transfer and zero-shot performance on an unknown language within the same language family.

Akshat Gupta, Sai Krishna Rallabandi, Alan W Black

Tremendous progress in speech and language processing has brought language technologies closer to daily human life. Voice technology has the potential to act as a horizontal enabling layer across all aspects of digitization. It is especially beneficial to rural communities in scenarios like a pandemic. In this work we present our initial exploratory work towards one such direction -- building voice enabled banking services for multilingual societies. Speech interaction for typical banking transactions in multilingual communities involves the presence of filled pauses and is characterized by Code Mixing. Code Mixing is a phenomenon where lexical items from one language are embedded in the utterance of another. Therefore speech systems deployed for banking applications should be able to process such content. In our work we investigate various training strategies for building speech based intent recognition systems. We present our results using a Naive Bayes classifier on approximate acoustic phone units using the Allosaurus library.

Akshat Gupta, Milan Desai, Wusheng Liang et al.

Spatiotemporal action recognition is the task of locating and classifying actions in videos. Our project applies this task to analyzing video footage of restaurant workers preparing food, for which potential applications include automated checkout and inventory management. Such videos are quite different from the standardized datasets that researchers are used to, as they involve small objects, rapid actions, and notoriously unbalanced data classes. We explore two approaches. The first approach involves the familiar object detector You Only Look Once, and another applying a recently proposed analogue for action recognition, You Only Watch Once. In the first, we design and implement a novel, recurrent modification of YOLO using convolutional LSTMs and explore the various subtleties in the training of such a network. In the second, we study the ability of YOWOs three dimensional convolutions to capture the spatiotemporal features of our unique dataset

Manuel Ladron de Guevara, Christopher George, Akshat Gupta et al.

Language is ambiguous; many terms and expressions can convey the same idea. This is especially true in creative practice, where ideas and design intents are highly subjective. We present a dataset, Ambiguous Descriptions of Art Images (ADARI), of contemporary workpieces, which aims to provide a foundational resource for subjective image description and multimodal word disambiguation in the context of creative practice. The dataset contains a total of 240k images labeled with 260k descriptive sentences. It is additionally organized into sub-domains of architecture, art, design, fashion, furniture, product design and technology. In subjective image description, labels are not deterministic: for example, the ambiguous label dynamic might correspond to hundreds of different images. To understand this complexity, we analyze the ambiguity and relevance of text with respect to images using the state-of-the-art pre-trained BERT model for sentence classification. We provide a baseline for multi-label classification tasks and demonstrate the potential of multimodal approaches for understanding ambiguity in design intentions. We hope that ADARI dataset and baselines constitute a first step towards subjective label classification.

Akshat Gupta, Prasad N R

We describe an alternative learning method for neural networks, which we call Blind Descent. By design, Blind Descent does not face problems like exploding or vanishing gradients. In Blind Descent, gradients are not used to guide the learning process. In this paper, we present Blind Descent as a more fundamental learning process compared to gradient descent. We also show that gradient descent can be seen as a specific case of the Blind Descent algorithm. We also train two neural network architectures, a multilayer perceptron and a convolutional neural network, using the most general Blind Descent algorithm to demonstrate a proof of concept.