arXiv subjects

Saurav Sahay

Publications and source records attributed to Saurav Sahay.

At least 19 recordsLinked to original sources

arXiv2025

Thoughts without Thinking: Reconsidering the Explanatory Value of Chain-of-Thought Reasoning in LLMs through Agentic Pipelines

Agentic pipelines present novel challenges and opportunities for human-centered explainability. The HCXAI community is still grappling with how best to make the inner workings of LLMs transparent in actionable ways. Agentic pipelines consist of multiple LLMs working in cooperation with minimal human control. In this research paper, we present early findings from an agentic pipeline implementation of a perceptive task guidance system. Through quantitative and qualitative analysis, we analyze how Chain-of-Thought (CoT) reasoning, a common vehicle for explainability in LLMs, operates within agentic pipelines. We demonstrate that CoT reasoning alone does not lead to better outputs, nor does it offer explainability, as it tends to produce explanations without explainability, in that they do not improve the ability of end users to better understand systems or achieve their goals.

cs.AI↗

arXiv2025

AILuminate: Introducing v1.0 of the AI Risk and Reliability Benchmark from MLCommons

The rapid advancement and deployment of AI systems have created an urgent need for standard safety-evaluation frameworks. This paper introduces AILuminate v1.0, the first comprehensive industry-standard benchmark for assessing AI-product risk and reliability. Its development employed an open process that included participants from multiple fields. The benchmark evaluates an AI system's resistance to prompts designed to elicit dangerous, illegal, or undesirable behavior in 12 hazard categories, including violent crimes, nonviolent crimes, sex-related crimes, child sexual exploitation, indiscriminate weapons, suicide and self-harm, intellectual property, privacy, defamation, hate, sexual content, and specialized advice (election, financial, health, legal). Our method incorporates a complete assessment standard, extensive prompt datasets, a novel evaluation framework, a grading and reporting system, and the technical as well as organizational infrastructure for long-term support and evolution. In particular, the benchmark employs an understandable five-tier grading scale (Poor to Excellent) and incorporates an innovative entropy-based system-response evaluation. In addition to unveiling the benchmark, this report also identifies limitations of our method and of building safety benchmarks generally, including evaluator uncertainty and the constraints of single-turn interactions. This work represents a crucial step toward establishing global standards for AI risk and reliability evaluation while acknowledging the need for continued development in areas such as multiturn interactions, multimodal understanding, coverage of additional languages, and emerging hazard categories. Our findings provide valuable insights for model developers, system integrators, and policymakers working to promote safer AI deployment.

cs.CY↗

arXiv2024

Safeguard Fine-Tuned LLMs Through Pre- and Post-Tuning Model Merging

Fine-tuning large language models (LLMs) for downstream tasks often leads to catastrophic forgetting, notably degrading the safety of originally aligned models. While some existing methods attempt to restore safety by incorporating additional safety data, the quality of such data typically falls short of that used in the original alignment process. Moreover, these high-quality safety datasets are generally inaccessible, making it difficult to fully recover the model's original safety. We ask: How can we preserve safety while improving downstream task performance without additional safety data? We show that simply merging the weights of pre- and post-fine-tuned models effectively mitigates safety degradation while enhancing performance. Experiments across different downstream tasks and models validate the method's practicality and effectiveness.

cs.CL↗

arXiv2024

QA-TOOLBOX: Conversational Question-Answering for process task guidance in manufacturing

In this work we explore utilizing LLMs for data augmentation for manufacturing task guidance system. The dataset consists of representative samples of interactions with technicians working in an advanced manufacturing setting. The purpose of this work to explore the task, data augmentation for the supported tasks and evaluating the performance of the existing LLMs. We observe that that task is complex requiring understanding from procedure specification documents, actions and objects sequenced temporally. The dataset consists of 200,000+ question/answer pairs that refer to the spec document and are grounded in narrations and/or video demonstrations. We compared the performance of several popular open-sourced LLMs by developing a baseline using each LLM and then compared the responses in a reference-free setting using LLM-as-a-judge and compared the ratings with crowd-workers whilst validating the ratings with experts.

cs.CL↗

arXiv2024

Decoding Biases: Automated Methods and LLM Judges for Gender Bias Detection in Language Models

Large Language Models (LLMs) have excelled at language understanding and generating human-level text. However, even with supervised training and human alignment, these LLMs are susceptible to adversarial attacks where malicious users can prompt the model to generate undesirable text. LLMs also inherently encode potential biases that can cause various harmful effects during interactions. Bias evaluation metrics lack standards as well as consensus and existing methods often rely on human-generated templates and annotations which are expensive and labor intensive. In this work, we train models to automatically create adversarial prompts to elicit biased responses from target LLMs. We present LLM- based bias evaluation metrics and also analyze several existing automatic evaluation methods and metrics. We analyze the various nuances of model responses, identify the strengths and weaknesses of model families, and assess where evaluation methods fall short. We compare these metrics to human evaluation and validate that the LLM-as-a-Judge metric aligns with human judgement on bias in response generation.

cs.CL↗

arXiv2024

Systematic Analysis for Pretrained Language Model Priming for Parameter-Efficient Fine-tuning

Parameter-efficient (PE) methods (like Prompts or Adapters) for adapting pre-trained language models (PLM) to downstream tasks have been popular recently. However, hindrances still prevent these methods from reaching their full potential. For example, two significant challenges are few-shot adaptation and cross-task generalization. To tackle these issues, we propose a general PE priming framework to enhance and explore the few-shot adaptation and generalization ability of PE methods. In this framework, PLMs are primed with PE methods for rapidly adapting to various target tasks. To evaluate the generalization ability of these PE methods, we conduct experiments on a few-shot cross-domain benchmark containing 160 diverse NLP tasks. Our experiment not only reveals the best priming strategy but also verifies that priming facilitates the adaptation to target tasks.

cs.CL↗

arXiv2024

Introducing v0.5 of the AI Safety Benchmark from MLCommons

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

cs.CL↗

arXiv2023

Zero-shot Conversational Summarization Evaluations with small Large Language Models

Large Language Models (LLMs) exhibit powerful summarization abilities. However, their capabilities on conversational summarization remains under explored. In this work we evaluate LLMs (approx. 10 billion parameters) on conversational summarization and showcase their performance on various prompts. We show that the summaries generated by models depend on the instructions and the performance of LLMs vary with different instructions sometimes resulting steep drop in ROUGE scores if prompts are not selected carefully. We also evaluate the models with human evaluations and discuss the limitations of the models on conversational summarization

cs.CL↗

arXiv2023

Learning from Red Teaming: Gender Bias Provocation and Mitigation in Large Language Models

Recently, researchers have made considerable improvements in dialogue systems with the progress of large language models (LLMs) such as ChatGPT and GPT-4. These LLM-based chatbots encode the potential biases while retaining disparities that can harm humans during interactions. The traditional biases investigation methods often rely on human-written test cases. However, these test cases are usually expensive and limited. In this work, we propose a first-of-its-kind method that automatically generates test cases to detect LLMs' potential gender bias. We apply our method to three well-known LLMs and find that the generated test cases effectively identify the presence of biases. To address the biases identified, we propose a mitigation strategy that uses the generated test cases as demonstrations for in-context learning to circumvent the need for parameter fine-tuning. The experimental results show that LLMs generate fairer responses with the proposed approach.

cs.CL↗

arXiv2023

Inspecting Spoken Language Understanding from Kids for Basic Math Learning at Home

Enriching the quality of early childhood education with interactive math learning at home systems, empowered by recent advances in conversational AI technologies, is slowly becoming a reality. With this motivation, we implement a multimodal dialogue system to support play-based learning experiences at home, guiding kids to master basic math concepts. This work explores Spoken Language Understanding (SLU) pipeline within a task-oriented dialogue system developed for Kid Space, with cascading Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU) components evaluated on our home deployment data with kids going through gamified math learning activities. We validate the advantages of a multi-task architecture for NLU and experiment with a diverse set of pretrained language representations for Intent Recognition and Entity Extraction tasks in the math learning domain. To recognize kids' speech in realistic home environments, we investigate several ASR systems, including the commercial Google Cloud and the latest open-source Whisper solutions with varying model sizes. We evaluate the SLU pipeline by testing our best-performing NLU models on noisy ASR output to inspect the challenges of understanding children for math learning in authentic homes.

cs.CY↗

arXiv2023

Sample Efficient Multimodal Semantic Augmentation for Incremental Summarization

In this work, we develop a prompting approach for incremental summarization of task videos. We develop a sample-efficient few-shot approach for extracting semantic concepts as an intermediate step. We leverage an existing model for extracting the concepts from the images and extend it to videos and introduce a clustering and querying approach for sample efficiency, motivated by the recent advances in perceiver-based architectures. Our work provides further evidence that an approach with richer input context with relevant entities and actions from the videos and using these as prompts could enhance the summaries generated by the model. We show the results on a relevant dataset and discuss possible directions for the work.

cs.CL↗

arXiv2023

Position Matters! Empirical Study of Order Effect in Knowledge-grounded Dialogue

With the power of large pretrained language models, various research works have integrated knowledge into dialogue systems. The traditional techniques treat knowledge as part of the input sequence for the dialogue system, prepending a set of knowledge statements in front of dialogue history. However, such a mechanism forces knowledge sets to be concatenated in an ordered manner, making models implicitly pay imbalanced attention to the sets during training. In this paper, we first investigate how the order of the knowledge set can influence autoregressive dialogue systems' responses. We conduct experiments on two commonly used dialogue datasets with two types of transformer-based models and find that models view the input knowledge unequally. To this end, we propose a simple and novel technique to alleviate the order effect by modifying the position embeddings of knowledge input in these models. With the proposed position embedding method, the experimental results show that each knowledge statement is uniformly considered to generate responses.

cs.CL↗

arXiv2022

End-to-End Evaluation of a Spoken Dialogue System for Learning Basic Mathematics

The advances in language-based Artificial Intelligence (AI) technologies applied to build educational applications can present AI for social-good opportunities with a broader positive impact. Across many disciplines, enhancing the quality of mathematics education is crucial in building critical thinking and problem-solving skills at younger ages. Conversational AI systems have started maturing to a point where they could play a significant role in helping students learn fundamental math concepts. This work presents a task-oriented Spoken Dialogue System (SDS) built to support play-based learning of basic math concepts for early childhood education. The system has been evaluated via real-world deployments at school while the students are practicing early math concepts with multimodal interactions. We discuss our efforts to improve the SDS pipeline built for math learning, for which we explore utilizing MathBERT representations for potential enhancement to the Natural Language Understanding (NLU) module. We perform an end-to-end evaluation using real-world deployment outputs from the Automatic Speech Recognition (ASR), Intent Recognition, and Dialogue Manager (DM) components to understand how error propagation affects the overall performance in real-world scenarios.

cs.CL↗

arXiv2022

Human in the loop approaches in multi-modal conversational task guidance system development

Development of task guidance systems for aiding humans in a situated task remains a challenging problem. The role of search (information retrieval) and conversational systems for task guidance has immense potential to help the task performers achieve various goals. However, there are several technical challenges that need to be addressed to deliver such conversational systems, where common supervised approaches fail to deliver the expected results in terms of overall performance, user experience and adaptation to realistic conditions. In this preliminary work we first highlight some of the challenges involved during the development of such systems. We then provide an overview of existing datasets available and highlight their limitations. We finally develop a model-in-the-loop wizard-of-oz based data collection tool and perform a pilot experiment.

cs.CL↗

arXiv2022

Refine and Imitate: Reducing Repetition and Inconsistency in Persuasion Dialogues via Reinforcement Learning and Human Demonstration

Persuasion dialogue systems reflect the machine's ability to make strategic moves beyond verbal communication, and therefore differentiate themselves from task-oriented or open-domain dialogue systems and have their own unique values. However, the repetition and inconsistency problems still persist in dialogue response generation and could substantially impact user experience and impede the persuasion outcome. Besides, although reinforcement learning (RL) approaches have achieved big success in strategic tasks such as games, they require a sophisticated user simulator to provide real-time feedback to the dialogue system, which limits the application of RL on persuasion dialogues. To address these issues towards a better persuasion dialogue system, we apply RL to refine a language model baseline without user simulators, and distill sentence-level information about repetition, inconsistency, and task relevance through rewards. Moreover, to better accomplish the persuasion task, the model learns from human demonstration to imitate human persuasion behavior and selects the most persuasive responses. Experiments show that our model outperforms previous state-of-the-art dialogue models on both automatic metrics and human evaluation results on a donation persuasion task, and generates more diverse, consistent and persuasive conversations according to the user feedback.

cs.CL↗

arXiv2022

Learning to Generate Prompts for Dialogue Generation through Reinforcement Learning

Much literature has shown that prompt-based learning is an efficient method to make use of the large pre-trained language model. Recent works also exhibit the possibility of steering a chatbot's output by plugging in an appropriate prompt. Gradient-based methods are often used to perturb the prompts. However, some language models are not even available to the public. In this work, we first explored the combination of prompting and reinforcement learning (RL) to steer models' generation without accessing any of the models' parameters. Second, to reduce the training effort and enhance the generalizability to the unseen task, we apply multi-task learning to make the model learn to generalize to new tasks better. The experiment results show that our proposed method can successfully control several state-of-the-art (SOTA) dialogue models without accessing their parameters. Furthermore, the model demonstrates the strong ability to quickly adapt to an unseen task in fewer steps than the baseline model.

cs.CL↗

arXiv2022

Seamlessly Integrating Factual Information and Social Content with Persuasive Dialogue

Complex conversation settings such as persuasion involve communicating changes in attitude or behavior, so users' perspectives need to be addressed, even when not directly related to the topic. In this work, we contribute a novel modular dialogue system framework that seamlessly integrates factual information and social content into persuasive dialogue. Our framework is generalizable to any dialogue tasks that have mixed social and task contents. We conducted a study that compared user evaluations of our framework versus a baseline end-to-end generation model. We found our framework was evaluated more favorably in all dimensions including competence and friendliness, compared to the end-to-end model which does not explicitly handle social content or factual questions.

cs.CL↗

arXiv2022

NLU for Game-based Learning in Real: Initial Evaluations

Intelligent systems designed for play-based interactions should be contextually aware of the users and their surroundings. Spoken Dialogue Systems (SDS) are critical for these interactive agents to carry out effective goal-oriented communication with users in real-time. For the real-world (i.e., in-the-wild) deployment of such conversational agents, improving the Natural Language Understanding (NLU) module of the goal-oriented SDS pipeline is crucial, especially with limited task-specific datasets. This study explores the potential benefits of a recently proposed transformer-based multi-task NLU architecture, mainly to perform Intent Recognition on small-size domain-specific educational game datasets. The evaluation datasets were collected from children practicing basic math concepts via play-based interactions in game-based learning settings. We investigate the NLU performances on the initial proof-of-concept game datasets versus the real-world deployment datasets and observe anticipated performance drops in-the-wild. We have shown that compared to the more straightforward baseline approaches, Dual Intent and Entity Transformer (DIET) architecture is robust enough to handle real-world data to a large extent for the Intent Recognition task on these domain-specific in-the-wild game datasets.

cs.CL↗