arXiv subjects

Utkarsh Soni

Publications and source records attributed to Utkarsh Soni.

12 recordsLinked to original sources

arXiv2026

When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM Agents

LLM agents increasingly rely on external skills retrieved at runtime, making skill selection from large repositories a critical challenge. We present a production skill router over 34,396 skills and a large-scale study of skill retrieval using limited real supervision and synthetic data. We found that the synthetic-data fine-tuning improves in-distribution retrieval but it causes catastrophic forgetting on real and out-of-distribution (OOD) data. We evaluate several forgetting mitigation fine-tuning approaches inspired by continual learning, including embedding-anchor regularization, Learning without Forgetting (LwF), Elastic Weight Consolidation (EWC), and L2-initialization. The results show that these approaches not only retain the performance on OOD skills retrieval but also improve the retrieval on synthetic in-distribution skills by 13.98\% for 0.6B Qwen retriever and reranker. Our results provide a practical benchmark and a robust fine-tuning recipe for scarce, multi-positive supervision.

cs.IR↗

arXiv2026

Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs

Prompt optimization can improve multi-agent LLM systems, but the prompts being optimized often serve two entangled roles: generating task-relevant content and specifying execution-critical protocols, such as message routing, output formatting, and termination signals, on which the underlying code relies. As a result, a prompt edit intended to improve content generation can inadvertently corrupt the protocol and cause the entire agent pipeline to fail. Our key observation is that these two roles have different representations: execution protocols are typically structured, while task-relevant content is usually expressed in unstructured language. Based on this, we propose control-data flow separation, where execution-critical control is represented as typed, validated program objects, while task-relevant language remains the optimizable data flow for agent communication. This design allows optimizers to improve multi-agent behavior without exposing the routing or formatting interface to prompt drift. Across synthetic reasoning, collaborative review generation, and insurance rating workflows, our framework empirically achieves 100% eventual protocol validity while consistently improving task performance.

cs.AI↗

arXiv2023

PromptAid: Prompt Exploration, Perturbation, Testing and Iteration using Visual Analytics for Large Language Models

Large Language Models (LLMs) have gained widespread popularity due to their ability to perform ad-hoc Natural Language Processing (NLP) tasks with a simple natural language prompt. Part of the appeal for LLMs is their approachability to the general public, including individuals with no prior technical experience in NLP techniques. However, natural language prompts can vary significantly in terms of their linguistic structure, context, and other semantics. Modifying one or more of these aspects can result in significant differences in task performance. Non-expert users may find it challenging to identify the changes needed to improve a prompt, especially when they lack domain-specific knowledge and lack appropriate feedback. To address this challenge, we present PromptAid, a visual analytics system designed to interactively create, refine, and test prompts through exploration, perturbation, testing, and iteration. PromptAid uses multiple, coordinated visualizations which allow users to improve prompts by using the three strategies: keyword perturbations, paraphrasing perturbations, and obtaining the best set of in-context few-shot examples. PromptAid was designed through an iterative prototyping process involving NLP experts and was evaluated through quantitative and qualitative assessments for LLMs. Our findings indicate that PromptAid helps users to iterate over prompt template alterations with less cognitive overhead, generate diverse prompts with help of recommendations, and analyze the performance of the generated prompts while surpassing existing state-of-the-art prompting interfaces in performance.

cs.HC↗

arXiv2023

Methods and Mechanisms for Interactive Novelty Handling in Adversarial Environments

Learning to detect, characterize and accommodate novelties is a challenge that agents operating in open-world domains need to address to be able to guarantee satisfactory task performance. Certain novelties (e.g., changes in environment dynamics) can interfere with the performance or prevent agents from accomplishing task goals altogether. In this paper, we introduce general methods and architectural mechanisms for detecting and characterizing different types of novelties, and for building an appropriate adaptive model to accommodate them utilizing logical representations and reasoning methods. We demonstrate the effectiveness of the proposed methods in evaluations performed by a third party in the adversarial multi-agent board game Monopoly. The results show high novelty detection and accommodation rates across a variety of novelty types, including changes to the rules of the game, as well as changes to the agent's action capabilities.

cs.AI↗

arXiv2023

Towards customizable reinforcement learning agents: Enabling preference specification through online vocabulary expansion

There is a growing interest in developing automated agents that can work alongside humans. In addition to completing the assigned task, such an agent will undoubtedly be expected to behave in a manner that is preferred by the human. This requires the human to communicate their preferences to the agent. To achieve this, the current approaches either require the users to specify the reward function or the preference is interactively learned from queries that ask the user to compare behavior. The former approach can be challenging if the internal representation used by the agent is inscrutable to the human while the latter is unnecessarily cumbersome for the user if their preference can be specified more easily in symbolic terms. In this work, we propose PRESCA (PREference Specification through Concept Acquisition), a system that allows users to specify their preferences in terms of concepts that they understand. PRESCA maintains a set of such concepts in a shared vocabulary. If the relevant concept is not in the shared vocabulary, then it is learned. To make learning a new concept more feedback efficient, PRESCA leverages causal associations between the target concept and concepts that are already known. In addition, we use a novel data augmentation approach to further reduce required feedback. We evaluate PRESCA by using it on a Minecraft environment and show that it can effectively align the agent with the user's preference.

cs.AI↗

arXiv2022

Bridging the Gap: Providing Post-Hoc Symbolic Explanations for Sequential Decision-Making Problems with Inscrutable Representations

As increasingly complex AI systems are introduced into our daily lives, it becomes important for such systems to be capable of explaining the rationale for their decisions and allowing users to contest these decisions. A significant hurdle to allowing for such explanatory dialogue could be the vocabulary mismatch between the user and the AI system. This paper introduces methods for providing contrastive explanations in terms of user-specified concepts for sequential decision-making settings where the system's model of the task may be best represented as an inscrutable model. We do this by building partial symbolic models of a local approximation of the task that can be leveraged to answer the user queries. We test these methods on a popular Atari game (Montezuma's Revenge) and variants of Sokoban (a well-known planning benchmark) and report the results of user studies to evaluate whether people find explanations generated in this form useful.

cs.AI↗

arXiv2021

Why? Why not? When? Visual Explanations of Agent Behavior in Reinforcement Learning

Reinforcement learning (RL) is used in many domains, including autonomous driving, robotics, stock trading, and video games. Unfortunately, the black box nature of RL agents, combined with legal and ethical considerations, makes it increasingly important that humans (including those are who not experts in RL) understand the reasoning behind the actions taken by an RL agent, particularly in safety-critical domains. To help address this challenge, we introduce PolicyExplainer, a visual analytics interface which lets the user directly query an autonomous agent. PolicyExplainer visualizes the states, policy, and expected future rewards for an agent, and supports asking and answering questions such as: Why take this action? Why not take this other action? When is this action taken? PolicyExplainer is designed based upon a domain analysis with RL researchers, and is evaluated via qualitative and quantitative assessments on a trio of domains: taxi navigation, a stack bot domain, and drug recommendation for HIV patients. We find that PolicyExplainer promotes trust and understanding of agent decisions better than a state-of-the-art text-based explanation approach. Interviews with domain practitioners provide further validation for PolicyExplainer as applied to safety-critical domains. Our results help demonstrate how visualization-based approaches can be leveraged to decode the behavior of autonomous RL agents, particularly for RL non-experts.

cs.HC↗

arXiv2021

Integrating Planning, Execution and Monitoring in the presence of Open World Novelties: Case Study of an Open World Monopoly Solver

The game of monopoly is an adversarial multi-agent domain where there is no fixed goal other than to be the last player solvent, There are useful subgoals like monopolizing sets of properties, and developing them. There is also a lot of randomness from dice rolls, card-draws, and adversaries' strategies. This unpredictability is made worse when unknown novelties are added during gameplay. Given these challenges, Monopoly was one of the test beds chosen for the DARPA-SAILON program which aims to create agents that can detect and accommodate novelties. To handle the game complexities, we developed an agent that eschews complete plans, and adapts it's policy online as the game evolves. In the most recent independent evaluation in the SAILON program, our agent was the best performing agent on most measures. We herein present our approach and results.

cs.AI↗

arXiv2021

Not all users are the same: Providing personalized explanations for sequential decision making problems

There is a growing interest in designing autonomous agents that can work alongside humans. Such agents will undoubtedly be expected to explain their behavior and decisions. While generating explanations is an actively researched topic, most works tend to focus on methods that generate explanations that are one size fits all. As in the specifics of the user-model are completely ignored. The handful of works that look at tailoring their explanation to the user's background rely on having specific models of the users (either analytic models or learned labeling models). The goal of this work is thus to propose an end-to-end adaptive explanation generation system that begins by learning the different types of users that the agent could interact with. Then during the interaction with the target user, it is tasked with identifying the type on the fly and adjust its explanations accordingly. The former is achieved by a data-driven clustering approach while for the latter, we compile our explanation generation problem into a POMDP. We demonstrate the usefulness of our system on two domains using state-of-the-art POMDP solvers. We also report the results of a user study that investigates the benefits of providing personalized explanations in a human-robot interaction setting.

cs.AI↗

arXiv2020

Let Me At Least Learn What You Really Like: Dealing With Noisy Humans When Learning Preferences

Learning the preferences of a human improves the quality of the interaction with the human. The number of queries available to learn preferences maybe limited especially when interacting with a human, and so active learning is a must. One approach to active learning is to use uncertainty sampling to decide the informativeness of a query. In this paper, we propose a modification to uncertainty sampling which uses the expected output value to help speed up learning of preferences. We compare our approach with the uncertainty sampling baseline, as well as conduct an ablation study to test the validity of each component of our approach.

cs.LG↗

arXiv2019

Same Stats, Different Graphs: Exploring the Space of Graphs in Terms of Graph Properties

Data analysts commonly utilize statistics to summarize large datasets. While it is often sufficient to explore only the summary statistics of a dataset (e.g., min/mean/max), Anscombe's Quartet demonstrates how such statistics can be misleading. We consider a similar problem in the context of graph mining. To study the relationships between different graph properties, we examine low-order non-isomorphic graphs and provide a simple visual analytics system to explore correlations across multiple graph properties. However, for larger graphs, studying the entire space quickly becomes intractable. We use different random graph generation methods to further look into the distribution of graph properties for higher order graphs and investigate the impact of various sampling methodologies. We also describe a method for generating many graphs that are identical over a number of graph properties and statistics yet are clearly different and identifiably distinct.

cs.CG↗

arXiv2019

Same Stats, Different Graphs (Graph Statistics and Why We Need Graph Drawings)

Data analysts commonly utilize statistics to summarize large datasets. While it is often sufficient to explore only the summary statistics of a dataset (e.g., min/mean/max), Anscombe's Quartet demonstrates how such statistics can be misleading. Graph mining has a similar problem in that graph statistics (e.g., density, connectivity, clustering coefficient) may not capture all of the critical properties of a given graph. To study the relationships between different graph properties and statistics, we examine all low-order (<= 10) non-isomorphic graphs and provide a simple visual analytics system to explore correlations across multiple graph properties. However, for graphs with more than ten nodes, generating the entire space of graphs becomes quickly intractable. We use different random graph generation methods to further look into the distribution of graph statistics for higher order graphs and investigate the impact of various sampling methodologies. We also describe a method for generating many graphs that are identical over a number of graph properties and statistics yet are clearly different and identifiably distinct.

cs.CG↗