Pulkit Verma

Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi et al. · allen-ai, amazon-science

How well can NLP models generalize to a variety of unseen tasks when provided with task instructions? To address this question, we first introduce Super-NaturalInstructions, a benchmark of 1,616 diverse NLP tasks and their expert-written instructions. Our collection covers 76 distinct task types, including but not limited to classification, extraction, infilling, sequence tagging, text rewriting, and text composition. This large and diverse collection of tasks enables rigorous benchmarking of cross-task generalization under instructions -- training models to follow instructions on a subset of tasks and evaluating them on the remaining unseen ones. Furthermore, we build Tk-Instruct, a transformer model trained to follow a variety of in-context instructions (plain language task definitions or k-shot examples). Our experiments show that Tk-Instruct outperforms existing instruction-following models such as InstructGPT by over 9% on our benchmark despite being an order of magnitude smaller. We further analyze generalization as a function of various scaling parameters, such as the number of observed tasks, the number of instances per task, and model sizes. We hope our dataset and model facilitate future progress towards more general-purpose NLP models.

Alexander K. Saeri, Jess Graham, Michael Noetel et al.

Artificial intelligence poses many risks, ranging from familiar present-day harms to unprecedented and potentially catastrophic ones. Effective risk management requires prioritization: we must understand which risks are most severe, who is most vulnerable, and who is most responsible for addressing them. We report results from a three-round Delphi study conducted late 2025 with 272 international AI experts. Experts rated 24 AI risks on harm probability and severity, sector and actor vulnerability, actor responsibility, and overall concern. Experts estimated the five most severe harms in the next 5 years were likely to come from dangerous capabilities, competitive dynamics, weapons & cyberattacks (including CBRNE), power centralization, and false information. In a business-as-usual scenario, experts judged 18 of 24 risks as having a more than 10% probability of catastrophic outcomes (e.g., more than 1 million deaths or more than USD 100B in financial loss) in the next 5 years (2025-2030). In a scenario where pragmatic mitigations are implemented, experts still judged five risks as having a more than 10% probability of catastrophic outcomes: dangerous capabilities, weapons & cyberattacks, environmental harm, inequality & unemployment, and power centralization. All 24 risks were judged as being more than 5% likely to cause catastrophic outcomes. AI users and the general public were judged the most vulnerable to these risks, but experts assigned the highest responsibility for addressing them to general-purpose AI developers and governance actors (including governments, regulators, and standards bodies). Across most risks, experts identified information, finance, and national security as the most vulnerable sectors. These findings can guide AI risk prioritization and clarify expert expectations about who should bear responsibility for mitigation.

Rashmeet Kaur Nayyar, Pulkit Verma, Siddharth Srivastava

Much of the research on learning symbolic models of AI agents focuses on agents with stationary models. This assumption fails to hold in settings where the agent's capabilities may change as a result of learning, adaptation, or other post-deployment modifications. Efficient assessment of agents in such settings is critical for learning the true capabilities of an AI system and for ensuring its safe usage. In this work, we propose a novel approach to "differentially" assess black-box AI agents that have drifted from their previously known models. As a starting point, we consider the fully observable and deterministic setting. We leverage sparse observations of the drifted agent's current behavior and knowledge of its initial model to generate an active querying policy that selectively queries the agent and computes an updated model of its functionality. Empirical evaluation shows that our approach is much more efficient than re-learning the agent model from scratch. We also show that the cost of differential assessment using our method is proportional to the amount of drift in the agent's functionality.

Pulkit Verma, Rushang Karia, Siddharth Srivastava

It is essential for users to understand what their AI systems can and can't do in order to use them safely. However, the problem of enabling users to assess AI systems with sequential decision-making (SDM) capabilities is relatively understudied. This paper presents a new approach for modeling the capabilities of black-box AI systems that can plan and act, along with the possible effects and requirements for executing those capabilities in stochastic settings. We present an active-learning approach that can effectively interact with a black-box SDM system and learn an interpretable probabilistic model describing its capabilities. Theoretical analysis of the approach identifies the conditions under which the learning process is guaranteed to converge to the correct model of the agent; empirical evaluations on different agents and simulated scenarios show that this approach is few-shot generalizable and can effectively describe the capabilities of arbitrary black-box SDM agents in a sample-efficient manner.

Daniel Bramblett, Rushang Karia, Adrian Ciotinga et al.

Black-box AI (BBAI) systems such as foundational models are increasingly being used for sequential decision making. To ensure that such systems are safe to operate and deploy, it is imperative to develop efficient methods that can provide a sound and interpretable representation of the BBAI's capabilities. This paper shows that PDDL-style representations can be used to efficiently learn and model an input BBAI's planning capabilities. It uses the Monte-Carlo tree search paradigm to systematically create test tasks, acquire data, and prune the hypothesis space of possible symbolic models. Learned models describe a BBAI's capabilities, the conditions under which they can be executed, and the possible outcomes of executing them along with their associated probabilities. Theoretical results show soundness, completeness and convergence of the learned models. Empirical results with multiple BBAI systems illustrate the scope, efficiency, and accuracy of the presented methods.

Vaidehi Bagaria, Nikshep Grampurohit, Pulkit Verma

Reinforcement learning (RL) allows vision-language-action (VLA) policies to generalize beyond their training distribution by optimizing directly for task success, but post-training is computationally expensive. A natural response has been to speed rollout collection through faster simulators and world models. In GRPO-based VLA RL, we find that the dominant cost lies elsewhere: gradient computation accounts for approximately 78% of wall-clock time per step in our runs, while rollout collection accounts for only 21%. Gradient cost dominates because much of this computation is spent on phases that contribute little to learning. GRPO's learning signal is driven by advantage variance: only phases where successful and failed rollouts diverge produce learning signal. However, GRPO assigns the same advantage to every chunk in a rollout. As a result, actor-update compute is spent uniformly across the trajectory, including phases the policy already handles after pre-training and supervised fine-tuning. This paper presents Probabilistic Chunk Masking (PCM), a drop-in modification to GRPO that allocates gradient computation to a small, probabilistically selected subset of chunks per trajectory. PCM scores semantic phases using success-failure action variance, a rollout-derived proxy for per-phase gradient variance, and samples a fixed chunk budget with online-updated phase-level keep probabilities. We formalize per-phase gradient variance as the quantity determines where gradient computation is useful and show that success-failure action variance provides a measurable proxy for it. PCM requires no reward model or learned critic. On three LIBERO benchmarks, PCM matches the final success rate of standard GRPO while achieving 2.38 times wall-clock speedup, 4.8 times faster gradient updates, and 60% lower peak activation memory, while backpropagating through fewer than 20% of trajectory chunks.

Rushang Karia, Daniel Bramblett, Daksh Dobhal et al.

This paper presents $\forall$uto$\exists$val, a new approach for scaling LLM assessment in translating formal syntax -- such as first-order logic, regular expressions, etc -- to natural language (interpretation) or vice versa (compilation), thereby facilitating their use in applications such as generating/explaining logic and control flow for programs etc. Existing approaches for LLM assessment in these areas require labor-intensive ground-truth creation, the availability of which undermines the separation of training and test sets. Furthermore, such datasets typically include relatively few hand-coded test cases over which LLM accuracy is determined, thus making them inadequate for determining the safety or correctness of their generated outputs. We introduce a new approach that utilizes context-free grammars (CFGs) to generate out-of-distribution datasets on the fly and perform closed-loop testing of LLM capabilities using formal verifiers to guarantee the correctness of LLM outputs without any human intervention. We release our dataset and benchmark as open-source code at \url{https://github.com/AAIR-lab/auto-llm-assessment}. We also conduct an assessment of several SOTA closed and open-source LLMs to showcase the feasibility and scalability of this paradigm. Our experiments reveal that SOTA LLMs are unable to solve the formal translation task adequately.

Dillon Z. Chen, Pulkit Verma, Siddharth Srivastava et al.

Automated decision-making is a fundamental topic that spans multiple sub-disciplines in AI: reinforcement learning (RL), AI planning (AP), foundation models, and operations research, among others. Despite recent efforts to ``bridge the gaps'' between these communities, there remain many insights that have not yet transcended the boundaries. Our goal in this paper is to provide a brief and non-exhaustive primer on ideas well-known in AP, but less so in other sub-disciplines. We do so by introducing the classical AP problem and representation, and extensions that handle uncertainty and time through the Markov Decision Process formalism. Next, we survey state-of-the-art techniques and ideas for solving AP problems, focusing on their ability to exploit problem structure. Lastly, we cover subfields within AP for learning structure from unstructured inputs and learning to generalise to unseen scenarios and situations.

Rushang Karia, Pulkit Verma, Alberto Speranzon et al.

This paper introduces a new approach for continual planning and model learning in relational, non-stationary stochastic environments. Such capabilities are essential for the deployment of sequential decision-making systems in the uncertain and constantly evolving real world. Working in such practical settings with unknown (and non-stationary) transition systems and changing tasks, the proposed framework models gaps in the agent's current state of knowledge and uses them to conduct focused, investigative explorations. Data collected using these explorations is used for learning generalizable probabilistic models for solving the current task despite continual changes in the environment dynamics. Empirical evaluations on several non-stationary benchmark domains show that this approach significantly outperforms planning and RL baselines in terms of sample complexity. Theoretical results show that the system exhibits desirable convergence properties when stationarity holds.

Pulkit Verma, Ngoc La, Anthony Favier et al.

Large language models (LLMs) have demonstrated impressive capabilities across diverse tasks, yet their ability to perform structured symbolic planning remains limited, particularly in domains requiring formal representations like the Planning Domain Definition Language (PDDL). In this paper, we present a novel instruction tuning framework, PDDL-Instruct, designed to enhance LLMs' symbolic planning capabilities through logical chain-of-thought reasoning. Our approach focuses on teaching models to rigorously reason about action applicability, state transitions, and plan validity using explicit logical inference steps. By developing instruction prompts that guide models through the precise logical reasoning required to determine when actions can be applied in a given state, we enable LLMs to self-correct their planning processes through structured reflection. The framework systematically builds verification skills by decomposing the planning process into explicit reasoning chains about precondition satisfaction, effect application, and invariant preservation. Experimental results on multiple planning domains show that our chain-of-thought reasoning based instruction-tuned models are significantly better at planning, achieving planning accuracy of up to 94% on standard benchmarks, representing a 66% absolute improvement over baseline models. This work bridges the gap between the general reasoning capabilities of LLMs and the logical precision required for automated planning, offering a promising direction for developing better AI planning systems.

Naman Shah, Pulkit Verma, Trevor Angle et al.

This paper presents JEDAI, an AI system designed for outreach and educational efforts aimed at non-AI experts. JEDAI features a novel synthesis of research ideas from integrated task and motion planning and explainable AI. JEDAI helps users create high-level, intuitive plans while ensuring that they will be executable by the robot. It also provides users customized explanations about errors and helps improve their understanding of AI planning as well as the limits and capabilities of the underlying robot system.

Pulkit Verma, Siddharth Srivastava

One of the several obstacles in the widespread use of AI systems is the lack of requirements of interpretability that can enable a layperson to ensure the safe and reliable behavior of such systems. We extend the analysis of an agent assessment module that lets an AI system execute high-level instruction sequences in simulators and answer the user queries about its execution of sequences of actions. We show that such a primitive query-response capability is sufficient to efficiently derive a user-interpretable causal model of the system in stationary, fully observable, and deterministic settings. We also introduce dynamic causal decision networks (DCDNs) that capture the causal structure of STRIPS-like domains. A comparative analysis of different classes of queries is also presented in terms of the computational requirements needed to answer them and the efforts required to evaluate their responses to learn the correct model.

Pulkit Verma, Shashank Rao Marpally, Siddharth Srivastava

Several approaches have been developed for answering users' specific questions about AI behavior and for assessing their core functionality in terms of primitive executable actions. However, the problem of summarizing an AI agent's broad capabilities for a user is comparatively new. This paper presents an algorithm for discovering from scratch the suite of high-level "capabilities" that an AI system with arbitrary internal planning algorithms/policies can perform. It computes conditions describing the applicability and effects of these capabilities in user-interpretable terms. Starting from a set of user-interpretable state properties, an AI agent, and a simulator that the agent can interact with, our algorithm returns a set of high-level capabilities with their parameterized descriptions. Empirical evaluation on several game-based scenarios shows that this approach efficiently learns descriptions of various types of AI agents in deterministic, fully observable settings. User studies show that such descriptions are easier to understand and reason with than the agent's primitive actions.

Pulkit Verma, Shashank Rao Marpally, Siddharth Srivastava

This paper develops a new approach for estimating an interpretable, relational model of a black-box autonomous agent that can plan and act. Our main contributions are a new paradigm for estimating such models using a minimal query interface with the agent, and a hierarchical querying algorithm that generates an interrogation policy for estimating the agent's internal model in a vocabulary provided by the user. Empirical evaluation of our approach shows that despite the intractable search space of possible agent models, our approach allows correct and scalable estimation of interpretable agent models for a wide class of black-box autonomous agents. Our results also show that this approach can use predicate classifiers to learn interpretable models of planning agents that represent states as images.

Rahul Tallamraju, Pulkit Verma, Venkatesh Sripada et al.

In this work, we consider a multi-wheeled payload transport system. Each of the wheels can be selectively actuated. When they are not actuated, wheels are free moving and do not consume battery power. The payload transport system is modeled as an actuated multi-agent system, with each wheel-motor pair as an agent. Kinematic and dynamic models are developed to ensure that the payload transport system moves as desired. We design optimization formulations to decide on the number of wheels to be active and which of the wheels to be active so that the battery is conserved and the wear on the motors is reduced. Our multi-level control framework over the agents ensures that near-optimal number of agents is active for the payload transport system to function. Through simulation studies we show that our solution ensures energy efficient operation and increases the distance traveled by the payload transport system, for the same battery power. We have built the payload transport system and provide results for preliminary experimental validation.

Yahnit Sirineni, Pulkit Verma, Kamalakar Karlapalem

In this work, we address traffic management of multiple payload transport systems comprising of non-holonomic robots. We consider loosely coupled rigid robot formations carrying a payload from one place to another. Each payload transport system (PTS) moves in various kinds of environments with obstacles. We ensure each PTS completes its given task by avoiding collisions with other payload systems and obstacles as well. Each PTS has one leader and multiple followers and the followers maintain a desired distance and angle with respect to the leader using a decentralized leader-follower control architecture while moving in the traffic. We showcase, through simulations the time taken by each PTS to traverse its respective trajectory with and without other PTS and obstacles. We show that our strategies help manage the traffic for a large number of PTS moving from one place to another.

Pulkit Verma, Rahul Tallamraju, Abhay Rawat et al.

In this work, we present an algorithm for robot replacement to increase the operational time of a multi-robot payload transport system. Our system comprises a group of nonholonomic wheeled mobile robots traversing on a known trajectory. We design a multi-robot system with loosely coupled robots that ensures the system lasts much longer than the battery life of an individual robot. A system level optimization is presented, to decide on the operational state (charging or discharging) of each robot in the system. The charging state implies that the robot is not in a formation and is kept on charge whereas the discharging state implies that the robot is a part of the formation. Robot battery recharge hubs are present along the trajectory. Robots in the formation can be replaced at these hub locations with charged robots using a replacement mechanism. We showcase the efficacy of the proposed scheduling framework through simulations and experiments with real robots.