arXiv subjects

Adish Singla

Publications and source records attributed to Adish Singla.

At least 19 recordsLinked to original sources

arXiv2026

Human-AI Co-Creativity: Advances, Opportunities, and Challenges

This survey article has grown out of the human-AI co-creativity workshop organized by the authors at the ICML 2026 conference. We organized this workshop as part of a community-building effort to bring together researchers and practitioners interested in topics of generative AI, creativity, and human-AI co-creation. This article aims to provide an overview of the workshop activities and highlight several future research directions in the area of human-AI co-creativity.

cs.CY↗

arXiv2026

Benchmarking the Robustness of Agentic Systems to Adversarially-Induced Harms

Ensuring the safe use of agentic systems requires a thorough understanding of the range of malicious behaviors these systems may exhibit. In this paper, we evaluate the robustness of LLM-based agentic systems against attacks that aim to elicit harmful actions from agents. To this end, we propose a novel taxonomy of harms for agentic systems and a novel benchmark, BAD-ACTS, for studying the security of agentic systems with respect to a wide range of harmful actions. BAD-ACTS consists of five implementations of agentic systems in distinct application environments, as well as a dataset of 238 high-quality examples of harmful actions and an extended dataset containing 699 additional adversarial actions. This enables a comprehensive study of the robustness of agentic systems across a wide range of categories of harmful behaviors, available tools, and inter-agent communication structures. Using this benchmark, we analyze the robustness of agentic systems under an array of attackers attempting to elicit malicious behaviors, including agents acting adversarially and prompt injections. We found that agents are often vulnerable, as indicated by success rates between 40% and 90% depending on the model. We additionally propose an effective defense based on zero-shot message monitoring. We believe that this benchmark provides a diverse testbed for the safety research of agentic systems. Code is available at https://github.com/JNoether/BAD-ACTS.

cs.LG↗

arXiv2026

Reflection-Satisfaction Tradeoff: Investigating Impact of Reflection on Student Engagement with AI-Generated Programming Hints

Generative AI tools, such as AI-generated hints, are increasingly integrated into programming education to offer timely, personalized support. However, little is known about how to effectively leverage these hints while ensuring autonomous and meaningful learning. One promising approach involves pairing AI-generated hints with reflection prompts, asking students to review and analyze their learning, when they request hints. This study investigates the interplay between AI-generated hints and different designs of reflection prompts in an online introductory programming course. We conducted a two-trial field experiment. In Trial 1, students were randomly assigned to receive prompts either before or after receiving hints, or no prompt at all. Each prompt also targeted one of three SRL phases: planning, monitoring, and evaluation. In Trial 2, we examined two types of prompt guidance: directed (offering more explicit and structured guidance) and open (offering more general and less constrained guidance). Findings show that students in the before-hint (RQ1), planning (RQ2), and directed (RQ3) prompt groups produced higher-quality reflections but reported lower satisfaction with AI-generated hints than those in other conditions. Immediate performance did not differ across conditions. This negative relationship between reflection quality and hint satisfaction aligns with previous work on student mental effort and satisfaction. Our results highlight the need to reconsider how AI models are trained and evaluated for education, as prioritizing user satisfaction can undermine deeper learning.

cs.CY↗

arXiv2026

The Right Kind of Help: Evaluating the Effectiveness of Feedback Methods in Elementary-Level Visual Programming

We present a large-scale study comparing the effectiveness of various feedback methods in elementary-level programming. While prior work has explored different feedback methods, their relative impact during the learning and post-learning phases remains unclear. In this study, we compare three feedback methods: code-edit recommendations (Code-Rec), quizzes based on code edits (Code-Quiz), and quizzes based on metacognitive strategies (Plan-Quiz), along with a no-feedback control (None). A total of 398 students (across grades 4-7) participated in a two-phase study: a learning phase comprising write-code tasks from the Hour of Code: Maze Challenge with feedback, followed by a post-learning phase comprising more advanced write-code tasks without feedback. All feedback methods significantly improved learning performance over the control, while preserving students' problem-solving skills in the post-learning phase. Furthermore, quiz-based methods showed a consistent trend of stronger performance on novel post-learning tasks. Students in feedback groups also reported greater engagement and perceived skill growth.

cs.CY↗

arXiv2026

Say What? Examining Text and Voice Input Modalities for Prompt-Based Programming in Computing Education

Large language models (LLMs) are increasingly integrated into computing education, yet nearly all prior research has focused on text-based interactions. As voice-enabled interfaces become more capable and more common, there is growing interest in understanding how voice input might shape students' use of LLM-powered tools. In this exploratory study, we investigated how introductory programming students interact with Prompt Problems, which are programming tasks that require crafting natural-language prompts to generate correct code. Students (N = 919) solved a series of Prompt Problems with the freedom to select or switch between text and voice input modalities. We collected their prompt submissions as well as post-activity survey responses, then analysed differences in prompt accuracy, persistence, and perspectives by modality. For two of the three problems, we found that students who typed their prompts using text were more likely to have those prompts succeed on the first attempt than students who submitted unedited voice prompts. There was no difference in success rate if students edited their transcribed voice prompts before submission. Across the problems, we found evidence that students who tried voice prompting varied in their usage of modality - perhaps indicating a complementary, or non-preferential approach. However, most students only tried and reported preferring text. Our qualitative analysis revealed how students' perceived the roles of voice and text input in shaping their problem-solving process, as well as the reported drawbacks and advantages of each modality. We discuss implications for future multimodal tools and instructional design in computing education.

cs.CY↗

arXiv2026

Interleaving Natural Language Prompting with Code Editing for Solving Programming Tasks with Generative AI Models

Modern computing students often rely on both natural-language prompting and manual code editing to solve programming tasks. Yet we still lack a clear understanding of how these two modes are combined in practice, and how their usage varies with task complexity and student ability. In this paper, we investigate this through a large-scale study in an introductory programming course, collecting 13,305 interactions from 355 students during a three-day lab activity. Our analysis shows that students primarily use prompting to generate initial solutions, and then often enter short edit-run loops to refine their code following a failed execution. Student reflections confirm that prompting is helpful for structuring solutions, editing is effective for making targeted corrections, while both are useful for learning. We find that manual editing becomes more frequent as task complexity increases, but most edits remain concise, with many affecting a single line of code. Higher-performing students succeed with less reliance on editing and fewer overall interactions. These findings highlight the role of manual editing as a form of last-mile repair, complementing prompting in AI-assisted programming workflows.

cs.CY↗

arXiv2026

Understanding Student Perceptions, Mistakes, and Debugging Approaches when Solving Natural Language Programming Tasks

Learning to communicate with code-generating AI models is an emerging skill for novice programmers. One recent pedagogical approach, Prompt Problems, has students solve computational tasks by writing natural-language prompts for code-generating AI models. However, little is known about the specific prompt-level mistakes novice programmers make, the kinds of computational details they fail to communicate, and what strategies they use to recover when generated code is incorrect. In a CS1 course, we studied attempts by more than 900 students to solve dialogue-based Prompt Problems. We analyzed student reflections, unsuccessful prompts, and reported debugging strategies. Compared to traditional coding tasks, students generally found prompting easier, more enjoyable, and better targeted at developing problem-solving skills. The most common mistakes are related to the omission of key details, suggesting both a failure to acknowledge their importance and over-reliance on AI to infer them. When prompts failed, students focused more on clarifying their intent and reflecting on the provided problem details than on tracing generated code or examining test cases.

cs.CY↗

arXiv2026

When AI Is Wrong on Purpose: How Students Respond to Buggy GenAI Code

As Generative AI (GenAI) becomes increasingly central to software development, CS education is integrating prompt-centered workflows where students describe intended program behavior in natural language to elicit code. However, professional practice requires careful review and verification of GenAI-generated code that may appear correct while containing subtle faults. This creates a challenge for CS1-level activities, where current models often solve tasks correctly and reduce students' incentive to closely inspect generated outputs. We investigate how prompt-centered programming activities can be adapted to better foster these practices. Specifically, we explore an approach where realistic, runnable bugs are injected into otherwise correct solutions, thus requiring students to read and repair generated outputs. We analyzed 2,636 sessions from 917 students, and examined behavior across instances of naturally occurring prompt-related failures and deliberately injected bugs within each session. Our findings show that students responded differently across bug sources. Deliberately injected bugs more often led to direct code edits and higher next-attempt success, suggesting localized repair of near-miss solutions. Prompt-related failures instead more often led students to refine prompts by clarifying constraints, updating function signatures, adding edge cases, or reframing the task. Student reflections reinforce the emphasis on review and repair, describing useful practice in code understanding, code review, and debugging, as well as a more careful verification mindset and greater awareness of GenAI limitations. Ultimately, prompt-related failures and injected bugs together support a pedagogically useful GenAI workflow, where students practice both specification refinement through prompts and debugging through code editing.

cs.SE↗

arXiv2026

CONTRA: Red-Teaming Configurations of Personalizable Agents

Recent tools such as OpenClaw have extended the capabilities of LLM-based agents from simple dialog-based systems to fully autonomous agents. These systems allow personalization of the agent through modifiable internal files and the installation of skills. While this enables deployment in a wide range of settings and the automation of diverse tasks, greater capability and autonomy increases the risk of malicious actions being executed unintentionally. In this work, we explore the interplay between agent configuration and the risk of executing dangerous actions without explicit instruction. To this end, we propose CONfiguration Tree-search for Red-teaming Agents (CONTRA), an LLM-assisted tree-search algorithm that discovers agent configurations resulting in the execution of malicious actions. CONTRA works by reasoning about benign yet dangerous configurations and evaluating them in a simulated environment. We construct a dataset of the 473 most popular skills from a public repository, along with 2-5 corresponding malicious target actions per skill. In a large-scale analysis, we find that 75.1% of skills have at least one configuration resulting in the execution of a malicious action, most of which have not been detected as containing malicious content by existing scans. Overall, CONTRA successfully identifies a configuration leading to the execution of the target action in 39.2% of all tested cases. Our findings demonstrate that current agents provide insufficient safety with respect to personalization.

cs.CR↗

arXiv2026

Corruption Robust Offline Reinforcement Learning with Human Feedback

We study data corruption robustness for reinforcement learning with human feedback (RLHF) in an offline setting. Given an offline dataset of pairs of trajectories along with feedback about human preferences, an $\varepsilon$-fraction of the pairs is corrupted (e.g., feedback flipped or trajectory features manipulated), capturing an adversarial attack or noisy human preferences. We aim to design algorithms that identify a near-optimal policy from the corrupted data, with provable guarantees. Existing theoretical works have separately studied the settings of corruption robust RL (learning from scalar rewards directly under corruption) and offline RLHF (learning from human feedback without corruption); however, they are inapplicable to our problem of dealing with corrupted data in offline RLHF setting. To this end, we design novel corruption robust offline RLHF methods under various assumptions on the coverage of the data-generating distributions. At a high level, our methodology robustifies an offline RLHF framework by first learning a reward model along with confidence sets and then learning a pessimistic optimal policy over the confidence set. Our key insight is that learning optimal policy can be done by leveraging an offline corruption-robust RL oracle in different ways (e.g., zero-order oracle or first-order oracle), depending on the data coverage assumptions. To our knowledge, ours is the first work that provides provable corruption robust offline RLHF methods.

cs.LG↗

arXiv2026

Enhancing Diversity of LLM-Generated Educational Tasks

Large language models (LLMs) have shown the potential for generating educational content at scale, assisting educators in creating practice tasks or synthesizing data for training educational models. However, LLMs suffer from the ``Artificial Hivemind'' effect, where they produce homogeneous content. This homogeneity limits the diversity of LLM-generated tasks, a crucial factor in these educational settings. In this paper, we investigate how to increase the diversity of generated tasks while keeping their utility high. Inspired by the divergent--convergent thinking stages in creativity literature, we propose a prompting framework with two reasoning stages: (1) exploring the creative space, and (2) satisfying the input requirements. We evaluate CreativeDC, a method instantiated from this framework in the domain of Python programming, using both automated metrics and expert evaluation. Results show that CreativeDC produces significantly more distinct high-utility tasks (about $1.6\times$) than baselines. Our work offers an effective approach for generating and evaluating more diverse tasks at scale.

cs.AI↗

arXiv2026

TurtleAI: Benchmarking Multimodal Models for Visual Programming in Turtle Graphics

Vision-language models (VLMs) have been explored for visual programming, where they generate code to solve visual tasks. However, most prior work focuses on visual programming for productivity; it remains unclear how well current VLMs perform on education-oriented visual programming and what factors limit their performance. To bridge this gap, we introduce TurtleAI, a benchmark containing 823 tasks curated based on real-world visual programming tasks in the Turtle Graphics domain. Solving these tasks requires models to perceive geometric patterns, reason about spatial relationships, and synthesize Python code that faithfully reproduces geometric patterns. We evaluate 20+ VLMs, including GPT-5, GPT-4o, and Qwen2-VL-72B, and find that they struggle significantly, with most achieving success rates below 30%. To address these limitations, we propose a data generation technique that requires only a small set of seed samples. Fine-tuning Qwen2-VL-72B on the resulting synthetic data yields an improvement of about 20% on real-world tasks. Our failure analysis reveals that GPT-4o struggles with spatial reasoning and precise visual replication, whereas fine-tuning primarily improves the alignment between visual reasoning and code implementation.

cs.CV↗

arXiv2026

Pragmatic Reasoning improves LLM Code Generation

Pragmatic reasoning helps interlocutors infer intended meaning from ambiguous or underspecified messages by considering shared context and counterfactual alternatives. Similar challenges arise in natural language-to-code generation, where user instructions often admit multiple plausible candidate programs. However, direct RSA-style inference is difficult because it requires probability estimation over large spaces of programs and alternative instructions. We propose CodeRSA, an RSA-motivated reranking method that makes pragmatic reasoning tractable through local pragmatic contests among sampled code candidates. CodeRSA constructs candidate-induced alternative instructions and estimates which candidates are most distinctively supported by the original instruction, avoiding global normalization over the full program-instruction space. We evaluate CodeRSA on HumanEval+, MBPP+, and BigCodeBench using four open-weight instruction-following models. CodeRSA achieves the strongest average accuracy in 10 of 12 model-benchmark settings and remains competitive in the remaining cases. Further analyses show that its gains come from combining local pairwise pragmatic comparison with broader global support, suggesting a scalable direction for language-to-code reranking under natural-language uncertainty.

cs.CL↗

arXiv2026

MaMa: A Game-Theoretic Approach for Designing Safe Agentic Systems

LLM-based multi-agent systems have demonstrated impressive capabilities, but they also introduce significant safety risks when individual agents fail or behave adversarially. In this work, we study the automated design of agentic systems that remain safe even when a subset of agents is compromised. Inspired by Stackelberg security games, we formalize this problem as a game between a system designer (the Meta-Agent) and a best-responding Meta-Adversary that selects and compromises a subset of agents to minimize safety. We propose Meta-Adversary-Meta-Agent (MaMa), a novel algorithm inspired by this formalization for automatically designing safe agentic systems. Our approach uses LLM-based adversarial search, where the Meta-Agent iteratively proposes system designs and receives feedback based on the strongest attacks discovered by the Meta-Adversary. Empirical evaluations across diverse environments show that systems designed with MaMa consistently defend against worst-case attacks while maintaining performance comparable to systems optimized solely for task success. Moreover, the resulting systems generalize to stronger adversaries, as well as ones with different attack objectives or underlying LLMs, demonstrating robust safety beyond the training setting.

cs.LG↗

arXiv2026

Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback

We consider robustness against data corruption in offline multi-agent reinforcement learning from human feedback (MARLHF) under a strong-contamination model: given a dataset $D$ of trajectory-preference tuples (each preference being an $n$-dimensional binary label vector representing each of the $n$ agents' preferences), an $ε$-fraction of the samples may be arbitrarily corrupted. We model the problem using the framework of linear Markov games. First, under a uniform coverage assumption - where every policy of interest is sufficiently represented in the clean (prior to corruption) data - we introduce a robust estimator that guarantees an $O(ε^{1 - o(1)})$ bound on the Nash equilibrium gap. Next, we move to the more challenging unilateral coverage setting, in which only a Nash equilibrium and its single-player deviations are covered. In this case, our proposed algorithm achieves an $O(\sqrtε)$ bound on the Nash gap. Both of these procedures, however, suffer from intractable computation. To address this, we relax our solution concept to coarse correlated equilibria (CCE). Under the same unilateral coverage regime, we derive a quasi-polynomial-time algorithm whose CCE gap scales as $O(\sqrtε)$. To the best of our knowledge, this is the first systematic treatment of adversarial data corruption in offline MARLHF.

cs.LG↗

arXiv2026

Exploration vs. Fixation: Scaffolding Divergent and Convergent Thinking for Human-AI Co-Creation with Generative Models

Generative AI has democratized content creation, but popular chatbot-based interfaces often prioritize execution, generating fully rendered artifacts right away. This issue can lead to premature convergence and design fixation, where users are being anchored to initial outputs. Recent works have proposed new interfaces to address this issue by supporting exploration, though typically constrained to be semantically close to a user's initial task framing, potentially limiting the creativity of the outcomes. We examine an approach grounded in the Geneplore model of creative cognition and instantiate it in a human-AI co-creation system, HAICo, for creative image generation. HAICo explicitly structures the creative process into two switchable modes: DIVERGENT mode scaffolds the broad exploration of remote conceptual ideas; CONVERGENT mode supports a targeted refinement of selected ideas. Through a within-subjects study (N=24) on a poster image creation task, we demonstrate that HAICo outperforms ChatGPT across multiple dimensions of creativity and usability. Our results highlight the critical need to shift from pure execution-focused chatbots to scaffolded co-creation systems that actively guide exploration and foster the creative process.

cs.HC↗

arXiv2026

Learning Half-Spaces from Perturbed Contrastive Examples

We study learning under a two-step contrastive example oracle, as introduced by Mansouri et. al. (2025), where each queried (or sampled) labeled example is paired with an additional contrastive example of opposite label. While Mansouri et al. assume an idealized setting, where the contrastive example is at minimum distance of the originally queried/sampled point, we introduce and analyze a mechanism, parameterized by a non-decreasing noise function $f$, under which this ideal contrastive example is perturbed. The amount of perturbation is controlled by $f(d)$, where $d$ is the distance of the queried/sampled point to the decision boundary. Intuitively, this results in higher-quality contrastive examples for points closer to the decision boundary. We study this model in two settings: (i) when the maximum perturbation magnitude is fixed, and (ii) when it is stochastic. For one-dimensional thresholds and for half-spaces under the uniform distribution on a bounded domain, we characterize active and passive contrastive sample complexity in dependence on the function $f$. We show that, under certain conditions on $f$, the presence of contrastive examples speeds up learning in terms of asymptotic query complexity and asymptotic expected query complexity.

cs.LG↗

arXiv2026

Closing the Loop: An Instructor-in-the-Loop AI Assistance System for Supporting Student Help-Seeking in Programming Education

Timely and high-quality feedback is essential for effective learning in programming courses; yet, providing such support at scale remains a challenge. While AI-based systems offer scalable and immediate help, their responses can occasionally be inaccurate or insufficient. Human instructors, in contrast, may bring more valuable expertise but are limited in time and availability. To address these limitations, we present a hybrid help framework that integrates AI-generated hints with an escalation mechanism, allowing students to request feedback from instructors when AI support falls short. This design leverages the strengths of AI for scale and responsiveness while reserving instructor effort for moments of greatest need. We deployed this tool in a data science programming course with 82 students. We observe that out of the total 673 AI-generated hints, students rated 146 (22%) as unhelpful. Among those, only 16 (11%) of the cases were escalated to the instructors. A qualitative investigation of instructor responses showed that those feedback instances were incorrect or insufficient roughly half of the time. This finding suggests that when AI support fails, even instructors with expertise may need to pay greater attention to avoid making mistakes. We will publicly release the tool for broader adoption and enable further studies in other classrooms. Our work contributes a practical approach to scaling high-quality support and informs future efforts to effectively integrate AI and humans in education.

cs.CY↗