Sreecharan Sankaranarayanan

Conrad Borchers, Bahar Shahrokhian, Francesco Balzan et al.

Large Language Models (LLMs) enable new possibilities for qualitative research at scale, including coding and data annotation. While multi-agent systems (MAS) can emulate human coding workflows, their benefits over single-agent coding remain poorly understood. We conducted an experimental study of how agent persona and temperature shape consensus-building and coding accuracy of dialog segments based on a codebook with 8 codes. Our open-source MAS mirrors deductive human coding through structured agent discussion and consensus arbitration. Using six open-source LLMs (with 3 to 32 billion parameters) and 18 experimental configurations, we analyze over 77,000 coding decisions against a gold-standard dataset of human-annotated transcripts from online math tutoring sessions. Temperature significantly impacted whether and when consensus was reached across all six LLMs. MAS with multiple personas (including neutral, assertive, or empathetic), significantly delayed consensus in four out of six LLMs compared to uniform personas. In three of those LLMs, higher temperatures significantly diminished the effects of multiple personas on consensus. However, neither temperature nor persona pairing lead to robust improvements in coding accuracy. Single agents matched or outperformed MAS consensus in most conditions. Only one model (OpenHermesV2:7B) and code category showed above-chance gains from MAS deliberation when temperature was 0.5 or lower and especially when the agents included at least one assertive persona. Qualitative analysis of MAS collaboration for these configurations suggests that MAS may nonetheless aid in narrowing ambiguous code applications that could improve codebooks and human-AI coding. We contribute new insight into the limits of LLM-based qualitative methods, challenging the notion that diverse MAS personas lead to better outcomes. We open-source our MAS and experimentation code.

Sreecharan Sankaranarayanan

The democratization of Large Language Models (LLMs) has given rise to ``Vibe Coding," a workflow where novice programmers prioritize semantic intent over syntactic implementation. While this lowers barriers to entry, we hypothesize that without pedagogical guardrails, it is fundamentally misaligned with cognitive skill acquisition. Drawing on the distinction between Cognitive Offloading and Cognitive Outsourcing, we argue that unrestricted AI encourages novices to outsource the Intrinsic Cognitive Load required for schema formation, rather than merely offloading Extraneous Load. This accumulation of ``Epistemic Debt" creates ``Fragile Experts" whose high functional utility masks critically low corrective competence. To quantify and mitigate this debt, we conducted a between-subjects experiment (N=78) using a custom Cursor IDE plugin backed by Claude 3.5 Sonnet. Participants represented "AI-Native" learners across three conditions: Manual (Control), Unrestricted AI (Outsourcing), and Scaffolded AI (Offloading). The Scaffolded condition utilized a novel ``Explanation Gate," leveraging a real-time LLM-as-a-Judge framework to enforce a ``Teach-Back" protocol before generated code could be integrated. Results reveal a ``Collapse of Competence": while Unrestricted AI users matched the productivity of the Scaffolded group (p < .001 vs. Manual), they suffered a 77% failure rate in a subsequent AI-Blackout maintenance task, compared to only 39% in the Scaffolded group. Qualitative analysis suggests that successful vibe coders naturally engage in self-scaffolding, treating the AI as a consultant rather than a contractor. We discuss the implications for the maintainability of AI-generated software and propose that future learning systems must enforce Metacognitive Friction to prevent the mass production of unmaintainable code.

Denis Shchepakin, Sreecharan Sankaranarayanan, Dawn Zimmaro

Bayesian Knowledge Tracing (BKT) is a probabilistic model of a learner's state of mastery corresponding to a knowledge component. It considers the learner's state of mastery as a "hidden" or latent binary variable and updates this state based on the observed correctness of the learner's response using parameters that represent transition probabilities between states. BKT is often represented as a Hidden Markov Model and the Expectation-Maximization (EM) algorithm is used to infer these parameters. However, this algorithm can suffer from several issues including producing multiple viable sets of parameters, settling into a local minima, producing degenerate parameter values, and a high computational cost during fitting. This paper takes a "from first principles" approach to deriving constraints that can be imposed on the BKT parameter space. Starting from the basic mathematical truths of probability and building up to the behaviors expected of the BKT parameters in real systems, this paper presents a mathematical derivation that results in succinct constraints that can be imposed on the BKT parameter space. Since these constraints are necessary conditions, they can be applied prior to fitting in order to reduce computational cost and the likelihood of issues that can emerge from the EM procedure. In order to see that promise through, the paper further introduces a novel algorithm for estimating BKT parameters subject to the newly defined constraints. While the issue of degenerate parameter values has been reported previously, this paper is the first, to our best knowledge, to derive the constrains from first principles while also presenting an algorithm that respects those constraints.

Sreecharan Sankaranarayanan, Siddharth Reddy Kandimalla, Christopher Bogart et al.

Contributing to the literature on aptitude-treatment interactions between worked examples and problem-solving, this paper addresses differential learning from the two approaches when students are positioned as domain experts learning new concepts. Our evaluation is situated in a team project that is part of an advanced software engineering course. In this course, students who possess foundational domain knowledge but are learning new concepts engage alternatively in programming followed by worked example-based reflection. They are either allowed to finish programming or are curtailed after a pre-specified time to participate in a longer worked example-based reflection. We find significant pre- to post-test learning gains in both conditions. Then, we not only find significantly more learning when students participated in longer worked example-based reflections but also a significant performance improvement on a problem-solving transfer task. These findings suggest that domain experts learning new concepts benefit more from worked example-based reflections than from problem-solving.

Sreecharan Sankaranarayanan, Raghuram Mandyam Annasamy, Katia Sycara et al.

Q-Ensembles are a model-free approach where input images are fed into different Q-networks and exploration is driven by the assumption that uncertainty is proportional to the variance of the output Q-values obtained. They have been shown to perform relatively well compared to other exploration strategies. Further, model-based approaches, such as encoder-decoder models have been used successfully for next frame prediction given previous frames. This paper proposes to integrate the model-free Q-ensembles and model-based approaches with the hope of compounding the benefits of both and achieving superior exploration as a result. Results show that a model-based trajectory memory approach when combined with Q-ensembles produces superior performance when compared to only using Q-ensembles.

Gaurav Singh Tomar, Sreecharan Sankaranarayanan, Xu Wang et al.

An earlier study of a collaborative chat intervention in a Massive Open Online Course (MOOC) identified negative effects on attrition stemming from a requirement for students to be matched with exactly one partner prior to beginning the activity. That study raised questions about how to orchestrate a collaborative chat intervention in a MOOC context in order to provide the benefit of synchronous social engagement without the coordination difficulties. In this paper we present a careful analysis of an intervention designed to overcome coordination difficulties by welcoming students into the chat on a rolling basis as they arrive rather than requiring them to be matched with a partner before beginning. The results suggest the most positive impact when experiencing a chat with exactly one partner rather than more or less. A qualitative analysis of the chat data reveals differential experiences between these configurations that suggests a potential explanation for the effect and raises questions for future research.