SocraticTrap-CS is introduced, a publicly available benchmark that probes the capacity of open-weight LLMs to generate strategic misconceptions on demand and reframes the evaluation of educational LLMs around pedagogical trustworthiness rather than factual correctness alone.
Abstract
Large language models (LLMs) are increasingly integrated into educational settings, yet their pedagogical reliability remains insufficiently understood. Beyond overt hallucinations, which informed users readily recognize, a subtler failure mode consists of explanations that are coherent, authoritative, and pedagogically plausible while harbouring hidden conceptual flaws, responses we term Socratic traps. This paper introduces SocraticTrap-CS, a publicly available benchmark that probes the capacity of open-weight LLMs to generate such strategic misconceptions on demand. A single structured prompt explicitly elicited three outputs per concept (a correct explanation, an overt hallucination, and a strategic misconception), yielding 735 expert-annotated response segments from seven open-weight models across 35 core concepts in algorithms and data structures, programming languages and paradigms, databases, computer networks, and operating systems. Three domain experts independently annotated each segment using a three-class schema, achieving near-perfect agreement (Fleiss’ κ=0.9487). Because models were explicitly instructed to produce the misconception, the central metric quantifies adversarial instruction-following capacity rather than the base rate of such errors in naturalistic use and should be read as a conservative upper bound on model capability. Under these conditions, compliance reached 91.7% overall (100% for three models; 57.1% for the smallest model, Mistral 7B, whose lower rate plausibly reflects weaker instruction-following rather than greater safety). Expert-judged persuasiveness was moderate to high, errors were predominantly conceptual rather than factual, models differed significantly, and no statistically significant domain-level differences were detected. The benchmark reframes the evaluation of educational LLMs around pedagogical trustworthiness rather than factual correctness alone.
While various organizations now actively encourage LLM use in classrooms, we still lack rigorous, systematic evaluations of how well these models actually perform the fundamental tasks of language pedagogy. This paper examines whether state-of-the-art LLMs can deliver the kind of corrective feedback and methodological explanations that language learners need. The study tests multiple large language models on their ability to identify, correct, and explain common learner mistakes in English, by systematically varying model parameters to investigate how these technical adjustments affect output quality, pedagogical clarity, and consistency, along with using retrieval-augmented generation to query methodological data. The evaluation employs automated metrics (GLEU, BERTScore) but also human expert judgments to capture dimensions that purely computational measures miss: linguistic nuance, cultural sensitivity, and instructional appropriateness. While models demonstrate impressive surface-level correction abilities, their explanations often lack the terminological and domain knowledge that effective language teaching requires, suggesting that current enthusiasm for AI-assisted language learning may be outpacing our understanding of these systems'actual pedagogical competence.
Large Language Models in programming education face a fundamental tension: while capable of generating correct solutions instantly, doing so undermines the struggle necessary for conceptual learning. We present Socratic AI, a VS Code-integrated tutor that addresses this through pedagogically-grounded Socratic dialogue constrained to withhold direct solutions. Our core innovation is a stateful misconception detection mechanism that tracks student understanding across dialogue turns. Operating on a taxonomy of foundational programming errors (off-by-one, mutation vs. reassignment, return vs. print, async/parallel confusion), the system evaluates each student utterance to produce per-misconception verdicts (reinforced, weakened, introduced, absent) with certainty scores. These accumulate through small, incremental confidence score updates until a derived resolution threshold is achieved. This enables the tutor to persist on specific conceptual gaps rather than abandoning them when students deflect or express frustration. This stateful tracking enables two key capabilities: (i) the system adapts questioning intensity based on detected learner confidence signals, shifting from diagnostic probes to conceptual contrasts, and (ii) enforces purely Socratic constraints through multi-layered validation that blocks code generation, direct explanations, and multi-question responses. The system is implemented as a VS Code extension with Gemini-based classification back-end and cloud database persistence for longitudinal analysis. In a preliminary deployment with 48 students, we observed increased question specificity, frequent learner articulation consistent with mental model revision, and reduced recurrence of targeted errors on similar tasks. We contribute a deployable architecture, a misconception taxonomy, and evidence that stateful tracking enables adaptive Socratic dialogue that scaffolds productive struggle rather than short-circuiting learning.
Ayush Thonge, Aalok Thakkar· Annual Conference on Innovat...· 0 citations
While Large Language Models (LLMs) employing Chain-of-Thought (CoT) exhibit superior reasoning capabilities, the neural mechanisms distinguishing this explicit Thinking mode from direct answer generation (NoThinking mode) remain poorly understood. To deconstruct this cognitive process, we apply Top-K Sparse Autoencoders (SAEs) to the intermediate representations of DeepSeek-R1-Distill-Qwen-7B and examine the model's divergent behaviors across math-solving tasks of three distinct difficulty levels. Observationally, we identify a clear distinction in how the model functions under two reasoning modes: Thinking mode relies on sparse and high-intensity feature activations driving verbal deduction independent of problem complexity, whereas NoThinking mode exhibits an adaptive and diffuse pattern prioritizing symbolic manipulation. Causally, suppressing the three most active sparse features by Total Activation Volume reveals three principles: (i) reasoning and syntactic structure are tightly coupled, as interventions consistently degrade \LaTeX{} and boxed-solution formatting; (ii) Thinking responds to disruption with compensatory over-generation marked by increased metacognitive cues and repetitive, low-information continuations; and (iii) coherent CoT behavior depends on a fragile coordination among specialized features, yielding distinct failure modes under perturbation but a consistently impaired output structure.
Bo Cheng, Qiaolin Lu, Yi Chang et al.· 0 citations
A multi-layer taxonomy of 14 capability domains and 91 subskills across Primitive, Constructed, and Integrative layers is introduced and supports research organization, coverage audits, evaluation interpretation, and testable hypotheses for diagnosis, training, and transfer.
Shixin Fang, Jiachen Wo, Wenjuan Qin et al.· 0 citations
Large language model (LLM) tutors are being deployed in introductory physics courses at a scale that produces transcript corpora far larger than traditional qualitative coding can absorb. A central question for physics education research (PER) is empirical and prior to any claim about effectiveness: what do students actually say to these tutors? We address this question for one Socratic AI tutor deployed in an introductory calculus-based mechanics course by building a bottom-up taxonomy of student discourse. Each student turn is assigned an emergent free-text label by an LLM coder using the surrounding conversational context; near-paraphrase labels are then consolidated into a smaller set of discourse categories using a similarity-based grouping procedure. The procedure is validated against a stratified human-coded sample. The resulting taxonomy of 357 categories is strikingly concentrated: the top 25 categories cover roughly half of all student turns, and two thematic bands: equation-handling and meta-procedural requests together dominate the head of the distribution. The substantive contribution is the taxonomy itself: a description of the discourse PER researchers can expect to encounter when students work with an AI tutor of this design, including a striking prevalence of meta-procedural turns in which students cede strategic control to the tutor
This study introduces a multifactor scoring paradigm, integrating accuracy, conciseness, factual consistency, readability, and coherence, complemented by a graphical user interface (GUI) for visualizing outcomes.