Context masking is established as necessary for attributing early answer availability to an explanation rather than its hidden input when early-prefix evidence disappears after masking.
Abstract
Large language model (LLM) tutors may have access to teacher notes, answer keys, rubrics, or retrieved solutions while producing student-facing explanations. We study whether truncated reasoning probes can distinguish direct access to such private context from answer information carried by the written explanation. Using Truncated Reasoning AUC Evaluation (TRACE), we evaluate 1000 GSM8K problems under question-only, correct answer-key, and wrong answer-key contexts. When forced-answer probes retain the private key, answer-key TRACE AUC rises from 0.375 to 0.900, and the gold answer is recoverable with no explanation at all in 998 of 1000 cases. We then introduce a context-masked replay: answer-key-generated prefixes are probed under the corresponding question-only prompt. Masking reduces 10\% prefix accuracy from 0.997 to 0.126 and median AUC from 0.900 to 0.375, nearly matching question-only values of 0.113 and 0.375. On 746 pairs where both explanations end correctly, the masked mean AUC difference is $-0.0086$ with a 95\% bootstrap interval spanning zero. Wrong keys still account for 272 of 387 incorrect final responses, showing that private artifacts can influence outputs even when early-prefix evidence disappears after masking. These results establish context masking as necessary for attributing early answer availability to an explanation rather than its hidden input.
These results establish an auditable release boundary and failure attribution under declared contracts, not universal semantic safety or learning gains, not universal semantic safety or learning gains.
Agentic tutoring systems increasingly rely on large language models (LLMs) to generate real-time feedback, yet risky outputs, such as answer leakage, contradiction, and hallucination, can directly affect learning. Answer leakage is a central concern in this setting. It exposes problem solutions to students and can be verified through human annotation. This article presents a pedagogical style-conditioned audit protocol for automated evaluators that detect answer leakage in LLM-generated tutoring feedback before deployment. Using 1100 mathematics problems from the MATH dataset and 4400 feedback items generated under Baseline, Direct, Socratic, and Encouragement styles, we compare rule-based, embedding-based, natural language inference-based, and two LLM-as-a-Judge evaluators from different model families (Qwen2.5-32B and Gemma3-27B) and validate them against a 314-item human-annotated gold standard. The results show that pedagogical style changes how answer-leakage signals appear in feedback and how reliably automated evaluators detect them. Rule-based and LLM-Judge evaluators that perform well under Baseline, Direct, and Encouragement show consistent degradation under Socratic feedback. The primary LLM-Judge (Qwen2.5-32B) achieves overall F1 = 0.726 but drops to 0.427 on Socratic items, and a second LLM-Judge (Gemma3-27B) reproduces this degradation with Socratic F1 = 0.342. The four factorial prompt-sensitivity variants reach overall F1 up to 0.844, but their Socratic F1 remains at or below 0.390. These findings show that automated tutoring-feedback auditing must be style-specific, prompt-calibrated, and human-grounded before deployment.
Wen-Zhuo Song, Xuechen Wang, Shuang Liang et al.· IEEE Transactions on Learnin...· 0 citations
Programming tutors should support learners'own explanations rather than immediately providing model answers. We present ExplainRoute, a pre-deployment audit framework for non-answer-giving programming tutors. Given a code line and a learner explanation, it estimates the explanation state and selects one of two bounded responses: a Feynman-style self-explanation prompt or a Socratic scaffold. The framework exposes its state, strategy, cited code fragment, and leakage risk through a machine-checkable contract. Unlike benchmarks that rank tutors by fluency alone, ExplainRoute audits information boundaries, response polarity, failure closure, and the value of learner-explanation visibility before classroom deployment. We evaluate it offline on the 1,770-pair SelfCode corpus using a code-group split, with 443 pairs reserved in 11 untouched holdout groups. The evaluation compares direct answers, fixed open self-explanation, fixed Socratic scaffolding, adaptive routing, and an adaptive no-state ablation. Contract validity reaches 100% for all pedagogical conditions. Adaptive routing matches the frozen reference rule on 60.5% of records, with state macro-F1 of 0.238 (Open: 0.229; Socratic: 0.246), showing no reliable adaptive advantage. An independent language-model judge scores adaptive responses 4.516/5, outperforming the no-state ablation (2.819/5) but slightly below fixed open self-explanation (4.598/5) and Socratic scaffolding (4.658/5). A blinded rubric evaluation on a stratified 40-row subset confirms that visible learner explanations improve information value while adaptive routing does not outperform fixed strategies. The contribution is a validated audit protocol and a boundary finding, rather than evidence of improved learning, retention, or causal instructional effectiveness.
Y. Gai, Ying-Ying Zhang, Xue-Fei Huang· 0 citations
Factual question answering (QA) typically assumes a single canonical answer, obscuring whether large language models (LLMs) retain divergent accounts of long-tail facts. To address this gap, we introduce ElephantBench, a closed-book knowledge probe comprising 1,094 questions generated through an auditable graph-based pipeline. The pipeline retrieves related documents from a low-exposure web corpus, identifies naturally occurring disagreements, and converts them into multi-account QA records. Each answer is verified against the originating documents and authoritative public web sources and is then reviewed by human annotators. Across 32 models, even the strongest model recovers both accounts on only 52.4% of questions, while on nearly all remaining questions it recalls one account but omits the other. Scaling model size and inference-time reasoning improve recall but do not eliminate this incompleteness. Corpus analysis further shows that exposure imbalance favors the dominant account, whereas greater minority-side exposure is associated with more complete recall. These findings establish ElephantBench as a reproducible knowledge probe for diagnosing epistemic myopia in parametric memory. More broadly, our graph-based benchmark construction pipeline provides an efficient and scalable way to turn long-tail corpora into source-traceable knowledge probes, supporting efforts to evaluate and advance the epistemic rigour of next-generation LLMs. Code is available at https://github.com/Tencent/ElephantBench.
Zhuo-Shi Pan, Jun-Ru Lu, Yan-Fei Qian et al.· 0 citations
This work measures the capability that role assumes and finds it lacking under the protocol the role is usually deployed with, one-shot greedy authoring with no test-time reasoning.
Wen-Hui Chen, Jianlin Chen, Ziyao Lin et al.· 0 citations
A post hoc robustness audit of the same 1,179 confirmatory answer-phase tutor turns under the frozen helpfulness and pedagogy rubrics finds that general-purpose helpfulness is not a reliable pedagogy signal.
