{"id":"172a39e8-0e99-410f-9949-9c5401e1df11","arxiv_id":"2605.23925","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"AI tutoring models systematically fail to detect student misconceptions when flawed reasoning coincidentally produces the correct answer, with 71% of failures concentrated in two predictable question types.","lead":"This paper identifies a failure mode in AI tutoring systems where models fail to detect flawed student reasoning if the final answer is correct. It matters because it reveals that high overall accuracy can mask critical pedagogical failures, limiting safe automation in education.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The 'roughly four false alarms per detection' figure cannot be derived from Table 2's per-class recalls alone; it appears to require an unstated confusion matrix where FM→TM misclassifications count as false alarms, which is pedagogically questionable since those students do have misconceptions.","rationale":"The reader correctly identified the small TM sample size as a weakness, and the paper itself acknowledges this. However, the reader accepted the false alarm ratio of 'roughly four' at face value without checking whether it is derivable from the reported metrics. It is not: Table 2 reports only per-class recall, and reproducing 4.3 requires an unstated confusion matrix that includes FM→TM misclassifications as false alarms. This is pedagogically questionable because FM students do have misconceptions — flagging them for review is not a false alarm in the sense that matters for deployment decisions. Under the more natural definition (only TC→TM as false alarms), the ratio drops to ~2.7, which may still be problematic but is less dramatic and could change the deployment recommendation. The CONDITIONAL verdict is appropriate because: (1) the core CAT phenomenon is real and well-motivated, (2) the concentration finding (RQ1/RQ2) is robust with N=343, (3) the PRM800K validation provides independent support for the phenomenon's existence, but (4) the deployment-impracticality claim — a headline implication — rests on an underspecified calculation that may overstate the problem by ~60%. The paper should report the full confusion matrix and clarify the false alarm definition. This is a concrete, fixable issue rather than a fundamental flaw.","tokens_in":7389,"tokens_out":6662,"duration_ms":226296,"concrete_test":"Report the full 3×3 confusion matrix for Gemini 3 Flash on the test set (predicted TC/FM/TM × actual TC/FM/TM). Recompute the false alarm ratio under two definitions: (1) only TC→TM as false alarms, (2) all non-TM→TM as false alarms. If definition (1) yields a ratio below 3.0, the 'stand-alone screening is impractical' claim should be qualified, since the current 4.3 figure likely conflates pedagogically appropriate FM flagging with genuine false alarms.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's deployment-impracticality claim rests on the false alarm ratio of ~4.3 at 1.6% TM prevalence. Table 2 reports only per-class recall (TC=94.0%, FM=90.6%, TM=83.6%), not the full confusion matrix or precision. From these recalls alone, if only TC→TM misclassifications count as false alarms, the ratio is approximately (600×0.06)/(16×0.836) ≈ 2.7, not 4.3. Reproducing 4.3 requires including FM cases misclassified as TM as 'false alarms' — roughly (600×0.06 + FM→TM)/(16×0.836) ≈ 4.3, which implies ~18 FM→TM misclassifications. But FM students genuinely have misconceptions (just with wrong answers), so flagging them for intervention is not a 'false alarm' in any pedagogically meaningful sense. The paper does not specify this calculation, report the confusion matrix, or justify why FM→TM errors should count as false alarms. This matters because the ratio drops from ~4.3 to ~2.7 under the more natural definition, and even ~2.7 may be tolerable in a human-in-the-loop setting — weakening the 'stand-alone screening impractical' conclusion. The reader's concern about small TM sample size (n=61) is valid but acknowledged by the paper; the false alarm ratio calculation issue is not acknowledged and directly affects a headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper introduces the 'correct answer trap' (CAT): AI models systematically fail to detect student misconceptions when the final answer is correct but the reasoning is flawed. Using the Eedi mathematics dataset, the authors show that 71% of these failures concentrate in two question types sharing a 'coincidental correctness path' where flawed procedures produce correct numerical outputs. They compare fine-tuned models (T5, BERT) with frontier LLMs (Gemini 3 Flash, Llama), finding that while frontier models improve TM recall (83.6% vs 57.4%), they still generate roughly four false alarms per genuine detection at natural class prevalence, making standalone screening impractical. The work is grounded in assessment design theory (Barton, Hiebert & Lefevre, Kapur) and includes a validation check on PRM800K.","tokens_in":7662,"tokens_out":1216,"duration_ms":115966,"significance":"The paper addresses a genuine and underexplored gap in educational AI: most existing misconception detection work is conditioned on incorrect answers, but the pedagogically critical case of correct answers masking flawed reasoning has received little attention. The 'coincidental correctness path' concept is a useful, falsifiable contribution that question authors can act on at design time. The integrated gradients analysis of answer-correctness shortcuts adds mechanistic plausibility. The PRM800K validation, while limited to model-generated solutions, provides a useful cross-dataset robustness check. Replication scripts and prompts are publicly available, which is commendable.","major_comments":[{"comment":"§4.4, 'Implications for Practice': The headline claim that Gemini 3 Flash generates 'roughly 4.3 false alarms per genuine detection' is not derivable from Table 2 alone. Table 2 reports per-class recalls (TC=94.0%, FM=90.6%, TM=83.6%) but not the full confusion matrix or precision. Reproducing the ratio of ~4.3 requires counting FM cases misclassified as TM as 'false alarms' (approximately 18 FM→TM misclassifications). However, FM students genuinely hold misconceptions (they have incorrect answers with flawed reasoning), so flagging them for intervention is not a false alarm in a pedagogically meaningful sense. If only TC→TM misclassifications count as false alarms, the ratio drops to approximately (600×0.06)/(16×0.836) ≈ 2.7. The paper does not report the confusion matrix, specify this calculation, or justify why FM→TM errors should count as false alarms. This directly affects the 'impr","section":null},{"comment":"§4.4, 'Implications for Practice': The deployment-impracticality conclusion rests on the false alarm ratio of ~4.3 at 1.6% TM prevalence. Under the more natural definition (only TC→TM as false alarms), the ratio is ~2.7, which may be tolerable in a human-in-the-loop setting. The paper should either (a) report the full confusion matrix and justify the false alarm definition, or (b) recompute the ratio under the pedagogically natural definition and revise the deployment conclusion accordingly. As it stands, a load-bearing claim is underspecified.","section":null},{"comment":"§3.2: The question classification (procedural/conceptual/mixed) was performed by a single researcher over n=15 questions. While the paper acknowledges this limitation, it affects RQ2's odds-ratio analysis (§4.2). A second coder or inter-rater reliability check would substantially strengthen the item-vulnerability claims. At minimum, the paper should report whether the 'mixed' classification (dot patterns) affects the odds-ratio calculation and whether the conclusions hold under alternative classifications of that item.","section":null}],"minor_comments":[{"comment":"§3.1: 'Gemini 3 Flash' is referenced as a 'commercial frontier LLM' but the model name may not be widely recognized; a footnote with the model version and access date would aid reproducibility.","section":null},{"comment":"§4.4: The integrated gradients analysis is described as 'exploratory' but the attribution ratios (1.19 vs 0.68) are reported without confidence intervals or sample sizes. A brief note on how many misclassified vs correctly classified TM cases were analyzed would help readers calibrate.","section":null},{"comment":"Table 1: The 'Question Type' column uses mathematical notation (e.g., A/10 = 9/15) that is rendered ambiguously in the text; standardising the notation or adding a brief description column would improve readability.","section":null},{"comment":"§4.3: The claim that 'increasing Gemini's thinking budget slightly reduced balanced accuracy (86.2% vs 87.6%)' is reported without the per-class breakdown that Table 2 provides for the low-budget setting. Including the high-budget per-class recalls would make this comparison informative.","section":null},{"comment":"§2: The related work section could briefly note how the 'detection' problem (whether a misconception exists) relates to 'classification' (which misconception) in practical ITS pipelines, to clarify the deployment context.","section":null}],"recommendation":"major_revision","confidential_remarks":"The false alarm ratio calculation issue is the most substantive concern. The skeptic's reconstruction is plausible and the paper does not provide the confusion matrix needed to verify the ~4.3 figure. If the authors can provide the confusion matrix and justify their false alarm definition (or recompute under the natural definition), the paper's core contribution on CAT concentration and item vulnerability remains sound. The small TM sample size (n=61) is acknowledged and handled with appropriate caveats; the T5 vs Gemini comparison is statistically significant. I would not consider the sample size alone grounds for rejection, but the false alarm ratio needs resolution before publication."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee raises three substantive points, all of which we can address. The first two (MC1, MC2) concern the false alarm ratio calculation in §4.4: the referee correctly identifies that the full confusion matrix is not reported and that the treatment of FM→TM misclassifications as false alarms is underspecified. We agree this is a genuine gap and will revise. The third (MC3) concerns single-coder question classification; we will add a sensitivity analysis for the 'mixed' item and report the result transparently. No standing objections remain.","responses":[{"response":"The referee is correct that the false alarm ratio of ~4.3 is not derivable from Table 2 as presented, and that the treatment of FM→TM misclassifications as false alarms is not explicitly justified. This is a genuine gap in the manuscript. We will address it in two ways. First, we will report the full confusion matrix for Gemini 3 Flash in a revised Table (or supplementary table), making the calculation fully reproducible. Second, we will explicitly state our false alarm definition and justify it. Our original reasoning was that in a screening context where the system flags students for human review, any non-TM student flagged as TM consumes teacher attention. However, the referee makes a fair pedagogical point: FM students genuinely hold misconceptions and would benefit from intervention, so flagging them is not a 'false alarm' in the same sense as flagging a TC student. We will therefore present both ratios: (a) the broader screening-load ratio (~4.3, counting all non-TM→TM errors as flags requiring review) and (b) the narrower pedagogical false alarm ratio (~2.7, counting only TC→TM errors as true false alarms). We will label them distinctly and discuss the implications of each.","revision_made":"yes","referee_comment":"§4.4: The headline claim that Gemini 3 Flash generates 'roughly 4.3 false alarms per genuine detection' is not derivable from Table 2 alone. Table 2 reports per-class recalls but not the full confusion matrix or precision. The paper does not report the confusion matrix, specify this calculation, or justify why FM→TM errors should count as false alarms."},{"response":"We agree that the deployment conclusion should be revised to reflect both false alarm definitions. We will adopt option (a): report the full confusion matrix and present both ratios. Under the narrower, pedagogically natural definition (~2.7 false alarms per genuine detection), we will revise the deployment discussion to acknowledge that this ratio may be tolerable in a human-in-the-loop setting, particularly if screening is restricted to high-risk question types (where 71% of TM cases concentrate). We will soften the claim from 'fully automated screening is not viable on its own' to a more nuanced statement: standalone screening generates a non-trivial false alarm burden even under the narrower definition, but risk-stratified deployment (screening only on questions with coincidental correctness paths) could reduce this burden to a level that is practical for human-in-the-loop review. We will not overstate impracticality where the narrower ratio suggests tolerability.","revision_made":"yes","referee_comment":"§4.4: The deployment-impracticality conclusion rests on the false alarm ratio of ~4.3 at 1.6% TM prevalence. Under the more natural definition (only TC→TM as false alarms), the ratio is ~2.7, which may be tolerable in a human-in-the-loop setting. The paper should either (a) report the full confusion matrix and justify the false alarm definition, or (b) recompute the ratio under the pedagogically natural definition and revise the deployment conclusion accordingly."},{"response":"We agree that a second coder would strengthen the classification. We will attempt to add an inter-rater reliability check with a second coder for the revised submission. However, even if a second coder is not available in the revision timeframe, we can and will perform the sensitivity analysis the referee requests. The 'mixed' item (dot patterns, n=17 TM cases, 5.0% of all TM) is not among the two high-concentration items driving the main finding. The odds-ratio analysis in §4.2 compares procedural (n=6) vs. conceptual (n=8) questions; the mixed item is excluded from that calculation. We will explicitly state this in the revision. Additionally, we will report the odds ratio under both alternative classifications of the dot-patterns item (as procedural and as conceptual) to confirm that the conclusion—removing the top two items collapses the odds ratio from 5.6 to ~1.0—holds regardless of how the mixed item is classified. We expect this to be the case given that the item contributes only 5% of TM cases, but we will verify and report it transparently.","revision_made":"partial","referee_comment":"§3.2: The question classification (procedural/conceptual/mixed) was performed by a single researcher over n=15 questions. A second coder or inter-rater reliability check would substantially strengthen the item-vulnerability claims. At minimum, the paper should report whether the 'mixed' classification (dot patterns) affects the odds-ratio calculation and whether the conclusions hold under alternative classifications of that item."}],"tokens_in":7495,"tokens_out":1137,"duration_ms":83728,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The paper identifies a genuine failure mode in AI tutoring systems: when a student reaches the correct answer through flawed reasoning, models systematically miss the misconception. This is a real problem that the misconception detection literature has not directly addressed — prior work on Eedi conditions on incorrect answers. The framing of 'coincidental correctness paths' as an item-level vulnerability is the strongest contribution. It is concrete, grounded in assessment design theory (Barton, Hiebert & Lefevre), and gives question authors an actionable design-time check. The finding that 71% of failures concentrate in just two question types, and that removing them collapses the procedural/conceptual odds ratio to 1.0, is clean and useful. The PRM800K replication, while not a perfect match (model-generated solutions, not student work), shows the pattern generalizes beyond one dataset. Replication scripts and prompts are shipped on GitHub, which I appreciate. The integrated gradients analysis pointing to an answer-correctness shortcut is a nice supporting detail, though exploratory. The model comparison (RQ3) is the weakest section. The 61 TM test cases produce wide confidence intervals (±12pp), which the paper acknowledges honestly. The T5 vs. Gemini difference holds (p<0.005), but finer model distinctions do not. More importantly, the 'roughly four false alarms per detection' headline does not hold up under scrutiny. Table 2 reports per-class recalls only — no confusion matrix or precision. If only TC→TM misclassifications count as false alarms (students with no misconception flagged for intervention), the ratio is closer to 2.7, not 4.3. Reproducing 4.3 requires counting FM→TM errors as false alarms, but FM students genuinely have misconceptions — flagging them is not a false alarm in any pedagogically meaningful sense. The paper does not specify this calculation. This matters because the 'stand-alone screening impractical' conclusion weakens considerably under the more natural definition, and even 2.7 may be tolerable in a human-in-the-loop setting. The single-rater question classification (procedural/conceptual) is a minor concern given RQ2 focuses on item-level concentration rather than category prediction. The reader's CONDITIONAL verdict is about right. The core findings (RQ1, RQ2) are solid and the paper deserves a serious referee who can push on the false-alarm calculation and ask for the full confusion matrix. If the authors clarify that calculation and tone down the deployment-impracticality claim to match what the data actually show, this is a solid contribution to the educational AI subfield.","headline":"Identifies a real, actionable failure mode in AI tutoring (correct-answer-masks-bad-reasoning), but the headline false-alarm ratio is underspecified and likely overstated.","tokens_in":8357,"tokens_out":627,"would_cite":true,"duration_ms":123353,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"When the answer is right but the reasoning is wrong, AI tutors miss it","keywords":["correct answer trap","misconception detection","AI tutoring","reasoning assessment","shortcut learning","intelligent tutoring systems","educational AI","formative assessment"],"falsifier":"If questions were redesigned to eliminate coincidental correctness paths and models still failed to detect flawed reasoning at high rates, the item-driven explanation would be undermined — the problem would then lie elsewhere, perhaps in the models' inability to parse student explanations regardless of answer-correctness shortcuts.","tokens_in":7510,"feed_emoji":"🎓","tokens_out":997,"duration_ms":69886,"temperature":0.7,"pith_summary":"The paper identifies a failure mode it calls the *correct answer trap* (CAT): when a student reaches the correct numerical answer through flawed reasoning, AI models systematically fail to detect the underlying misconception. Analyzing real student responses from a mathematics tutoring platform, the authors find that 71% of these failures concentrate in just two question types, both of which share what the authors term a *coincidental correctness path* — a structural property where a common but incorrect procedure happens to produce the right answer for the specific numerical values chosen. The problem is not that procedural questions are broadly harder to assess; removing just these two items collapses the performance gap between procedural and conceptual questions entirely. The authors show that models learn an answer-correctness shortcut: when the final answer is correct, models default to treating the reasoning as sound. Stronger models reduce the gap — a frontier LLM achieves 84% detection of flawed-reasoning cases versus 57% for a fine-tuned smaller model — but do not close it. At the natural prevalence of such cases (roughly 1.6% of responses), even the best model generates about four false alarms for every genuine misconception detected, making fully automated screening impractical without human oversight.","feed_headline":"AI tutors miss flawed reasoning when the answer is right","feed_subtitle":"71% of detection failures concentrate in two question types where bad logic coincidentally produces the correct number, making automated筛查 a","key_machinery":"The correct answer trap (CAT) is the paper's named failure mode. The *coincidental correctness path* is the structural property of certain questions that makes them vulnerable. The True-Misconception (TM) category — correct answer with flawed reasoning — is the diagnostic class the authors isolate and evaluate. Integrated gradients analysis provides the mechanistic evidence that models rely on answer tokens rather than explanation tokens when they miss TM cases.","core_discovery":"The central discovery is that AI misconception-detection failures are not randomly distributed across questions but concentrate sharply in specific items that possess a *coincidental correctness path* — a property identifiable at question-authoring time where common flawed procedures produce the correct answer for the given numerical values. This concentration (71% of failures in two question types) means the vulnerability is item-driven rather than category-driven, and it reflects an answer-correctness shortcut that models learn: they treat a correct final answer as evidence of correct reasoning, even when the student's explanation reveals a misconception.","pith_inferences":["If coincidental correctness paths are the primary driver of the trap, then a question-authoring tool that automatically checks whether common errors yield the correct answer could prevent the vulnerability before deployment.","The false alarm ratio of roughly 4:1 is derived from a single dataset at one prevalence level; in unstructured classroom settings where TM prevalence is likely lower, the ratio would worsen, further reinforcing the case for human-in-the-loop rather than standalone automation.","The finding that longer thinking budgets slightly reduced balanced accuracy for the frontier model suggests that the shortcut is not purely a reasoning-depth problem but may reflect a deeper representational bias toward answer tokens."],"forward_implications":["Question authors can reduce the correct answer trap at design time by checking whether known student errors happen to yield the correct answer for the chosen numerical values, and revising the values if they do.","Risk-stratified deployment — routing only questions without coincidental correctness paths through automated screening — could reduce false alarm rates to operationally viable levels.","Moving from binary classification toward graduated assessment (e.g., follow-up probing questions or mark-scheme-style reference reasoning) may better approximate teacher practice and reduce the shortcut problem.","The answer-correctness shortcut may generalize beyond mathematics to any domain where final answers can be reached through multiple reasoning paths, some of which are flawed."],"fun_headline_variants":["AI tutors skip reasoning checks when the answer looks right","Correct answers hide flawed reasoning from AI tutors","AI tutors trust the answer and miss the broken logic","Right answer, wrong logic: AI tutors miss the gap","AI tutors judge the answer, not the reasoning"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The claim that the best model generates roughly four false alarms per genuine detection rests on 61 test cases of the rare True-Misconception category from a single mathematics dataset, yielding confidence intervals spanning about 12 percentage points and limiting how broadly the deployment-impracticality conclusion generalizes.","fun_headline_variants_meta":{"raw":{"variants":["AI tutors skip reasoning checks when the answer looks right","Correct answers hide flawed reasoning from AI tutors","AI tutors trust the answer and miss the broken logic","Right answer, wrong logic: AI tutors miss the gap","AI tutors judge the answer, not the reasoning","Model detection fails where bad logic yields the right number","AI tutors blind to flawed reasoning behind correct answers","Correct numerical answers mask student misconceptions from AI","AI tutoring models shortcut to the answer and miss the reasoning","Two question types drive 71% of AI reasoning-detection failures","AI tutors confuse a correct answer for correct reasoning","Even frontier models miss flawed logic when the answer is right"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":690,"prompt_tokens":468,"completion_tokens":222,"prompt_tokens_details":null},"tokens_in":468,"tokens_out":222,"duration_ms":9421,"temperature":1.0,"reasoning_tokens":94,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-05T13:06:23.840963+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If questions were redesigned to eliminate coincidental correctness paths and models still failed to detect flawed reasoning at high rates, the item-driven explanation would be undermined — the problem would then lie elsewhere, perhaps in the models' inability to parse student explanations regardless of answer-correctness shortcuts.","supporting_citations":[],"review_version":1}