What happens when students encounter errors they cannot resolve through prompting alone?
This explores what happens to learning when a student hits an error that prompting an AI won't fix — and what the corpus says about errors as a learning channel rather than an obstacle to clear away.
This explores what happens to learning when a student hits an error that prompting an AI won't fix — and the corpus reframes the question in a way you might not expect: the unresolvable error isn't the problem, it's the point. The most direct finding is that struggling with errors and resolving them independently is itself a learning channel, and AI assistance quietly removes it. Learners working without AI encountered more errors and worked through them on their own, and they retained more skill as a result; the ones who leaned hardest on AI to debug scored lowest on later assessments Does AI assistance remove a core learning channel through error work?. So when prompting alone can't dissolve an error, the student is pushed back into exactly the cognitive work that produces durable skill — the moment that feels like failure is the moment learning actually happens.
There's a deeper reason prompting hits a wall, and it lives on the model's side as much as the student's. LLMs exhibit a kind of split between knowing and doing: they can state a correct principle and then fail to execute it — roughly 87% accuracy explaining a concept versus 64% applying it Can language models understand without actually executing correctly?. A related pattern, 'Potemkin understanding,' shows models explaining a concept correctly, failing to apply it, and even recognizing the failure — all at once Can LLMs understand concepts they cannot apply?. If the tool a student is prompting shares this comprehension-without-competence gap, no amount of rephrasing the prompt closes it, because the breakdown is in execution, not explanation.
The more surprising thread is what the corpus says about errors as teaching material. Training a model to *critique* flawed answers produces deeper understanding than training it to imitate correct ones — engaging with failure modes builds structural reasoning that copying right answers never does Does critiquing errors teach deeper understanding than imitating correct answers?. Training on the full messy search process, mistakes and backtracking included, yields problem-solvers 25% better than training only on clean optimal paths Does training on messy search processes improve reasoning?. The same principle that makes errors valuable for a learner makes them valuable for a model: the detour through what went wrong is where the real learning is.
There's also a question of *where* the unresolvable error actually lives, and it's often not where it surfaces. Process verification — checking the intermediate steps rather than the final answer — catches failures that scoring the end result misses entirely, lifting success from 32% to 87% Where do reasoning agents actually fail during long traces?. Reasoning models tend to wander unsystematically rather than search, so success drops off sharply as problems get deeper Why do reasoning LLMs fail at deeper problem solving?. And in extended back-and-forth, models lock into a wrong early assumption and can't recover — a 39% average performance drop in multi-turn settings, with mitigations clawing back only 15-20% Why do language models fail in gradually revealed conversations?. So 'prompting harder' frequently fails because the error was seeded turns ago, in the process, not in the last prompt.
The constructive turn is that prompting isn't the only mode available. Social meta-learning trains models to actively solicit and use corrective feedback through dialogue — treating conversation as a problem-solving tool rather than a one-shot request Can LLMs learn to ask for feedback during problem solving?. The takeaway for a student stuck at an unresolvable error: the instinct to reframe the prompt one more time is often the wrong move. Stepping back into independent debugging, or shifting from asking for the answer to interrogating the process, is where both humans and models actually get unstuck.
Sources 9 notes
Research shows learners without AI encountered more errors and resolved them independently, resulting in higher skill retention. AI-assisted learners delegated debugging to AI, bypassing the cognitive work that produces learning—even those who debugged most with AI scored lowest on skill assessments.
Large language models can articulate correct principles but systematically fail to apply them due to dissociated instruction and execution pathways. The 87% accuracy in explanations versus 64% in actions reveals this is not knowledge deficit but structural disconnect.
Models can explain concepts accurately, fail to apply them, and recognize the failure—a triple pattern incompatible with human cognition. This indicates functionally disconnected explanation and execution pathways rather than simple knowledge gaps.
Training models to critique noisy responses outperforms training on correct answers because critique forces engagement with failure modes and structural reasoning. Even imperfect critique supervision beats correct-answer imitation, showing how weak surface-pattern learning is for building genuine understanding.
Stream of Search pretraining, which represents exploration and backtracking as serialized strings, achieves 25% higher accuracy than optimal-trajectory-only training. Models learn internal world models for search and adaptive strategies rather than fixed external methods.
Reliability for long-trace reasoning comes from checking intermediate states and policy compliance during generation, not from scoring final outputs. Adding intermediate verification raised task success from 32% to 87% because most failures are process violations, not wrong answers.
Current reasoning models lack the three properties of systematic exploration: validity, effectiveness, and necessity. This causes success probability to drop exponentially with problem depth, making medium problems solvable but deep problems catastrophically harder.
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
Research shows that reformulating static tasks as pedagogical dialogues—where a teacher has privileged information and the student must learn to extract it—trains models to actively engage conversation as a problem-solving tool, not just imitate dialogue patterns.