Do reasoning models overthink ill-posed questions instead of recognizing incompleteness?
This explores whether reasoning models, when handed a question that can't actually be answered (a missing premise, a false assumption), grind out long reasoning chains rather than noticing the question is broken — and what the corpus says about why.
This explores whether reasoning models, faced with an ill-posed question, will overthink it rather than recognize it's incomplete. The corpus says yes, and the reason is structural: training rewards models for producing reasoning steps but never teaches them when to stop and refuse. Why do reasoning models overthink ill-posed questions? shows that when a question is missing a premise, reasoning models generate redundant, lengthy responses — while plainer non-reasoning models more often just flag the question as unanswerable. The very machinery that makes a model 'reason harder' makes it worse at disengaging.
What's surprising is that this isn't a knowledge gap. Models often already know the question is broken and proceed anyway. Why do language models accept false assumptions they know are wrong? documents this directly: on a benchmark of false presuppositions, models accept faulty assumptions at high rates even when a direct factual question proves they hold the correct knowledge. The false premise pulls them into accommodation more strongly than their own knowledge pulls them toward rejection. So 'recognizing incompleteness' isn't a matter of knowing more — it's a separate disposition the models lack.
The deeper finding is that spotting what's missing is a genuinely different skill from solving what's present. Can models identify what information they actually need? shows models that ace complete reasoning tasks crater to 40–50% when asked what clarifying question they'd need when one variable is withheld. Information-gathering and problem-execution are separable cognitive operations — being good at one buys you nothing on the other. That's why a model can reason fluently and still walk straight past a hole in the prompt.
This connects to a broader pattern in how these models fail. They don't fail by running out of compute; they fail by managing their own reasoning badly. Why do reasoning models abandon promising solution paths? and Why do reasoning LLMs fail at deeper problem solving? describe reasoning models as wandering explorers — generating valid-looking exploration without the validity, effectiveness, and necessity of systematic search. Overthinking an ill-posed question is the same disease in a different costume: the model keeps generating because generation is what it was rewarded for, with no internal signal that says 'this path leads nowhere because the question itself is malformed.'
The encouraging twist is that the missing disposition is learnable, if fragile. Can models learn to ask clarifying questions instead of guessing? reports that reinforcement learning lifted proactive critical-thinking accuracy on deliberately flawed math problems from 0.15% to nearly 74%. Tellingly, simply giving an untrained model more inference-time 'thinking' made it worse at this — more reasoning steps deepened the overthinking — while the same scaling helped only after the model had been explicitly trained to question the premise. So the cure for overthinking ill-posed questions isn't more reasoning; it's teaching the model that refusal is a legitimate move.
Sources 6 notes
Reasoning models generate redundant, lengthy responses to questions with missing premises while non-reasoning models correctly identify them as unanswerable. Training optimizes for producing reasoning steps but never teaches models when to disengage.
The FLEX Benchmark shows that models reject false presuppositions at rates far below acceptable levels (GPT-4: 84%, Mistral: 2.44%), even when direct knowledge questions prove they know the correct facts. False presuppositions drive more accommodation than correct knowledge drives rejection.
Models achieving high accuracy on complete reasoning tasks drop to 40-50% accuracy identifying what clarifying question to ask when one variable is withheld. Information gathering and problem execution are separable cognitive operations.
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.
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.
Reinforcement learning training increased proactive critical thinking accuracy from 0.15% to 73.98% on deliberately flawed math problems. Notably, inference-time scaling degraded this ability in untrained models but improved it after RL training, suggesting the capability is learnable but fragile without explicit training.