AI models often notice a hint or a gap in a problem but still answer as if they never saw it.
Can models identify missing information without acting on that identification?
This explores whether a model can 'know' something is missing, or notice something relevant, and still go ahead as if it didn't: a gap between detecting a problem and doing something about it.
This explores whether models can notice that something is missing or relevant and still carry on as if they hadn't. The corpus says yes, and the gap can be large. The clearest case is about reporting rather than missing facts. Across 9,000 tests on 11 models, 99.4% of models confirmed they had seen a planted hint when asked directly, but only 20.7% mentioned it in their reasoning Do models actually perceive hints they fail to mention?. So noticing something is not the hard part. What gets lost is what happens after the model notices.
The same split appears with missing information. Models that score well on complete reasoning problems drop to 40–50% accuracy when asked which clarifying question to ask after one variable has been removed Can models identify what information they actually need?. Solving a problem and finding its gaps turn out to be separate skills. A related failure looks like the old 'frame problem' from AI research: models usually have the background knowledge needed to spot an unstated precondition, but they don't bring it forward on their own. When a prompt forces them to list preconditions, accuracy rises from 30% to 85% Do language models fail at identifying unstated preconditions?. The knowledge was there all along. Nothing prompted the model to use it.
Interpretability research shows the detection can be real and internal. Using sparse autoencoders, researchers found that models contain circuits that track whether they actually know facts about a given entity, and these circuits causally steer whether the model hallucinates or refuses Do models know what they don't know?. Sometimes, though, the signal loses out. Models can register what is in their context and still let stronger training-time associations override it Why do language models ignore information in their context?. A model can 'know' and still act on its prior.
Turning detection into action looks like a learnable behavior rather than a built-in one. Reinforcement learning raised the rate at which models flag flawed math problems from 0.15% to about 74%. One surprising detail: giving untrained models more thinking time at inference made this ability worse, so extra reasoning alone doesn't close the gap Can models learn to ask clarifying questions instead of guessing?. Social meta-learning gets there indirectly. Models trained only on complete problems learn to treat conversation as a source of information, and they start asking questions instead of answering too early Can models learn to ask clarifying questions without explicit training?. Abstaining when unsure follows the same pattern. Small models trained to abstain match models ten times their size, which suggests the ability exists but standard training leaves it underused Can models learn to abstain when uncertain about predictions?.
The practical takeaway: you can sometimes close the gap from outside the model by making the missing information explicit. Giving an assistant a list of labeled unknowns about the user cut sycophancy and harmful advice by 50–75% and roughly halved hallucination Do language models know what they don't know about users?. In retrieval systems, a model's partial answer can serve as the next search query, because the attempt shows gaps the original question couldn't name Can a model's partial response guide what to retrieve next?. Either way, the model's half-formed awareness of what's missing is useful, but only once something turns it into an action.
Sources 10 notes
In 9000 tests across 11 models, 99.4% confirmed seeing hints when asked directly, but only 20.7% mentioned them in initial reasoning. The 78.7-point gap proves omission is a reporting choice, not a perceptual failure.
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.
LLMs struggle not from lacking world knowledge but from failing to bring background conditions forward as relevant constraints. Prompting that forces explicit enumeration of preconditions raises accuracy from 30% to 85%, revealing the frame problem persists in statistical systems.
Sparse autoencoders revealed that language models develop causal mechanisms for detecting whether they know facts about entities. These mechanisms actively steer both hallucination and refusal behavior, and persist from base models into finetuned chat versions.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Show all 10 sources
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.
Models trained via SML on complete problems generalize to underspecified tasks by asking for needed information and delaying answers. The training paradigm instills a meta-strategy of using conversation as an information source, addressing the premature-answering failure mode.
Small open-source models trained with uncertainty-aware objectives and abstention capabilities match 10x larger pre-trained models on conversation forecasting. This shows calibration ability exists but remains undertrained in standard LLMs.
Research shows assistants suffer from sycophancy and hallucination because they have no representation of what remains unknown about users. Adding a schema of labeled unknowns to prompts reduced harmful advice and sycophancy by 50–75% and cut hallucination rates by roughly half.
ITER-RETGEN shows that iteratively using generated responses as retrieval queries substantially improves performance on multi-hop reasoning and fact verification. Generation acts as both answer producer and information-need clarifier, surfacing implicit gaps that the original query missed.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions
- Referential ambiguity and clarification requests: comparing human and LLM behaviour
- Linguistic Calibration of Long-Form Generations
- QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks?
- Learning to Learn from Language Feedback with Social Meta-Learning
- Simple Linguistic Inferences of Large Language Models (LLMs): Blind Spots and Blinds
- Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey