The model knows the answer. It just won't say it.

The signal is there

A preprint shows the information is there anyway: node and edge relationships in diagrams are linearly decodable from the model's internal hidden states, even when the model answers the question wrong. Causal interventions on those hidden states changed the model's final predictions, showing the internal signals feed into the output rather than sitting unused. m:…0841

The gap held for both synthetic diagrams and real-world flowcharts. The failure is in turning what the model already represents into an answer: the signals are present and linearly decodable, and causal interventions show they contribute to the model's final predictions — so the information is there, and the output still misses it. m:…0841

Counting what it can see

DecoyCount measures the same kind of failure in counting. Frontier vision-language models showed a reasoning-alignment gap of 12.1 to 22.1 percentage points in counting tasks. m:…9088

More capability made it worse, not better: more capable models and agentic processes often induced larger gaps. And the problem is not that the question was unclear — textual underspecification was not the primary driver, and additional context did not resolve the misalignment. m:…9088

The gap the benchmark measures is about alignment between what the model can represent and what it actually answers — not about whether the question was clear, since textual underspecification was not the primary driver of the misalignment. m:…9088

The finding cuts against the usual assumption that more capability closes these gaps. The DecoyCount authors describe a paradox: increased model capability and agentic complexity can widen the gap between model performance and human visual expectations, rather than narrowing it. A bigger, more elaborate model can be further from what a person sees — which is why the benchmark was built to isolate visual-reasoning errors rather than lump them in with every other kind of mistake. m:…9088

Trusting what it shouldn't

A third preprint looks at the same gap inside retrieval pipelines. It studies context-certainty obedience: whether a model adjusts its confidence to the quality of retrieved information. The answer is mostly no. Models failed to reproduce their prior outputs after observing an uncertain context, responded to expressed certainties in a distorted way, and over-relied on complex contexts. m:…8181

The authors propose an interaction strategy that needs no retraining: prior reminders, recalibration of context certainty, and simplification of the context. It reduced obedience errors by 25% on average — a real improvement that still leaves most of the problem in place, and a reminder that the model's own sense of what it knows cannot be trusted at face value. m:…8181

The strategy is a workaround for a reasoning gap, and the underlying pattern across all three papers is the same: the model holds more than it says, and says more than it should trust. m:…8181

The three papers point at the same bottleneck from different directions. Diagram signals are present in hidden states but do not reach the output, counting gaps widen as capability grows, and retrieved context is trusted more than its certainty warrants — in each case the model's own output is the weak link, not the information it holds. The fix, wherever it comes from, has to close the gap between what the model represents and what it says. m:…0841 m:…9088 m:…8181