LLMs encode framing effects as distinct hidden states, researchers show

The paper's core contribution isn't just that framing affects answers (known), but that these effects operate through discrete, swappable computational channels in intermediate layers rather than diffuse reasoning differences. This specificity matters because it suggests the problem may be addressable through targeted interventions rather than wholesale retraining.

This connects directly to the BenchMIRT finding that benchmarks often measure narrow task performance without capturing what's actually happening inside the model. StateSwap reveals a concrete mechanism behind one class of benchmark fragility: identical logical content produces different outputs based on hidden state activation patterns, not surface reasoning quality. It also echoes the Latent Recurrent Thoughts work from the same day, which showed that refinement in representation space (rather than token space) can unlock better inference. Here, the discovery that support and elimination framings activate separable latent pathways suggests similar internal structure exists for other reasoning modes. Together, these papers point toward a shift in how we think about model robustness: not as a property of the model itself, but as a property of how its internal representations respond to input framing.

If researchers can reliably suppress framing effects by applying adversarial activation swaps during inference on held-out benchmarks (MMLU, HellaSwag) without degrading baseline accuracy, that confirms the mechanism is both real and controllable. If the effect disappears in larger models (70B+) or newer architectures, that would suggest it's a scaling artifact rather than a fundamental property worth defending against.