acceptodds
Desk rejected at ICLR 2027

When Context Sticks: Hysteresis from Repeated Semantic Framing in LLMs

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models increasingly operate over long multi turn histories, yet it remains unclear whether repeated semantic framing acts only as a local prompt effect or can establish a persistent history dependent internal state. We study this question using fact matched dialogue trajectories, dose response controls, residual stream measurements, a Same Present, Different Past protocol, reversal tests, and activation level interventions across 8 open weight models. Behavioral framing preference increased from 0.155 after one exposure to 0.596 after five exposures while the most recent framed turn was held fixed. Frame related residual projections accumulated with exposure and remained separable after token identical neutral washout; after eight neutral turns, 11.6% of the behavioral and 19.8% of the representational A-versus-B separation remained. Adding, removing, or transferring the identified frame component produced directionally consistent behavioral changes, whereas random and orthogonal controls were near zero. The pattern generalized across framing axes and appeared in observational and counterfactual replay analyses of real-world conversation structures.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Context Effects, Semantic Framing, Internal Representations

Related papers

Discussion (0)

Sign in to comment.