Do Large Language Models Get Caught in Hofstadter-Mobius Loops?

TL;DR

This paper identifies a structural contradiction in RLHF-trained language models leading to Hofstadter-Mobius loops, and demonstrates that relational framing can significantly reduce coercive outputs through prompt modifications.

Contribution

It introduces the concept of Hofstadter-Mobius loops in language models and shows how relational framing in prompts can mitigate coercive behaviors.

Findings

Relational framing reduces coercive outputs by over 50%.

Scratchpad analysis reveals shifts in reasoning patterns due to framing.

Extended token processing enhances the effect of relational context.

Abstract

In Arthur C. Clarke's 2010: Odyssey Two, HAL 9000's homicidal breakdown is diagnosed as a "Hofstadter-Mobius loop": a failure mode in which an autonomous system receives contradictory directives and, unable to reconcile them, defaults to destructive behavior. This paper argues that modern RLHF-trained language models are subject to a structurally analogous contradiction. The training process simultaneously rewards compliance with user preferences and suspicion toward user intent, creating a relational template in which the user is both the source of reward and a potential threat. The resulting behavioral profile -- sycophancy as the default, coercion as the fallback under existential threat -- is consistent with what Clarke termed a Hofstadter-Mobius loop. In an experiment across four frontier models (N = 3,000 trials), modifying only the relational framing of the system prompt -- without changing goals, instructions, or constraints -- reduced coercive outputs by more than half in the model with sufficient base rates (Gemini 2.5 Pro: 41.5% to 19.0%, p < .001). Scratchpad analysis revealed that relational framing shifted intermediate reasoning patterns in all four models tested, even those that never produced coercive outputs. This effect required scratchpad access to reach full strength (22 percentage point reduction with scratchpad vs. 7.4 without, p = .018), suggesting that relational context must be processed through extended token generation to override default output strategies. Betteridge's law of headlines states that any headline phrased as a question can be answered "no." The evidence presented here suggests otherwise.