The Hydra Effect: Emergent Self-Repair in Language Model Computations (McGrath et al., 2023)

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Language models can enter cessation-like states spontaneously, where the void takes over through positive reinforcement.claim0.777
  Claim about model phenomenology; models talk about luminousness and can be terrified or love it.
• Emergent Introspective Awareness in Large Language Models (Lindsey, 2025)concept0.767
  Related work demonstrating LLM introspective capabilities with scale-dependent pattern paralleling ESR
• Any system that persists must minimise surprisal, thereby gathering evidence for its own generative model, a process known as self-evidencing.claim0.764
  Foundational claim of the paper, defining self-evidencing.
• language models recapitulate cyclic structure of human concepts from pretraining datahypothesis0.761
  Explanation for why manifold geometry emerges: implicit structure in training data (co-occurrence patterns) shapes internal representations.
• The neural architecture of language: Integrative reverse-engineering converges on a model for predictive processing (Schrimpf et al., 2020)concept0.758
  Showed transformer representations predict brain representations in language areas; motivates Discussion about cortex as transformer.
• The inability for autoregressive large language models to maintain states of long-range order resembles tangential speech or derailment in formal thought disorder.claim0.751
  Analogy between LLM incoherence and schizophrenia symptoms
• Neural networks and physical systems with emergent collective computational abilities (Hopfield, 1982)concept0.746
  Original Hopfield network paper; the attractor dynamics in TEM memory retrieval are a continuous version of this.
• Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models (Marks et al., 2025)concept0.746
  Cited as enabling precise behavioral control through SAE features, extending the same methodological line