concept
active
concept:introspection

Introspection

The ability of a model to observe its own past internal states or computations; claimed to be architecturally permitted by transformers.

Neighborhood — ranked by edge-count

Communities (1)

community

Methods (1)

method
  • KV caching
    implements
    Caching of key-value pairs to avoid recomputation; also provides a mechanism for introspection of earlier computations.

Concepts (7)

concept
  • Key gap identified in the literature; systematic self-examination processes for machine consciousness development.
  • The central concept: the ability of a model to access and report on its internal states, as defined by the paper's criteria.
  • The capacity of a model to self-report on its internal emotional state when its SAE features are steered, used here as a measurement tool
  • Pearson-Vogel et al.'s finding that models can detect prior concept injections; introspective signals exist in middle layers suppressed by post-training
  • The authors' characterization of genuine but limited introspective capability found only in early-layer injection regimes
  • World Models
    associated_with
    Theme issue context: relates to internal models of environment, central to consciousness and cognition across substrates.

Artifacts (1)

artifact

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

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

  • Stage 3 of character training: SFT on synthetic introspective data generated by post-distillation checkpoint
  • The capacity to detect and report one's own internal states, measured via the five-adjective task and paradox reflection
  • Tracking of functional/computational cognitive states, distinguished from phenomenal introspection.
  • Direct introspection into phenomenal consciousness; its correlation with functional introspection is an open question.
  • Identified gap; methods for enabling machine consciousness development through self-examination.
  • Spearman ρ measuring rank-order agreement between logit-based self-report and probe score; the paper's primary monotonic association metric
  • The novel framework introduced in the paper: an HMM-based pain-belief signal integrated into the reward function to drive exploration
  • Training data generated by the post-distillation model through self-reflection and self-interaction, capturing character nuances beyond the constitution