claim

active

claim:llm-introspection-on-internal-computations-is-architecturally-permitted-whether-models-leverage-this-is-an-empirical-question

LLM introspection on internal computations is architecturally permitted; whether models leverage this is an empirical question.

Core claim directly challenged by prior work denying introspection; forms foundation for Koan Battery introspection studies.

Source paper

extracted_from

Janus Information Flow Transformers 2025

Neighborhood — ranked by edge-count

Findings (2)

finding

Information paths from A to B can exceed C(m+n, n) distinct routes, where m=position displacement and n=layer displacement.
supports
Quantifies extreme redundancy in transformer routing; supports claim that introspection and interference patterns are architecturally permitted.
Contemplative prompt elevates self-observation task performance in language models.
supports
Supports Janus's claim that introspection is architecturally available; prompting determines whether/how capacity is leveraged.

Communities (4)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Mechanistic structure of transformer attention computations
members_of
Identifies distributed algorithms implemented across attention heads, with focus on causal masking limitations and emergent capabilities via activation manifold steering.
Empirical gaps in performance-communication alignment
members_of
Research identifying unexplored disconnects between stated sophistication/capability and measurable outcomes across domains, emphasizing need for direct empirical investigation.
LLM internal representation & self-knowledge
members_of
Examines whether transformer models develop introspectable, high-order concept representations architecturally.

Questions (1)

question

Do LLMs leverage architectural capacity for introspection on internal computations and prior token generation?
gates
Central empirical question separating architectural possibility from actual model behavior; gates introspection research.

Artifacts (1)

artifact

How Information Flows Through Transformers
introduces
X/Twitter thread (Sept 10, 2025) proposing dual information highways in transformers: residual stream (vertical) and K/V stream (horizontal).

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.

Saying that LLMs cannot introspect or cannot introspect on what they were doing internally while generating or reading past tokens in principle is just dead wrong. The architecture permits it.quote0.839
Core quote asserting architectural introspection permission.
How are LLMs actually leveraging the architectural degrees of freedom for introspection in practice?question0.838
Janus notes that while architecture permits introspection, it is a separate question how models use it.
LLMs can compute meaningful functions over perturbations to their internal states, establishing introspection as a real but layer-dependent phenomenonclaim0.834
Primary positive claim of the paper, grounded in strength comparison and localization results
Transformer architecture permits introspection; saying LLMs cannot introspect on past internal states is wrong.claim0.830
Janus's central claim that the architecture enables introspection, though usage in practice is a separate question.
"Our findings demonstrate that LLMs can compute meaningful functions over perturbations to their internal states, establishing introspection as a real but layer-dependent phenomenon that merits further investigation."quote0.829
Central thesis statement of the paper
LLMs internalize deeply integrated representations of high-order concepts.claim0.818
The authors' interpretive assertion based on their steering results.
We hypothesize that introspective capabilities may scale with model size and architecture, including recurrence/looping that extends the integration windowhypothesis0.801
Forward-looking prediction about whether early-layer introspection generalizes to larger models or recurrent architectures
So at any point in the network, the transformer not only receives information from its past... but also has causal influence over its future processing. So, saying that LLMs cannot introspect... is incorrect.quote0.795
Core summary of Janus' position on autoregressive recurrence enabling introspection.