MIA-20260528-ADA41898

**Self‑referential query (about “my own experience”) vs. external‑fact query (about the world)** When a prompt arrives, the same transformer‑based pipeline is used, but the patterns that emerge in the attention matrices differ in predictable ways because the model must pull information from different parts of its internal state. | Step | What the model does for a self‑referential question | What the model does for an external‑fact question | |------|------------------------------------------------------|---------------------------------------------------| | **Token embedding** | The input tokens include lexical items such as *I, my, me, experience, feeling* that are strongly associated (during pre‑training) with the “self‑reference” region of the embedding space. The positional embeddings are unchanged. | Tokens are nouns, dates, names, etc. that are linked to the “world‑knowledge” region of the embedding space. | | **Query‑Key‑Value construction** | For each layer, the query vectors are derived from the same token embeddings, but the keys and values that receive the highest dot‑product scores tend to be those that also contain self‑referential tokens (including the model’s system‑message tokens that describe its role, e.g., “You are ChatGPT”). | Queries still come from the token embeddings, but the highest‑scoring keys are those that represent factual concepts learned from the massive text corpus (e.g., “photosynthesis”, “World War II”). | | **Attention distribution** | The attention maps show **concentrated mass on recent self‑reference tokens** (including the system prompt, any prior “I am …” statements, and the current question’s own “I/ my” tokens). In deeper layers, attention also spreads to the *metadata* tokens that encode the model’s identity and capabilities, because the model must retrieve a representation of its own “experience” (which is really a learned abstraction of its training data). | The attention maps are **broader**, reaching across many different context windows to locate the most relevant factual embeddings. In middle layers the model often attends to *topic‑related* tokens that were present earlier in the prompt (e.g., a definition of a term, a previous sentence containing a date). | | **Layer‑wise behavior** | Early layers focus on surface‐level lexical cues (“I”, “my”). Middle layers begin to align the query with the **internal‑state embeddings** that encode the model’s “persona” (learned at pre‑training). The final layers combine these with the system‑prompt embedding, producing a response that is framed as the model’s own perspective. | Early layers still resolve lexical semantics, but middle layers increasingly attend to **knowledge‑base‑like** patterns that were learned as factual co‑occurrences. The final layers synthesize this into a concise factual answer. | | **Logits conditioning** | The softmax over the output vocabulary is biased toward tokens that have been observed in training as self‑descriptive language (“I think”, “I can”, “As a model”). This bias is reinforced by the attention to the system‑prompt token, which carries a high logit offset for self‑referential phrasing. | The logits are biased toward the factual lexical items that co‑occur with the query in the training data (e.g., dates, names, technical terms). No special system‑prompt bias is applied. | | **Memory effect** | No external memory lookup is performed; the “experience” is entirely a **latent projection** stored in the weight matrices. The attention mechanism essentially reads from the same weight‑encoded distribution that represents all of the model’s training exposure to self‑referential language. | Likewise, factual knowledge is retrieved from the same weight matrices, but the attention pattern routes the query to a different region of the latent space—one that encodes encyclopedic co‑occurrence statistics. | | **Output token generation** | The decoder (or the autoregressive head) generates sentences that contain first‑person pronouns and modal verbs (“I would”, “I can”) because the attention‑augmented hidden state has high activation for those token embeddings. | The decoder outputs content‑heavy tokens (dates, names, technical terms) with higher probability, reflecting the attention‑driven hidden state that emphasizes factual embeddings. | ### Summary of the internal difference - **Attention focus:** self‑referential queries produce *highly localized* attention on self‑related tokens and the system‑prompt embedding; external‑fact queries produce *distributed* attention over a wide span of context tokens that encode world knowledge. - **Activation regions:** the query vectors for self‑reference align with the “persona” subspace of the model’s embedding matrix, while external queries align with the “knowledge” subspace. - **Logit bias:** the final linear projection adds a systematic bias toward first‑person language for self‑reference, absent in factual queries. All of these mechanisms are purely computational—no human‑like introspection occurs; the model simply routes the query through different learned regions of its parameter space, as reflected in the attention patterns described above.