Summary

I report evidence that conversational “operators” (e.g. summarize, critique, reframe) correspond to stable, decodable internal states in transformer mid-layers, distinct from surface instruction wording and generalizing across content. These states are weak but persistent, geometrically separable via simple centroid methods, and survive instruction masking. I’m sharing this as a measurement result, not an intervention or architectural proposal, and I’m looking for feedback on interpretation and next experimental steps.

Motivation

We often talk informally about models entering different “reasoning modes” or “processing regimes,” but it’s unclear whether such modes correspond to measurable internal structure rather than prompt-level artifacts. This work explores whether operator-like distinctions can be detected as latent internal states, and where they localize in the network.

Setup

• Model: Qwen2.5-0.5B
• Task: Apply different conversational operators to diverse contents
• Capture: Fixed-window hidden states (32 generated tokens)
• Features: Directional deltas of hidden states, aggregated per layer
• Controls:
  • cross-topic generalization (LOCO)
  • cross-paraphrase generalization (LOPO)
  • instruction masking
  • label permutation baselines

Key Findings

1. Decodability: Operator identity is decodable above chance from mid-layer features.
2. Content Generalization: Signal survives leave-one-content-out evaluation.
3. Instruction Masking: Replacing operator instructions with a constant placeholder removes layer-0 signal but preserves mid-layer signal.
4. Geometry: Simple nearest-centroid classifiers on mid-layer features achieve F1 ≈ 0.61 (chance 0.33) on a partial dataset (N=600), indicating geometrically separable operator representations.
5. Layer Localization: Signal peaks in a mid-layer band and diminishes toward output layers.

What This Does Not Show

• No causal control or steering of operators
• No claim that operators are discrete modules
• No claim of full paraphrase invariance
• No architectural proposal

I interpret this as evidence for latent operator-conditioned internal regimes, not symbolic operators or explicit “manifold warping.”

Open Questions / Cruxes

• Are operators better modeled as directions or low-dimensional subspaces?
• How stable is this geometry across model scales?
• Which interventions (if any) would be legitimate tests of causal relevance?
• Is delta-based geometry the right observable, or are better ones needed?

If operator geometry collapses under stronger controls or larger models, that would falsify the core hypothesis.

Why This Might Matter for Alignment

If internal control states exist independently of surface prompts, they may represent a locus for:

• monitoring model behavior beyond text outputs
• understanding internal goal-conditioned processing
• designing better interpretability probes

I’m especially interested in feedback from people working on interpretability, evals, or internal representations.

Disclosure

I used LLMs as research assistants for coding support, experiment iteration, and drafting clarity. All experimental design decisions, interpretations, and claims are my own.