Self-description
Describes its own responses
Its summary is trained against updates measured in the same loop.
—selected next-state distance
Starting-state number → next-state distance
Same text. Different starting states.
Pause at one recorded text position and replace the loop’s control state with each of 12 test states. If they all lead to the same next state, the current state makes little difference at this moment. Separated next states show that it matters.
1. Keep text and trained weights fixed2. Try 12 starting states3. Compare their next states
Each bar measures how far a next state is from the reference’s next state. Zero means the same result; 2 is the greatest possible separation. All eight state coordinates are used.
Select a bar to compare that test with the reference.
Control
Describes another loop’s recordings
The training target comes from a separate recorded run.
—selected next-state distance
Same starting-state tests · same distance scale
Loading the measured responses…
Inspect other recorded text positions and copies
The initial view is text position 600 in copy 1 of a selected strong run. All 60 sampled positions and four copies are available below. Each copy has its own text stream; the two interfaces are evaluated on matching external text. At each context, the model cache is held fixed while the starting state is varied.
What can this comparison tell us?
The bars show measured one-step responses of trained interfaces. At the default moment the self-description run produces widely separated next states, often close to two groups; the control’s next states remain close together. These are numerical control states, without demonstrated word labels or psychological meanings.
This selected pair does not establish a general success rate or isolate every training difference. In the initial small-model comparison, 2 of 5 self-description runs sustained substantial dependence without persistent two-state oscillation. The path cuts in the next tab test which feedback route causes the movements after training.
Saved measurements; no model inference runs in this page.
Which feedback path causes the large movements?
The control state can change the small reader directly. It can also change the language model’s attention through the shared bias. Hold either route fixed and measure how often the loop still takes a large step.
These buttons select saved intervention results. They do not retrain or simulate a new loop.
Steps with distance greater than 0.5
—
0% of steps100%
The reader is the main return path
Large movements almost disappear when the state can no longer change the reader’s query. They remain when the separate model-attention route is held fixed. This supports a causal role for feedback through the reader in this trained run.
The reading still changes model predictions
Across five self-description runs, applying the reader’s attention inside the model changes its highest-probability next token at 11–16% of positions on matched texts. Prediction loss changes by −0.14% to +1.51%. A control describing another loop changes predictions too.
Influence on the model and feedback from the model are separate measurements. Interventions restricting the state to the model route produce less than 1% of the full starting-state-dependent variation. Most of the loop’s measured dependence remains in the small reader.
All six final interfaces and measurement definitions
| Interface | Dependence ratio | Pattern correlation | Top prediction changed |
|---|
The dependence ratio compares centered next-state variation across starting states with displacement variation. It is not a probability or fraction of steps. Pattern correlation compares predicted and measured response shapes. The predictions have large scale errors despite positive correlations. The field-prediction network shapes training but is not used to select the next state during operation.
Path cuts: strongest developmental run, with learned weights fixed. Large steps use Euclidean distance between unit control states. Sources and definitions.
Train a description of how the loop responds
The language model’s weights stay fixed. A small network learns how to read its recent activity and carry a state between readings.
- 1 · READ
Weight recent activity
The current eight-number state sets a query over the model’s activity at 16 recent text positions.
- 2 · SUMMARIZE
Compress the reading
The weighted reading becomes a 32-number summary. “Summary” means a learned vector, not a sentence or image.
- 3 · RETURN
Set the next state
A decoder maps the summary to eight control numbers. Those numbers change the next query, closing the loop.
A description used during training
A second network takes the summary and candidate starting states and predicts their updates. Training asks the summary to distinguish its copy’s response pattern from other copies’ patterns. The four copies share weights while reading different texts. In the control, the target is another run’s recorded responses.
Remove a common-destination shortcut
Different starting states can all lead to one next state. Naming that destination can identify the copy without describing any dependence on the starting state. Subtracting the common destination leaves the differences between possible next states for the summary to distinguish. This is the centered-field objective.
The description and the loop it describes can both change under this training loss. The prediction network is not consulted during operation: the summary decoder still supplies the next state directly.
Let the reading affect the model
The same attention weights can also bias attention inside the language model. This shared channel, called the spotlight, lets prediction loss reach the small reader. The path-cut experiment distinguishes this route from feedback directly through the reader.
Dimensions, code and source recordings
The control vector has eight coordinates and length one. The summary has 32 coordinates and length √32. The main model is Qwen3-0.6B; only the surrounding interface is trained. Methods and reproduction code specify numerical settings and configurations.
Displayed response arrays come from fields_iface_3000.npz in the developmental self-description and recorded-other runs. The 12 unit probe states use seed 4242. The exporter includes source-file hashes and all sampled contexts. Download displayed data · Data exporter.