Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2025
|
| Online Access: | https://doi.org/10.5281/zenodo.17850797 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <p>Most modern large models are trained once in a massive offline phase and then deployed<br>with nearly fixed parameters. They are brittle in non-stationary environments: they suffer catastrophic forgetting, are corrupted by noisy data, and do not manage their own plasticity. Biological agents appear to behave differently. They maintain multiple internal “worlds” or contexts and use fast value-like signals—“feelings” such as confusion, familiarity, and risk—to decide when and how to change themselves. This unified paper collects and synthesizes five previous installments on Emotion-Guided<br>World Models (EGWM). Across Parts I–V we build a sequence of toy, yet fully implemented, agents in simple 2D classification environments with multiple hidden worlds and noisy, nonstationary streams of experience. In Part I, we introduce EGWM and show that even a single logistic regression with a simple phase-consistency feeling can refuse self-inconsistent phases and slightly improve robustness to label noise. We also add an Emotion-Guided Experiment Planner (EGEP) that uses uncertainty to choose which unlabeled points to query, achieving better sample efficiency than random querying. In Part II, we move from one model to a world bank of specialist heads, governed by a<br>small set of value channels (self-consistency, familiarity, mismatch, disagreement, and a crude elegance proxy). A hand-designed emotion-gated policy that chooses when to ignore, spawn, or update worlds cleanly recovers multiple regimes in a continual stream where a monolithic learner plateaus around 0.60–0.70 accuracy. An emotion-gated world bank instead reaches per-world accuracies in the 0.94–0.97 range. We show that the key discovery feeling is not inter-model disagreement, but mismatch between a fresh scratch model and the best existing world; “spawnhappy” gates that over-segment the world early and compress later perform best. Tiny learned governors trained with REINFORCE can match hand-crafted policies for accuracy and fairness, but naively elegance-weighted rewards often drive the controller toward inaction. In Part III, we fix this “elegance collapse” by separating growth and compression. A confusion-driven growth phase creates a rich, redundant world bank; a competence-constrained compression phase merges heads only when it can do so without harming accuracy. In a stationary 3-world environment, staged grow–compress reduces the number of heads from ∼ 3–4 down to about 2 while maintaining near-100% per-world accuracy and sharply reducing overlap. In a non-stationary 4-world environment where worlds appear, disappear, and reappear, a<br>∗Independent researcher. 1monolithic model exhibits catastrophic interference, while the grow–compress bank maintains high accuracy and a bounded number of heads. Sweeping the compression target reveals a competence–elegance trade-off; an annealed schedule (lenient early, strict late) allows exploration first and compact structure later. In Part IV, we synthesize the architecture and extend it with multi-time-scale value signals.<br>Fast feelings (confusion, uncertainty) operate per batch; medium-scale feelings (competence, elegance) operate over segments; slow feelings (relevance) integrate age and importance. In segmented lives with “alive” and “dead” worlds, age-only pruning deletes both rare-but-important and truly dead worlds; reward-weighted relevance prunes unimportant heads while preserving high-importance worlds and reduces the number of heads by roughly 40% relative to no-prune. Feature growth allows confused heads to add quadratic features, and tiny MLP heads on nonlinear worlds achieve 95–99% per-world accuracy when capacity matches the task. Uncertaintydriven querying then attains essentially the same accuracy as an always-query baseline while using only about 18–25% of the labels. In Part V, we test a harder question: can the same feelings also solve the routing problem<br>when multiple worlds share the same input distribution but differ in their labeling rules? We compare a monolithic model, an oracle multi-head world bank, and several routers that see only the current input and per-head feelings (probabilities and value channels), including a teachertrained router with perfect experts. In this regime, all such routers remain far below the oracle multi-head performance and behave similarly to the monolithic baseline. Single-step feelings are not sufficient statistics for the latent world identity; this strongly motivates introducing a slower, persistent context code zt as the next step in EGWM. Taken together, these five parts describe a small but concrete learning core: a capacitymatched world bank whose growth, compression, forgetting, and information-seeking are governed by a handful of multi-time-scale value signals. Feelings are most powerful when they govern structure—when to create, reuse, merge, and retire worlds—rather than being just scalar learning rates. The toy setting is far from AGI, but it illustrates how a system can begin to “feel” when its internal theory of the world is too crude, too fragmented, or beautifully simple.</p>