I’ve recently uploaded three interrelated preprints to TechRxiv, each exploring new paradigms for how LLMs can develop memory, alignment, and behavioral intuition through cognitively inspired mechanisms.
Below is a brief overview of each paper, with key ideas and open questions for feedback.
1. ForecastLLMs: Overconfidence-Inspired Foresight in Language Models
Link to Paper →
Concept:
Inspired by the Dunning-Kruger effect, this paper proposes that LLMs may benefit from a structured form of overconfidence — not in hallucinating facts, but in proactively forecasting what the user might ask next, or what knowledge gaps may soon emerge.
Rather than waiting for instructions, the model speculatively offers follow-ups or anticipatory scaffolding. The argument is that this could:
- Reduce user friction
- Improve token efficiency
- Open a pathway to agentic, curiosity-driven models
Prompt for Discussion:
Could deliberate overconfidence in LLMs act as a regularizer or alignment aid — or is it a slippery slope toward confabulation?
2. Dream-Augmented Language Models (DLLMs): Personalization via Off-Session Memory Compression
Concept:
DLLMs are based on the idea that LLMs, like humans, could benefit from “dreaming” — background sessions that compress recent interactions into long-term personalized memory.
Instead of repeatedly prompting an LLM with your history, it remembers you between sessions using scheduled, low-resource memory updates. Inspired by cognitive consolidation during sleep.
Potential Benefits:
- Token and compute savings
- Enhanced personalization without vector bloat
- Energy-efficient long-term user modeling
Prompt for Discussion:
Could scheduled “dreams” replace or augment fine-tuning for personal assistants or longitudinal reasoning agents?
3. LLM-Wide Dream: Ambient and Societal Memory Formation in LLMs
Link to Paper →
Concept:
This generalizes DLLMs to the collective level. LLM-Wide Dream is a framework for passive, ambient learning — where LLMs gather anonymized behavioral summaries across users and regions during idle times.
The result is a societal memory graph, which enables insights like:
- Regional gaps in vaccine awareness
- Demographic trends in educational confusion
- Emergent public sentiment shifts
It’s partly inspired by Jung’s “collective unconscious” and modern federated learning.
Prompt for Discussion:
What are the risks and ethical implications of collective cognition in LLMs? Could ambient learning be essential for alignment at scale?