– Harvard & MIT
Cognitive Architectures Paper: https://arxiv.org/abs/2309.02427
– Princeton University
Conversational Swarm Intelligence: https://arxiv.org/abs/2309.03220
– Carnegie Mellon
Underappreciated AI news you missed.
CSI: Chatbots in Chatrooms
A new method for enabling large groups to hold networked conversations and converge on solutions.
Method: Breaks large groups into small chatrooms with AI agents sharing insights between rooms.
Pilot Study: Compared contribution in standard chatroom vs conversational swarm.
Results: 30% more contributions and 7% less variance in swarms.
Conclusion: Conversational swarms increased participation and evenness of contribution.
Future Work: Test with larger groups to further validate benefits.
Similarity: Working groups, sociocracy
ToM: Artificial Neurons
Evaluated if hidden embeddings in LLMs exhibit neural correlates of theory of mind (ToM) similar to single neurons recorded from the human brain. Tested multiple LLMs on false belief tasks using human task materials. Found that some embeddings responded selectively to true vs false belief trials, reminiscent of human prefrontal cortex neurons. Could accurately decode true vs false beliefs from the population of embeddings. The proportion of selective embeddings correlated with models’ task accuracy. These results suggest the embeddings facilitate the models’ ToM capabilities. Overall, reveals a striking parallel between the artificial and biological neural correlates supporting theory of mind.
Methods: Tested LLMs on human false belief tasks; analyzed model embeddings.
Selective Embeddings: Some responded significantly to true/false belief trials.
Decoding Beliefs: Could accurately predict beliefs from all embeddings.
Correlates with Performance: More selective embeddings in higher performing models.
Parallel to Human Neurons: Embeddings exhibit neural correlates of ToM like human brain.
CoALA
The paper proposes a conceptual framework called CoALA to systematically design and understand language agents powered by large language models (LLMs). CoALA draws inspiration from decades of research on cognitive architectures that organize perception, memory, reasoning, learning, and decision-making to achieve human-like intelligence.
Similarity: Structurally similar to SOAR
Missing: Never mentions morality or mission
Limited: Does not add much to the field
Alternative: Look for my ACE framework coming out soon
