AM
A. Marcu
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
Multi-agent systems increasingly rely on communication and orchestration to coordinate, yet their interactions are often transient rather than accumulated into persistent relationships. We ask whether persistent relational structure can instead emerge from interaction itself, and formalise a neuron–agent isomorphism that models a MAS as a social graph evolving through reward-modulated Hebbian plasticity. The graph is coupled back into agent behaviour through reward diffusion and weight-gated experience sharing for RL-fine-tuned agents, and through an inference-time social module for VLM-based agents. We evaluate the framework in WIRE, a five-stage embodied Craftium environment with increasing coordination demands. We characterise when social plasticity supports coordination, showing that its effect depends on the underlying agent and the choice of co-firing signal. On Gemma-E4B, Hebbian coupling increases mean task return from 521 to 570 and cooperative milestone completion from 8.5% to 12.1%, with longer-timescale reward modulation further increasing cooperative completion to 14.4%. Decentralised plasticity-based coordination is also competitive with centralised orchestration at matched compute. Different social cues induce distinct relational structures, with observation yielding the highest cooperative completion (16.3%) despite communication producing substantially stronger bonds. Finally, when initial bond strength is controlled, agents preferentially collaborate with teammates with whom they had previously co-fired (0.56 vs. 0.40 under random re-pairing) and retain stronger bonds, showing that agents that co-fire together wire together even after team recomposition. The code can be found at: https://github.com/tapri-lab/wired-together.
...
Multi-agent systems increasingly rely on communication and orchestration to coordinate, yet their interactions are often transient rather than accumulated into persistent relationships. We ask whether persistent relational structure can instead emerge from interaction itself, and formalise a neuron–agent isomorphism that models a MAS as a social graph evolving through reward-modulated Hebbian plasticity. The graph is coupled back into agent behaviour through reward diffusion and weight-gated experience sharing for RL-fine-tuned agents, and through an inference-time social module for VLM-based agents. We evaluate the framework in WIRE, a five-stage embodied Craftium environment with increasing coordination demands. We characterise when social plasticity supports coordination, showing that its effect depends on the underlying agent and the choice of co-firing signal. On Gemma-E4B, Hebbian coupling increases mean task return from 521 to 570 and cooperative milestone completion from 8.5% to 12.1%, with longer-timescale reward modulation further increasing cooperative completion to 14.4%. Decentralised plasticity-based coordination is also competitive with centralised orchestration at matched compute. Different social cues induce distinct relational structures, with observation yielding the highest cooperative completion (16.3%) despite communication producing substantially stronger bonds. Finally, when initial bond strength is controlled, agents preferentially collaborate with teammates with whom they had previously co-fired (0.56 vs. 0.40 under random re-pairing) and retain stronger bonds, showing that agents that co-fire together wire together even after team recomposition. The code can be found at: https://github.com/tapri-lab/wired-together.
Bachelor thesis
(2024)
-
Ana Cristiana Marcu, L. Cavalcante Siebert, A. Homayounirad, E. Liscio, J. Yang
Public deliberations play a crucial role in democratic systems. However, the unstructured nature of deliberations leads to challenges for moderators to analyze the large volume of data produced. This paper aims to solve this challenge by automatically identifying subjective topics behind public discourse by leveraging Large Language Models (LLMs). The study is structured around two core objectives: Identifying Gold Labels and Exploring Subjective Human Labels. The results highlight that fine-tuning the LLaMa-2 model with QLoRa outperforms other methods for Identifying Gold Labels, while the Few-Shot Chain of Thoughts method, enhanced with EmotionPrompt, is particularly effective in capturing subjective variations in human annotations. However, the study also underscores significant limitations, such as the dependency on large, high-quality annotated datasets and the tendency of models to produce hallucinations. These findings highlight the potential of LLMs to identify subjective topics behind public discourse, while also emphasizing the need for further research to address these challenges.
...
Public deliberations play a crucial role in democratic systems. However, the unstructured nature of deliberations leads to challenges for moderators to analyze the large volume of data produced. This paper aims to solve this challenge by automatically identifying subjective topics behind public discourse by leveraging Large Language Models (LLMs). The study is structured around two core objectives: Identifying Gold Labels and Exploring Subjective Human Labels. The results highlight that fine-tuning the LLaMa-2 model with QLoRa outperforms other methods for Identifying Gold Labels, while the Few-Shot Chain of Thoughts method, enhanced with EmotionPrompt, is particularly effective in capturing subjective variations in human annotations. However, the study also underscores significant limitations, such as the dependency on large, high-quality annotated datasets and the tendency of models to produce hallucinations. These findings highlight the potential of LLMs to identify subjective topics behind public discourse, while also emphasizing the need for further research to address these challenges.