Due to its powerful representational capabilities, Transformers have gradually become the mainstream model in the field of machine vision.

Research on artificial consciousness lacks the equivalent of the perceptron: a small, trainable module that can be copied, benchmarked, and iteratively improved. We introduce the Reflexive Integrated Information Unit (RIIU), a recurrent cell that augments its hidden state $h$ with two additional vectors: (i) a meta-state $μ$ that records the cell's own causal footprint, and (ii) a broadcast buffer $B$ that exposes that footprint to the rest of the network. A sliding-window covariance and a differentiable Auto-$Φ$ surrogate let each RIIU maximize local information integration online. We prove that RIIUs (1) are end-to-end differentiable, (2) compose additively, and (3) perform $Φ$-monotone plasticity under gradient ascent. In an eight-way Grid-world, a four-layer RIIU agent restores $>90\%$ reward within 13 steps after actuator failure, twice as fast as a parameter-matched GRU, while maintaining a non-zero Auto-$Φ$ signal. By shrinking "consciousness-like" computation down to unit scale, RIIUs turn a philosophical debate into an empirical mathematical problem.

The increasing polarization in democratic societies is an emergent outcome of political opinion dynamics. Yet, the fundamental mechanisms behind the formation of political opinions, from individual beliefs to collective consensus, remain unknown. Understanding that a causal mechanism must account for both bottom-up and top-down influences, we conceptualize political opinion dynamics as hierarchical coarse-graining, where microscale opinions integrate into a macro-scale state variable. Using the CODA (Continuous Opinions Discrete Actions) model, we simulate Bayesian opinion updating, social identity-based information integration, and migration between social identity groups to represent higher-level connectivity. This results in coarse-graining across micro, meso, and macro levels. Our findings show that higher-level connectivity shapes information integration, yielding three regimes: independent (disconnected, local convergence), parallel (fast, global convergence), and iterative (slow, stepwise convergence). In the iterative regime, low connectivity fosters transient diversity, indicating an informed consensus. In all regimes, time-scale separation leads to downward causation, where agents converge on the aggregate majority choice, driving consensus. Critically, any degree of coherent higher-level information integration can overcome misalignment via global downward causation. The results highlight how emergent properties of the causal mechanism, such as downward causation, are essential for consensus and may inform more precise investigations into polarized political discourse.

This paper explores the hypothesis that the OpenAI-o1 model--a transformer-based AI trained with reinforcement learning from human feedback (RLHF)--displays characteristics of consciousness during its training and inference phases. Adopting functionalism, which argues that mental states are defined by their functional roles, we assess the possibility of AI consciousness. Drawing on theories from neuroscience, philosophy of mind, and AI research, we justify the use of functionalism and examine the model's architecture using frameworks like Integrated Information Theory (IIT) and active inference. The paper also investigates how RLHF influences the model's internal reasoning processes, potentially giving rise to consciousness-like experiences. We compare AI and human consciousness, addressing counterarguments such as the absence of a biological basis and subjective qualia. Our findings suggest that the OpenAI-o1 model shows aspects of consciousness, while acknowledging the ongoing debates surrounding AI sentience.

Understanding how humans process natural language has long been a vital research direction. The field of natural language processing (NLP) has recently experienced a surge in the development of powerful language models. These models have proven to be invaluable tools for studying another complex system known to process human language: the brain. Previous studies have demonstrated that the features of language models can be mapped to fMRI brain activity. This raises the question: is there a commonality between information processing in language models and the human brain? To estimate information flow patterns in a language model, we examined the causal relationships between different layers. Drawing inspiration from the workspace framework for consciousness, we hypothesized that features integrating more information would more accurately predict higher hierarchical brain activity. To validate this hypothesis, we classified language model features into two categories based on causal network measures: 'low in-degree' and 'high in-degree'. We subsequently compared the brain prediction accuracy maps for these two groups. Our results reveal that the difference in prediction accuracy follows a hierarchical pattern, consistent with the cortical hierarchy map revealed by activity time constants. This finding suggests a parallel between how language models and the human brain process linguistic information.

We introduce the notion of Morton-style factorial coding and illustrate how it may help understand information integration and perceptual cod- ing in the brain. We show that by focusing on average responses one may miss the existence of factorial coding mechanisms that become only apparent when analyzing spike count histograms. We show evidence suggesting that the classical/non-classical receptive field organization in the cortex effectively enforces the development of Morton-style factorial codes. This may provide some cues to help understand perceptual cod- ing in the brain and to develop new unsupervised learning algorithms. While methods like ICA (Bell & Sejnowski, 1997) develop independent codes, in Morton-style coding the goal is to make two or more external aspects of the world become independent when conditioning on internal representations.

For an empirical science, physics can be remarkably dismissive of some of our most basic observations. We see objects existing in definite locations, but the wave nature of matter washes that away. We perceive time to flow, but how could it, really? We feel ourselves to be free agents, and that's just quaint. Physicists like nothing better than to expose our view of the universe as parochial. But when asked why our impressions are so off, they mumble some excuse and slip out the side door of the party. Physicists, in other words, face the same hard problem of consciousness as neuroscientists do: the problem of bridging objective description and subjective experience. To relate fundamental theory to what we actually observe in the world, they must explain what it means "to observe"--to become conscious of. And they tend to be slapdash about it. They divide the world into "system" and "observer," study the former intensely, and take the latter for granted--or, worse, for a fool.

This work considers the problem of learning cooperative policies in multi-agent settings with partially observable and non-stationary environments without a communication channel. We focus on improving information sharing between agents and propose a new multi-agent actor-critic method called \textit{Multi-Agent Cooperative Recurrent Proximal Policy Optimization} (MACRPO). We propose two novel ways of integrating information across agents and time in MACRPO: First, we use a recurrent layer in critic's network architecture and propose a new framework to use a meta-trajectory to train the recurrent layer. This allows the network to learn the cooperation and dynamics of interactions between agents, and also handle partial observability. Second, we propose a new advantage function that incorporates other agents' rewards and value functions. We evaluate our algorithm on three challenging multi-agent environments with continuous and discrete action spaces, Deepdrive-Zero, Multi-Walker, and Particle environment. We compare the results with several ablations and state-of-the-art multi-agent algorithms such as QMIX and MADDPG and also single-agent methods with shared parameters between agents such as IMPALA and APEX. The results show superior performance against other algorithms. The code is available online at https://github.com/kargarisaac/macrpo.

We know that they have the same sensors – called nociceptors – that cause us to flinch or cry when we are hurt. And they certainly behave like they are sensing something unpleasant. When a chef places them in boiling water, for instance, they twitch their tails as if they are in agony. But are they actually "aware" of the sensation? Or is that response merely a reflex?