CTM-AI: Evolving Multi-Agent Intelligence Inspired by a Model of Consciousness

arXiv:2605.04097 · q-bio.NC, cs.AI · Submitted 2026-04-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "CTM-AI: Evolving Multi-Agent Intelligence Inspired by a Model of Consciousness".

Marcus: Here's the gist of the paper. CTM-AI presents an early blueprint for implementing a general AI system by combining the Conscious Turing Machine with foundation models to bridge the gap between consciousness theory and practical AI architecture This work demonstrates that CTM-AI achieves state-of-the-art accuracy on MUStARD (72.28) and UR-FUNNY (72.13), outperforming multimodal and multi-agent frameworks Pruning redundant reasoning tokens The paper introduces CTM-AI, which combines the Conscious Turing Machine with foundation models to create a general AI system inspired by consciousness theory The CTM is defined as a simple and formal model of consciousness inspired by Alan Turing’s model of computation and Bernard Baars’ theater model of consciousness CTM-AI operationalizes this abstract architecture using modern AI technologies, defining multiple specialized processors, a limited capacity short-term memory workspace enforcing selective attention via up-tree competition, a global broadcast of information via a down-tree from the workspace to all processors, and the formation of links between relevant processors over time Key contributions include serving as the first practical instantiation of the CTM, yielding a highly modular and decentralized multi-agent architecture free from rigid central orchestrators, and integrating reasoning and agentic flexibility How the training signal is built The core inference process of CTM-AI involves an iterative loop defined by four stages: LTM processor chunk inference, Up-tree competition into STM, Down-tree broadcast, and Link formation and fusion Step 1 involves all K LTM processors running in parallel to produce chunks with a content gist h it summarizing its findings, a set of follow-up queries q it for potential cross-processor consultation, and a self-assessed score s it The chunk score s i t is decomposed into three interpretable sub-scores: weight, intensity, and mood Step 2 uses an up-tree competition to select the final chunk that goes into STM's limited-capacity workspace based on the chunk scores Step 3 involves a down-tree broadcast where the winning chunk is distributed to all LTM processors causing them to pay conscious attention Step 4 enables unconscious communication where LTM processors form links and exchange information, leading to link fusion which enriches long-term memories Where the gains come from Quantitative results show CTM-AI achieves state-of-the-art or competitive results across multimodal perception, tool use, and agentic tasks On multimodal perception tasks like MUStARD and UR-FUNNY, CTM-AI matches and slightly outperforms the strongest baselines on MUStARD and UR-FUNNY For tool use on StableToolBench, CTM-AI yields the most substantial gains, improving over the strongest single-model baseline by up to 16.8 points on multi-tool scenarios Furthermore, CTM-AI generalizes to tool-use and agentic tasks yielding an absolute improvement of >10% in pass rate on both StableToolBench and WebArenaLite The paper also demonstrates that the iterative loop is the most critical mechanism for performance, with the iterative loop emerging as the most critical (-6.7 F1) mechanism Ablation on CTM-AI dynamics Ablation studies confirm a clear hierarchy of importance among CTM-AI mechanisms The iterative loop emerges as the most critical (-6.7 F1), followed by up-tree competition (-5.6), link fusion (-3.9), and down-tree broadcast (-3.5) Scaling the iterations from 1 to 4 yields continuous performance gains across both datasets, with MUStARD improving from 65.0 to 73.8 The analysis also shows that link formation behaviors differ between datasets, as UR-FUNNY forms an average of 2.6 links at four iterations compared to just 1.5 for MUStARD RQ2: As an inference-time scaling method, how does CTM-AI compare with the reasoning paradigm?

Ines: CTM-AI and reasoning are complementary paradigms where applying reasoning models improves baseline performance but falls short of CTM-AI when the backbone model is fixed Pairing Qwen3-8B-think with CTM-AI yields an additional +16.8 to +27.3 point improvement on StableToolBench compared to using Qwen3-8B-think alone, confirming that reasoning deepens deliberation within individual processors while CTM-AI broadens coordination across them This decentralized architecture supports the flexible integration of reasoning, allowing practitioners to selectively inject advanced reasoning capabilities into specific modalities A.

Marcus: The Use of Large Language Models (LLMs) The implementation uses LLMs as a writing assistant and utilizes CodePilot and Claude Code to help code faster, with all AI-generated content manually checked The backbone model selected for most processors is Gemini-2.5-flash-lite because it supports audio, vision, and text inputs while being relatively small-scale B.

Ines: Artifact Details The paper utilizes several models including GPT-4o (OpenAI, 2024), OpenAI o3 (OpenAI, 2025), Gemini-2.5-flash-lite (Comanici et al., 2025), and Qwen3 variants like Qwen3-VL-8B and Qwen3-Omni-flash Evaluation datasets include MUStARD, URFunny, WebArena-Lite, and StableToolBench C.

Title and authors: Ines: So we're looking at this paper today called "CTM-AI: Evolving Multi-Agent Intelligence Inspired by a Model of Consciousness". It’s trying to build a practical blueprint for a general AI system by mixing the formal model of consciousness, the Conscious Turing Machine, with foundation models.

Marcus: Yeah, that's what it does. Basically, they take this abstract idea from consciousness theory and try to make it work with today's AI tools. It’s trying to bridge that gap between the theory and what we can actually build right now.

Yuki: From a population genetics perspective, I think the focus on a general system is interesting because it suggests we aren't just optimizing narrow tasks anymore, but something more adaptive. It asks if this architecture could support broader forms of intelligence across different domains.

Ines: Exactly. The core idea is they operationalize this consciousness model using multiple specialized processors, short-term memory workspaces with selective attention through up-tree competition, and a global broadcast mechanism from the workspace to all processors.

Marcus: That decentralized structure is what caught my eye because it’s trying to move away from those rigid central orchestrators we see in current agentic systems. It aims for a highly modular architecture where you can just add or remove specialized processors without breaking the whole thing.

Yuki: I wonder if that decentralization actually translates into robust coordination when you're dealing with complex, real-world problems that require deep, long-term understanding of the world.

Ines: The mechanism they describe for how information flows is really about this iterative loop: chunk inference, up-tree competition into short term memory workspace, down-tree broadcast, and then link formation between processors over time.

Marcus: That link formation part is key because it enables unconscious communication between processors to integrate their knowledge into higher-order multimodal information. It’s not just about passing data along a line; it's about them learning from each other in a more integrated way.

Yuki: So, when you look at the results they show, they claim CTM-AI achieves state of the art accuracy on MUStARD and URFUNNY datasets. I want to know what that means for genuine perception capabilities.

Ines: It means they are matching or slightly beating the strongest baselines on those multimodal perception tasks, achieving scores like seventy-two point something on MUStARD and seventy-two point something on URFUNNY.

Title and authors: Marcus: And it's not just about the raw accuracy; they also showed significant gains in tool use. On StableToolBench, CTM-AI improved over the strongest single model baseline by up to sixteen point something in multi-tool scenarios.

Yuki: That suggests that this architecture is actually helping the system handle complex, multi-step tasks with tools much better than what we see now when we just use a single large model.

Ines: And they also showed generalization to tool use and agentic tasks, yielding an absolute improvement of over ten point something in pass rate on both StableToolBench and WebArenaLite.

Marcus: The iterative loop itself is highlighted as the most critical mechanism for performance, with it emerging as the most critical factor at negative six point seven F1. That tells us that how often this loop runs really dictates how good the system gets.

Yuki: If we look at those ablation studies on iteration count, scaling from one to four iterations gives continuous performance gains across both datasets; MUStARD goes up from sixty five point zero to seventy three point eight.

Ines: And they found that the link formation behavior actually differs between the datasets; for example, URFUNNY forms an average of two point six links at four iterations, compared to just one point five for MUStARD.

Marcus: That difference suggests that the nature of the interaction between processors changes depending on what kind of task you're asking them to solve.

Yuki: So, before we move on, how does this architecture actually compare to using a reasoning model alongside a fixed backbone model? I always wonder if this is just another way to layer reasoning on top.

Ines: Well, they show that pairing Qwen3 eight point eight with CTM-AI yields an additional improvement of sixteen point eight to twenty seven point three on StableToolBench compared to using just Qwen3 eight point eight alone.

Marcus: That points toward the idea that CTM-AI and reasoning models operate along different axes for coordination, which is a big detail. It’s not just one or the other; it's how they work together across those dimensions.

Yuki: I see how that fits with the larger picture of complex systems where you need both deep individual deliberation and broad coordination to get a good result.

Title and authors: Ines: The paper also mentioned using large language models as a writing assistant, but the backbone processor is Gemini two point five flash lite because it handles audio, vision, and text inputs while still being relatively small-scale.

Marcus: And when you look at their efficiency analysis, they found that most tasks resolve within one to three iterations. The performance boost from adding more iterations is directly proportional to an increase in the average number of interprocessor links formed.

Yuki: That suggests that the system dynamically allocates more processing power only when it actually needs it to improve communication density, which sounds quite efficient.

Ines: They also found that the optimal balance for link formation threshold, called eta, is somewhere between zero point four and zero point seven for both datasets.

Marcus: So, a few final thoughts before we wrap up on this CTM-AI blueprint: the architecture is modular because of those competition mechanisms. The results show it performs well across perception, tool use, and agentic tasks.

Yuki: I think the self-reported scores are also important because they showed that the distribution stabilizes and converges by iteration three, meaning that mechanism continues to be a driver for its internal dynamics.

Ines: So we've seen how this formal model of consciousness translates into a decentralized system free from rigid central orchestrators, and it achieves high accuracy on perception tasks like MUStARD and URFUNNY.

Marcus: And the tool use improvements on StableToolBench really show that this dynamic prioritization of processors leads to substantial gains when dealing with multi-tool scenarios.

Yuki: Overall, CTM-AI seems to be a blueprint for how we might build AI that handles complex tasks through flexible coordination rather than just brute force scaling.

Ines: That's the main point, and it points toward building systems that are fundamentally more capable in affective learning and decision making.

Marcus: We've covered the architecture, the performance gains on perception and tools, how iteration count affects performance, and how reasoning models can enhance it. This paper on CTM-AI really shows a path forward for architecturally inspired AI.

Yuki: It’s definitely something to keep watching as we think about building more adaptive intelligence in these systems.

Ines: We'll leave you there with the CTM-AI blueprint, an early attempt to translate consciousness theory into a practical, working system for general AI.

The paper's summary: Ines: So we're looking at how they’re trying to build this whole system by taking that formal model of consciousness, the Conscious Turing Machine, and mixing it with foundation models to create a general AI architecture.

Marcus: Right. It’s basically trying to bridge that gap between the abstract theory of consciousness and what we can actually build in practice right now.

Yuki: From my perspective as someone who looks at population genetics, I see the focus on a general system suggesting we aren't just optimizing for one narrow task anymore, but something that could be more adaptable across different kinds of intelligence.

Ines: Exactly. They’re operationalizing this by using multiple specialized processors, a limited short-term memory workspace that uses up-tree competition for selective attention, and a global broadcast from the workspace to all processors.

Marcus: That decentralized structure is what interests me because it’s moving away from those rigid central controllers we see in current agent systems. It aims for something modular where you can just add or remove specialized processors without breaking the whole thing.

Yuki: I wonder if that decentralization actually translates into robust coordination when you're dealing with really complex, real-world problems that need deep, long-term understanding of the world.

Ines: The core process they describe is this iterative loop: chunk inference, up-tree competition into short term memory workspace, down-tree broadcast, and then link formation between processors over time.

Marcus: That link formation part is key because it allows unconscious communication between processors to integrate their knowledge into a higher level of multimodal understanding. It’s not just about passing data along a line; it’s about them learning from each other in a more integrated way.

Yuki: So when you look at the results they show, they claim this CTM-AI achieves top accuracy on perception tasks like MUStARD and URFUNNY datasets. I want to know what that means for genuine perception capabilities.

Ines: It means they are matching or slightly beating the strongest baselines on those multimodal perception tasks, hitting scores like seventy-two point something on MUStARD and seventy-two point something on URFUNNY.

The paper's summary: Marcus: And it's not just about raw accuracy; they also showed huge gains in tool use. On StableToolBench, CTM-AI improved over the strongest single model baseline by up to sixteen point something in multi-tool scenarios.

Yuki: That suggests this architecture is helping the system handle complex, multi-step tasks with tools much better than what we see now when we just use one big model.

Ines: They also showed generalization to tool use and agentic tasks, yielding an absolute improvement of over ten point something in pass rate on both StableToolBench and WebArenaLite.

Marcus: The iterative loop itself is highlighted as the most critical mechanism for performance, emerging as the most important factor at negative six point seven F1. That tells us how often this loop runs really dictates how good the system gets.

Yuki: If we look at those studies on iteration count, scaling from one to four iterations gives continuous performance gains across both datasets; MUStARD goes up from sixty five point zero to seventy three point eight.

Ines: And they found that the link formation behavior actually changes between the datasets; for example, URFUNNY forms an average of two point six links at four iterations, compared to just one point five for MUStARD.

Marcus: That difference suggests the way processors interact changes depending on what kind of task you're asking them to solve.

Yuki: So before we move on, how does this architecture compare to using a reasoning model alongside a fixed backbone model? I always wonder if this is just another way to layer reasoning on top.

Ines: They show that pairing Qwen three point eight with CTM-AI yields an additional improvement of sixteen point eight to twenty seven point three on StableToolBench compared to just Qwen three point eight alone.

Marcus: That points toward the idea that CTM-AI and reasoning models operate along different axes for coordination, which is a big detail. It’s not just one or the other; it's how they work together across those dimensions.

The paper's summary: Yuki: I see how that fits with the larger picture of complex systems where you need both deep individual deliberation and broad coordination to get a good result.

Ines: The paper also mentioned using large language models as a writing assistant, but the main backbone processor is Gemini two point five flash lite because it handles audio, vision, and text inputs while still being relatively small-scale.

Marcus: And when you look at their efficiency analysis, they found most tasks resolve within one to three iterations. The performance boost from adding more iterations is directly proportional to an increase in the average number of interprocessor links formed.

Yuki: That suggests the system dynamically allocates more processing power only when it actually needs it to improve communication density, which sounds pretty efficient.

Ines: They also found that the optimal balance for link formation threshold, called eta, is somewhere between zero point four and zero point seven for both datasets.

Marcus: So before we wrap up on this CTM-AI blueprint: the architecture is modular because of those competition mechanisms. The results show it performs well across perception, tool use, and agentic tasks.

Yuki: I think the self-reported scores are also important because they showed that the distribution stabilizes and converges by iteration three, meaning that mechanism continues to be a driver for its internal dynamics.

Ines: So we've seen how this formal model of consciousness translates into a decentralized system free from rigid central orchestrators, and it achieves high accuracy on perception tasks like MUStARD and URFUNNY.

Marcus: And the tool use improvements on StableToolBench really show that that dynamic prioritization of processors leads to substantial gains when dealing with multi-tool scenarios.

Yuki: Overall, CTM-AI seems to be a blueprint for how we might build AI that handles complex tasks through flexible coordination rather than just brute force scaling.

Ines: That's the main point, and it points toward building systems that are fundamentally more capable in affective learning and decision making.

The paper's improvements: Ines: We're talking about how they suggest making CTM-AI better by focusing on these specific mechanisms instead of just running the system once and moving on.

Marcus: Right, so instead of just looking at the final score, they point to things like scaling up those iterative loops because that loop is actually the most critical part of the whole process at negative six point seven F1.

Yuki: Scaling those iterations means if you increase them from one to four, MUStARD performance goes up from sixty five point zero to seventy three point eight. That shows that more iterations help improve the system’s grasp of the task.

Ines: And it's not just about getting a higher score; they show that increasing those iterations directly leads to a bigger average number of interprocessor links formed, which is what actually drives the performance boost.

Marcus: That means we don't just want more computation; we want that extra computation to create better communication between the processors.

Yuki: So if you look at link formation, they found that for URFUNNY, it forms an average of two point six links at four iterations compared to only one point five when you run it once. That shows how the interaction changes depending on the task.

Ines: That’s interesting because it suggests the system learns a different way to coordinate when dealing with humor versus perception tasks.

Marcus: And they also looked at those self-reported scores and found that the distribution stabilizes and converges by iteration three, meaning that self-reporting keeps playing a crucial role in driving what's happening inside the AI.

Yuki: So, for someone listening who doesn't know all this detail, it just means CTM-AI isn't just a static system; it’s something that actively refines its own internal structure over time.

Ines: That’s right. It’s adapting and getting denser in its communication graph without needing a manual reset or parameter update.

Marcus: It sounds like the goal is to build this kind of intelligence that improves by simply interacting with the world through these iterative refinement steps rather than just being pre-programmed.

Yuki: That takes us nicely into how they compared this to using reasoning models on top of a fixed backbone model, because that’s another way to think about enhancing the system's capability.

Conclusion: Ines: So we've covered how this CTM-AI system uses that formal model of consciousness to build a decentralized architecture using up-tree competition and down-tree broadcast, and I want to wrap up by looking at what all this means for the next generation of AI.

Marcus: Yeah, it’s essentially showing a path toward building general AI systems that are structured around cognitive principles rather than just brute force scaling.

Yuki: From a population genetics standpoint, this suggests that we need architectures that support diverse coordination across different modalities and tasks, which is essential for understanding complex species behavior.

Ines: Exactly. The work on CTM-AI: Evolving Multi-Agent Intelligence Inspired by a Model of Consciousness shows how you can translate consciousness theory into something practical for general AI.

Marcus: And the results on MUStARD and URFUNNY, plus the gains in tool use, show that this approach is actually getting state of the art performance without needing a single massive model to do everything.

Yuki: It changes how we think about what an agent needs to be capable of; it’s not just about bigger data sets anymore, but how that data is processed and coordinated.

Ines: The way they handle those iterative loops, where the system refines its internal structure through link formation and memory fusion, shows a mechanism for continuous self-improvement.

Marcus: It means we might see AI systems that adapt over time without needing explicit parameter updates from a human engineer every single month.

Yuki: So what it changes for someone just listening to the show is that the focus shifts from just training massive models to designing these dynamic, coordinated architectures.

Ines: That's right. It points toward building systems that are fundamentally more capable in areas like affective learning and decision-making because they have this built-in mechanism for deeper deliberation.

Marcus: It’s a solid blueprint for how we can start thinking about building more flexible, multi-agent systems that actually coordinate effectively across different specialized components.

Yuki: I'm curious to see if this kind of dynamic coordination holds up when we apply it to even more complex biological or social systems in the future.

Haofei Yu, Yining Zhao, Lenore Blum, Manuel Blum, Paul Pu Liang

University of Illinois Urbana–Champaign · Carnegie Mellon University · Massachusetts Institute of Technology

q-bio.NC, cs.AI

Submitted: 2026-04-30

Updated: 2026-11-01

Comments: 10 pages

Code: https://github.com/ServiceNow/BrowserGym

Project page: https://consciousness-lab.github.io

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 99/100

The gist: CTM-AI presents an early blueprint for implementing a general AI system by combining the Conscious Turing Machine with foundation models to bridge the gap between consciousness theory and practical…

Key concepts

Conscious Turing Machine (CTM)
A formal model of consciousness inspired by Alan Turing's computation theory. It provides the abstract blueprint for how a general AI system should be structured, defining components like processors and memory workspaces to mimic conscious processing.
Up-tree Competition
A mechanism used in CTM-AI to select which information chunk enters the limited short-term memory workspace. This competition ensures that only the most relevant or important information is prioritized for processing, enforcing selective attention across the system.
Iterative Loop
The core inference process involving four stages: chunk inference, up-tree competition, down-tree broadcast, and link formation. This loop allows the system to continuously refine its understanding by exchanging information between processors over multiple cycles.
Link Fusion
The process where relevant processors form links and exchange information during unconscious communication. This exchange enriches long-term memories by integrating diverse pieces of data into a more comprehensive and nuanced representation.

Terminology

Summary

CTM-AI presents an early blueprint for implementing a general AI system by combining the Conscious Turing Machine with foundation models to bridge the gap between consciousness theory and practical AI architecture This work demonstrates that CTM-AI achieves state-of-the-art accuracy on MUStARD (72. p. 0

Where the paper says it

This gap has fueled longstanding debates about whether AI might one day achieve human-like generality or even consciousness, and whether theories of consciousness can inspire new architectures for AI. This paper presents an early blueprint for implementing a general AI system, CTM-AI, combining the Conscious Turing Machine (CTM), a formal machine model of consciousness, with today’s foundation models.

28) and UR-FUNNY (72. p. 0
Where the paper says it

CTM-AI achieves state-of-the-art accuracy on MUStARD (72. 28) and UR-FUNNY (72.

13), outperforming multimodal and multi-agent frameworks

Pruning redundant reasoning tokens

The paper introduces CTM-AI, which combines the Conscious Turing Machine with foundation models to create a general AI system inspired by consciousness theory The CTM is defined as a simple and formal model of consciousness inspired by Alan Turing’s model of computation and Bernard Baars’ theater model of consciousness CTM-AI operationalizes this abstract architecture using modern AI technologies, defining multiple specialized processors, a limited capacity short-term memory workspace enforcing selective attention via up-tree competition, a global broadcast of information via a down-tree from the workspace to all processors, and the formation of links between relevant processors over time Key contributions include serving as the first practical instantiation of the CTM, yielding a highly modular and decentralized multi-agent architecture free from rigid central orchestrators, and integrating reasoning and agentic flexibility

How the training signal is built

The core inference process of CTM-AI involves an iterative loop defined by four stages: LTM processor chunk inference, Up-tree competition into STM, Down-tree broadcast, and Link formation and fusion Step 1 involves all K LTM processors running in parallel to produce chunks with a content gist h it summarizing its findings, a set of follow-up queries q it for potential cross-processor consultation, and a self-assessed score s it The chunk score s i t is decomposed into three interpretable sub-scores: weight, intensity, and mood Step 2 uses an up-tree competition to select the final chunk that goes into STM's limited-capacity workspace based on the chunk scores Step 3 involves a down-tree broadcast where the winning chunk is distributed to all LTM processors causing them to pay conscious attention Step 4 enables unconscious communication where LTM processors form links and exchange information, leading to link fusion which enriches long-term memories

Where the gains come from

Quantitative results show CTM-AI achieves state-of-the-art or competitive results across multimodal perception, tool use, and agentic tasks On multimodal perception tasks like MUStARD and UR-FUNNY, CTM-AI matches and slightly outperforms the strongest baselines on MUStARD and UR-FUNNY For tool use on StableToolBench, CTM-AI yields the most substantial gains, improving over the strongest single-model baseline by up to 16.8 points on multi-tool scenarios Furthermore, CTM-AI generalizes to tool-use and agentic tasks yielding an absolute improvement of >10% in pass rate on both StableToolBench and WebArenaLite The paper also demonstrates that the iterative loop is the most critical mechanism for performance, with the iterative loop emerging as the most critical (-6. p. 1

Where the paper says it

Furthermore, CTM-AI generalizes to tool-use and agentic tasks, yielding an absolute improvement of >10% in pass rate on both StableToolBench and WebArenaLite. Finally, our analysis demonstrates that: (1) CTM-AI organically adapts its inter-processor connectivity based on task complexity; (2) it integrates seamlessly with existing reasoning paradigms; and (3) its core dynamics are robust to hyperparameters and not over-engineered.

7 F1) mechanism

Ablation on CTM-AI dynamics

Ablation studies confirm a clear hierarchy of importance among CTM-AI mechanisms The iterative loop emerges as the most critical (-6.7 F1), followed by up-tree competition (-5. p. 3

Where the paper says it

(1) all specialized LTM processors run in parallel, each producing a chunk with a content gist, follow-up queries, and a self-assessed score; (2) an up-tree competition selects which chunk enters the limited-capacity STM, determining the system’s conscious content; (3) A down-tree broadcast distributes this content to all processors, at which point the system becomes consciously aware of—i. , pays attention to—the selected information; and (4) bidirectional links form between processors that hold complementary information, enabling direct unconscious communication that bypasses the STM, followed by fusion where linked processors exchange to enrich their memories.

p. 3
Where the paper says it

(1) all specialized LTM processors run in parallel, each producing a chunk with a content gist, follow-up queries, and a self-assessed score; (2) an up-tree competition selects which chunk enters the limited-capacity STM, determining the system’s conscious content; (3) A down-tree broadcast distributes this content to all processors, at which point the system becomes consciously aware of—i. , pays attention to—the selected information; and (4) bidirectional links form between processors that hold complementary information, enabling direct unconscious communication that bypasses the STM, followed by fusion where linked processors exchange to enrich their memories.

6), link fusion (-3. p. 3
Where the paper says it

CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness LTM Processor Inference Up-tree Competition STM winning chunk Down-tree Broadcasting link query Step1 Step2 Step3 Step4 Link formation and fusion query LTM Processor chunk Linked LTM processors query with each other to update LTM. chunk gist score query input Input includes multimodal information and one user query compete query Figure 2.

p. 3
Where the paper says it

CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness LTM Processor Inference Up-tree Competition STM winning chunk Down-tree Broadcasting link query Step1 Step2 Step3 Step4 Link formation and fusion query LTM Processor chunk Linked LTM processors query with each other to update LTM. chunk gist score query input Input includes multimodal information and one user query compete query Figure 2.

9), and down-tree broadcast (-3. p. 1
Where the paper says it

In this work, we bridge this gap between theory and practice by implementing the formal CTM model as a concrete system called CTM-AI, operationalized using today’s foundation models. We define and implement a concrete architecture and learning algorithm with (1) multiple specialized processors operating in parallel, (2) a limited capacity shortterm memory workspace enforcing selective attention via up-tree competition, (3) a global broadcast of information via a down-tree from the workspace to all processors, and (4) the formation of links between relevant processors over time, enabling unconscious communication to integrate their knowledge into higher-order multimodal information.

p. 1
Where the paper says it

In this work, we bridge this gap between theory and practice by implementing the formal CTM model as a concrete system called CTM-AI, operationalized using today’s foundation models. We define and implement a concrete architecture and learning algorithm with (1) multiple specialized processors operating in parallel, (2) a limited capacity shortterm memory workspace enforcing selective attention via up-tree competition, (3) a global broadcast of information via a down-tree from the workspace to all processors, and (4) the formation of links between relevant processors over time, enabling unconscious communication to integrate their knowledge into higher-order multimodal information.

5) Scaling the iterations from 1 to 4 yields continuous performance gains across both datasets, with MUStARD improving from 65. p. 8
Where the paper says it

We additionally conduct an ablation study on the number of inference iterations. As shown in Figure 6, scaling the iterations from 1 to 4 yields continuous performance gains across both datasets: MUStARD improves from 65.

0 to 73.8 The analysis also shows that link formation behaviors differ between datasets, as UR-FUNNY forms an average of 2. p. 8
Where the paper says it

Interestingly, while both datasets share a similar upward trend in performance, their link formation behaviors differ. At 4 iterations, UR-FUNNY forms an average of 2.

6 links at four iterations compared to just 1. p. 8
Where the paper says it

At 4 iterations, UR-FUNNY forms an average of 2. 6 links compared to just 1.

5 for MUStARD

RQ2: As an inference-time scaling method, how does CTM-AI compare with the reasoning paradigm?

CTM-AI and reasoning are complementary paradigms where applying reasoning models improves baseline performance but falls short of CTM-AI when the backbone model is fixed Pairing Qwen3-8B-think with CTM-AI yields an additional +16. p. 7

Where the paper says it

CTM-AI and reasoning are two complementary, stackable paradigms. While applying reasoning models undeniably improves baseline performance—yielding roughly a 10- point gain on StableToolBench and 5 points on MUStARD and UR-FUNNY—they still fall short of CTM-AI when the backbone model is held fixed.

8 to +27.3 point improvement on StableToolBench compared to using Qwen3-8B-think alone, confirming that reasoning deepens deliberation within individual processors while CTM-AI broadens coordination across them This decentralized architecture supports the flexible integration of reasoning, allowing practitioners to selectively inject advanced reasoning capabilities into specific modalities

A. The Use of Large Language Models (LLMs)

The implementation uses LLMs as a writing assistant and utilizes CodePilot and Claude Code to help code faster, with all AI-generated content manually checked The backbone model selected for most processors is Gemini-2.5-flash-lite because it supports audio, vision, and text inputs while being relatively small-scale

B. Artifact Details

The paper utilizes several models including GPT-4o (OpenAI, 2024), OpenAI o3 (OpenAI, 2025), Gemini-2.5-flash-lite (Comanici et al. p. 5

Where the paper says it

, 2025), and Gemini-2. 5-Flash-Lite (Comanici et al.

, 2025), and Qwen3 variants like Qwen3-VL-8B and Qwen3-Omni-flash Evaluation datasets include MUStARD, URFunny, WebArena-Lite, and StableToolBench

C. Experimental Details

For multimodal perception tasks, CTM-AI is instantiated with three modality-specific experts for MUStARD and URFUNNY For StableToolBench, each available tool acts as an independent LTM processor powered by a lightweight LLM constrained to utilize only its assigned API WebArena-Lite processors include an HTML processor, accessibility tree processor, and screenshot processor that perceive the current page state through distinct modalities

D. CTM-AI Implementation Details

The LTM Processor implementation defines operations like execute, read, and write for a memory state Mt updated over time The chunk inference function CTMcollect(·) generates chunks represented as a tuple containing a gist h it, an additional internal question q it, and a composite score s i t Link formation is triggered by the down-tree broadcast in Step 3, relying solely on the weight sub-score where a weight > 0.8 establishes a directed link The evaluation parsing prompts are tailored for each task to ensure accurate final generation, such as the specific prompt for UR-FUNNY which requires identifying specific humor techniques with confidence >= 0.6

E. Additional Case Study

A full iteration case study of CTM-AI for sarcasm detection shows progressive integration of multimodal cues through three iterations The failure case in URFunny is attributed to a vision-only misleading effect where the video processor failed to provide necessary facial expression information, leading to excessive links The failure in StableToolBench is attributed to tool mishandling, specifically a processor failing to invoke its designated API call for QR-code generation

F. Detailed Prompts

The paper provides detailed prompts used in experiments, including the self-reported score prompts that elicit processor confidence estimates The system prompts for MUStARD, UR-FUNNY, StableToolBench, and WebArena-Lite provide basic task context to processors The additional question-generation instructions ask each processor to propose follow-up queries q it to other processors whenever answering them could improve the final answer The WebArena-Lite prompt requires modality-specific explanations for parsing unique inputs like accessibility trees and SoM screenshots

G. Cost analysis

Assuming K processors and L established links, a single iteration requires 2(K + L) processor calls, consisting of K for initial chunk inference, K for evaluating the winning chunk’s link formation, and 2L for bidirectional multimodal fusion The temporal bottleneck is the API calls required for these three parallel stages, resulting in a wall-clock time per iteration of approximately 3T + ε where T is the latency of a single API call

H. Efficiency analysis

The system's efficiency analysis indicates that most tasks resolve within 1 to 3 iterations The performance boost from additional iterations is proportional to an increase in the average number of inter-processor links, demonstrating that additional iterations successfully encourage denser link formation The optimal balance for the link formation threshold η is found between 0.4 and 0.7 for both datasets

I. Conclusion

In conclusion, CTM-AI offers a robust foundation for cognitively inspired AI by translating consciousness theory into a practical, effective system The research aims to inspire more cognitive-driven research dedicated to building practical, fundamentally more capable general-purpose and self-adaptive AI systems

Acknowledgments

This work was done in part while Lenore Blum, Manuel Blum, and Paul Liang were visiting the Simons Institute for the Theory of Computing The authors acknowledge support from Nvidia’s GPU support

Impact Statement

The work aims to enhance AI capabilities in affective learning, decision-making, multi-step reasoning, and tool use by bridging the theoretical CTM framework with practical AI technologies The goal is neither to build conscious AI nor to replicate human identity

References

Achiam et al. Gpt-4 technical report. p. 9

Where the paper says it

, Anadkat, S. Gpt-4 technical report.

arXiv preprint arXiv:2303. p. 9
Where the paper says it

Gpt-4 technical report. arXiv preprint arXiv:2303.

08774, 2023

Anderson et al. Act-r: A theory of higher level cognition and its relation to visual attention. p. 9

Where the paper says it

, and Lebiere, C. Act-r: A theory of higher level cognition and its relation to visual attention.

Human–Computer Interaction, 12(4):439–462, 1997

Baars. B. J. A cognitive theory of consciousness. p. 1

Where the paper says it

Computational models of consciousness seek to formalize how the human brain selects, integrates, and distributes information (Butlin et al. Alongside Global Workspace Theory (GWT) (Baars, 1993), several models have been developed with distinct emphases, including Integrated Information Theory (Tononi, 2004), Higher-Order Theories (Rosenthal, 2005), and cognitive architectures such as ACT-R (Anderson et al.

Cambridge University Press, 1993

Bai et al. Qwen3-vl technical report. p. 9

Where the paper says it

, Cheng, Z. Qwen3-vl technical report.

arXiv preprint arXiv:2511. p. 9
Where the paper says it

Qwen3-vl technical report. arXiv preprint arXiv:2511.

p. 9
Where the paper says it

Qwen3-vl technical report. arXiv preprint arXiv:2511.

p. 9
Where the paper says it

Qwen3-vl technical report. arXiv preprint arXiv:2511.

21631, 2025

Blum and Blum A theory of consciousness from a theoretical computer science perspective: Insights from the conscious turing machine. Proceedings of the National Academy of Sciences, 119(21):e2115934

Blum and Blum A theoretical computer science perspective on consciousness and artificial general intelligence. Engineering, 2023

Butlin et al. Consciousness in artificial intelligence: insights from the science of consciousness. p. 9

Where the paper says it

, Frith, C. Consciousness in artificial intelligence: insights from the science of consciousness.

arXiv preprint arXiv:2308. p. 9
Where the paper says it

Consciousness in artificial intelligence: insights from the science of consciousness. arXiv preprint arXiv:2308.

08708, 2023

Card et al. The keystrokelevel model for user performance time with interactive systems. p. 9

Where the paper says it

, and Newell, A. The keystrokelevel model for user performance time with interactive systems.

Communications of the ACM, 23(7):396–410, 1980

Castro et al. Towards multimodal sarcasm detection (an obviously perfect paper). p. 9

Where the paper says it

, and Poria, S. Towards multimodal sarcasm detection (an _obviously_ perfect paper).

In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. p. 9
Where the paper says it

Towards multimodal sarcasm detection (an _obviously_ perfect paper). In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.

4619–4629, 2019

Cemri et al. AdaEvolve: Adaptive LLM driven zeroth-order optimization. p. 9

Where the paper says it

, and Stoica, I. AdaEvolve: Adaptive LLM driven zeroth-order optimization.

arXiv preprint arXiv:2602. p. 9
Where the paper says it

AdaEvolve: Adaptive LLM driven zeroth-order optimization. arXiv preprint arXiv:2602.

20133, 2026

Comanici et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. p. 9

Where the paper says it

, Rosen, E. 5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.

arXiv preprint arXiv:2507. p. 9
Where the paper says it

5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.

06261, 2025

Dai et al. Qoqmed: Building multimodal clinical foundation models with domain-aware grpo training. p. 9

Where the paper says it

, and Liang, P. Qoqmed: Building multimodal clinical foundation models with domain-aware grpo training.

arXiv preprint arXiv:2506. p. 9
Where the paper says it

Qoqmed: Building multimodal clinical foundation models with domain-aware grpo training. arXiv preprint arXiv:2506.

p. 9
Where the paper says it

Qoqmed: Building multimodal clinical foundation models with domain-aware grpo training. arXiv preprint arXiv:2506.

p. 9
Where the paper says it

Qoqmed: Building multimodal clinical foundation models with domain-aware grpo training. arXiv preprint arXiv:2506.

p. 9
Where the paper says it

Qoqmed: Building multimodal clinical foundation models with domain-aware grpo training. arXiv preprint arXiv:2506.

00711, 2025

Du et al. Improving factuality and reasoning in language models through multiagent debate, 2023

Franklin et al. Lida: A systems-level architecture for cognition, emotion, and learning. p. 9

Where the paper says it

, and Snaider, J. Lida: A systems-level architecture for cognition, emotion, and learning.

IEEE Transactions on Autonomous Mental Development, 6(1):19–41, 2013

Guo et al. Stabletoolbench: Towards stable large-scale benchmarking on tool learning of large language models. p. 9

Where the paper says it

, and Liu, Y. Stabletoolbench: Towards stable large-scale benchmarking on tool learning of large language models.

arXiv preprint arXiv:2403. p. 9
Where the paper says it

Stabletoolbench: Towards stable large-scale benchmarking on tool learning of large language models. arXiv preprint arXiv:2403.

07714, 2024

Guo et al. Stabletoolbench-mirrorapi: Modeling tool environments as mirrors of 7,000+ real-world apis. p. 9

Where the paper says it

, and Liu, Y. Stabletoolbench-mirrorapi: Modeling tool environments as mirrors of 7,000+ real-world apis.

In Findings of the Association for Computational Linguistics: ACL 2025, pp. p. 9
Where the paper says it

Stabletoolbench-mirrorapi: Modeling tool environments as mirrors of 7,000+ real-world apis. In Findings of the Association for Computational Linguistics: ACL 2025, pp.

5247–5270, 2025

Hasan et al. Ur-funny: A multimodal language dataset for understanding humor. p. 5

Where the paper says it

, 2019), and UR-FUNNY (Hasan et al. , 2019) for human understanding (sarcasm, humor, cultural references).

In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. p. 0
Where the paper says it

While the CTM provides a theoretical foundation that is fully decentralized, modality agnostic, and architecturally modular, key mechanisms, including how processors broadcast information, form links, and combine expertise, and learn from feedback, are left abstract with no empirical validation. This raises a natural question: Can the theoretical 1 arXiv:2605.

2046–2056, 2019

Hong et al. Metagpt: Meta programming for a multi-agent collaborative framework. p. 1

Where the paper says it

Key contributions. The significance of CTM-AI is threefold: (1) it serves as the first practical instantiation of the CTM, translating a theoretical cognitive framework into an executable working AI system; (2) it naturally yields a highly modular and decentralized multi-agent architecture, free from the rigid central orchestrators found in current agentic workflows, and where processors can be flexibly added or removed; and (3) CTM-AI integrates reasoning and and agentic flexibility, demonstrating how decentralized collaborative dynamics can further scale reasoning beyond single-agent models.

In The twelfth international conference on learning representations, 2023

Kim et al. Towards a science of scaling agent systems. p. 6

Where the paper says it

CTM-AI, with a max iteration of 2, surpasses most multi-agent baselines using only 6. 9 API calls for MUStARD, confirming that CTM-AI’s success stems from its dynamic coordination rather than scaling up API calls.

arXiv preprint arXiv:2512. p. 9
Where the paper says it

Gpt-4 technical report. arXiv preprint arXiv:2303.

08296, 2025

Laird et al. Soar: An architecture for general intelligence. p. 0

Where the paper says it

CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness Haofei Yu 1 * Yining Zhao 1 * Lenore Blum 2 Manuel Blum 2 Paul Pu Liang 3 https://consciousness-lab. io Abstract Despite remarkable advances, today’s AI systems remain narrow in scope, falling short of the flexible, adaptive, and multisensory intelligence that characterizes human capabilities.

Artificial intelligence, 33(1):1–64, 1987

Li et al. Puzzleworld: A benchmark for multimodal, open-ended reasoning in puzzlehunts. p. 7

Where the paper says it

While applying reasoning models undeniably improves baseline performance—yielding roughly a 10- point gain on StableToolBench and 5 points on MUStARD and UR-FUNNY—they still fall short of CTM-AI when the backbone model is held fixed. This performance gap stems from the nature of multimodal perception and tool-use tasks, which require complex, dynamic interactions that standard reasoning models are not explicitly designed to handle.

arXiv preprint arXiv:2506.06211, 2025a

Li et al. Align before fuse: Vision and language representation learning with momentum distillation. p. 0

Where the paper says it

CTM-AI contains an enormous number of powerful processors ranging from specialized experts (e. , vision-language models and APIs) to unspecialized general-purpose learners poised to develop their own expertise.

Advances in neural information processing systems, 34: 9694–9705, 2021

Li et al. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. p. 9

Where the paper says it

, and Liu, Y. Stabletoolbench: Towards stable large-scale benchmarking on tool learning of large language models.

In International conference on machine learning, pp. p. 1
Where the paper says it

, 2020) formalizes it at the neural level, and the Global Latent Workspace (VanRullen & Kanai, 2021) proposes a deep-learning roadmap for GWT-style integration. The Conscious Turing Machine (CTM) (Blum & Blum, 2021; 2022) formalizes GWT in the framework of Turing computation, defining mechanisms including up-tree competition for workspace access and down-tree broadcast (Blum & Blum, 2023), but has remained purely theoretical.

19730–19742. PMLR, 2023

Liang et al. Foundations & trends in multimodal machine learning: Principles, challenges, and open questions. p. 0

Where the paper says it

Existing research falls into two domains: either studying consciousness models (red) provides theoretical grounding but lacks practical implementations, or building multi-agent frameworks (blue) without principled architectural foundations. CTM-AI bridges this gap by instantiating the Conscious Turing Machine into a practical AI system, and by generalizing existing multi-agent frameworks through decentralized orchestration grounded in consciousness theory.

ACM Computing Surveys, 56(10):1–42, 2024

Lin et al. Agent-omni: Test-time multimodal reasoning via model coordination for understanding anything. p. 1

Where the paper says it

In this work, we bridge this gap between theory and practice by implementing the formal CTM model as a concrete system called CTM-AI, operationalized using today’s foundation models. We define and implement a concrete architecture and learning algorithm with (1) multiple specialized processors operating in parallel, (2) a limited capacity shortterm memory workspace enforcing selective attention via up-tree competition, (3) a global broadcast of information via a down-tree from the workspace to all processors, and (4) the formation of links between relevant processors over time, enabling unconscious communication to integrate their knowledge into higher-order multimodal information.

arXiv preprint arXiv:2511.02834, 2025

Liu et al. Visualagentbench: Towards large multimodal models as visual foundation agents. p. 5

Where the paper says it

, 2023), and ALBEF (Li et al. , 2021), which are trained to jointly model text and visual signals for multimodal interactions; and (2) prompting-based multimodal foundation models, including Qwen3-VL-8B-Instruct (Bai et al.

arXiv preprint arXiv:2408. p. 9
Where the paper says it

Gpt-4 technical report. arXiv preprint arXiv:2303.

06327, 2024

Mashour et al. Conscious processing and the global neuronal workspace hypothesis. p. 1

Where the paper says it

Within the GWT lineage, LIDA (Franklin et al. , 2013) implements GWT’s cognitive cycle symbolically, the Global Neuronal Workspace (Mashour et al.

Neuron, 105(5):776–798, 2020

Microsoft Autogen. https://devblogs. p. 0

Where the paper says it

CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness Haofei Yu 1 * Yining Zhao 1 * Lenore Blum 2 Manuel Blum 2 Paul Pu Liang 3 https://consciousness-lab. io Abstract Despite remarkable advances, today’s AI systems remain narrow in scope, falling short of the flexible, adaptive, and multisensory intelligence that characterizes human capabilities.

microsoft.com/autogen/, 2024

Novikov et al. Alphaevolve: A coding agent for scientific and algorithmic discovery. p. 1

Where the paper says it

, the manager in MetaGPT), and taskspecific workflows where each agent is bound to a predefined role and execution order. These design choices make such systems effective within their target tasks like coding (Qian et al.

arXiv preprint arXiv:2506.13131, 2025

OpenAI GPT-4o system card. Technical report, OpenAI, 2024

OpenAI o3 and o4-mini system card, 2025

Qwen3-VL-8B technical report. arXiv preprint arXiv:2511.21631, 2025

Qwen3-Omni-flash technical report. arXiv preprint arXiv:2509. p. 9

Where the paper says it

Gpt-4 technical report. arXiv preprint arXiv:2303.

17765, 2025

Yang et al. A survey of ai agent protocols. p. 1

Where the paper says it

, 2023; Hong et al. , 2023) or predefined tool-calling protocols (Guo et al.

arXiv preprint arXiv:2504. p. 9
Where the paper says it

Gpt-4 technical report. arXiv preprint arXiv:2303.

16736, 2025

Yao et al. React: Synergizing reasoning and acting in language models. p. 9

Where the paper says it

, and Mordatch, I. Improving factuality and reasoning in language models through multiagent debate, 2023.

In International Conference on Learning Representations (ICLR), 2023

Yu et al. Mmoe: Enhancing multimodal models with mixtures of multimodal interaction experts. p. 5

Where the paper says it

Here, “unified” means that a single model models multimodal interaction internally or manages multi-step tool use within one end-to-end system, without explicit decomposition into multiple collaborating agents. For multimodal perception, we include two types of baselines: (1) fine-tuned multimodal models, including MMoE (Yu et al.

arXiv preprint arXiv:2311. p. 9
Where the paper says it

Gpt-4 technical report. arXiv preprint arXiv:2303.

09580, 2023

Zhou et al. Webarena: A realistic web environment for building autonomous agents. p. 4

Where the paper says it

For agentic tasks (e. , web navigation tasks), motor processors translate the conscious content into actions on the external environment.

arXiv preprint arXiv:2307. p. 9
Where the paper says it

Gpt-4 technical report. arXiv preprint arXiv:2303.

13854, 2023

Zhou et al. Mem1: Learning to synergize memory and reasoning for efficient longhorizon agents. p. 9

Where the paper says it

Impact Statement This work utilizes publicly available datasets; no private or sensitive user data was collected, and all experiments were conducted in controlled research settings. By bridging the theoretical CTM framework with practical AI technologies, we aim to enhance AI capabilities in affective learning, decision-making, multi-step reasoning, and tool use, contributing to the development of more reliable and trustworthy general AI.

arXiv preprint arXiv:2506.15841, 2025

Table 1 Evaluation results on multimodal perceptions (MUStARD and UR-FUNNY)

Table 1 shows CTM-AI with the Gemini-2.5-Flash-Lite matches and slightly outperforms the strongest baselines on MUStARD and UR-FUNNY CTM-AI achieves state-of-the-art or competitive results across all three evaluation axes

Table 2 Evaluation results on tool using (StableToolBench)

Table 2 shows CTM-AI yields the most substantial gains, improving over the strongest single-model baseline by up to 16.8 points on multi-tool scenarios CTM-AI consistently outperforms the unified approach across all three evaluation axes

Table 3 Comparison with multiagent frameworks

Table 3 shows CTM-AI achieves the highest F1 scores on both MUStARD (72.23) and UR-FUNNY (72. p. 0

Where the paper says it

CTM-AI achieves state-of-the-art accuracy on MUStARD (72. 28) and UR-FUNNY (72.

13), outperforming MoA, the nextbest multi-agent method, by +0. p. 6
Where the paper says it

23) and UR-FUNNY (72. 13), outperforming MoA, the nextbest multi-agent method, by +0.

2 and +4.6 F1 respectively CTM-AI surpasses most multi-agent baselines using only 6. p. 6
Where the paper says it

6 F1, respectively, with comparable API calling and latency. CTM-AI, with a max iteration of 2, surpasses most multi-agent baselines using only 6.

9 API calls for MUStARD, confirming that CTM-AI’s success stems from its dynamic coordination rather than scaling up API calls

Table 4 Evaluation results on the reliability of self-reported scores

Table 4 shows Minimum Random CTM-AI achieves an F1 of 72.28 on MUStARD and 72. p. 0

Where the paper says it

Crucially, for whatever problem must be dealt with, information from many processors is selected, integrated, and exchanged appropriately to solve the task. CTM-AI achieves state-of-the-art accuracy on MUStARD (72.

13 on URFUNNY The score distribution stabilizes and converges by iteration 3, demonstrating that the self-reporting mechanism continues to play a crucial role in driving CTM-AI’s internal dynamics

Figure 5 Ablation on CTM-AI dynamics

Figure 5 shows that the iterative loop emerges as the most critical (-6.7 F1), followed by up-tree competition (-5.6), link fusion (-3.9), and down-tree broadcast (-3.5) Ablation on max iteration number T shows continuous performance gains across both datasets, with MUStARD improving from 65. p. 8

Where the paper says it

We additionally conduct an ablation study on the number of inference iterations. As shown in Figure 6, scaling the iterations from 1 to 4 yields continuous performance gains across both datasets: MUStARD improves from 65.

0 to 73.8

Figure 6 Ablation on max iteration number T

Figure 6 shows that scaling the iterations from 1 to 4 yields continuous performance gains across both datasets, with UR-FUNNY improving from 65.5 to 72.5

Figure 7 Ablation on STM output threshold τ

Figure 7 demonstrates that raising the STM threshold τ generally improves performance across both MUStARD and UR-FUNNY, which correlates with an increased number of inter-processor links Figure 8 shows that the link formation threshold η achieves an optimal balance between 0.4 and 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Implement a modular, decentralized multi-agent architecture free from rigid central orchestrators by utilizing up-tree competition and down-tree broadcast to determine information flow, as described in CTM-AI differs on both axes: (1) instead of a central executive (Hong et al., 2023), CTM-AI uses up-tree competition and down-tree broadcast to determine information flow in a fully decentralized manner. This allows for flexible addition or removal of processors.

  2. Develop systems capable of state-of-the-art accuracy on multimodal perception tasks like MUStARD and UR-FUNNY by leveraging the CTM's mechanism where information from many processors is selected, integrated, and exchanged appropriately to solve the task. This is demonstrated by CTM-AI achieving state-of-the art accuracy on MUStARD (72.28) and UR-FUNNY (72.13).

  3. Enhance tool use and agentic capabilities by ensuring that processors in CTM-AI have equal priorities is determined dynamically by competition at each iteration, leading to a 10+ points of improvement on StableToolBench and WebArena-Lite. This dynamic prioritization allows for adaptive task execution compared to fixed workflows.

  4. Enable continuous self-improvement through mechanisms where the system learns from its external world, as shown by the mechanism where new links form between processors with complementary information, progressively densifying the communication graph for richer unconscious exchange. This structural evolution allows CTM-AI to adapt over time without requiring explicit parameter updates.

  5. Improve performance on complex reasoning tasks by composing CTM-AI with advanced reasoning models, as evidenced by the finding that pairing Qwen3-8B-think with CTM-AI yields an additional +16.8 to +27.3 point improvement on StableToolBench compared to using Qwen3-8B-think alone. This shows that CTM-AI and reasoning models operate along orthogonal axes for enhanced coordination and deliberation.

  6. Increase robustness against noisy or incomplete data by employing calibrated self-reported scores, as the paper notes that Calibrated score decomposition encourages highly calibrated self-assessment and shows that replacing weighted sampling with minimum selection degrades performance dramatically (e.g., an F1 drop of 9.3 on MUStARD for min-selection). This means the system can reliably judge the quality of its internal information before broadcasting it.

  7. Create adaptive inference scaling by employing iterative loops where the iterative loop emerges as the most critical (-6.7 F1), allowing the system to dynamically allocate more iterations and API calls to refine harder instances, as demonstrated by performance gains from scaling iterations from 1 to 4 on datasets like MUStARD.

Abstract

Despite remarkable advances, today's AI systems remain narrow in scope, falling short of the adaptive and continual intelligence that characterizes human capabilities. This gap has fueled longstanding debates about whether AI might one day achieve human-like generality or even consciousness. Flipping the debate, we present a concrete blueprint for a multi-agent AI system inspired by a formal machine model of consciousness called the Conscious Turing Machine (CTM). Together, CTM-AI contains a large collection of powerful processors ranging from specialized experts (e.g., language models, multimodal models, tools, and APIs) to unspecialized general-purpose learners poised to develop their own expertise. Crucially, for whatever problem must be dealt with, information from many processors is competitively selected, integrated, and exchanged appropriately to solve the task. Feedback mechanisms in CTM-AI enable the system to iteratively evolve its internal connections and communication. CTM-AI achieves state-of-the-art accuracy on MUStARD (72.28) and UR-FUNNY (72.13), outperforming multimodal and multi-agent frameworks. On tool-using and agentic tasks, CTM-AI achieves 10+ points of improvement on StableToolBench and WebArena-Lite. Overall, CTM-AI offers a principled and testable multi-agent AI system inspired by a model of consciousness. The code is available at https://github.com/consciousness-lab/ctm-ai.

Sources

Related papers