The Neural Bifrost: Bridging Deep Brain Signals and the Surface of Consciousness

20/04/2026

•

23 min listen

•

Bartosz Lenart

AIconsciousnessneuroscienceEEGintracranialglobal-neuronal-workspaceagentic-AI+1
Get Instant Insight

Right now, your eyes are seeing things you'll never know about.

A 2026 experiment caught the moment your brain decides which ones you'll get to notice. Ten epilepsy patients, electrodes already threaded inside their brains for surgery, watched a faint flicker on a screen and pressed a button when they saw it. The team recorded activity from the inside and the outside at the same time, then asked the question consciousness research has been stuck on for thirty years: where in the brain does "I saw it" actually happen?

This changes what "understands" means for a machine, and gives us a blueprint for building one that does.

The Core Idea

  • Two timelines, two jobs. Early on (~⅛ s after the flash), the back of the brain, the part that sees, fires the same way whether you noticed or not. Later (¼ to ½ s), a different network at the front and top of the brain fires only when you consciously saw it. The first network sees. The later one best predicts whether you'll know you saw.
  • A bridge, not a root. Conscious experience does not live in the eyes or in visual cortex. It lives in the network that broadcasts what the visual cortex computed. You see with the back of your brain; you know you saw with the front.
  • A 25-year-old theory just got its sharpest evidence. Global Neuronal Workspace predicted exactly this two-step pattern in 2001. Until now, no one had the resolution to watch it happen inside a living human brain.
  • The AI takeaway. If reportable conscious access is "gating plus broadcasting," multi-agent AI systems already have the skeleton. The hard question, and what the rest of this article is about, is what gets promoted past the gate.

The Architecture Lesson

The brain solves the "everyone is talking at once" problem with a two-stage filter and a hard gate. Specialists work in parallel below awareness. One network decides what crosses. The crossing is sequential, selective, almost on/off. Engineering inference: AI systems that weight every input by relevance get muddier the longer they run. Systems built around a threshold-and-broadcast gate stay coherent.

[ NEURAL COMPRESSION COMPLETE ]

80% signal retained.
Full depth below.

The Neural Bifrost

A patient lies on an operating-suite cot in Paris, wires inside her brain. A laptop flickers a fuzzy striped circle for sixteen milliseconds. She presses a button. "Seen." Consciousness has not lit up in one place; a signal has crossed a bridge, and this paper finds out which one.

The Problem With Watching the Brain From the Outside

Imagine understanding a city's traffic from a rooftop. You see flow, bottlenecks, the choke around an accident. You never see which van carries perishable cargo, or which silent decision in a single cab will shape the night. Neuroscientists have lived on that rooftop for sixty years.

The instrument is electroencephalography (EEG): a cap of metal discs recording the summed field of millions of neurons firing centimetres below.

  • Scalp precision: centimetres
  • Cortical columns that actually compute: sub-millimetre1

The cap sees flow, not vehicles.

Watching the brain from the outside

The Wave at the Centre of the Debate

For thirty years, one squiggle on the EEG trace has been the main character of the consciousness debate: the P3b, a broad bump peaking between a third and two-thirds of a second after you notice something, strongest over the back of the top of the head.12 Report "I saw it" and the bump is large. Report "I didn't" (with the same physical stimulus) and it shrinks or vanishes.

But what is it? A memory being updated? A context being revised? Or the signature of the moment a piece of information becomes available to your entire brain at once?3 You can see the wave. You cannot see where it is born.

The 2026 study by Lozito et al.4 (Sapienza University of Rome / Paris Brain Institute) found a way to look inside the wall. But before we go inside, we need a map: the brain is a layered architecture, and you cannot interpret an electrode without knowing which layer it sits in.

The Brain as a Tree Network

The rest of this article tracks a signal as it climbs from one part of the brain to another, so the route needs a picture. If Yggdrasil is a useful metaphor for parallel reasoning, the brain is the biological original, and the Norse cosmology fits more cleanly than it has any right to.

The brain as a tree network

The Roots: The Senses

Your sensory cortices, the parts of your brain that handle sight, sound, and touch, are the roots. They sit toward the back and sides of your head, dense, specialised, tireless. They fire whether you are conscious of what they process or not: visual cortex extracts edges, the fusiform identifies "face" or "letter." Most of that work never reaches awareness.

The Crown: Prefrontal and Parietal Cortex

Your prefrontal and parietal cortex, at the front of your head and at the top toward the back, sit at the crown. They do not see the world. They receive compressed reports and, in the right conditions, broadcast them outward. When they fire reliably, you can report what you saw. When they stay quiet, the stimulus passes through without leaving a trace you can act on.

The Bifrost: The Bridge Between Them

Between roots and crown runs the frontoparietal network: long, fast cables of white matter connecting the back of the brain to the front. Sensory signals climb toward it. Most do not cross. The ones that do become conscious experience.

Lozito built an instrument to watch that bridge in a living human brain. The team needed three things at once: electrodes deep enough to see individual cortical regions firing, a stimulus weak enough that the brain genuinely had to decide whether to promote it, and a signal that tracks local computation rather than diffuse field effects.

Why You Can Only Run This Experiment On People Already Wired Up

The experimental architecture

You cannot drill into a healthy human brain to study consciousness. The window exists because of clinical necessity: surgeons implant depth electrodes for weeks before operating on people with drug-resistant epilepsy, to map which patch of cortex is starting their seizures. While the electrodes are in, the patient can volunteer for an experiment.

Ten agreed.4 Average age thirty-four. Together they contributed 541 usable contacts inside the brain, plus up to seven scalp derivations, all recording at the same time.

The Stimulus: Almost Not There

The experiment lives or dies on one choice: the flash has to sit right on the edge of being visible. Too clear and the patient sees every trial; too faint and they miss every trial. Either way you learn nothing about the moment of crossing.

The team used Gabor patches (fuzzy striped circles, the standard vision-research test pattern) flashed for sixteen thousandths of a second at low contrast, tuned per patient until the same flash produced "Seen" on roughly half the trials and "Unseen" on the other half.4 Same retina, same flash, same brain. The only thing that changed between trials was what the brain did about it.

The Signal They Listened To

To see what each electrode was doing, they tracked high-frequency broadband activity (70-140 Hz), the closest proxy we have for local neuronal population activity. It works on intracranial electrodes because it sits beneath the muscle noise that drowns it out at the scalp.

When the signal rises, nearby neurons are firing harder. Trial by trial, millisecond by millisecond, you can watch where in the brain the work is happening. That gave the team 541 channels of millisecond-resolution firing data, on every trial, simultaneously with the scalp wave. The right raw material. The harder problem was what to do with it.

541 Curves, And Only Two Of Them Mattered

The analytical engine

The team had 541 channels of brain activity, on every trial, from inside ten skulls. Rather than average the signals (which would smear early flash and late ignition into one blur), they let the computer sort each electrode by the shape of its response over time. Out of the resulting groupings, two dominant patterns carried the signal: one early and transient, one late and gradual.4

The Two Groups That Mattered

The Sensory cluster (the paper's Visual cluster) lit up early (roughly an eighth to a third of a second after the flash) in the visual parts at the back of the brain. Its response differentiated Seen from Unseen only weakly: its job was to register the flash and pass it on.

The Decision cluster (the paper's Accumulation cluster) lit up later (a quarter to half a second) in the frontoparietal network at the front and top. On trials the patient missed it barely climbed; on trials they saw, it rose from around 230 ms and peaked between 350 and 500 ms, exactly when the wave on the scalp was strongest.

When the team asked which cluster drove the scalp wave, the Decision cluster consistently outpredicted the Sensory cluster at frontal electrodes.4 The frontoparietal network's activity, not the sensory response, was the dominant contributor to the scalp P3b.

The study offers unusually direct evidence linking intracranial frontoparietal activity to the scalp P3b. And this exact architecture, two timelines and a late frontoparietal driver, was predicted twenty-five years before anyone could measure it. Lozito did not discover the pattern. They confirmed it at a resolution the theory had never been tested at.

The Pattern Was Predicted Twenty-Five Years Before Anyone Could See It

Global Neuronal Workspace theory

In 2001, Dehaene and Naccache formalised Global Neuronal Workspace (GNW)3 as a framework for consciousness (building on Dehaene, Kerszberg & Changeux's 1998 neuronal model): a two-phase architecture in which a signal first climbs through sensory cortex (mostly never reaching awareness), and then, if strong enough, ignites a sudden all-or-nothing amplification in the frontoparietal network that broadcasts it globally at once. The P3b is the predicted scalp signature.5 GNW also predicts that only one representation occupies the workspace at a time, a claim largely confirmed across two decades of attention-blink experiments.3

The Lozito findings line up almost one-to-one. The Sensory cluster is the climb. The Decision cluster is the ignition, and it drives the scalp wave, which is the broadcast signature. The wiring matches too: the brain's main long-distance cables between front and back, the superior longitudinal fasciculus, physically connect the Decision regions.4 The Bifrost has an actual cable plant.

To make this concrete, watch one trial unfold in real time.

A Single Trial in Slow Motion

Two distinct timelines

Back to the patient in Paris. One trial. A Gabor patch is about to flash for sixteen thousandths of a second. She does not yet know whether she will see it. Neither does her brain.

0 to 120 ms. Before the flash. Her brain is not idle. The rhythm of activity over her visual cortex, the alpha phase, partly determines what happens next. The roots are already disposed, slightly more excitable or slightly more suppressed, before the stimulus arrives.6

120 to 340 ms. The Sensory cluster fires. The Gabor flashes. The visual parts at the back light up. They would have lit up the same way if she were not going to notice. Her visual cortex knows. Nothing else does, yet.

Around 230 ms. The fork. Her brain decides. The Decision cluster either lights up or stays quiet. She will not be able to tell you afterwards when it happened, but this is the instant she becomes a person who saw the flash, or one who did not.

350 to 500 ms. Peak ignition. The Decision cluster peaked. The scalp wave swelled. The flash is now available to every part of her brain that needs it. She can report it, remember it, act on it.

490 ms onward. She presses the button. "Seen."

On missed trials, the first two steps played out almost identically. The third did not. No promotion. No wave. No report. Her visual cortex saw the flash; she did not. "I didn't see it" tracks, in this paradigm, what crossed the bridge.

The link is statistical, not causal; the sample is small (ten patients with focal epilepsy); and a counter-tradition shows the P3b largely disappears in no-report paradigms,7 making it the signature of reportable access rather than perception itself.

From Brain to Blueprint

Several of GNW's specific predictions were challenged by the 2025 Cogitate collaboration,8 notably that conscious content is richly encoded in prefrontal cortex and that ignition recurs at stimulus offset. What did survive, in exploratory form, was content-specific frontal-to-visual coupling, and the front-to-back broadcast fragment Lozito sharpened a year later is consistent with it: when the brain reports "I saw it," what fires under the wave is frontoparietal gating and broadcasting. If that is how the only working example of general intelligence solves selective broadcasting, what does it tell us about how to build the artificial ones?

Building the Bifrost in Software

Building the Bifrost in software

Lozito only maps visual gating in ten human brains. What follows treats it as a design blueprint, not a brain-computer interface argument or a claim that LLMs are conscious. The empirical hook is narrow; the engineering bet is broader.

The shared constraint is informational, not biological. Any system where many specialists write into one downstream consumer (motor cortex, attention head, context window, user) faces the same finite-bandwidth arithmetic, and evolution's cheapest curator is a thresholded broadcast. That is the pattern worth reverse-engineering in software, for the same reason agentic systems should amplify thought rather than replace it (AI as cognitive prosthetic). The orchestrator-as-Decision-cluster doesn't think for you. It decides what is worth broadcasting.

The Mapping

  • Specialist agents = Sensory cluster. Domain processors (browsing, code execution, RAG) whose outputs mostly should not reach the "conscious" layer. A browsing agent retrieving twenty results should not insist all twenty appear in working memory.
  • Orchestrator = Decision cluster. Operates on compressed reports, decides what gets promoted.
  • Shared context window = broadcast medium. What reaches it, every downstream process sees; what doesn't, disappears. The same constraint I explored in The Event Horizon of Thought: working memory and AI context are hard attention limits, and the brain handles them by making the gate selective.
  • Ignition threshold = promotion gate. A tangentially relevant retrieval, a redundant calculation, a browsing result that adds nothing: these are Unseen trials. Specialist fired, orchestrator didn't promote. The system working correctly, not failing.

Serial and Sharp: How Promotion Should Work

Two properties of the brain's gate carry over. First, GNW predicts only one representation occupies the global workspace at a time, a coherence mechanism rather than a bug: broadcasting everything simultaneously prevents downstream modules from holding a stable shared representation. Second, the gate in Lozito is sharp, close to off on Unseen trials and on on Seen ones, not a smooth slider. Routers built on soft attention tend toward smoothness and drift; routers built around an explicit commit step stay coherent over long horizons.

Specialists run in parallel. Promotions should be sequential, sharp, and curated.

Three Stacks, Same Triplet

Several recent systems are converging on this shape from different starting points:

  • At the orchestrator layer: routing only one specialist into the shared workspace at a time.9
  • At the token layer: gating on the model's own logits before it emits the next token.10
  • At the tool-call layer: separating planning from execution behind multi-variable authorisation checks instead of one fuzzy policy.11

Different stacks, same triplet: threshold, commit, broadcast.

A harder gate produces more coherent long-horizon behaviour, and fewer of the smooth, agreeable, slightly hollow outputs that fall out of systems mirroring the user's tone rather than committing to a position (why mirror-mode AI sounds conscious without being so). A bridge is only as useful as what arrives at it.

What Arrives At The Bridge Is Not Raw

What about below the gate? Are the specialists passive sensors, or something more?

A 2024 Nature paper by El-Gaby and colleagues12 looked inside the brains of mice running a repeated four-goal reward task. The "specialist" regions did not just pass raw signals upward. They held something more like a mental scoreboard: small, modular loops tracking where the mouse was in the task, not where it was in space. Eighty percent of recorded medial-frontal neurons were tuned to abstract task progress; the same loop fired at the same step of the task whether the mouse was in one corner of the room or the other. A map of structure, not a map of space.

Three things about the scoreboard mattered:

1. Modular. Activity was arranged into parallel rings, each anchored to a different goal, each tracking a distinct lag from its anchor.12 Specialists should each keep their own scoreboard indexed by progress, not share one giant scratchpad.

2. Circular. Each ring was closed, and the same rings replayed during sleep.12 For recurring workflows, memory should wrap: the end of one cycle feeds the next.

3. Forward-looking. Activity in the scoreboard predicted choices the mouse had not made yet, including behaviour an entire trial ahead, tens of seconds before the decision.12 Specialists should hand up predictions, not just observations.

These lessons are no longer hypothetical. Two recent agent-memory systems1314 are converging on the same shape: modular episodic buffers paired with longer-term knowledge, with a scheduled "replay" step that prunes low-value traces. Roots are starting to look like roots, not pages in a vector store.

Lozito showed the bridge decides what crosses. El-Gaby showed what arrives at the bridge is already structured, modular, and forward-looking.

The roots are not passive sensors. They are planning engines.

The Lesson of the Roots

On Unseen trials the sensory cortex fired correctly. The fusiform encoded the Gabor patch properly. What failed was not the sensing but the promotion.

The same pattern appears in multi-agent systems. A specialist retrieves the correct answer, but the context window is saturated, or the orchestrator's confidence falls below threshold, or three other specialists fire at once and crowd it out. The answer was there. The bridge stayed closed. The user gets a worse response than the system actually computed.

Better orchestration is not better specialists. It is a better promotion mechanism: sensitive, selective, and fast enough to get the right representations to the global workspace before the decision window closes.

Same Pattern, Different Machines

Inside one year, four groups starting from completely different points have built systems sharing the same underlying shape:

  • From theory: One team15 derived a multi-agent architecture directly from Global Neuronal Workspace, replacing passive shared memory with an active broadcast hub.
  • From orchestration benchmarks: Another16 organised specialists as nodes in a graph that rewires itself per task, with a single orchestrator routing on the full shared state, beating the most popular reasoning frameworks across six frontier models.
  • From biological simulation: A third17 arrived from the opposite direction, recovering ignition out of a connectome-constrained simulation of cortex.
  • From parallel reasoning: My own Yggdrasil argument puts specialists in parallel branches and lets a single trunk decide what gets promoted.

What keeps showing up: an active broadcast hub, specialists as nodes, a commit-style promotion step, one shared workspace accessed in sequence. Mechanisms differ; the pattern recurs: any piece can appear without the others, suggesting a broader coordination problem that brains and agent systems both solve under bandwidth limits.

Passive message-passing tends to produce stagnation. Active selective broadcasting tends to produce coherence.

Where This Goes Next

The pieces still live in separate papers. The interesting question is what they look like merged.

The gate is moving below the text layer, into the model's own probabilities. The roots are starting to hand up predictions rather than raw retrievals. And the missing primitive, the oldest trick in biology's book and among the newest in ours, is sleep:18 a state in which the bridge stays closed, the roots replay the day, and what survived the night decides the next morning's promotions.

The natural endpoint: an agent whose visible output is the small minority of its computation that survived a hard threshold, sustained by a recurrent loop, broadcast once, then released. Everything else stays unseen, the way most of what your visual cortex computes never reaches you.

Most of what such a system "thinks" will, by design, never be shown. Not a deficit. The architecture working.

The brain figured this out with a few hundred million years to converge. We are reverse-engineering the same answer in less than a decade.

Closing

We started inside a Paris operating suite and ended at the design of agentic systems that are emerging. One architectural claim connects them.

Consciousness is the crossing

The brain's wiring for what you can report seeing looks like a two-stage filter with a hard gate.

  • Sensory cortex computes the world.
  • A frontoparietal network appears to gate which of those computations become reportable.
  • The P3b is the scalp signature of that gating, not consciousness itself.

Everything else in the EEG trace (the shimmer that never becomes a clean late wave) is the story of what did not cross. In the brain, the gating window opens around 230 ms and closes by roughly 490 ms. In an agentic system, the analogous gate is whatever decides which specialist outputs reach the shared workspace.

The "Bifrost" shape is older than the field that named it. We are learning to build it on purpose.

The concrete brief:

  • Drop the soft filter. Systems that let every retrieval and tool call through by relevance tend to drift over long horizons.19
  • Add a hard commit step. Routers with a single-promotion gate hold their shape better.91011
  • Treat specialists as planners, not sensors. Have them hand up forward simulations, not raw observations.20
  • Add "sleep" (agentic RL). An offline replay of the day's trajectories; what survives the night updates tomorrow's policy.18

None of this is about building conscious AI. It is about borrowing a routing pattern biology had a few hundred million years to refine.

What's Next

This piece sits inside a longer arc on consciousness, attention, and how to build systems that respect the architecture of mind rather than fight it.

◆

Bartosz Lenart

I write blog.ckpt to think out loud about AI reasoning, agentic systems, and what it actually feels like to build cognitive architectures that don't fall over.

Enjoyed the article?
Get new essays on AI reasoning, cognition, and system design.
No noise.

Found this useful?

Share it with others who might benefit, or save the citation for your research.

Share
Citation

Bartosz Lenart (2026). The Neural Bifrost: Bridging Deep Brain Signals and the Surface of Consciousness. Retrieved April 21, 2026, from https://bartoszlenart.com/blog/neural-bifrost

References

License

This work is licensed under a Creative Commons Attribution 4.0 International License.

You are free to: Share copy and redistribute the material in any medium or format Adapt remix, transform, and build upon the material for any purpose, even commercially

Under the following terms: Attribution You must give appropriate credit, provide a link to the license, and indicate if changes were made.

Footnotes

  1. Polich, J. (2007). Updating P300: An integrative theory of P3a and P3b. Clinical Neurophysiology, 118(10), 2128-2148. Foundational review establishing the P3b as a content-independent correlate of conscious report (centro-parietal, 300-700ms). ↩ ↩2

  2. Sergent, C., Baillet, S., & Dehaene, S. (2005). Timing of the brain events underlying access to consciousness during the attentional blink. Nature Neuroscience, 8(10), 1391-1400. Demonstrated the all-or-none character of the P3b in attentional blink: trials are either Seen with full P3b or Unseen with none, not graded. ↩

  3. Dehaene, S., & Changeux, J.-P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200-227. Canonical synthesis of GNW: two-phase architecture (feedforward climb, late nonlinear ignition) and NMDA-mediated recurrent amplification. ↩ ↩2 ↩3

  4. Lozito, S., Lasaponara, S., Liu, J., Navarro, V., Lehongre, K., Frazzini, V., Malkinson, T. S., Doricchi, F., & Bartolomeo, P. (2026). Towards a bridge between intracerebral and surface EEG signatures of conscious report. Neuroscience of Consciousness, 2026(1), niag011. Primary study: simultaneous intracranial + scalp EEG in 10 patients; trajectory clustering and ridge regression link frontoparietal HFBB to scalp P3b. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  5. Dehaene, S., Lau, H., & Kouider, S. (2017). What is consciousness, and could machines have it? Science, 358(6362), 486-492. Distinguishes global access (C1) from self-monitoring (C2); argues current AI instantiates neither. ↩

  6. Mathewson, K. E., Gratton, G., Fabiani, M., Beck, D. M., & Ro, T. (2009). To see or not to see: Prestimulus alpha phase predicts visual awareness. Journal of Neuroscience, 29(9), 2725-2732. Pre-stimulus alpha in occipital cortex partially predicts whether the next near-threshold stimulus is consciously perceived. ↩

  7. Pitts, M. A., Metzler, S., & Hillyard, S. A. (2014). Isolating neural correlates of conscious perception from neural correlates of reporting one's perception. Frontiers in Psychology, 5, 1078. The post-perceptual challenge to the P3b: in no-report paradigms the P3b largely disappears while earlier markers persist. ↩

  8. Cogitate Consortium, Ferrante, O., et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642(8066), 133-142. Seven-year, 256-subject adversarial GNW vs IIT collaboration. Both failed key predictions; content-specific frontal-to-visual synchronisation survived. ↩

  9. Patel, N. (2026). Dynamic Attentional Context Scoping: Agent-Triggered Focus Sessions for Isolated Per-Agent Steering in Multi-Agent LLM Orchestration. arXiv preprint, arXiv:2604.07911. Two-mode (Registry / Focus) orchestrator: 90-98% steering accuracy vs 21-60% flat-context baseline; gap widens with N. ↩ ↩2

  10. Son, D. (2026). ProbeLogits: Kernel-Level LLM Inference Primitives for AI-Native Operating Systems. arXiv preprint, arXiv:2604.11943. Hard gating as an OS-kernel primitive: a single forward pass reads token logits before generation, zero learned parameters, tunable strictness α. Deployed inside Anima OS. ↩ ↩2

  11. Guerin, C., & Guerin, F. (2026). KAIJU: An Executive Kernel for Intent-Gated Execution of LLM Agents. arXiv preprint, arXiv:2604.02375. Intent-Gated Execution (IGX): tool calls authorised on four independent variables (scope, intent, impact, clearance) rather than one learned policy. ↩ ↩2

  12. El-Gaby, M., Harris, A. L., Whittington, J. C. R., Dorrell, W., Walton, M. E., & Behrens, T. E. J. (2024). A cellular basis for mapping behavioural structure. Nature, 635, 702-711. State-Memory Buffers in mouse medial frontal cortex: modular, circular sequences encoding abstract task progress, generalising across layouts, predicting choices 30-90° ahead. ↩ ↩2 ↩3 ↩4

  13. CraniMem (2026). Bounded Episodic Memory with Goal-Conditioned Gating for Long-Horizon LLM Agents. ICLR 2026. Bounded episodic buffer + long-term knowledge graph + scheduled consolidation that replays high-utility traces and prunes low-utility ones. ↩

  14. Liu, Y., et al. (2025). RoboMemory: A Brain-Inspired Multi-Memory Agentic Framework for Interactive Environmental Learning in Physical Embodied Systems. arXiv preprint, arXiv:2508.01415. Unifies spatial, temporal, episodic, and semantic memory inside a parallelised embodied-agent loop. ↩

  15. Shang, W. (2026). "Theater of Mind" for LLMs: A cognitive architecture based on Global Workspace Theory. arXiv preprint, arXiv:2604.08206. Proposes Global Workspace Agents (GWA): event-driven broadcast hub, entropy-based intrinsic drive, dual-layer memory bifurcation. ↩

  16. Hao, G., Dai, Y., Qin, X., & Yu, S. (2026). Brain-Inspired Graph Multi-Agent Systems for LLM Reasoning. arXiv preprint, arXiv:2603.15371. Specialists as nodes in a dynamically constructed directed graph with a GraphDesigner and global Orchestrator. Outperforms ReAct and Tree-of-Thoughts on Game24, Six Fives, Tower of London across six frontier LLMs. ↩

  17. Klatzmann, U., Froudist-Walsh, S., Bliss, D. P., Theodoni, P., Mejías, J., Niu, M., Rapan, L., Palomero-Gallagher, N., Sergent, C., Dehaene, S., & Wang, X.-J. (2025). A dynamic bifurcation mechanism explains cortex-wide neural correlates of conscious access. Cell Reports, 44(3), 115372. Connectome-constrained macaque-cortex model: ignition emerges as an all-or-none bifurcation from a hierarchical AMPA/NMDA receptor gradient (validated against autoradiography). ↩

  18. 2026 sleep-replay primitives for LLMs. SleepGate augments transformers with periodic sleep cycles over the KV cache (conflict-aware temporal tagging plus selective eviction reduces interference horizon from O(n) to O(log n)). Sleeping LLM implements an explicit wake-sleep cycle in which MEMIT edits dissolve into LoRA consolidation during an 8-step maintenance phase, reporting 100% recall on 60 facts in Llama 3.1-70B. Evolving Memory / Cognitive Trajectory Engine implements explicit SWS/REM/Consolidation phases over agent execution traces. ↩ ↩2

  19. Two convergent lines on long-horizon drift in agent systems. NeurIPS 2025: Towards Trajectory-Level Alignment: Detecting Intent Drift in Long-Horizon LLM Dialogues. Introduces the Intent Drift Score (IDS), a prefix-monotone metric for trajectory-level instability in multi-turn agents, scaling to million-token contexts with human-rating correlation above 0.82. Anthropic's Effective context engineering for AI agents describes the same failure mode as context rot: model performance degrades as context windows grow, treating context as a finite resource with diminishing returns. ↩

  20. Forward-simulating specialists in 2025-2026 multi-agent systems. LlamaAgent (SPRE) uses Strategic Planning & Resourceful Execution with hierarchical task decomposition five levels deep, reporting 77.8% accuracy and 40% token reduction versus ReAct baselines. Orchestrator: Active Inference for Multi-Agent Systems in Long-Horizon Tasks (arXiv:2509.05651) tracks agent-environment dynamics with active inference benchmarks to optimise global task performance under partial observability. ↩