Large language models are typically trained by predicting the next token in vast collections of text. Although this process produces impressive abilities in language generation, mathematical problem-solving and logical inference, it does not necessarily reproduce the mechanisms used by the human brain. Human language and reasoning appear to rely on partly distinct neural systems, creating a fundamental question for AI researchers: when a model solves a reasoning problem, is it drawing on internal representations that resemble those formed in the brain, or is it reaching the correct answer through fundamentally different computational routes?

To investigate this question, the researchers focused on deductive reasoning and compared the internal activity of LLMs with task-based functional magnetic resonance imaging, or fMRI, signals collected from human participants. fMRI does not measure individual neurons directly. Instead, it tracks changes in blood oxygenation that serve as an indirect indicator of neural activity across brain regions. The study examined areas associated with reasoning and assessed how well patterns in model representations could predict the measured brain responses while participants engaged with reasoning tasks.

The researchers used a neural predictivity metric to estimate the relationship between model activity and brain activity. In this framework, a model is considered more neurally predictive when its internal representations can explain a larger portion of the reliable, or explainable, variation in recorded neural signals. At the aggregate level, the tested language models accounted for a substantial fraction of explainable variance in reasoning-related brain regions. This does not mean that the models think like humans, but it does indicate that some of their representational structure overlaps with neural patterns involved in higher-order cognition.

The picture became more complicated when the researchers examined individual reasoning types. Predictivity was lower within specific categories of reasoning than it was across the aggregate dataset. This gap suggests that LLMs may capture broad computational patterns shared by multiple reasoning tasks while diverging from the brain in the detailed strategies used for particular forms of deduction. A model might therefore appear broadly aligned with reasoning-related neural activity while still relying on shortcuts, abstractions or processing sequences that differ significantly from those used by human thinkers.

The study’s central advance was to turn this representational relationship into a practical intervention. The researchers developed a brain-guided framework that identifies directions created by the joint structure of model representations and brain representations. In simplified terms, the method searches for patterns that are simultaneously meaningful in the model and predictive of neural activity, then uses those patterns to steer the model’s internal state. Rather than merely measuring whether a model resembles the brain, the approach attempts to move its representations toward a region of computational space associated with human reasoning signals.

This steering can occur during inference, when a model is generating an answer, or during training, when the model’s parameters are fine-tuned. Inference-time intervention modifies the model’s processing without necessarily retraining all of its weights, making it potentially useful for testing whether neural signals can influence reasoning on demand. Fine-tuning, by contrast, allows the brain-derived information to become incorporated into the model’s longer-term behavior. Together, the two approaches provide a way to examine whether neural guidance produces genuine reasoning improvements rather than superficial changes in wording.

Across ten language models ranging from 1.5 billion to 72 billion parameters, the researchers report that task-evoked brain signals improved reasoning performance. The gains were described as orthogonal to those produced by language-only supervision, meaning that the neural guidance contributed information not already captured by conventional text-based training. The improvements also transferred across reasoning types, suggesting that brain-derived directions may influence more general reasoning processes instead of helping only on the exact tasks used to collect the fMRI data. In the strongest reported cases, accuracy increased by as much as 13 percentage points.

The findings do not establish that LLMs possess human-like consciousness, nor do they show that fMRI signals contain a complete blueprint for reasoning. Brain activity is noisy, indirect and shaped by many interacting processes, while model representations depend heavily on architecture, training data and optimization. Nevertheless, the results suggest that neural data can serve as a functional guide for AI systems, offering a source of supervision that is fundamentally different from additional text. If confirmed and expanded with larger datasets, more precise neural measurements and broader cognitive tasks, brain-guided training could become a new route toward models that are not only more capable, but also more robust and closely aligned with the computational organization of human cognition.

Subject of Research: Brain-guided language models and the relationship between LLM representations and human neural mechanisms underlying deductive reasoning.

Article Title: Beyond representational alignment with brain-guided language models for robust reasoning.

Article References: Xiao, M., Du, K. & Lin, Z. “Beyond representational alignment with brain-guided language models for robust reasoning.” Nature Machine Intelligence (2026). https://doi.org/10.1038/s42256-026-01278-w

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s42256-026-01278-w

Keywords: large language models, brain-guided AI, deductive reasoning, fMRI, neural predictivity, brain–computer alignment, representation learning, cognitive alignment, artificial intelligence, robust reasoning

At the heart of the method is a computational framework that couples a cognitive-map representation with neural sampling. Instead of treating memory as a static database, the system learns a map-like structure that captures relationships between locations, contexts, and trajectories. This representation becomes the substrate for imagining what might happen next.

The researchers show that their model can sample plausible future scenarios from the learned cognitive map. Those samples are then evaluated according to a target objective, enabling the agent to choose actions that are not merely feasible but directed toward achieving specific goals. In effect, the AI performs “what-if” rollouts internally, guided by learned spatial structure.

Technically, the model’s imagination process is driven by sampling dynamics embedded in a neural architecture. By drawing from the map-informed distribution, the system can explore multiple candidate futures, improving robustness when environments are uncertain or partially observed. Rather than producing a single deterministic prediction, it generates a set of possibilities and selects among them.

To demonstrate effectiveness, the study evaluates performance on planning tasks that require navigating toward goals through environments where naive strategies would struggle. The results indicate that neural sampling from cognitive maps improves both planning quality and flexibility, outperforming approaches that lack goal-conditioned imagination.

The findings also suggest a path toward more general-purpose agents. Cognitive maps are often viewed as a bridge between neuroscience and AI; here, they serve as a practical mechanism for decision-making. By enabling goal-directed imagination, the approach could help AI systems transfer planning skills across tasks that share underlying spatial or relational structure.

Importantly, this work reframes planning as an inference problem: the agent infers which internal futures are most consistent with the goal. This perspective could make planning faster and more scalable, particularly in domains where enumerating every possible route is impractical.

As viral interest grows, the study’s core promise is clear: AI that can think ahead using map-like memory could move planning from static algorithms toward adaptive, simulation-based intelligence.

Subject of Research: Neural sampling, cognitive maps, goal-directed imagination and planning
Article Title: Neural sampling from cognitive maps enables goal-directed imagination and planning.
Article References: Lin, H., Yang, Y., Zhao, R. et al. Neural sampling from cognitive maps enables goal-directed imagination and planning. Nat Mach Intell 8, 1045–1065 (2026). https://doi.org/10.1038/s42256-026-01254-4
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s42256-026-01254-4
Keywords:

A team of researchers has introduced a groundbreaking synaptic device fully controlled and modulated by photons, diverging fundamentally from prior designs that hinge on electric signal transduction stages. This novel optical synapse harnesses a rare-earth-doped long-afterglow crystal, whose persistent luminescence properties enable information storage and processing purely through photonic pathways. This evasion of electrical intermediaries is not merely a conceptual novelty — it can substantially curtail noise levels, slash energy requirements, and accelerate operations, particularly in tasks grounded in visual data processing that naturally involve photons at inception.

The core material at the heart of this device is a doped crystalline matrix exhibiting both immediate photon emission and delayed luminescent afterglow, mediated by trapped charge carriers. When illuminated, some carriers relax promptly to emit photons, while others become temporarily ensnared within defect-induced traps, releasing their energy over extended periods. Crucially, the population dynamics of these traps depend intricately on the device’s illumination history, effectively encoding temporal patterns of input signals. This history-dependent modulation mirrors the synaptic plasticity observed in neural networks, where synaptic strength adapts dynamically based on prior activity, forging a physical instantiation of short-term memory.

To accurately characterize and predict the photophysical behavior of this device, the team formulated a comprehensive kinetic model. This framework accounts for the generation, capture, storage, and release of photoexcited carriers, integrating the competing pathways that determine the balance between instantaneous and persistent luminescence. The model reveals that prior light exposure alters the availability of traps, thereby modulating subsequent emission efficiency. Such nonlinear temporal dependencies provide an elegant mechanistic basis for bidirectional synaptic plasticity—all achieved without recourse to any electrical stimulation or control circuitry.

Experimental validation of the device’s synaptic functionalities was performed using dual-wavelength optical stimulation protocols. Under ultraviolet excitation, the device exhibited paired-pulse facilitation: a second light pulse closely following the first produced an enhanced luminescent response. This enhancement arises because initial excitation partially saturates trap states, thereby biasing subsequent carriers towards faster, direct recombination pathways. Conversely, near-infrared stimulation induced paired-pulse depression. Here, the initial pulse emptied previously trapped carriers, causing a diminished response to the following pulse. The coexistence of these opposing plasticity modes—excitatory and inhibitory—imbues the device with the versatility necessary for emulating complex neural processing.

Moreover, the experimental results corresponded impeccably with the theoretical model’s predictions, underscoring a robust understanding of the underlying physical processes. The research team demonstrated fine-tuned control over device response via modulation of key parameters such as light intensity, pulse duration, and inter-pulse timing. Importantly, they confirmed that the synaptic behaviors observed stemmed authentically from trap dynamics rather than simply residual luminescence, reinforcing the physical legitimacy and repeatability of their design.

Pushing the boundaries towards practical application, the researchers integrated the photon-modulated synaptic crystal atop a commercial silicon imaging sensor, creating a prototype neuromorphic vision system. In this hybrid device, the photonic synapse layer processes incoming images in situ, effectively performing early-stage data interpretation. Notably, strong optical signals persist longer in the crystal’s afterglow, while weak or noisy signals decay rapidly. This intrinsic temporal filtering acts as a form of in-sensor contrast enhancement and noise suppression, circumventing the need for conventional post-processing steps and streamlining the entire image acquisition pipeline.

Leveraging this innovative sensor, the team evaluated performance on image recognition tasks, particularly handwritten digit classification. A simulated neural network employing the measured synaptic device responses achieved an impressive 95.99% accuracy following noise reduction—a stark improvement over approximately 78% accuracy without integrated optical denoising. This milestone not only validates the concept of merging sensing, memory, and processing but also showcases the potential for enhanced computational efficacy and reduced complexity in real-world AI systems.

While current operational speeds of the device range across milliseconds to seconds, slower than typical electronic components, they align closely with biological timescales relevant to visual processing. This temporal congruence suggests the potential for biologically inspired computational timing regimes, rather than simply faster hardware clocks. The authors envisage that scaling the device dimensions and refining the doped crystalline material properties could yield significant increments in speed and energy efficiency, opening pathways to broader applicability.

This research exemplifies a visionary stride toward fully optical neuromorphic computing platforms. By combining optical sensing, information storage, and processing within a single crystal device, it bypasses longstanding bottlenecks inherent in electronic systems. Such all-photonic architectures hold promise for diverse sectors such as robotic vision, autonomous vehicles, wearable electronics, and edge computing, where limited power budgets and rapid data interpretation are paramount.

Further developments may include integrating arrays of these photon-modulated synapses to form complex optical neural networks and exploring materials with even longer-lived trap states or tunable spectral responses. The ability to engineer synaptic plasticity through tailored photonic stimuli provides a rich toolbox for crafting adaptive, intelligent machines that operate closer to the efficiency and sophistication of biological brains.

Overall, this pioneering work sets a new benchmark for neuromorphic device design, suggesting a future where light itself not only conveys information but also processes and remembers it—ushering in a new era of energy-frugal, high-speed, and biologically plausible artificial intelligence.

Subject of Research:
Not applicable

Article Title:
Fully photon-modulated synaptic devices with bidirectional plasticity for neuromorphic vision and recognition

News Publication Date:
25-May-2026

Web References:
https://www.spiedigitallibrary.org/journals/advanced-photonics/volume-8/issue-04/046001/Fully-photon-modulated-synaptic-devices-with-bidirectional-plasticity-for-neuromorphic/10.1117/1.AP.8.4.046001.full

References:
Y. Yan et al., “Fully photon-modulated synaptic devices with bidirectional plasticity for neuromorphic vision and recognition,” Adv. Photon., vol. 8, no. 4, p. 046001, 2026. doi: 10.1117/1.AP.8.4.046001

Image Credits:
Y. Yan et al.

The human brain is an extraordinarily efficient organ, containing about 100 billion neurons interconnected by approximately 100 trillion synapses. This intricate and densely packed architecture enables the brain to store and process vast amounts of information seamlessly and efficiently. Current AI marvels, while capable of impressive feats, often fall short of this biological benchmark largely due to their bulky circuitry and significant energy demands. The quest for brain-like AI chips has become a pivotal goal in AI research, highlighting the limitations of existing semiconductor technologies.

In contrast, memristors present themselves as an attractive alternative to conventional semiconductor components. Capable of retaining a memory of the current that has passed through them, memristors efficiently perform memory and computation in a single device. This dual functionality permits a significantly denser configuration than traditional semiconductor devices, providing the potential to store substantially more information in a similar physical area—up to dozens of times more than SRAM technologies.

Despite their promise, the integration of memristors into larger systems has experienced obstacles that have curtailed their widespread adoption. Historically, challenges such as process complexity, low manufacturing yield, and issues related to voltage loss and current leakage have hindered their transition from small-scale laboratory prototypes to large-scale wafer production. The transition involves not just technological advancements but also a comprehensive understanding of the interactions between materials, circuit designs, and operational algorithms.

Professor Choi’s team, in collaboration with Dr. Dmitri Strukov from the University of California, Santa Barbara, has unveiled an innovative methodology that emphasizes the co-design of materials, components, circuits, and algorithms. This strategic synthesis has allowed for the creation of a memristor crossbar circuit that achieved an astonishing yield of approximately 95% on a four-inch wafer. This achievement is particularly notable for its comparatively simple fabrication process, contrasting sharply with previous methods that relied on overly complex procedures.

In their latest research, the team also successfully demonstrated a three-dimensional vertical stacking structure using memristors. This breakthrough signifies more than just dimensional expansion; it opens the door to the realization of large-scale AI systems powered by memristor technology. By stacking these devices, the potential for dramatically increasing the computational capability of artificial intelligence systems emerges, as they can utilize vastly greater amounts of memory and processing resources without the significant spatial requirements that traditional circuitry entails.

When the researchers applied their cutting-edge technology within a spiking neural network framework, they observed notable enhancements in both efficiency and stability during AI computations. This observation underscores the transformative potential of memristor technology to drive performance improvements in AI applications, which could lead to systems that are not only faster but also require far less energy than their traditional counterparts.

The implications of Professor Choi’s research extend beyond mere academic curiosity; they represent a pivotal shift in how future semiconductor devices may be designed and integrated. As the field moves toward increasingly sophisticated models of artificial intelligence that mimic human cognitive processes, the advancements introduced by this research could catalyze the development of high-performance computing environments that operate with unprecedented efficiency.

In his reflections on the research, Professor Choi articulated the significance of their findings by stating, “This study proposed a method for improving memristor integration technology, which had been limited in the past.” The optimism for the potential evolution of a next-generation semiconductor platform seems well-grounded, as the technological landscape of AI continues to evolve rapidly in tandem with these advancements.

The research has garnered substantial support from key funding bodies, including the U.S. National Science Foundation and various programs under the Korea Institute for Advancement of Technology and the National Research Foundation of Korea. With Professor Choi as the lead author and Professor Dmitri Strukov as a co-author, the findings were published in October in the esteemed journal Nature Communications, marking a significant milestone in the ongoing discourse within the semiconductor and AI research community.

As industries across the spectrum begin to recognize the transformative potential of memristor technology, the horizon for next-generation computing appears increasingly bright. The implications not only hold promise for the field of AI but could also redefine how interconnected devices operate in the era of the Internet of Things (IoT) and beyond, shaping the future of technology for generations to come.

The exploration of memristors has only just begun, but the findings from this research present a substantial leap toward practical applications that mimic biological systems. The potential is vast, and as researchers continue to refine their techniques and enhance the functionality of these devices, we may soon find ourselves on the brink of a technological renaissance fueled by brain-inspired computing.

As we look forward to the continued evolution of this research, it is evident that the intersection of materials science, neuroscience, and engineering is leading us toward potentially unimaginable advancements in artificial intelligence and computing technologies. The future may hold not only faster computers but systems that learn, adapt, and evolve in ways reminiscent of human thought processes, heralding a new era in both AI and human-computer interaction.

Subject of Research: Memristive Passive Crossbar Circuits for Neuromorphic Computing
Article Title: Wafer-scale Fabrication of Memristive Passive Crossbar Circuits for Brain-scale Neuromorphic Computing
News Publication Date: 1-Oct-2025
Web References: DOI
References: Nature Communications
Image Credits: Design and Process of Scalable Manual Crossbar Circuit

Keywords

Applied sciences and engineering, Engineering, Materials engineering, Fabrication