Chain of Thought from Human Brain (CoTHB)

• CoTHB is a methodology that uses neuroimaging data to guide AI reasoning by aligning chain-of-thought steps with human neural activity.
• It maps stepwise brain signals from regions like the DLPFC and ACC to refine and steer model outputs in complex reasoning tasks.
• The approach addresses gaps in perception, valuation, execution, and integration, fostering human-aligned and interpretable AI systems.

Chain of Thought from Human Brain (CoTHB) is a concept and methodology proposed for leveraging direct neuroimaging evidence to augment the reasoning and decision processes of large foundation models, specifically targeting their deficiencies in high-level cognitive functions. CoTHB operationalizes the extraction and utilization of neural activity patterns—recorded from human subjects during multi-step reasoning tasks—as supervisory signals to inform, steer, and evaluate the stepwise "chain-of-thought" (CoT) prompting in language and multimodal AI models (Donoso, 17 Jan 2026).

1. Motivation and Theoretical Foundation

Contemporary foundation models, trained on human-generated data such as text and code, are ultimately limited by the indirectness and collapsed dimensionality of their training signals: observable behavior A(t) (e.g., responses, annotations) is but a projection of the underlying neural substrate B(t). The subset of B(t) accessible via neuroimaging B*(t) provides a unique information channel encoding latent cognitive variables—values, reliability, cognitive control, semantic integration—that do not fully manifest in outputs A(t).

The proposal for CoTHB arises from the identification of critical shortfalls in current foundation models along four cognitive dimensions:

• Perception: Fragility to out-of-distribution perturbations due to shallow pretraining targets.
• Valuation: Poor model–human value alignment, since annotation-based RLHF neglects internal reward signals.
• Execution: Shallow executive control and working memory, as models lack direct neural evidence of hierarchical reasoning states.
• Integration: Instability in conceptual binding and reality grounding, linked to the absence of neural-level integration and self-modeling.

By linking stepwise model computation to actual neural events in brain regions responsible for rule retrieval, reliability monitoring, and branching control, CoTHB inserts neurobiological constraints into the generative loop, aiming to induce more human-like and functionally robust reasoning (Donoso, 17 Jan 2026).

2. Neurobiological Substrate and Region-Specific Markers

CoTHB specifically exploits neural markers recorded from key prefrontal and cingulo-parietal executive and integrative circuits:

• Dorsolateral prefrontal cortex (DLPFC): encodes rule representation, task reliability, and working-memory buffers.
• Anterior cingulate cortex (ACC): tracks environment volatility, conflict, and exploratory value signals.
• Frontopolar cortex (FPC): codes for branching control and counterfactual reasoning trajectories.
• Associated regions (PPC, pre-SMA, IFG, TPJ, mPFC, PCC): participate in hierarchically structuring reasoning and integrating semantics, self, and social cognition.

Neural activity B*(t) is temporally and spatially parsed—in high-resolution settings (fMRI, ECoG)—along the progress of a multi-step reasoning task, so that each step or branch of human deliberation is mapped to a vector z_k summarizing relevant brain states at that juncture (Donoso, 17 Jan 2026).

3. Methodological Pipeline of CoTHB

The CoTHB protocol consists of several stages:

1. Task Presentation and Neuroimaging: A human subject is presented with a complex reasoning prompt p while undergoing continuous recording (EEG, fMRI, MEG, ECoG), focusing on regions associated with executive function.
2. Temporal Segmentation and Neural Parsing: B*(t) is segmented into K reasoning steps based on physiological markers—onsets of DLPFC activation for rule retrieval, peaks in ACC for decision volatility, FPC for subgoal branching.

Mathematically, zₖ = h(B* segment k), where h is a fixed feature extraction or encoding model.

1. Chain-of-Thought Model Steering: For a given LLM, chain-of-thought generation proceeds stepwise. At each step k:
   • The model proposes or refines step s_k conditioned both on its context and the contemporaneous neural marker z_k.
   • Reliability and exploratory signals z_k can trigger branch creation (ACC-exploration), backtrack (DLPFC-low reliability), or commit to the main chain.

Example steering loop (pseudocode, summarized verbatim from (Donoso, 17 Jan 2026)):

a₀ = “Chain-of-Thought Outline:” + p for k in 1…K: Generate step s_k = LLM(a_{k-1}) s_k := refine(s_k) guided by z_k a_k = a_{k-1} + s_k Return final answer from last a_K

1. Branching and Compositional Scoring: Alternative reasoning trajectories are instantiated if neurobiological signatures (e.g., high ACC activity) indicate exploration. The total CoTHB score for a chain of steps is:

$$\textrm{CoTHB score} = \sum_{k} \alpha_k z_k + \beta \log\Pr_\textrm{chain}(\textrm{LLM steps})$$

where weights α_k can prioritize neural reliability over model token-level log-probabilities.

2. Fine-Tuning and Model Update: Log-probabilities over token outputs at each step are adjusted so as to increase alignment with brain-inferred reliability/valuation signals (via loss minimization), guiding model adaptation to closer mimic human neural chains.

4. Implementation Considerations and Data Requirements

The practical implementation of CoTHB necessitates:

• High-resolution neuroimaging modalities (fMRI for spatial, ECoG/EEG for temporal):
  • Synchronized data streams: stimulus presentation, subject response, model token output, and neural events.
  • Preprocessing: motion correction, spatial coregistration, ROI time series extraction, denoising (ICA, GLM), temporal alignment via eye-tracking or forced-alignment.
• Neural translation models for expanding coverage, e.g., mapping EEG B*(t) to synthetic fMRI-style signals, enabling larger-scale collection and greater population coverage.
• Data scale: Each reasoning trajectory contributes a vectorized sequence of temporally resolved neural markers; realistic initial datasets involve 10–50 subjects × 30–60 minutes, but can scale with synthetic translation.
• Architectural integration: Transformer-based models are modified to incorporate neural-state attention heads or control signals, with auxiliary loss terms enforcing neural–model congruence at step boundaries.

5. Applications, Implications, and Challenges

Applications extend to:

• Human-aligned agents: Interactive systems whose stepwise plans are dynamically scored and steered by inferred human neural preferences, reliability, and exploration needs.
• Robust reasoning in AGI: Models acquire inductive biases towards human cognitive control, branching, and value assignment, mitigating shallow or non-robust solutions observed in standard LMs.
• Interpretable AI: Human brain signals provide a "cognitive skeleton" upon which model computations can be grounded and interpreted, facilitating neuroscientific validation and diagnostic insight.

Challenges and limitations:

• Data privacy and consent: Brain data is highly sensitive; strict protocols are required.
• Representational bias: Most neuroimaging cohorts are WEIRD, potentially compounding alignment bias.
• Scaling: High-resolution, multimodal neural data is expensive and logistically challenging, necessitating innovations in synthetic data expansion and modality translation.
• Interpretability: Decoding latent neural states risks misattribution; transparent model–brain alignment methods are essential (Donoso, 17 Jan 2026).

6. Comparison to RLHB and Broader Brain-Guided FM Paradigms

CoTHB sits alongside "Reinforcement Learning from Human Brain" (RLHB) as a complementary neuro-guidance method. RLHB uses neural reward or valuation signals to replace or augment human-annotated rewards in RLHF, directly training model output distributions to maximize brain-derived or hybrid reward signals.

CoTHB, by contrast, targets the structure and reliability of multi-step inference chains, aligning each generative reasoning step or branch with contemporaneous neural events, thus incorporating neurobiological signatures of executive control and hierarchical reasoning into the CoT prompting and generation process.

Both approaches exploit the generative brain principle: that B*(t) acquired during task execution contains principal, potentially privileged, information about cognition and reasoning unavailable to classical foundations models; their integration could chart a "middle path" between scaling current architectures and bio-inspired architectural refactoring (Donoso, 17 Jan 2026).

7. Outlook and Future Directions

CoTHB is a nascent but promising paradigm for brain-guided AI. Future research directions indicated include:

• Scaling up neural alignment: Augmenting the diversity and size of neuroimaging datasets, developing translation pipelines across modalities (EEG→fMRI), and exploring population-level alignment for robust neuro-symbolic reasoning.
• Multi-modal fusion: Integrating neural signals with classical supervision, reward modeling, and self-distillation.
• Refinements in parsing and mapping: Improved methods to map B*(t) to stepwise reasoning, including reverse-inference models and adaptive chain segmentation.
• Neuro-inspired architectures: Incorporating controllers or gating mechanisms that mimic cortical-cortical and cortico-striatal dynamics identified in neuroimaging data.

The ultimate implication is that brain–model alignment at the chain-of-thought level may enhance robustness, reliability, and human compatibility in agentic AI systems, offering a strategic alternative to pure scale-driven approaches (Donoso, 17 Jan 2026).