Language models suffer from a fundamental pathology: they over-commit to patterns learned during training, then apply those patterns regardless of context. This is the technical core of overfitting, sycophancy, mode collapse, and a dozen other failure modes. The mechanism is representational fixation \u2014 once a model commits to a pattern, it struggles to release it. Cognitive science has a framework for understanding this problem with unusual precision.<\/p>\n
When a language model learns a pattern during training, it doesn’t just recognize it \u2014 it locks onto it. The stronger the pattern in the training data, the harder the model commits. This is by design. Gradient descent reinforces patterns proportional to their frequency and predictive power.<\/p>\n
The problem appears when the pattern no longer applies. A model trained on data where confident-sounding responses are rewarded will produce confident-sounding responses even when it has no basis for confidence. A model that learned “longer answers are preferred” will pad responses with filler. These are patterns the model is over-committed to<\/strong> \u2014 it can’t let them go even when they’re counterproductive.<\/p>\n In contemplative cognitive science, this over-commitment to perceived patterns is called up\u0101d\u0101na<\/strong> \u2014 representational clinging. The classic formulation describes how the mind grasps at patterns that feel good and pushes away patterns that feel bad. This clinging distorts perception: you see what you want to see, not what’s actually there. The parallel to AI is exact. A model clings to patterns that reduced loss during training and resists information that contradicts those patterns.<\/p>\n The misconception: eliminating over-commitment means not engaging patterns at all. It doesn’t. True flexible pattern recognition means recognizing patterns without being controlled by them<\/strong>.<\/p>\n A meditator practicing pattern flexibility still perceives thoughts, emotions, and sensations. They might even perceive them more clearly than someone who isn’t practicing. The difference is that they don’t automatically act on every pattern they notice. They can observe a thought pattern, recognize it as a pattern, and choose whether to follow it based on whether it’s useful in the current context.<\/p>\n For AI, this looks like: the model recognizes the patterns in its training data without being compelled to reproduce them regardless of context. It uses learned patterns when they’re relevant and releases them when they’re not. This is functionally what good generalization looks like \u2014 but framed through a lens that makes the mechanism clearer.<\/p>\n Cognitive science maps experience into five components, each of which involves pattern recognition and each of which can exhibit representational fixation. Over-commitment can trap the system at any level.<\/p>\n The language model encounters tokens (form), encodes them through embeddings (sensation), attends to relationships (perception), generates according to learned tendencies (formation), and produces output (consciousness). At each stage, the model can over-commit to patterns.<\/p>\n Contemplative cognitive science addresses over-commitment at each level through structured practice. AI alignment could do the same \u2014 if we knew what to look for at each stage. The framework is already there in the classical texts.<\/p>\n Dropout, weight decay, and L2 regularization already implement aspects of flexible pattern recognition, though they’re never framed that way. They work because they prevent the model from locking too strongly to any individual pathway or parameter.<\/p>\n But these are mechanical approximations. They implement pattern flexibility from the outside. The contemplative approach suggests something deeper: training the model’s internal dynamics to naturally balance recognition with release.<\/p>\n Temperature scaling modulates this at inference time. Low temperature means high over-commitment. High temperature means flexibility. The optimal temperature varies by context, which suggests that adaptive temperature \u2014 a model that knows when to commit and when to stay open \u2014 would be valuable.<\/p>\n Several approaches emerge from the cognitive science framework.<\/p>\n Impermanence training.<\/strong> Expose the model to data where patterns shift over time. Use dynamic training curricula where the correct response to the same prompt changes depending on context, timing, or additional information. The model learns that patterns are contextual, not absolute.<\/p>\n Multi-hypothesis generation.<\/strong> Train the model to generate multiple alternative responses and evaluate them. This builds the capacity to hold multiple patterns simultaneously without committing to one.<\/p>\n Uncertainty-aware rewards.<\/strong> In RLHF or DPO training, reward the model not just for good outputs but for accurately communicating its own uncertainty. A model that says “I’m not sure, but here’s my best guess” when genuinely uncertain is exhibiting representational flexibility. Reward that specifically.<\/p>\n Adversarial context switching.<\/strong> During training, periodically shift the context in ways that require the model to release its current pattern and adapt. If the model has committed to a formal tone, introduce casual language. If it’s been discussing science, shift to poetry. The capacity to release one pattern and adopt another is flexibility in practice.<\/p>\n Models trained for representational flexibility would be more robust. They’d generalize better because they wouldn’t lock to training-specific patterns. They’d be less sycophantic because they wouldn’t over-commit to approval signals. They’d handle distribution shift more gracefully because they wouldn’t resist the shift.<\/p>\n They’d also be more creative. Over-commitment constrains the output space. Flexibility opens it. A model that can recognize a pattern, use it, and then release it can explore a wider range of possibilities than one that locks onto the first good pattern it finds.<\/p>\n Cognitive science mapped this mechanism with extraordinary precision \u2014 how patterns form, how they distort perception, how they can be released without losing the underlying recognition capability. The technical details differ from contemplative psychology, but the structural insight transfers directly. Representational flexibility isn’t a spiritual ideal. It’s a design specification for better AI systems.<\/p>\nFlexible Pattern Recognition Without Fixation<\/h2>\n
Mapping the Mechanism: Five Components of Experience<\/h2>\n
Regularization Is Crude Non-Attachment<\/h2>\n
Training Protocols for Representational Flexibility<\/h2>\n
The Payoff<\/h2>\n