💬 NLP

Mechanistic Interpretability of Cognitive Complexity in LLMs via Linear Probing using Bloom's Taxonomy

This study demonstrates that Large Language Models encode cognitive complexity levels defined by Bloom's Taxonomy in a linearly separable subspace of their internal representations, achieving high classification accuracy and suggesting that cognitive difficulty is resolved early in the forward pass.

Original authors: Bianca Raimondi, Maurizio Gabbrielli

Published 2026-07-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bianca Raimondi, Maurizio Gabbrielli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out how a super-smart robot thinks. For a long time, scientists treated these robots, called Large Language Models (LLMs), like magic black boxes. You type a question in, and a perfect answer pops out, but nobody knows how the robot got there. It's like watching a chef cook a gourmet meal without ever seeing the kitchen or the ingredients. Recently, a field called "mechanistic interpretability" has started peeking inside the kitchen, trying to map the robot's brain circuits. One big question they ask is: Does the robot actually understand the difficulty of a task, or is it just guessing based on the words it sees? To answer this, researchers use a famous educational tool called "Bloom's Taxonomy." Think of this as a ladder of thinking skills. The bottom rung is simple memory (reciting a fact), and the top rung is high-level creativity (designing something new). If a robot is truly smart, its internal brain signals should change as you climb this ladder, just like a human's brain does when moving from memorizing a date to solving a complex mystery.

This paper takes a magnifying glass to the robot's brain to see if it organizes its thoughts according to this ladder of difficulty. The researchers, Bianca Raimondi and Maurizio Gabbrielli, didn't just ask the robot questions; they looked at the electrical signals (activations) inside the robot's brain while it was thinking. They tested four different popular robot brains (Llama, Qwen, Gemma, and DeepSeek) using 1,128 questions ranging from simple "define this" tasks to complex "design a new system" challenges. They wanted to know: Can we draw a simple straight line through the robot's brain data to tell if it is doing a simple memory task or a complex creative one?

The answer is a resounding yes, and it happens surprisingly fast. The team found that as soon as the robot starts processing a sentence (within the first few layers of its brain), it already knows how hard the task is. They call this moment the "Cognitive Separability Onset." It's like the robot has a special "difficulty switch" that flips on almost immediately. By layer 5 (out of 30 to 36 layers in these models), the robot's brain signals for "Remembering" are so clearly different from "Creating" that a simple math tool can tell them apart with about 90% accuracy. It's as if the robot sorts its thoughts into neat, separate piles based on how hard the job is, and it does this before it even starts writing the answer.

Even cooler, the robot makes mistakes in a very human way. When the robot gets confused, it usually mixes up steps that are right next to each other on the ladder. For example, it might confuse "Analyzing" with "Evaluating," but it almost never confuses "Remembering" with "Creating." This suggests the robot isn't just memorizing random labels; it has built a mental map where the difficulty levels are arranged in a smooth, ordered line, just like the educational theory intended.

To make sure the robot wasn't just cheating by looking at simple clues (like noticing that "design" usually means "hard" and "list" means "easy"), the researchers ran control tests. They tried to guess the difficulty using only the words on the page, without looking at the robot's brain. Those simple word-counting methods failed miserably, getting only about 73% accuracy and making random, scattered mistakes. This proves that the robot is actually doing something deeper than just reading the prompt; it is building a complex internal representation of the task's difficulty.

Finally, the team tried to "steer" the robot's brain. They found a specific direction in the robot's brain signals that points toward "higher difficulty." When they nudged the robot's brain in that direction, the robot's answers actually changed. If they pushed a simple question toward the "creative" direction, the robot started giving more complex, analytical answers. However, if they pushed too hard or too late in the process, the robot started stuttering and repeating words, showing that this "difficulty direction" is a real, functional part of how the robot thinks, not just a passive pattern.

In short, this paper suggests that these AI models have a built-in, measurable sense of cognitive complexity. They don't just process words; they organize their internal thoughts according to a ladder of difficulty, sorting simple facts from complex ideas in a structured, linear way that mirrors human learning. This discovery gives us a new way to peek inside the black box and understand that, at least in some ways, these models are organizing their "thoughts" much like we do.

Technical Summary: Mechanistic Interpretability of Cognitive Complexity in LLMs via Linear Probing using Bloom's Taxonomy

Problem Statement

Large Language Models (LLMs) have demonstrated capabilities in complex reasoning and generation, yet their internal mechanisms remain a "black box." Current evaluation frameworks rely heavily on surface-level performance metrics (e.g., answer correctness), which fail to distinguish between rote memorization and genuine multi-step reasoning. This paper addresses the need to assess the cognitive depth of LLMs by investigating whether internal neural representations encode a hierarchical structure of cognitive complexity, specifically aligned with Bloom's Taxonomy (Remember, Understand, Apply, Analyze, Evaluate, Create). The core research questions focus on whether these cognitive levels are linearly separable within the model's residual streams, at what depth they emerge, and whether these representations are causally active during generation.

Methodology

The study employs a mechanistic interpretability approach using linear probing on the residual stream activations of four state-of-the-art open-weight LLMs: Llama-3.1-8B-Instruct, Qwen3-4B-Instruct-2507, gemma-3-4b-it, and DeepSeek-R1-Distill-Llama-8B.

Dataset Construction: A balanced dataset of 1,128 natural language questions was curated, with 188 questions per Bloom's Taxonomy level (0–5). Sources included Computer Science course queries and the EduQG dataset. The dataset was downsampled to ensure class balance and prevent statistical priors from influencing probe accuracy.
Activation Extraction: For each model, the authors performed a forward pass without fine-tuning, extracting the hidden state vector ( $x_l$ ) at the final token position for every layer $l$ in the residual stream.
Linear Probing: A logistic regression classifier was trained at each layer to predict the Bloom level ( $y \in \{0, \dots, 5\}$ ) from the extracted activation vectors. The use of a linear probe ensures that successful classification reflects information that is linearly decodable from the representations, rather than relying on the expressive capacity of non-linear classifiers.
Control Experiments: To rule out the possibility that probe accuracy was driven solely by surface-level lexical cues (e.g., specific verbs like "define" vs. "design"), the authors compared their results against:
- Length Analysis: Verifying no significant length differences across levels.
- Shallow Baselines: Training logistic regression on TF-IDF vectors and sentence embeddings of the raw text.
Causal Validation (Activation Steering): To test if these representations are causally active, the authors defined steering vectors as the difference between the mean activations of the highest cognitive level (Create) and lower levels. These vectors were injected into the residual stream at various layers with varying strengths ( $\alpha$ ) to observe shifts in the model's predicted cognitive level and the qualitative nature of generated outputs.

Key Results

1. Linear Separability and Cognitive Separability Onset (CSO)

The linear probes achieved high accuracy (approaching ceiling performance) across all models, indicating that cognitive complexity is encoded in a linearly accessible subspace of the residual stream.

Emergence Depth: Cognitive depth is not linearly separable at the input or embedding level. Instead, a Cognitive Separability Onset (CSO) occurs consistently around Layer 5 across different architectures.
Stability: Once the CSO is reached, probe accuracy remains stable or improves slightly in deeper layers, suggesting that the representation of cognitive complexity is established early and preserved rather than re-computed at every step.

2. Geometric Structure and Error Patterns

Centroid Analysis: The Euclidean distance between centroids of adjacent Bloom levels increases monotonically with depth, showing a progressive geometric disentanglement of cognitive levels.
Structured Errors: Confusion matrices reveal that misclassifications occur predominantly between adjacent Bloom levels (e.g., Analyze vs. Evaluate). Long-range errors (e.g., Remember vs. Create) are rare. This suggests the model encodes cognitive depth as a continuous, ordered manifold rather than arbitrary discrete classes, mirroring the ordinal structure of human educational theory.

3. Control Experiments

Shallow text baselines (TF-IDF and sentence embeddings) achieved significantly lower accuracy (~~61–73%) compared to residual stream probes (~~90%). Furthermore, the error patterns in shallow baselines were noisy and non-ordinal, whereas residual stream errors were strictly adjacent. This confirms that the observed cognitive structure is not merely a reflection of surface-level lexical cues but represents an internal, deeper organization of the model's latent space.

4. Causal Validation via Steering

Activation steering experiments demonstrated that perturbing the residual stream along the identified cognitive direction induces a systematic shift in the model's behavior:

Layer Dependence: Steering in early/intermediate layers (e.g., Layer 5) resulted in coherent shifts in reasoning style (e.g., moving from listing facts to analyzing relationships). Steering in deeper layers often led to fluency degradation and repetition without meaningful reasoning changes.
Strength Dependence: Moderate steering strengths produced interpretable transitions in cognitive framing, while high strengths caused instability.
Shift Magnitude: Interventions induced an average shift of +1 to +1.5 Bloom levels in the probe-predicted cognitive level.

Significance and Claims

The paper claims to provide direct mechanistic evidence that LLMs spontaneously organize their internal representations to encode hierarchical cognitive demand, abstracted from specific domain content.

Beyond Surface Metrics: The findings suggest that cognitive complexity is not just a behavioral output but a structural property of the model's latent space, emerging early in the forward pass.
Bloom's Taxonomy as a Lens: The study validates Bloom's Taxonomy as a useful framework not only for human education but for analyzing the internal mechanics of AI, revealing that models align their internal geometry with human-defined cognitive hierarchies.
Causal Relevance: By demonstrating that steering these representations alters model output, the authors argue that these cognitive representations are causally active components of the generation process, not merely correlational artifacts.

The authors maintain a modest scope, noting that their analysis is currently limited to decoder-only transformers and English educational questions. They posit that these results lay the groundwork for future work in monitoring and regulating AI cognitive settings via their neural correlates, moving toward evaluation standards comparable to human intelligence.