🤖 machine learning

Representational Curvature Modulates Behavioral Uncertainty in Large Language Models

This paper demonstrates that the geometric curvature of a model's representational trajectory—how much it "bends" during processing—is directly correlated with and can be used to modulate the predictive uncertainty (entropy) of large language models.

Original authors: Jack King, Evelina Fedorenko, Eghbal A. Hosseini

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Jack King, Evelina Fedorenko, Eghbal A. Hosseini

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 watching a professional figure skater performing a routine.

When the skater is moving in a smooth, predictable, straight line across the ice, you can easily guess where they will be in three seconds. You aren't surprised; you are "certain." But if the skater suddenly starts performing sharp, jagged, zig-zagging turns, you can no longer predict their next move. You feel a sense of "uncertainty."

This paper, written by researchers at MIT, suggests that Large Language Models (like ChatGPT) work in a very similar way.

The Core Idea: The "Smooth Path" vs. The "Jagged Path"

When an AI reads a sentence, it doesn't just see words; it converts those words into a mathematical "path" (called a trajectory) through a high-dimensional space.

The researchers discovered a direct link between the shape of this mathematical path and how confident the AI is about the next word:

The Straight Path (Low Curvature = High Confidence): When the AI is processing a predictable sentence (like "The cat sat on the..."), its internal mathematical path is relatively straight. Because the path is smooth, the AI can easily "extrapolate" (predict) where the path is going. This leads to low entropy, meaning the AI is very sure the next word is "mat."
The Bending Path (High Curvature = Low Confidence): When the sentence becomes complex, surprising, or ambiguous, the AI's internal path starts to bend and twist sharply. These sharp turns are what the researchers call "curvature." When the path bends, the AI can't easily guess the next direction, leading to high entropy—the AI becomes "uncertain" and might struggle to choose between several different words.

How They Proved It (The "Steering Wheel" Test)

To make sure this wasn't just a coincidence, the researchers performed a "perturbation" experiment. Think of this like a driving instructor nudging the steering wheel of a car:

The Targeted Nudge: If they nudged the AI's "steering" specifically along its current path to make it bend more sharply, the AI immediately became more confused (uncertainty went up).
The Random Nudge: If they nudged the AI in a completely random, irrelevant direction, the AI’s confidence didn't change at all.

This proved that the geometry of the path isn't just a side effect; it is actually the "engine" that drives how the AI makes decisions.

Can We "Train" the AI to be Smoother?

Finally, the researchers tried to see if they could teach the AI to be a "smoother skater." They added a new rule to the AI's training process: "Try to keep your mathematical paths as straight as possible."

They found that this worked! The "untangled" models (the ones trained to be smooth) were more confident and had lower uncertainty, and—crucially—they didn't get any dumber. They were able to keep their high accuracy while becoming more "decisive" in their internal logic.

Why Does This Matter?

Right now, we know LLMs are incredibly smart, but we don't always know why they are certain about some things and hallucinate (make things up) about others.

This research suggests that uncertainty is written into the very shape of the AI's thoughts. By understanding the "curves" in an AI's mind, we might eventually be able to build models that are more reliable, more predictable, and better at knowing when they are actually confused.

Technical Summary: Representational Curvature Modulates Behavioral Uncertainty in Large Language Models

Problem Statement

The "temporal straightening" hypothesis suggests that autoregressive Large Language Models (LLMs) transform input sequences into geometrically predictable (straighter) trajectories across layers to facilitate next-token prediction via linear extrapolation. While previous research has established that these trajectories do indeed become straighter in middle layers, a critical gap remains: is this geometric transformation functionally relevant to the model's behavior? Specifically, does the "bendiness" (curvature) of a representational trajectory directly relate to the model's predictive uncertainty (entropy)?

Methodology

The researchers investigated this link using two open-source models (GPT-2 XL and Pythia-2.8B) and a custom-trained GPT-2 Small model. Their approach involved three distinct experimental pillars:

Correlation Analysis: They defined contextual curvature ( $C_p^k$ ) as a geometric measure of how sharply the representational trajectory (the path of hidden states in the residual stream) bends over a local window of tokens. They then used 10-fold cross-validated OLS regression to determine if this curvature could predict next-token entropy ( $H$ ), a measure of behavioral uncertainty.
Targeted Perturbation Experiments: To establish causality, they applied additive perturbations ( $\delta$ $δ$ ) to the residual stream at intermediate layers. They designed a "ladder" of five perturbation types to test geometric alignment:
- Trajectory-agnostic: Full-space, Random-subspace, and Activation-subspace (top principal components of data).
- Trajectory-aligned: Trajectory-subspace (top PCs of the sample's own displacement vectors) and Planar-subspace (the plane spanned by the two most recent displacement vectors).
Curvature Regularization (Training): They trained a GPT-2 Small model from scratch using an auxiliary loss function. They compared three conditions: Baseline (standard cross-entropy), Untangling (penalizing curvature to encourage straightness), and Tangling (rewarding curvature to encourage bends).

Key Contributions

Functional Linkage: The paper provides the first direct evidence linking the geometric properties of internal representations (curvature) to the statistical properties of model outputs (entropy).
Geometric Specificity: It demonstrates that the model's uncertainty is not just sensitive to any change in activation space, but specifically to changes made along the temporal trajectory of the representation.
Training Dynamics Insight: It shows that the coupling between geometry and behavior is not static but emerges as a learned feature during the pretraining process.

Results

Predictive Power: Contextual curvature is positively correlated with next-token entropy. This correlation is strongest in the middle layers of the network, which are also the layers where trajectories are the straightest.
Emergence during Training: Using Pythia-2.8B checkpoints, the authors showed that as training progresses, curvature decreases and its ability to predict entropy increases in tandem. This suggests the model "learns" to use geometry to manage uncertainty.
Selective Sensitivity: Perturbation experiments revealed that only trajectory-aligned interventions (Trajectory-subspace and Planar-subspace) reliably modulated entropy. Random or data-aligned (but trajectory-agnostic) perturbations had no significant effect, proving that the model's uncertainty is tied to the path of the representation.
Regularization Effects: Training with an "untangling" objective (lower curvature) successfully reduced token-level entropy without increasing validation loss. Conversely, "tangling" increased entropy.

Significance

This work advances the field of mechanistic interpretability by moving beyond static "snapshots" of activations toward a dynamic, temporal understanding of how LLMs process information. It suggests that:

Temporal straightening is a functional solution to the next-token prediction objective, not just a byproduct.
Behavioral control of LLMs (such as modulating confidence or uncertainty) can be achieved by intervening in the temporal geometry of the residual stream.
New training paradigms could potentially use geometric regularization to improve model calibration or reliability without sacrificing predictive accuracy.