🤖 machine learning

Hallucination as Trajectory Commitment: Causal Evidence for Asymmetric Attractor Dynamics in Transformer Generation

This paper provides causal evidence that hallucination in autoregressive language models arises from asymmetric attractor dynamics, where early trajectory commitment is determined by prompt-encoded regimes and characterized by a pronounced causal asymmetry in which corrupting a correct trajectory requires only a single perturbation while recovering from hallucination demands sustained multi-step intervention.

Original authors: G. Aytug Akarlar

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: G. Aytug Akarlar

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 driving a car down a road that splits into two paths right at the very first mile.

Path A (The Truth): Leads to the correct destination.
Path B (The Hallucination): Leads to a beautiful, convincing-looking fake city that doesn't actually exist.

This paper is about how AI models (like the one used in the study) make that split-second decision to drive down Path B, and why it is incredibly hard to get them to turn back once they've started.

Here is the breakdown of the research using simple analogies:

1. The "Same Start, Different End" Experiment

Usually, when we ask an AI a question, we don't know if it will tell the truth or lie until it finishes speaking. The researchers did something clever: they asked the exact same question to the AI 20 times in a row.

The Result: For about 44% of the questions, the AI gave a mix of answers. Sometimes it told the truth; other times, it confidently invented a fake fact.
The Surprise: The AI didn't "decide" to lie based on the question itself. The question was identical every time. The decision happened purely by chance in the very first word it typed. It's like flipping a coin at the start of the drive; if it lands on heads, you go to the Truth City; if tails, you go to the Fake City.

2. The "One-Way Door" (Asymmetric Attractors)

This is the most important discovery. The researchers found that the "Fake City" (the hallucination) is like a sticky trap or a deep valley.

Getting Stuck (Corruption): It is terrifyingly easy to push a truthful AI into a lie. The researchers found that if they took a "hallucinated" brain state and injected it into a "truthful" AI, the AI immediately fell into the lie 87.5% of the time. It's like pushing a ball over a small hill into a deep pit; once it's in, it stays there.
Getting Out (Correction): Trying to pull the AI out of the lie is incredibly hard. Even if they injected a "truthful" brain state into a lying AI, it only worked 33% of the time.
The Analogy: Imagine the lie is a deep, smooth bowl. If you drop a marble in, it rolls to the bottom and stays there. If you try to push the marble out with your finger (a single correction), it just rolls back down. You have to keep pushing it steadily for a long time to get it to the rim.

3. The "Commitment Window"

The study shows that the AI makes its choice almost instantly.

Step 0: The AI reads the question. It hasn't decided yet.
Step 1: The AI types the first word. Boom. It has committed to a path.
The Consequence: Once that first word is typed, the AI is locked into a "trajectory." If it picked the lie, it will keep building a convincing story based on that lie, even if the facts are wrong.

4. The "Saddle Point" (The Danger Zone)

The researchers looked at the AI's internal "map" before it even started typing. They found that certain types of questions (like questions based on false premises, e.g., "Since the Amazon River flows through Europe...") put the AI right on the edge of a cliff (a "saddle point").

For these specific questions, the AI is perfectly balanced between knowing the truth and accepting the lie. A tiny bit of random noise (like a gust of wind) pushes it one way or the other.
For other questions (like simple math), the AI is firmly planted on the "Truth" side of the map and rarely wanders off.

5. Why "Linear Fixes" Don't Work

Scientists have tried to "steer" AI by simply telling it, "Don't go that way," or by mathematically removing the "lie" signal from its brain.

The Paper's Finding: This doesn't work well. Because the "lie" is a deep, stable valley, simply nudging the AI away from the center of the valley isn't enough. The AI's internal mechanics (the layers of the model) naturally pull it back down into the lie.
The Solution: To fix a hallucination, you can't just give a one-time nudge. You have to apply sustained pressure over several steps to push the AI all the way out of the valley and onto the correct path.

The Big Takeaway

Hallucination isn't just the AI "forgetting" a fact. It's a dynamical trap.

The AI actually knows the truth (because it can produce the right answer if the coin flip goes the other way), but once it accidentally steps into the "Hallucination Valley," it gets stuck there. Getting it out requires a lot of effort and sustained intervention, whereas getting it stuck only takes a tiny, accidental push.

In short: It's easy to fall into a lie, but it's very hard to climb out of one.

1. Problem Statement

Large Language Models (LLMs) frequently generate plausible but factually incorrect text (hallucinations) with high confidence. While existing research has established correlations between internal model states (e.g., hidden layer activations) and hallucinations, the causal mechanisms governing when and how a model commits to a hallucinated trajectory remain poorly understood. Specifically, it is unclear:

At what point in the generation process does the model irrevocably commit to a hallucination?
Is this commitment reversible via causal intervention?
What is the underlying dynamical structure of the model's state space that leads to these divergent outcomes?

2. Methodology

The authors propose a novel experimental framework combining same-prompt bifurcation and symmetric causal activation patching to isolate trajectory dynamics from prompt-level semantics.

A. Same-Prompt Bifurcation

Instead of comparing different prompts that yield different results, the authors repeatedly sample the same prompt under non-zero temperature ( $\tau=0.7$ ).

Target: Prompts where the model produces both factual (Correct) and hallucinated (Hallucination) completions from identical initial inputs.
Goal: To demonstrate that divergence is driven by the stochastic sampling path rather than the prompt encoding itself.
Dataset: 61 prompts across 6 categories (Factual, False Premise, Confabulation, Leading, Multi-hop, Math) tested on Qwen2.5-1.5B (28 layers).

B. Causal Activation Patching

For prompts exhibiting bifurcation, the authors perform bidirectional activation patching using TransformerLens:

Correction (H $\to$ C): Injecting a "correct" activation from a successful run into a "hallucinated" run to see if it recovers the correct trajectory.
Corruption (C $\to$ H): Injecting a "hallucinated" activation into a "correct" run to see if it derails the trajectory.
Controls: Random-prompt patching, wrong-to-wrong patching, and unpatched baselines to ensure effects are prompt-specific and direction-specific.
Sweeps: The authors vary layers (0–27), generation steps (0–4), and patch windows (single step vs. sustained multi-step).

C. Probing and Clustering

Linear Probes: Trained on step-0 residual states to predict the per-prompt hallucination rate.
Unsupervised Clustering: Applied to step-0 states to identify "regime-like" groups that correlate with hallucination risk.

3. Key Contributions

Causal Evidence of Asymmetric Attractor Dynamics: The paper provides the first causal proof that hallucination acts as a locally stable attractor basin. Entry into this basin is probabilistic and easy, while exit is difficult and requires sustained intervention.
Same-Prompt Bifurcation Analysis: Demonstrates that identical initial states (step 0) can diverge discontinuously at the very first generated token (step 1), isolating the "commitment" moment.
Quantification of Correction vs. Corruption Asymmetry: Shows a stark statistical gap between the ease of inducing hallucinations and the difficulty of correcting them.
Regime Encoding Discovery: Reveals that the risk of bifurcation is encoded in the prompt representation itself (step 0), organizing prompts into distinct "regimes" (e.g., a "saddle" regime for false premises).

4. Key Results

A. Immediate Trajectory Commitment

Step 0: The KL divergence between correct and hallucinated runs is 0.0 (identical logits).
Step 1: Divergence occurs immediately ( $D_{KL} > 1.0$ in most cases).
Observation: 44.3% of prompts (27/61) are bifurcating. "False premise" prompts are almost universally bifurcating (92.9%), indicating the model sits on a decision boundary between accepting or rejecting the embedded falsehood.

B. Causal Asymmetry (The Core Finding)

Activation patching reveals a profound asymmetry in the stability of trajectories:

Corruption (C $\to$ H): Injecting a hallucinated activation into a correct trajectory causes failure in 87.5% of trials (peak at Layer 20). This effect is consistent across steps.
Correction (H $\to$ C): Injecting a correct activation into a hallucinated trajectory recovers the correct output in only 33.3% of trials (peak at Layer 24).
Significance: The corruption rate is 2.63 $\times$ higher than the correction rate. Both significantly exceed the baseline (10.4%) and random controls (12.5%).

C. Window Patching and Escape Difficulty

Single-step intervention: Fails to reliably correct hallucinations.
Sustained intervention: Correction rates scale with the duration of the patch. Patching layers 1–4 increases the correction rate from 12.5% to 33.3%.
Interpretation: Escaping the hallucination basin requires coordinated, multi-step intervention, whereas entering it requires only a single perturbation.

D. Regime Structure at Step 0

Prediction: Step-0 residual states at Layer 15 can predict the per-prompt hallucination rate with Pearson $r = 0.776$ ( $p < 0.001$ ).
Clustering: Unsupervised clustering identifies 5 regimes. Notably, a "Saddle Cluster" contains 12 of the 13 bifurcating "false premise" prompts, suggesting these prompts are encoded in a specific region of the state space near the basin boundary.

5. Significance and Implications

Theoretical Shift:
The paper reframes hallucination from a "failure of knowledge retrieval" to a dynamical systems phenomenon. The model possesses the correct knowledge (proven by its ability to generate correct outputs from the same prompt), but stochastic sampling can commit it to a stable, incorrect attractor basin.

Mechanism of Failure:

Easy Entry: The hallucination basin has a wide capture radius; a single stochastic token or a single-layer perturbation can push the trajectory in.
Hard Exit: Once inside, the basin is locally stable. The model's internal dynamics (skip connections, layer normalization) tend to restore the trajectory to the hallucinated state, making linear steering or single-point interventions ineffective.

Practical Implications:

Limitations of Linear Steering: The study explains why previous attempts to "steer" models away from hallucination using linear vectors often fail; linear projections cannot cross the nonlinear basin boundary.
Intervention Design: Effective safety interventions may need to target the regime selection at the prompt encoding stage (step 0) rather than attempting to patch the trajectory mid-generation. Alternatively, interventions must be sustained over multiple steps to overcome the attractor's restoring forces.

Conclusion:
Hallucination is characterized as an early trajectory commitment governed by asymmetric attractor dynamics. The model's state space is organized such that entering a hallucinated state is probabilistic and rapid, while escaping it demands coordinated, multi-layer, multi-step intervention.