💬 NLP

Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States

This paper demonstrates that while linear probes on LLM hidden states achieve high accuracy in distinguishing deductive, inductive, and abductive reasoning tasks, this separation is entirely driven by task format confounds rather than distinct computational representations of the reasoning modes themselves.

Original authors: Subramanyam Sahoo, Vinija Jain, Aman Chadha, Divya Chaudhary

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Subramanyam Sahoo, Vinija Jain, Aman Chadha, Divya Chaudhary

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

The Big Question: Do AI Brains Have Different "Modes"?

Imagine you are trying to figure out how a chef cooks. You notice that when they make a salad, they wear a green apron, and when they grill a steak, they wear a red apron. You might guess, "Ah! The green apron means they are using a 'salad strategy,' and the red apron means they are using a 'steak strategy.'"

This is exactly what researchers have been doing with Large Language Models (LLMs). They look at the AI's "hidden states" (its internal brain activity) and use a simple test called a linear probe to see if the AI switches between different "reasoning modes" (like Deductive, Inductive, and Abductive) just like a chef switches aprons.

For a long time, the data looked perfect. The "green apron" (deductive tasks) and "red apron" (inductive tasks) were in completely different parts of the AI's brain. Researchers thought, "Great! The AI has built distinct internal circuits for different types of thinking."

This paper says: "Wait a minute. You aren't seeing the thinking; you're seeing the uniform."

The Investigation: What's Actually Happening?

The researchers took a closer look at the AI (specifically a model called Qwen3-14B) using three different types of logic puzzles:

Deductive: Logic puzzles (like a math proof).
Inductive: Science questions (finding patterns).
Abductive: Mystery solving (guessing the best explanation).

They found that the AI's brain activity did look completely different for each type. But they suspected a trick.

The "Uniform" Confusion

Think of it like this:

The Deductive puzzles came from a website where every question has 4 possible answers and a long story.
The Inductive puzzles came from a different website where every question also has 4 possible answers but uses science words.
The Abductive puzzles came from a third website where every question only has 2 possible answers and is very short.

The researchers realized the AI wasn't necessarily switching its thinking style. Instead, it was just reacting to the format of the question. It was like the chef wearing the green apron not because they are making a salad, but because the kitchen where the salad is made is painted green.

The Three Tests (The "Detective Work")

To prove their theory, the researchers ran three specific tests:

1. The "Strip the Uniform" Test (Residual Analysis)
They took the AI's brain activity and mathematically "washed out" the format details (like how many answers there were, how long the text was, and which website the question came from).

Result: Once they removed the "uniform" (the format), the "green apron" and "red apron" disappeared. The brain activity for all three tasks looked exactly the same. The perfect separation was gone, dropping to random chance.
Analogy: If you take away the green and red aprons, the chef looks exactly the same whether they are making salad or steak.

2. The "Behavior Check" (Trace-Mode Agreement)
They looked at the actual words the AI wrote while solving the problems. Did the AI actually act differently when solving a logic puzzle versus a mystery?

Result: No. The AI solved all three types of problems correctly (86% accuracy), but the way it talked through the solution was almost identical for all of them. It didn't switch strategies; it used one "super-strategy" for everything.
Analogy: The chef uses the exact same knife skills and chopping motion whether they are cutting a tomato or a carrot. They aren't using a "tomato technique" vs. a "carrot technique."

3. The "Nudge" Test (Causal Steering)
Researchers tried to "push" the AI's brain in a specific direction to force it to act like it was in "Deductive Mode" or "Inductive Mode." They compared this to pushing it in a random direction.

Result: Pushing it in the "Deductive" direction didn't work any better than pushing it randomly. The geometry of the brain didn't actually control the thinking style.
Analogy: Trying to force the chef to switch aprons by pushing them didn't change what they cooked. The apron color was just a coincidence, not the cause of the cooking style.

The Conclusion

The paper concludes that high accuracy in these tests does not prove the AI has different internal reasoning circuits.

When researchers use different datasets for different types of reasoning (which is standard practice), the AI learns to recognize the format of the dataset (the "uniform") rather than the logic of the task. The "distinct geometry" researchers see in the AI's brain is just a reflection of the question format, not a reflection of a special thinking mode.

The Takeaway:
Just because an AI's brain looks different for different tasks doesn't mean it's thinking differently. It might just be wearing a different uniform. To truly understand how AI reasons, we need to strip away the format and see if the thinking actually changes. In this case, the AI seems to use one powerful, uniform strategy to solve all kinds of logic problems, rather than switching between specialized modes.

` blocks are stripped). The pipeline consists of five stages:

Layer-wise Linear Probing: A logistic regression probe is trained at each layer to predict the reasoning mode from hidden states. Manifold geometry (intrinsic dimensionality, curvature, hull contamination) is analyzed at the peak-performing layer.
Format Confound Decomposition: A four-stage residual analysis pipeline is introduced:
- Source Prediction: Probing for dataset source identity to check equivalence with mode prediction.
- Option Count Probe: Testing if the scalar number of options (2 vs. 4) alone separates modes.
- Format-Controlled Comparison: Restricting analysis to 4-choice tasks only (LogiQA + ARC).
- Residual Analysis: Using Ridge regression to predict hidden states from a format feature vector (source one-hot, option count, response length) and removing this linear information. The probe is then re-evaluated on the residuals.
Trace-Mode Agreement: The model's generated reasoning traces are embedded and compared against hand-crafted anchor descriptions for each mode. Agreement rates are measured to determine if the model's observable behavior adapts to the task type.
Causal Steering with Random Controls: Activation steering is applied using vectors derived from the centroid differences between modes. Crucially, this is compared against random-direction controls ( $N=20$ ) of equal magnitude to test if the geometric structure has a functional causal link to reasoning or if any perturbation of that magnitude produces similar effects.

Key Contributions

Format Confound Decomposition: The paper introduces a residual analysis pipeline demonstrating that when format features (source identity, option count, response length) are regressed out, probe accuracy collapses to chance levels.
Trace-Mode Agreement Analysis: The authors show that while the model achieves high overall accuracy (86%), its reasoning traces exhibit only 42.5% agreement with intended modes (vs. 33.3% chance), suggesting a lack of strategy adaptation.
Causal Controls with Random Baselines: The study implements random-direction controls for steering experiments, a method often absent in prior work, to establish that observed geometric shifts are not causally linked to reasoning mode selection.
Methodological Recommendations: The paper proposes that format deconfounding and random-direction controls should become standard practice in probing-based interpretability.

Results

1. Apparent Perfect Separation

At layer 32 (80% of network depth), linear probes achieve 100% cross-validated accuracy in distinguishing the three reasoning modes. The manifold geometry at this layer shows distinct clusters with varying intrinsic dimensionalities (20.6 for deductive, 28.5 for inductive, 33.6 for abductive) and minimal convex hull contamination ( $\le 1.5\%$ ).

2. Separation is Entirely Format-Driven

The residual analysis overturns the initial finding:

Source Equivalence: A probe predicting dataset source achieves 100% accuracy, identical to mode prediction.
Option Count: A probe using only the number of options achieves 33.3% accuracy (the prior for the 2-choice class), confirming structural features are exploitable.
Residual Collapse: After removing linear format information, the probe accuracy for both reasoning mode and dataset source drops to $\approx 33.5\%$ , indistinguishable from chance. The "reasoning-mode geometry" is entirely an artifact of task format.

3. Uniform Reasoning Strategy

The model achieves 86% accuracy across all task types but exhibits only 42.5% trace-mode agreement. This indicates that the model does not substantially alter its reasoning strategy when switching between deductive, inductive, and abductive tasks; it likely employs a general-purpose mechanism.

4. No Functional Link in Geometry

Causal steering experiments reveal that steering along the centroid-difference direction does not produce significantly different outcomes compared to random perturbations of equal magnitude. The empirical p-value is 0.286, and conflict injection causes coherence collapse for both targeted and random pairs. This confirms the geometric separation has no functional causal role in selecting reasoning modes.

Significance and Claims

The paper argues that high linear probe accuracy is not sufficient evidence of internal computational structure when reasoning labels are confounded with dataset sources. The "distinct representations" observed in previous literature are likely artifacts of the standard multi-source evaluation design rather than evidence of specialized reasoning circuits.

The authors claim that the model reasons well across all three types but appears to do so using a largely uniform strategy rather than distinct computational modes. Consequently, the paper calls for a shift in mechanistic interpretability practices:

Researchers must report source-prediction accuracy alongside mode-prediction accuracy.
Residual analysis to remove format features should be a standard control.
Steering experiments must include random-direction controls to verify directionality.
Future benchmarks should enforce identical surface formats across reasoning types to isolate genuine reasoning differences.

The study concludes that while LLMs possess reasoning capabilities, the assumption that they utilize distinct internal geometries for different reasoning modes is unsupported when format confounds are properly controlled.