💬 NLP

Characterizing Model-Native Skills

This paper proposes and validates a "model-native" approach to characterizing language model skills by extracting an interpretable orthogonal basis directly from sequence-level activations, demonstrating that this internal representation enables significantly more effective data selection and inference-time steering for reasoning and safety tasks compared to traditional human-defined taxonomies.

Original authors: Feiyang Kang, Mahavir Dabas, Myeongseob Ko, Ruoxi Jia

Published 2026-04-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Feiyang Kang, Mahavir Dabas, Myeongseob Ko, Ruoxi Jia

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 Idea: Stop Guessing, Start Listening

Imagine you are trying to teach a very smart, but slightly mysterious, robot how to solve math problems or behave safely.

The Old Way (Human-Defined Skills):
Currently, when we want to improve these robots, humans act like strict librarians. We create a massive, pre-written catalog of "skills." We say, "Okay, the robot needs to learn Calculus, Geometry, and Logic." We then go through thousands of examples and manually tag them with these labels.

The Problem: This is like trying to teach a dog to speak French by using a human dictionary. The dog doesn't think in French words; it thinks in smells, sounds, and feelings. Similarly, the robot doesn't organize its brain the way we do. It might group "Calculus" and "Logic" together in a way we never expected, or it might treat "Geometry" and "Storytelling" as the same thing. By forcing our human labels on the robot, we miss what's actually happening inside its brain.

The New Way (Model-Native Skills / AUTOSKILL):
This paper proposes a different approach: Let the robot tell us how it thinks.
Instead of imposing our own labels, the researchers developed a method called AUTOSKILL. They look directly at the robot's internal "electricity" (its neural activations) while it solves problems. They use a mathematical tool (PCA) to find the natural "axes" or directions the robot uses to organize information.

Think of it like this:

Human Approach: We look at a pile of rocks and say, "These are 'Red Rocks' and those are 'Smooth Rocks'."
AUTOSKILL Approach: We put the rocks in a special scanner that reveals they are actually organized by "Weight" and "Temperature." The scanner shows us that the robot naturally groups things by weight and temperature, not by color or smoothness. We then use those natural groups to teach the robot.

How It Works: The "Musical Instrument" Analogy

Imagine the robot's brain is a giant, complex musical instrument with thousands of strings.

The Old Way: We try to play a song by guessing which strings to pluck based on a sheet music written by a human composer. We often hit the wrong notes because the instrument was built differently than the sheet music assumed.
The New Way (AUTOSKILL): We listen to the instrument playing itself. We realize that certain strings vibrate together naturally to create a specific "sound" (a skill). We map out these natural vibrations. Now, instead of guessing, we know exactly which strings to pluck to get the sound we want.

The Two Big Wins

The researchers tested this on two very different tasks: Math Reasoning and Safety.

1. Teaching Math (The "Tutor" Analogy)

They wanted to teach the robot to get better at math competitions.

The Experiment: They took a huge pile of math problems.
- Group A (Human Method): They picked problems labeled "Calculus" or "Algebra" by humans.
- Group B (Model-Native Method): They used AUTOSKILL to find the problems that matched the robot's own internal "math directions."
The Result: The robot trained on the Model-Native data got significantly better at solving hard math problems (up to 41% better on some tests).
Why? The human labels were too rigid. The robot had its own unique way of seeing math. By feeding it data that matched its own internal map, the learning was much more efficient. It's like giving a student a textbook written in their native dialect rather than a foreign language they are struggling to translate.

2. Teaching Safety (The "Security Guard" Analogy)

They also wanted to teach the robot not to do bad things (like "jailbreaking" or ignoring safety rules).

The Problem: There are thousands of ways to trick a robot into being unsafe. Humans try to categorize these tricks (e.g., "Roleplay attacks," "Code attacks"). But the robot might see two very different "attacks" as the exact same threat in its brain.
The Experiment: They had a limited budget of "bad examples" to train the robot on.
- Human Method: Pick examples that look different on the surface (different words, different formats).
- Model-Native Method: Pick examples that trigger different internal reactions in the robot, even if they look similar on the surface.
The Result: The Model-Native method made the robot much safer with fewer examples.
Why? It avoided redundancy. If the robot thinks two different "jailbreak" prompts are the same thing, training on both is a waste of time. AUTOSKILL found the unique "threat directions" the robot actually cares about, making the training super efficient.

The "Steering Wheel" Bonus

Here is the coolest part: Because they found these skills inside the robot's brain, they can use them as steering wheels.

The Old Way: To make a robot be "honest," you have to retrain it from scratch or hope it learned it during training.
The New Way: Since they know the exact "direction" in the robot's brain that corresponds to a skill (like "Symbolic Math" or "Refusal"), they can just nudge the robot's brain in that direction while it is talking.
Analogy: Imagine driving a car. The old way is to rebuild the engine to go faster. The new way is to just turn the steering wheel slightly toward the "Fast Lane" while you are already driving. The robot can instantly become better at math or safer just by a tiny nudge in its internal "voltage."

Summary

The paper argues:
Stop trying to force human categories onto AI. AI has its own internal logic and structure. If we want to teach AI, fix its behavior, or make it safer, we should look at how the AI actually organizes information and use that as our guide.

The Takeaway:

Human Skills: Like trying to teach a fish to climb a tree because "climbing" is a useful skill for mammals.
Model-Native Skills: Like teaching the fish to swim faster by understanding how its fins actually work.
Result: The fish (AI) learns faster, swims better, and doesn't crash into the tree.

This approach leads to smarter, safer, and more efficient AI, without needing humans to guess what's going on inside the "black box."

1. Problem Statement

Current methods for characterizing and intervening in Large Language Model (LLM) behavior rely heavily on external, human-defined taxonomies. These include subject-based categories (e.g., "algebra," "logic"), textual skill descriptions, or manual profiling pipelines.

Limitations: These external hypotheses often fail to align with the model's internal representations. Different models may organize the same data differently, yet external taxonomies assign identical categories regardless of the model's specific architecture or training.
Goal: The authors argue that for effective intervention (such as data selection for fine-tuning or inference-time steering), skill characterization must be model-native: grounded in the model's own internal activation space rather than imposed via external ontologies.

2. Methodology: AUTOSKILL

The paper introduces AUTOSKILL, a framework that recovers a compact, orthogonal basis of "skills" directly from the model's sequence-level activations.

Core Mechanism

Activation Representation:
- For a dataset of text sequences, the model extracts hidden states for every token across all layers.
- These are concatenated to form a unified activation vector per token.
- Pooling Strategy:
  - Reasoning Tasks: Uses mean pooling over token activations to capture long-horizon patterns (e.g., step-by-step deduction).
  - Safety Tasks: Uses the activation of the first generated token, as safety decisions (refusal vs. compliance) are concentrated there.
Basis Recovery (PCA):
- The activation vectors are stacked into a matrix $A$ .
  Principal Component Analysis (PCA) is applied to the centered matrix to recover an orthogonal basis.
- Why PCA? It provides a compact set of directions that capture the dominant variance in the data. Each direction represents an independent axis of behavioral variation inherent to the model.
Interpretation (Post-hoc):
- The authors do not define skills in advance. Instead, they identify the "poles" of each principal component (PC) by finding examples with the highest and lowest cosine similarity to the direction vector.
- These extreme examples are summarized using an LLM (GPT-5) to generate human-interpretable labels (e.g., "Symbolic Calculus" vs. "Conceptual Discrete Reasoning").

Intervention Strategies

Once the basis is recovered, it supports two primary forms of intervention:

SFT Data Selection:
- Lightweight Proxy: To avoid expensive full fine-tuning for every direction, the authors use proxy interventions (pilot fine-tuning on 1k examples or inference-time steering) to identify which directions yield the best performance gains.
- Selection: Once a promising direction is identified, the top-k examples in the candidate pool (based on cosine similarity to that direction) are selected for full Supervised Fine-Tuning (SFT).
Inference-Time Steering:
- The recovered direction vectors serve directly as steering vectors.
- By adding a scaled vector to the model's hidden states (or bias parameters) during generation, the model's output can be steered toward a specific skill pole (e.g., more rigorous proofs or more conceptual reasoning) without retraining.

3. Key Contributions

Theoretical Framework: Formalizes the argument that skill characterization for intervention must be model-native, proposing the search for a compact, representation-grounded skill basis.
AUTOSKILL Implementation: A practical pipeline that uses PCA on sequence-level activations to recover orthogonal skill directions. It unifies data selection, inference-time steering, and safety coverage under a single characterization.
Empirical Validation: Demonstrates that model-native skills differ significantly across models (even on the same data) and that interventions based on these skills outperform human-defined baselines.

4. Experimental Results

Case Study I: Reasoning Post-Training

Setup: Evaluated on Llama3-8B and Qwen2.5-3B using mathematical reasoning datasets (GSM8K, MATH-500, AMC).
Model-Native Divergence: The recovered skill axes differed between models. For instance, while both models had a "Symbolic Calculus" axis, Llama contrasted it with "Conceptual Discrete Reasoning," whereas Qwen contrasted it with "Combinatorial Strategy Proofs."
Performance Gains:
- Data Selection: Selecting training data along the best-performing model-native directions improved Pass@1 by up to 20% on MATH-500 and 41% on AMC compared to random selection.
- Comparison: Outperformed curated datasets (s1.1, LIMO) and human-defined skill profiling (STAT). Human-curated datasets often failed to transfer across models, while AUTOSKILL adapted to each model's internal structure.
- Steering: Inference-time steering using the recovered vectors improved Pass@8 by up to 4.8% on MATH-500, a capability impossible with human-defined skills.

Case Study II: Safety Alignment

Setup: Tested on Mistral-7B under tight adversarial data budgets (50–100 examples).
Finding: Textually diverse jailbreak prompts often overlap significantly in activation space (latent redundancy).
Result: Selecting adversarial training data based on latent coverage (using AUTOSKILL's latent space selection) yielded more sample-efficient robustness than random sampling or selection based on textual diversity.
- At a budget of 50 examples, AUTOSKILL achieved 141 explicit refusals vs. 125 for random sampling on the WildJailbreak benchmark.
- Gains were most pronounced in low-budget regimes where redundancy is most costly.

5. Significance and Implications

Paradigm Shift: Moves the field from "labeling" model behavior with human concepts to "discovering" behavior through the model's own geometry.
Efficiency: Provides a data-efficient method for SFT and safety tuning by identifying the most impactful directions without exhaustive manual curation.
Generalizability: The approach is model-agnostic and can be applied to any LLM with access to activations, offering a unified way to understand and control model behavior across reasoning and safety domains.
Actionable Intervention: Unlike previous interpretability work that is purely descriptive, AUTOSKILL provides a direct mechanism for causal intervention (steering and data selection) that is proven to improve downstream performance.

Conclusion

The paper establishes that recovering skills from a model's internal representations is superior to imposing external taxonomies. By leveraging PCA on activation spaces, AUTOSKILL uncovers compact, orthogonal skill axes that are semantically interpretable, predictive of training outcomes, and directly usable for steering. This approach leads to significant performance improvements in reasoning and more efficient safety alignment, suggesting that the "true" skills of a model are best understood through its own latent geometry.