Characterizing Model-Native Skills

• URL: http://arxiv.org/abs/2604.17614v1
• Date: Sun, 19 Apr 2026 20:58:25 GMT
• Title: Characterizing Model-Native Skills
• Abstract summary: Skills are a natural unit for describing what a language model can do and how its behavior can be changed.<n>Existing characterizations rely on human-written, textual descriptions, or manual profiling pipelines.<n>We argue that when the goal is to intervene on model behavior, skill characterization should be *model-native*: grounded in the model's own representations.
• Score: 16.891026204025838
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Skills are a natural unit for describing what a language model can do and how its behavior can be changed. However, existing characterizations rely on human-written taxonomies, textual descriptions, or manual profiling pipelines--all external hypotheses about what matters that need not align with the model's internal representations. We argue that when the goal is to intervene on model behavior, skill characterization should be *model-native*: grounded in the model's own representations rather than imposed through external ontologies. We instantiate this view by recovering a compact orthogonal basis from sequence-level activations. The resulting basis is semantically interpretable but need not correspond to any predefined human ontology; instead, it captures axes of behavioral variation that the model itself organizes around. We validate this characterization on reasoning post-training, using the recovered basis for both SFT data selection and inference-time steering. We develop lightweight proxy interventions to identify which directions are most useful for a given model. Across Llama3-8B and Qwen2.5-3B, selecting data along those directions improves Pass@1 by up to 20% on MATH and 41% on AMC, outperforming data selection based on human-characterized skills. Because the basis lives in activation space, the same directions also serve as steering vectors at inference time, improving Pass@8 by up to 4.8% on MATH--an intervention that human-characterized skills cannot support. We further validate the characterization on safety alignment, where selecting adversarial training data for model-native skill coverage rather than textual diversity yields more sample-efficient learning. These results suggest that recovering skills from the model's own representations, rather than imposing them externally, provides a more effective foundation for intervening on model behavior. Codes are open-sourced.

Related papers

• Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal [29.196846051811576]
  We introduce a data-centric post-training pipeline that uses interpretability protocols to develop statistical hypotheses for the latent concepts separating preferred from dispreferred generations.<n>We show that our pipeline diagnoses undesirable signals in existing preference data, mitigates off-target learning, and can also help amplify or shape desired properties such as safeguards and model personality.
  arXiv  Detail & Related papers  (2026-06-10T17:31:16Z)
• Data Attribution in Large Language Models via Bidirectional Gradient Optimization [35.61634772862795]
  Large Language Models (LLMs) are increasingly deployed across diverse applications.<n>understanding which training data most influenced a model's output remains a fundamental open problem.<n>We address this challenge by expanding upon the inverse formulation: How would training data be affected if the model had seen the generated output during training?<n>Our method perturbs the base model using bidirectional gradient optimization (gradient ascent and descent) on a generated text sample and measures the resulting change in loss across training samples.
  arXiv  Detail & Related papers  (2026-06-03T14:21:53Z)
• From Data to Behavior: Predicting Unintended Model Behaviors Before Training [78.37660873165284]
  We introduce Data2Behavior, a new task for predicting unintended model behaviors prior to training.<n>We also propose Manipulating Data Features (MDF), a lightweight approach that summarizes candidate data through their mean representations.<n>Experiments on Qwen3-14B, Qwen2.5-32B-Instruct, and Gemma-3-12b-it confirm that MDF can anticipate unintended behaviors and provide insight into pre-training vulnerabilities.
  arXiv  Detail & Related papers  (2026-02-04T16:37:17Z)
• Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors [61.92704516732144]
  We show that the most robust features for correctness prediction are those that play a distinctive causal role in the model's behavior.<n>We propose two methods that leverage causal mechanisms to predict the correctness of model outputs.
  arXiv  Detail & Related papers  (2025-05-17T00:31:39Z)
• Entropy-Based Adaptive Weighting for Self-Training [15.089334734753677]
  We propose Entropy-Based Adaptive Weighting for Self-Training (EAST)<n>EAST is an adaptive weighting strategy designed to prioritize uncertain data during self-training.<n>We evaluate our approach on GSM8K and MATH benchmarks.
  arXiv  Detail & Related papers  (2025-03-31T10:04:35Z)
• Machine Unlearning on Pre-trained Models by Residual Feature Alignment Using LoRA [15.542668474378633]
  We propose a novel and efficient machine unlearning method on pre-trained models. We leverage LoRA to decompose the model's intermediate features into pre-trained features and residual features. The method aims to learn the zero residuals on the retained set and shifted residuals on the unlearning set.
  arXiv  Detail & Related papers  (2024-11-13T08:56:35Z)
• Secrets of RLHF in Large Language Models Part II: Reward Modeling [134.97964938009588]
  We introduce a series of novel methods to mitigate the influence of incorrect and ambiguous preferences in the dataset. We also introduce contrastive learning to enhance the ability of reward models to distinguish between chosen and rejected responses.
  arXiv  Detail & Related papers  (2024-01-11T17:56:59Z)
• AlignDiff: Aligning Diverse Human Preferences via Behavior-Customisable Diffusion Model [69.12623428463573]
  AlignDiff is a novel framework to quantify human preferences, covering abstractness, and guide diffusion planning. It can accurately match user-customized behaviors and efficiently switch from one to another. We demonstrate its superior performance on preference matching, switching, and covering compared to other baselines.
  arXiv  Detail & Related papers  (2023-10-03T13:53:08Z)
• Explain, Edit, and Understand: Rethinking User Study Design for Evaluating Model Explanations [97.91630330328815]
  We conduct a crowdsourcing study, where participants interact with deception detection models that have been trained to distinguish between genuine and fake hotel reviews. We observe that for a linear bag-of-words model, participants with access to the feature coefficients during training are able to cause a larger reduction in model confidence in the testing phase when compared to the no-explanation control.
  arXiv  Detail & Related papers  (2021-12-17T18:29:56Z)
• On the Transferability of Pre-trained Language Models: A Study from Artificial Datasets [74.11825654535895]
  Pre-training language models (LMs) on large-scale unlabeled text data makes the model much easier to achieve exceptional downstream performance. We study what specific traits in the pre-training data, other than the semantics, make a pre-trained LM superior to their counterparts trained from scratch on downstream tasks.
  arXiv  Detail & Related papers  (2021-09-08T10:39:57Z)
• Adversarial Infidelity Learning for Model Interpretation [43.37354056251584]
  We propose a Model-agnostic Effective Efficient Direct (MEED) IFS framework for model interpretation. Our framework mitigates concerns about sanity, shortcuts, model identifiability, and information transmission. Our AIL mechanism can help learn the desired conditional distribution between selected features and targets.
  arXiv  Detail & Related papers  (2020-06-09T16:27:17Z)
• Fairness-Aware Learning with Prejudice Free Representations [2.398608007786179]
  We propose a novel algorithm that can effectively identify and treat latent discriminating features. The approach helps to collect discrimination-free features that would improve the model performance.
  arXiv  Detail & Related papers  (2020-02-26T10:06:31Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.