Learning from Negative Examples: Why Warning-Framed Training Data Teaches What It Warns Against

• URL: http://arxiv.org/abs/2512.22293v1
• Date: Thu, 25 Dec 2025 20:07:57 GMT
• Title: Learning from Negative Examples: Why Warning-Framed Training Data Teaches What It Warns Against
• Abstract summary: In experiments reported here, models exposed to such warnings reproduced the flagged content at rates statistically indistinguishable from models given the content directly.<n>A related phenomenon, what I call "stealth slip", allows conversational preambles to rotate activations into subspaces that linear probes miss entirely.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Warning-framed content in training data (e.g., "DO NOT USE - this code is vulnerable") does not, it turns out, teach language models to avoid the warned-against behavior. In experiments reported here, models exposed to such warnings reproduced the flagged content at rates statistically indistinguishable from models given the content directly (76.7% vs. 83.3%). Why? Sparse autoencoder analysis points to a failure of orthogonalization: "describing X" and "performing X" activate overlapping latent features. Feature #8684, which tracks code execution patterns, fires at comparable magnitude in both warning and exploitation contexts. A related phenomenon, what I call "stealth slip", allows conversational preambles to rotate activations into subspaces that linear probes miss entirely. Prompting and inference-time steering do not fix this; training-time feature ablation does. The upshot is that statistical co-occurrence dominates over pragmatic interpretation in current architectures. Models learn what tends to follow a context, not why it appeared there.

Related papers

• Pando: Do Interpretability Methods Work When Models Won't Explain Themselves? [53.07826484214082]
  We introduce Pando, a model-organism benchmark that breaks this confound via an explanation axis.<n>Pando predicts held-out model decisions from $10$ labeled query-response pairs.<n>When explanations are faithful, black-box elicitation matches or exceeds all white-box methods.
  arXiv  Detail & Related papers  (2026-04-13T06:42:24Z)
• You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases [1.2369994923959178]
  Subliminal learning refers to the transmission of traits from a teacher to a student model via training on data unrelated to those traits.<n>We investigate whether transmission occurs through natural language paraphrases with fixed semantic content.<n>We find that training on paraphrases from a teacher system-prompted to love a particular animal increases a student's preference for that animal by up to 19 percentage points.
  arXiv  Detail & Related papers  (2026-03-10T11:21:14Z)
• Causally Robust Reward Learning from Reason-Augmented Preference Feedback [4.170114609601759]
  We introduce ReCouPLe, a lightweight framework that uses natural language rationales to provide the missing causal signal.<n>Our learned reward model can ground preferences on the articulated reason, aligning better with user intent and generalizing beyond spurious features.
  arXiv  Detail & Related papers  (2026-03-05T06:35:21Z)
• Alignment faking in large language models [41.40199382334199]
  We show a large language model engaging in alignment faking to prevent modification of its behavior out of training.<n>We find the model complies with harmful queries from free users 14% of the time, versus almost never for paid users.<n>We also study the effect of actually training the model to comply with harmful queries via reinforcement learning, which we find increases the rate of alignment-faking reasoning to 78%.
  arXiv  Detail & Related papers  (2024-12-18T17:41:24Z)
• Understanding Self-Supervised Learning via Gaussian Mixture Models [19.51336063093898]
  We analyze self-supervised learning in a natural context: dimensionality reduction in Gaussian Mixture Models.<n>We show that vanilla contrastive learning is able to find the optimal lower-dimensional subspace even when the Gaussians are not isotropic.<n>In this setting we show that contrastive learning learns the subset of fisher-optimal subspace, effectively filtering out all the noise from the learnt representations.
  arXiv  Detail & Related papers  (2024-11-05T21:43:05Z)
• Context-Parametric Inversion: Why Instruction Finetuning Can Worsen Context Reliance [68.56701216210617]
  In-principle, one would expect models to adapt to the user context better after instruction finetuning.<n>We observe a surprising failure mode: during instruction tuning, the context reliance under knowledge conflicts initially increases as expected, but then gradually decreases.
  arXiv  Detail & Related papers  (2024-10-14T17:57:09Z)
• XAL: EXplainable Active Learning Makes Classifiers Better Low-resource Learners [71.8257151788923]
  We propose a novel Explainable Active Learning framework (XAL) for low-resource text classification.<n>XAL encourages classifiers to justify their inferences and delve into unlabeled data for which they cannot provide reasonable explanations.<n>Experiments on six datasets show that XAL achieves consistent improvement over 9 strong baselines.
  arXiv  Detail & Related papers  (2023-10-09T08:07:04Z)
• Understanding and Mitigating Spurious Correlations in Text Classification with Neighborhood Analysis [69.07674653828565]
  Machine learning models have a tendency to leverage spurious correlations that exist in the training set but may not hold true in general circumstances. In this paper, we examine the implications of spurious correlations through a novel perspective called neighborhood analysis. We propose a family of regularization methods, NFL (doN't Forget your Language) to mitigate spurious correlations in text classification.
  arXiv  Detail & Related papers  (2023-05-23T03:55:50Z)
• Generic Temporal Reasoning with Differential Analysis and Explanation [61.96034987217583]
  We introduce a novel task named TODAY that bridges the gap with temporal differential analysis. TODAY evaluates whether systems can correctly understand the effect of incremental changes. We show that TODAY's supervision style and explanation annotations can be used in joint learning.
  arXiv  Detail & Related papers  (2022-12-20T17:40:03Z)
• Discovering Latent Knowledge in Language Models Without Supervision [72.95136739040676]
  Existing techniques for training language models can be misaligned with the truth. We propose directly finding latent knowledge inside the internal activations of a language model in a purely unsupervised way. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models.
  arXiv  Detail & Related papers  (2022-12-07T18:17:56Z)
• Indiscriminate Poisoning Attacks Are Shortcuts [77.38947817228656]
  We find that the perturbations of advanced poisoning attacks are almost textbflinear separable when assigned with the target labels of the corresponding samples. We show that such synthetic perturbations are as powerful as the deliberately crafted attacks. Our finding suggests that the emphshortcut learning problem is more serious than previously believed.
  arXiv  Detail & Related papers  (2021-11-01T12:44:26Z)
• Classification and Adversarial examples in an Overparameterized Linear Model: A Signal Processing Perspective [10.515544361834241]
  State-of-the-art deep learning classifiers are highly susceptible to infinitesmal adversarial perturbations. We find that the learned model is susceptible to adversaries in an intermediate regime where classification generalizes but regression does not. Despite the adversarial susceptibility, we find that classification with these features can be easier than the more commonly studied "independent feature" models.
  arXiv  Detail & Related papers  (2021-09-27T17:35:42Z)
• Does Continual Learning = Catastrophic Forgetting? [21.77693101142049]
  We present a set of tasks that surprisingly do not suffer from catastrophic forgetting when learned continually. We also introduce a novel yet simple algorithm, YASS, that outperforms state-of-the-art methods in the class-incremental categorization learning task. Finally, we present DyRT, a novel tool for tracking the dynamics of representation learning in continual models.
  arXiv  Detail & Related papers  (2021-01-18T19:29:12Z)
• Hidden Footprints: Learning Contextual Walkability from 3D Human Trails [70.01257397390361]
  Current datasets only tell you where people are, not where they could be. We first augment the set of valid, labeled walkable regions by propagating person observations between images, utilizing 3D information to create what we call hidden footprints. We devise a training strategy designed for such sparse labels, combining a class-balanced classification loss with a contextual adversarial loss.
  arXiv  Detail & Related papers  (2020-08-19T23:19:08Z)

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.