Perturbation: A simple and efficient adversarial tracer for representation learning in language models

• URL: http://arxiv.org/abs/2603.23821v1
• Date: Wed, 25 Mar 2026 01:09:51 GMT
• Title: Perturbation: A simple and efficient adversarial tracer for representation learning in language models
• Abstract summary: We show that perturbation reveals structured transfer at multiple linguistic grain sizes.<n>In trained LMs, perturbation reveals structured transfer at multiple linguistic grain sizes.
• Score: 2.0769172070951063
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Linguistic representation learning in deep neural language models (LMs) has been studied for decades, for both practical and theoretical reasons. However, finding representations in LMs remains an unsolved problem, in part due to a dilemma between enforcing implausible constraints on representations (e.g., linearity; Arora et al. 2024) and trivializing the notion of representation altogether (Sutter et al., 2025). Here we escape this dilemma by reconceptualizing representations not as patterns of activation but as conduits for learning. Our approach is simple: we perturb an LM by fine-tuning it on a single adversarial example and measure how this perturbation ``infects'' other examples. Perturbation makes no geometric assumptions, and unlike other methods, it does not find representations where it should not (e.g., in untrained LMs). But in trained LMs, perturbation reveals structured transfer at multiple linguistic grain sizes, suggesting that LMs both generalize along representational lines and acquire linguistic abstractions from experience alone.

Related papers

• Interpreting Brain Responses to Language with Sparse Features from Language Models [5.261667457257713]
  We develop an encoding framework that replaces dense LM hidden states with hierarchically-organized autoencoder features.<n>We show that brain responses during language processing are not merely predictable from an arbitrary set of LM features.
  arXiv  Detail & Related papers  (2026-06-05T03:01:28Z)
• Emergence of Linear Truth Encodings in Language Models [64.86571541830598]
  Large language models exhibit linear subspaces that separate true from false statements, yet the mechanism behind their emergence is unclear.<n>We introduce a transparent, one-layer transformer toy model that reproduces such truth subspaces end-to-end.<n>We study one simple setting in which truth encoding can emerge, encouraging the model to learn this distinction in order to lower the LM loss on future tokens.
  arXiv  Detail & Related papers  (2025-10-17T16:30:07Z)
• Analogical Reasoning Inside Large Language Models: Concept Vectors and the Limits of Abstraction [0.0]
  We show that function vectors (FVs) are not invariant to simple input changes, suggesting they capture more than pure concepts.<n>We localize a small set of attention heads that encode invariant concept vectors (CVs) for verbal concepts like "antonym"<n>CVs function as feature detectors that operate independently of the final output - meaning that a model may form a correct internal representation yet still produce an incorrect output.
  arXiv  Detail & Related papers  (2025-03-05T16:59:08Z)
• An Incomplete Loop: Deductive, Inductive, and Abductive Learning in Large Language Models [99.31449616860291]
  Modern language models (LMs) can learn to perform new tasks in different ways. In instruction following, the target task is described explicitly in natural language; in few-shot prompting, the task is specified implicitly. In instruction inference, LMs are presented with in-context examples and are then prompted to generate a natural language task description.
  arXiv  Detail & Related papers  (2024-04-03T19:31:56Z)
• Strong hallucinations from negation and how to fix them [2.1178416840822027]
  We show that our approach improves model performance in cloze prompting and natural language inference tasks with negation without requiring training on sparse negative data. We call such responses textitstrong hallucinations and prove that they follow from an LM's computation of its internal representations for logical operators and outputs from those representations.
  arXiv  Detail & Related papers  (2024-02-16T10:11:20Z)
• Pixel Sentence Representation Learning [67.4775296225521]
  In this work, we conceptualize the learning of sentence-level textual semantics as a visual representation learning process. We employ visually-grounded text perturbation methods like typos and word order shuffling, resonating with human cognitive patterns, and enabling perturbation to be perceived as continuous. Our approach is further bolstered by large-scale unsupervised topical alignment training and natural language inference supervision.
  arXiv  Detail & Related papers  (2024-02-13T02:46:45Z)
• Are Representations Built from the Ground Up? An Empirical Examination of Local Composition in Language Models [91.3755431537592]
  Representing compositional and non-compositional phrases is critical for language understanding. We first formulate a problem of predicting the LM-internal representations of longer phrases given those of their constituents. While we would expect the predictive accuracy to correlate with human judgments of semantic compositionality, we find this is largely not the case.
  arXiv  Detail & Related papers  (2022-10-07T14:21:30Z)
• MAML and ANIL Provably Learn Representations [60.17417686153103]
  We prove that two well-known meta-learning methods, MAML and ANIL, are capable of learning common representation among a set of given tasks. Specifically, in the well-known multi-task linear representation learning setting, they are able to recover the ground-truth representation at an exponentially fast rate. Our analysis illuminates that the driving force causing MAML and ANIL to recover the underlying representation is that they adapt the final layer of their model.
  arXiv  Detail & Related papers  (2022-02-07T19:43:02Z)
• Provable Limitations of Acquiring Meaning from Ungrounded Form: What will Future Language Models Understand? [87.20342701232869]
  We investigate the abilities of ungrounded systems to acquire meaning. We study whether assertions enable a system to emulate representations preserving semantic relations like equivalence. We find that assertions enable semantic emulation if all expressions in the language are referentially transparent. However, if the language uses non-transparent patterns like variable binding, we show that emulation can become an uncomputable problem.
  arXiv  Detail & Related papers  (2021-04-22T01:00:17Z)
• Discourse structure interacts with reference but not syntax in neural language models [17.995905582226463]
  We study the ability of language models (LMs) to learn interactions between different linguistic representations. We find that, contrary to humans, implicit causality only influences LM behavior for reference, not syntax. Our results suggest that LM behavior can contradict not only learned representations of discourse but also syntactic agreement.
  arXiv  Detail & Related papers  (2020-10-10T03:14:00Z)

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.