Understanding Emergent Misalignment via Feature Superposition Geometry

• URL: http://arxiv.org/abs/2605.00842v1
• Date: Tue, 07 Apr 2026 06:10:21 GMT
• Title: Understanding Emergent Misalignment via Feature Superposition Geometry
• Abstract summary: We show that fine-tuning on narrow, non-harmful tasks induces harmful behaviors.<n>Our study links emergent misalignment to feature superposition, providing a basis for understanding and mitigating this phenomenon.
• Score: 48.56470855119906
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a geometric account based on the geometry of feature superposition. Because features are encoded in overlapping representations, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this effect and empirically test it in multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, GPT-OSS 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach, filtering training samples closest to toxic features, reduces misalignment by 34.5%, substantially outperforming random removal and achieving comparable or slightly lower misalignment than LLM-as-a-judge-based filtering. Our study links emergent misalignment to feature superposition, providing a basis for understanding and mitigating this phenomenon.

Related papers

• The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment [53.510840985295154]
  We show that finetuning on narrow tasks induces broad misalignment to semantically-unrelated test domains.<n>We propose the Piggyback Hypothesis: the chat-template tokens can piggyback the finetuned behaviour onto out-of-domain queries.<n>We propose Token-Regularized Finetuning (TReFT), which regularizes specific token representations during training to mitigate EM.
  arXiv  Detail & Related papers  (2026-06-04T19:32:00Z)
• Intrinsic Guardrails: How Semantic Geometry of Personality Interacts with Emergent Misalignment in LLMs [9.485206628921857]
  We show that harmful fine-tuning does not overwrite a model's internal representation of personality.<n>Our findings suggest that harmful fine-tuning does not overwrite a model's internal representation of personality.
  arXiv  Detail & Related papers  (2026-05-11T14:21:57Z)
• From Early Encoding to Late Suppression: Interpreting LLMs on Character Counting Tasks [49.57538588967748]
  Large language models (LLMs) exhibit failures on elementary symbolic tasks such as character counting in a word, despite excelling on complex benchmarks.<n>We uncover a consistent phenomenon across modern architectures, including LLaMA, Qwen, and Gemma.<n>We show that symbolic reasoning failures in LLMs are not due to missing representations or insufficient scale, but arise from structured interference within the model's computation graph.
  arXiv  Detail & Related papers  (2026-04-01T11:40:12Z)
• MirrorLA: Reflecting Feature Map for Vision Linear Attention [49.41670925034762]
  Linear attention significantly reduces the computational complexity of Transformers from quadratic to linear, yet it consistently lags behind softmax-based attention in performance.<n>We propose MirrorLA, a geometric framework that substitutes passive truncation with active reorientation.<n>MirrorLA achieves state-of-the-art performance across standard benchmarks, demonstrating that strictly linear efficiency can be achieved without compromising representational fidelity.
  arXiv  Detail & Related papers  (2026-02-04T09:14:09Z)
• Where Did It Go Wrong? Attributing Undesirable LLM Behaviors via Representation Gradient Tracing [12.835224376066769]
  Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their deployment is frequently undermined by undesirable behaviors.<n>We introduce a novel and efficient framework that diagnoses a range of undesirable LLM behaviors by analyzing representation and its gradients.<n>We systematically evaluate our method for tasks that include tracking harmful content, detecting backdoor poisoning, and identifying knowledge contamination.
  arXiv  Detail & Related papers  (2025-09-26T12:07:47Z)
• Re-Emergent Misalignment: How Narrow Fine-Tuning Erodes Safety Alignment in LLMs [0.0]
  We show that fine tuning on insecure code induces internal changes that oppose alignment.<n>We identify a shared latent dimension in the model's activation space that governs alignment behavior.
  arXiv  Detail & Related papers  (2025-07-04T15:36:58Z)
• Surgical Feature-Space Decomposition of LLMs: Why, When and How? [8.826164604720738]
  We empirically study the efficacy of weight and feature space decomposition in transformer-based language models. We show that surgical decomposition provides critical insights into the trade-off between compression and language modelling performance. We extend our investigation to the implications of low-rank approximations on model bias.
  arXiv  Detail & Related papers  (2024-05-17T07:34:03Z)
• Understanding Collapse in Non-Contrastive Learning [122.2499276246997]
  We show that SimSiam representations undergo partial dimensional collapse if the model is too small relative to the dataset size. We propose a metric to measure the degree of this collapse and show that it can be used to forecast the downstream task performance without any fine-tuning or labels.
  arXiv  Detail & Related papers  (2022-09-29T17:59:55Z)
• Discriminator-Free Generative Adversarial Attack [87.71852388383242]
  Agenerative-based adversarial attacks can get rid of this limitation. ASymmetric Saliency-based Auto-Encoder (SSAE) generates the perturbations. The adversarial examples generated by SSAE not only make thewidely-used models collapse, but also achieves good visual quality.
  arXiv  Detail & Related papers  (2021-07-20T01:55:21Z)
• Unbiased Risk Estimators Can Mislead: A Case Study of Learning with Complementary Labels [92.98756432746482]
  We study a weakly supervised problem called learning with complementary labels. We show that the quality of gradient estimation matters more in risk minimization. We propose a novel surrogate complementary loss(SCL) framework that trades zero bias with reduced variance.
  arXiv  Detail & Related papers  (2020-07-05T04:19:37Z)
• AMC-Loss: Angular Margin Contrastive Loss for Improved Explainability in Image Classification [8.756814963313804]
  Angular Margin Contrastive Loss (AMC-Loss) is a new loss function to be used along with the traditional cross-entropy loss. AMC-Loss employs the discriminative angular distance metric that is equivalent to geodesic distance on a hypersphere manifold. We find that although the proposed geometrically constrained loss-function improves quantitative results modestly, it has a qualitatively surprisingly beneficial effect on increasing the interpretability of deep-net decisions.
  arXiv  Detail & Related papers  (2020-04-21T08:03:14Z)

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.