Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning

• URL: http://arxiv.org/abs/2507.16795v2
• Date: Sun, 09 Nov 2025 22:39:01 GMT
• Title: Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning
• Authors: Helena Casademunt, Caden Juang, Adam Karvonen, Samuel Marks, Senthooran Rajamanoharan, Neel Nanda,
• Abstract summary: Fine-tuning large language models can lead to unintended out-of-distribution generalization.<n>We introduce Concept Ablation Fine-Tuning (CAFT) to control how LLMs generalize from fine-tuning.<n>CAFT works by ablating concepts with linear projections during fine-tuning, steering the model away from unintended generalizations.
• Score: 12.179304379042401
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Fine-tuning large language models (LLMs) can lead to unintended out-of-distribution generalization. Standard approaches to this problem rely on modifying training data, for example by adding data that better specify the intended generalization. However, this is not always practical. We introduce Concept Ablation Fine-Tuning (CAFT), a technique that leverages interpretability tools to control how LLMs generalize from fine-tuning, without needing to modify the training data or otherwise use data from the target distribution. Given a set of directions in an LLM's latent space corresponding to undesired concepts, CAFT works by ablating these concepts with linear projections during fine-tuning, steering the model away from unintended generalizations. We successfully apply CAFT to three fine-tuning tasks, including emergent misalignment, a phenomenon where LLMs fine-tuned on a narrow task generalize to give egregiously misaligned responses to general questions. Without any changes to the fine-tuning data, CAFT reduces misaligned responses by 10x without degrading performance on the training distribution. Overall, CAFT represents a novel approach for steering LLM generalization without modifying training data.

Related papers

• Emergent Misalignment is Easy, Narrow Misalignment is Hard [10.936985574307736]
  Finetuning large language models on narrowly harmful datasets can cause them to become emergently misaligned.<n>We use emergent misalignment (EM) as a case study to investigate inductive biases governing learning and generalisation in LLMs.<n>We find a linear representation of the narrow solution also exists, and can be learned by introducing a KL divergence loss.
  arXiv  Detail & Related papers  (2026-02-08T07:50:04Z)
• Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models [57.20761595019967]
  We present Normalized Attention Guidance (NAG), an efficient, training-free mechanism that applies extrapolation in attention space with L1-based normalization and refinement.<n>NAG restores effective negative guidance where CFG collapses while maintaining fidelity.<n>NAG generalizes across architectures (UNet, DiT), sampling regimes (few-step, multi-step), and modalities (image, video)
  arXiv  Detail & Related papers  (2025-05-27T13:30:46Z)
• Generalization Capability for Imitation Learning [1.30536490219656]
  Imitation learning holds the promise of equipping robots with versatile skills by learning from expert demonstrations.<n>However, policies trained on finite datasets often struggle to generalize beyond the training distribution.<n>We present a unified perspective on the generalization capability of imitation learning, grounded in both information theorey and data distribution property.
  arXiv  Detail & Related papers  (2025-04-25T17:59:59Z)
• Aligning Instruction Tuning with Pre-training [81.4748965653345]
  We propose Aligning Instruction Tuning with Pre-training (AITP) to align instruction tuning with pre-training distributions.<n>We show consistent performance improvements with AITP on three fully open large language models (LLMs) across eight benchmarks.
  arXiv  Detail & Related papers  (2025-01-16T08:27:40Z)
• PACE: Marrying generalization in PArameter-efficient fine-tuning with Consistency rEgularization [35.922096876707975]
  PACE is a generalization of PArameter-efficient fine-tuning with Consistency rEgularization.<n>It implicitly regularizes gradients for enhanced generalization, but also implicitly aligns the fine-tuned and pre-trained models to retain knowledge.<n>It also improves LoRA in text classification (GLUE) and mathematical reasoning.
  arXiv  Detail & Related papers  (2024-09-25T17:56:00Z)
• Information Guided Regularization for Fine-tuning Language Models [11.831883526217942]
  We argue that a more surgical approach to regularization needs to exist for smoother transfer learning. We devise a novel approach to dropout for improved model regularization and better downstream generalization.
  arXiv  Detail & Related papers  (2024-06-20T05:18:37Z)
• The Right Time Matters: Data Arrangement Affects Zero-Shot Generalization in Instruction Tuning [86.19804569376333]
  We show that zero-shot generalization happens very early during instruction tuning.<n>We propose a more grounded training data arrangement framework, Test-centric Multi-turn Arrangement.
  arXiv  Detail & Related papers  (2024-06-17T16:40:21Z)
• Disperse-Then-Merge: Pushing the Limits of Instruction Tuning via Alignment Tax Reduction [75.25114727856861]
  Large language models (LLMs) tend to suffer from deterioration at the latter stage ofSupervised fine-tuning process. We introduce a simple disperse-then-merge framework to address the issue. Our framework outperforms various sophisticated methods such as data curation and training regularization on a series of standard knowledge and reasoning benchmarks.
  arXiv  Detail & Related papers  (2024-05-22T08:18:19Z)
• Can LLMs Separate Instructions From Data? And What Do We Even Mean By That? [60.50127555651554]
  Large Language Models (LLMs) show impressive results in numerous practical applications, but they lack essential safety features.<n>This makes them vulnerable to manipulations such as indirect prompt injections and generally unsuitable for safety-critical tasks.<n>We introduce a formal measure for instruction-data separation and an empirical variant that is calculable from a model's outputs.
  arXiv  Detail & Related papers  (2024-03-11T15:48:56Z)
• LEVI: Generalizable Fine-tuning via Layer-wise Ensemble of Different Views [28.081794908107604]
  Fine-tuning is used to leverage the power of pre-trained foundation models in new downstream tasks. Recent studies have observed challenges in the generalization of fine-tuned models to unseen distributions. We propose a novel generalizable fine-tuning method LEVI, where the pre-trained model is adaptively ensembled layer-wise with a small task-specific model.
  arXiv  Detail & Related papers  (2024-02-07T08:16:40Z)
• Mitigating the Alignment Tax of RLHF [76.4300447532456]
  aligning LLMs under Reinforcement Learning with Human Feedback can lead to forgetting pretrained abilities, also known as the alignment tax. We propose model averaging to maximize alignment performance while incurring minimal alignment tax. We validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B.
  arXiv  Detail & Related papers  (2023-09-12T14:16:54Z)
• Stochastic Training is Not Necessary for Generalization [57.04880404584737]
  It is widely believed that the implicit regularization of gradient descent (SGD) is fundamental to the impressive generalization behavior we observe in neural networks. In this work, we demonstrate that non-stochastic full-batch training can achieve strong performance on CIFAR-10 that is on-par with SGD.
  arXiv  Detail & Related papers  (2021-09-29T00:50:00Z)

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.