How Does Alignment Tuning Shape Representations of Sycophancy and Related Cue-Induced Biases in LLMs?

• URL: http://arxiv.org/abs/2607.18114v1
• Date: Mon, 20 Jul 2026 16:10:06 GMT
• Title: How Does Alignment Tuning Shape Representations of Sycophancy and Related Cue-Induced Biases in LLMs?
• Abstract summary: We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model.<n>We extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out transfer, and causal intervention.
• Score: 53.58941416780391
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer. We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model. Across five model families and seven BCT bias types, we extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out transfer, and causal intervention. The susceptibility is largely installed by alignment tuning rather than pretraining: pretrained base models barely cave to these biases, and their activations carry no cue-specific signal beyond question content. Within aligned models, each bias becomes a single coherent direction that we can both decode and steer along, recovering the unbiased answer across every family we test. The biases stay representationally distinct, however: cross-bias entanglement is model-specific rather than a property of the bias category, and even behaviorally similar biases occupy different directions. The same intervention also serves as a modest debiasing tool, recovering a meaningful share of bias-induced errors while preserving most correct answers across all instruct families. Cue-induced bias is therefore best understood not as a single flaw in LLMs but as a family of distinct, causally active directions that alignment tuning installs.

Related papers

• CogBias: Measuring and Mitigating Cognitive Bias in Large Language Models [8.81733256907103]
  Large Language Models (LLMs) are increasingly deployed in high-stakes decision-making contexts.<n>We show that cognitive biases emerge systematically across all four families of cognitive biases: Judgment, Information Processing, Social, and Response.<n>We apply activation steering to modulate biased behavior, achieving 26--32% reduction in bias score.
  arXiv  Detail & Related papers  (2026-04-01T20:22:14Z)
• 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)
• Making Bias Non-Predictive: Training Robust LLM Judges via Reinforcement Learning [91.8584139564909]
  Large language models (LLMs) increasingly serve as automated judges, yet they remain susceptible to cognitive biases.<n>We propose Epistemic Independence Training (EIT), a reinforcement learning framework grounded in a key principle.<n>EIT operationalizes this through a balanced conflict strategy where bias signals are equally likely to support correct and incorrect answers.
  arXiv  Detail & Related papers  (2026-02-02T01:43:48Z)
• Multi Attribute Bias Mitigation via Representation Learning [5.155573439009767]
  Real world images frequently exhibit multiple overlapping biases, including textures, watermarks, gendered makeup, scene object pairings, etc.<n>We tackle this multi bias problem with Generalized Multi Bias Mitigation (GMBM), a lean two stage framework that needs group labels only while training and minimizes bias at test time.<n>We validate GMBM on FB CMNIST, CelebA, and COCO, where we boost worst group accuracy, halve multi attribute bias amplification, and set a new low in SBA.
  arXiv  Detail & Related papers  (2025-09-03T18:08:59Z)
• Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs [51.00909549291524]
  Large language models (LLMs) exhibit cognitive biases.<n>These biases vary across models and can be amplified by instruction tuning.<n>It remains unclear if these differences in biases stem from pretraining, finetuning, or even random noise.
  arXiv  Detail & Related papers  (2025-07-09T18:01:14Z)
• CosFairNet:A Parameter-Space based Approach for Bias Free Learning [1.9116784879310025]
  Deep neural networks trained on biased data often inadvertently learn unintended inference rules. We introduce a novel approach to address bias directly in the model's parameter space, preventing its propagation across layers. We show enhanced classification accuracy and debiasing effectiveness across various synthetic and real-world datasets.
  arXiv  Detail & Related papers  (2024-10-19T13:06:40Z)
• Mitigating Bias for Question Answering Models by Tracking Bias Influence [84.66462028537475]
  We propose BMBI, an approach to mitigate the bias of multiple-choice QA models. Based on the intuition that a model would lean to be more biased if it learns from a biased example, we measure the bias level of a query instance. We show that our method could be applied to multiple QA formulations across multiple bias categories.
  arXiv  Detail & Related papers  (2023-10-13T00:49:09Z)
• Feature-Level Debiased Natural Language Understanding [86.8751772146264]
  Existing natural language understanding (NLU) models often rely on dataset biases to achieve high performance on specific datasets. We propose debiasing contrastive learning (DCT) to mitigate biased latent features and neglect the dynamic nature of bias. DCT outperforms state-of-the-art baselines on out-of-distribution datasets while maintaining in-distribution performance.
  arXiv  Detail & Related papers  (2022-12-11T06:16:14Z)
• Learning Debiased Models with Dynamic Gradient Alignment and Bias-conflicting Sample Mining [39.00256193731365]
  Deep neural networks notoriously suffer from dataset biases which are detrimental to model robustness, generalization and fairness. We propose a two-stage debiasing scheme to combat against the intractable unknown biases.
  arXiv  Detail & Related papers  (2021-11-25T14:50:10Z)
• Evading the Simplicity Bias: Training a Diverse Set of Models Discovers Solutions with Superior OOD Generalization [93.8373619657239]
  Neural networks trained with SGD were recently shown to rely preferentially on linearly-predictive features. This simplicity bias can explain their lack of robustness out of distribution (OOD) We demonstrate that the simplicity bias can be mitigated and OOD generalization improved.
  arXiv  Detail & Related papers  (2021-05-12T12:12:24Z)

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.