Cue-induced bias is best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.
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 bias types, we extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out (LODO) transfer, and causal intervention. The susceptibility is largely shaped by alignment tuning rather than pretraining: pretrained base models generally cave much less to these biases, and their activations carry much weaker cue-specific signal beyond question content. Within aligned models, each bias has a coherent linear direction that we can both decode and steer along, recovering the unbiased answer across every family we test. The biases do not collapse into a single shared representation, however: cross-bias overlap is model-specific, and even behaviorally similar biases occupy different directions. The same intervention also provides a proof-of-concept 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 causally effective linear directions that are largely shaped by alignment tuning.
Because the signal spans a contiguous layer band, LayerMix aggregates it to match oracle-layer performance without oracle access, and characterize the geometry within the controlled paired-example paradigm.
Existing bias auditing methods typically rely on model outputs, requiring costly benchmarks or judge models and potentially missing internal shifts that never appear in generated text. We propose a reference-based method that audits bias in hidden-state representations across related model variants, for example before and after fine-tuning. Because fine-tuning reshapes representation geometry, absolute hidden states are not directly comparable, so we encode each sentence by its similarities to a fixed set of anchor sentences, yielding relative representations in a shared comparison space. There we measure how target groups shift in their association with positive and negative attributes, a quantity we call the Representational Bias Shift $\Delta B$. Across three model families and the WildGuardMix, DecodingTrust and ToxiGen benchmarks, $\Delta B$ correlates with output-level bias change in 15 of the 18 settings we test, reaching $|r| = 0.84$ ($p<0.001$) under full fine-tuning and becoming more model-dependent under parameter-efficient adaptation. Thresholding $\Delta B$ detects checkpoints whose bias increased with ROC AUC between $0.65$ and $0.99$, and on WildGuardMix and DecodingTrust it separates them better than a SEAT-based baseline for all three families. $\Delta B$ is also stable under changes to the anchor set, attribute sets and target templates. Our method requires no task-specific evaluation data and audits a model in about three minutes, using $3$-$50\times$ less compute than the output-level benchmarks considered here. We view it as complementary to output-based auditing rather than a replacement for it.
M. Jeliński, Jan Dubinski, Maciej Chrabaszcz et al.· 0 citations
Multimodal Large Language Models (MLLMs) can assign similar confidence to answers that fail for different reasons. We propose HalluPrism, a behavioral diagnostic that re-runs an answer after visual degradation, blank-image replacement, and grounding or relation checks. These targeted probes yield a signature over visual-perturbation sensitivity (V ), image-removal confidence retention (L), and grounding/relation-probe instability (A). Across 58K+ examples from four benchmarks and four MLLMs, image-removal confidence retention is most prevalent, while grounding/relation-probe instability better separates failure families. Only 18 of 48 source-target checks are diagonally aligned, so the coordinates should be interpreted jointly rather than as independent causal sources. With the dataset fixed, the joint signature improves failure-family AUROC from 0.634 to 0.769 on HallusionBench and from 0.707 to 0.817 on VizWiz, with smaller gains on POPE and VSR. In pooled XGBoost analysis, AUROC rises from 0.78 with scalar confidence to 0.95 with (V, L, A) and 0.97 when confidence is added. The same signature does not automatically improve correctness ranking. The three tested direct scalarizations can harm it. These results separate failure diagnosis from abstention scoring: multimodal uncertainty should characterize failure structure before it is used to decide whether to abstain or correct.
Aman Prakash, Sourish Dasgupta, Tanmoy Chakraborty· 0 citations
GGSS---Geodesic-Gated Spherical Steering---a norm-preserving intervention that discovers a counterfactual bias subspace on the unit hypersphere, steers visual tokens along geodesic arcs, and uses an adaptive gate to focus correction on tokens that carry stronger demographic signal.
Yiqun Sun, Junyu Chen, Pengfei Wei et al.· 0 citations
Behavior manifold analysis is introduced, which isolates behavior-specific geometry by selecting sparse behavior-associated coordinates and lifting them into low-dimensional local charts, and provides a unified framework for understanding the mechanistic distinction between the two objectives.
Jun-Tong Wang, Shengkun Yang, Xiyuan Wang et al.· 0 citations
The framework, INFUSE, first stabilizes visual and textual representations around perturbation-averaged and ground-truth anchors, then aligns the stabilized representations across modalities with bidirectional contrastive objectives.
Aditi Sarker, Rafi Ibn Sultan, Hui Zhu et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
