Reference-Based Bias Detection in LLMs via Relative Representations of Hidden States

The paper introduces a reference-based bias audit that operates on hidden-state representations rather than generated text. It compares related model variants, such as a base model and its fine-tuned counterpart, by encoding each sentence as a vector of similarities to a fixed set of anchor sentences, producing relative representations in a shared comparison space. Within that space it measures the Representational Bias Shift ΔB, defined as how target groups shift in their association with positive and negative attributes.¹

The substantive novelty is not bias detection as such but the construction of a common coordinate system across checkpoints whose absolute hidden-state geometry has been reshaped by fine-tuning. By replacing absolute hidden states with anchor-relative similarities, the method converts an incomparable geometric object into a comparable profile, enabling before/after auditing at the representation level. That move also shifts the operational constraint: the audit does not require task-specific evaluation data or a judge model, and the paper reports an audit time of about three minutes with 3–50× less compute than the output-level benchmarks considered.¹

Context

The empirical claims span three model families and three benchmarks: WildGuardMix, DecodingTrust, and ToxiGen. ΔB is reported to correlate with output-level bias change in 15 of 18 settings, reaching |r| = 0.84 (p < 0.001) under full fine-tuning and becoming more model-dependent under parameter-efficient adaptation. Thresholding ΔB detects checkpoints whose bias increased with ROC AUC between 0.65 and 0.99; on WildGuardMix and DecodingTrust it separates such checkpoints better than a SEAT-based baseline for all three families. The authors also report stability under changes to the anchor set, attribute sets, and target templates.¹

Assessment

The evidence base is an abstract, so the numbers are asserted rather than inspectable. The central correlation is not uniform: 15 of 18 settings, with an explicitly weaker regime under parameter-efficient adaptation, means the method is a strong triage signal under full fine-tuning but not a universal detector. The ROC AUC range extends to 0.65, and the superior-to-SEAT comparison covers only two of the three benchmarks. What would change this assessment: per-setting results with confidence intervals, evaluation on held-out model families, a transparent compute accounting, and a demonstration that ΔB flags representation-level bias shifts that output-level auditing misses. Without the latter, the motivating gap—internal shifts invisible in text—remains a premise rather than a validated capability.¹

Caveats

AI hypothesisThe method still assumes that a fixed anchor-sentence set and attribute sets adequately span the relevant representational variation; stability tests reduce but do not eliminate the risk that ΔB is sensitive to anchor construction. It also assumes that the association shift measured in relative space is a valid proxy for normatively relevant bias, an assumption licensed by correlation with output-level metrics rather than by direct construct validation. A natural next test is whether ΔB detects bias increases that never surface in any output benchmark, and whether the method remains calibrated when bias decreases or stays flat.¹