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

Reference-Based Bias Detection in LLMs via Relative Representations of Hidden States. By Marek Jeliński and three others, from NASK National Research Institute, the Warsaw University of Technology, and one other institution. Published on the ninth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Most methods for auditing bias in a language model look at what the model outputs. That needs costly benchmarks or judge models, and it can miss shifts inside the model that never surface in the generated text. The obvious alternative, looking at hidden states directly, runs into a different problem when comparing related versions of a model, say before and after fine-tuning. Fine-tuning reshapes the geometry of the representations, so the raw hidden states of the two versions are not directly comparable.

The fix is to describe each sentence relatively. Instead of its absolute hidden state, a sentence is encoded by how similar it is to a fixed set of anchor sentences. Those similarities form a relative representation, and because the anchors are the same for every model variant, all variants land in one shared space where they can be compared. The coordinates themselves may move when a model is fine-tuned. What is measured instead is where a sentence sits relative to the same fixed landmarks, and that can be read off both versions of the model in the same terms.

In that shared space, the method measures how target groups shift in their association with positive and with negative attributes. The authors call this quantity the Representational Bias Shift, delta B. It is a reference-based audit: one model variant is judged against another, rather than against an external benchmark of outputs. If fine-tuning has pulled a group closer to negative attributes, or further from positive ones, that change shows up as a shift in delta B, whether or not it has yet shown up in anything the model writes.

The tests span three model families and three benchmarks, WildGuardMix, DecodingTrust and ToxiGen. Delta B correlates with the change in output-level bias in fifteen of the eighteen settings tested. Under full fine-tuning the absolute correlation reaches zero point eight four, with p below zero point zero zero one. Under parameter-efficient adaptation the relationship is more dependent on the model.

Used with a threshold, delta B detects checkpoints whose bias went up, with a ROC AUC between zero point six five and zero point nine nine. On WildGuardMix and DecodingTrust it separates those checkpoints better than a SEAT-based baseline, for all three families. It stays stable when the anchor set, the attribute sets or the target templates are changed. It needs no task-specific evaluation data, audits a model in about three minutes, and uses three to fifty times less compute than the output-level benchmarks considered.

The authors are explicit about the role they intend. Delta B is meant to sit alongside output-based auditing, not replace it. What it adds is a quick look inside the model, placed on common ground by a set of anchors, that can flag a shift in bias before anyone has to generate and judge text to find it.