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

Reward models (RMs) play a central role in aligning large language models (LLMs) with human preferences. However, RMs are often sensitive to spurious features such as response length. Existing inference-time approaches for mitigating these biases typically focus exclusively on response length, resulting in performance trade-offs. In this paper, we propose causally motivated intervention for mitigating multiple types of biases in RMs at inference time. Our method first identifies neurons whose activations are strongly correlated with predefined bias attributes, and applies neuron-level intervention that suppresses these signals. We evaluate our method on RM benchmarks and observe reductions in sensitivity to spurious features across diverse bias types, without inducing performance trade-offs. Moreover, when used for preference annotation, small RMs (2B and 7B) with our method, which edits less than 2% of all the neurons in RMs, enable LLMs to improve alignment, achieving performance comparable to that of a state-of-the-art 70B RM on AlpacaEval and MT-Bench. Further analysis reveals that bias signals are primarily encoded by neurons in early layers, shedding light on the internal mechanisms of bias exploitation in RMs.

Large language models (LLMs) are increasingly used as automatic judges for summarization and dialogue evaluation. Prior work has documented biases such as position, verbosity, and style preferences, but largely focuses on outcomes, leaving judge explanations underexplored. We instead ask whether LLM judges are cue-invariant, i.e., whether their rankings and explanations remain stable when non-evidential cues are perturbed while holding the underlying texts fixed. We introduce a suite of cue interventions (Blind, Truth, Flip, Placebo, Reveal-After) and tie-aware metrics that quantify outcome anchoring and rationale anchoring, including label-aligned rhetoric and explanation drift, alongside consistency and stereotype-intrusion checks. We design anchoring attacks using verbosity and confidence cues, and compare two mitigations: structured chain-of-thought prompting and PROOF-BEFORE-PREFERENCE (evidence lock, score, rank). Using a new dataset of 1,000 summaries from traditional extractive models and LLMs, we find substantial cue-anchored rationalization under label and placebo perturbations, while PROOF-BEFORE-PREFERENCE markedly improves cue invariance over baselines.

Bias audits of large language models now operate within governance frameworks such as the EU AI Act, making benchmark reliability a security concern in its own right. Many current benchmarks, however, collapse bias into a single scalar from one prompt format and one surface label. This design misses two failure modes that can be exploited without changing model weights. Across prompts, meaning-preserving format changes shift bias endorsement by more than $0.7$ on a fixed statement pool. Within a response, the discrete Selection and free-text Elaboration can take opposing stances, so an apparently clean aggregate may hide substantial internal inconsistency (a ``cancellation trap''). Selection-only and elaboration-only rankings are therefore nearly uncorrelated across eight LLMs (Spearman $ρ= 0.238$, $p = 0.570$): LLaMA3-70B ranks in the middle under selection-only scoring but highest under elaboration-only scoring on the same responses. We introduce \textsc{BiAxisAudit}, a protocol that reports each bias score together with a reliability estimate on two orthogonal axes. The across-prompt axis evaluates each statement under a factorial grid of task format, perspective, role, and sentiment, treating bias as a distribution rather than a point estimate. The within-response axis uses Split Coding to recover Selection and Elaboration as separate signals, measured by the Inconsistency Rate and Divergence Net Imbalance. Across eight LLMs with $80{,}200$ coded responses each, task format alone explains as much variance as model choice; $63.6\%$ of pooled bias signals (up to $85.2\%$ per model) appear in only one coding layer, and prompt-dimension interactions exceed main effects. The instrument also separates real bias reductions from apparent reductions caused by cross-layer redistribution: some prompt configurations reduce both BER and IR, whereas others suppress only selection-layer bias.