Large language models are increasingly used for ordinal classification, yet semantically equivalent changes to prompt organization can alter their predictions. We conduct systematic experiments to characterize positional bias from label order, demonstration order, and demonstration placement. First, we apply the three probes to ten frontier LLMs on a common ordinal-classification task; every model is sensitive to all three positional sources, showing that the problem is pervasive. Second, we vary eight prompt-, task-, and model-level factors across five datasets; accuracy and stability are often misaligned, and only lower scale cardinality consistently improves both. Third, we compare pointwise, pairwise, and listwise inference, alternative aggregation and debiasing methods, and joint configurations; the tested corrections do not provide a reliable remedy, while a comparison-based listwise formulation offers the best balance but transfers unevenly across models and bias sources. These findings show that positional robustness depends on the full system configuration rather than the model alone. Ordinal-classification systems should therefore be selected jointly for predictive performance and stability.
Position bias in multiple-choice LLM evaluation is widely cited as a confound in capability comparisons, but published measurements rely on single answer-order shuffles whose results confound the bias signal with content-level noise and sampling stochasticity. I introduce inspect_permute, an open-source extension to the inspect_ai evaluation framework that runs exhaustive answer-order permutations per question and reports the chi-squared / Cramer V signature of position bias with bootstrap confidence intervals. I apply the tool across four vendors (gpt-4o-mini, claude-haiku-4-5, gemini-2.5-flash, grok-3) on five MMLU subjects, 24,000 API calls under temperature-0 generation, with falsifier predictions pre-registered via a public SHA-256 hash before half the data was observed. Position bias turns out to be statistically detectable only within a roughly 60-95% base-accuracy Goldilocks zone. Below it, processing-load dominance swamps subject-specific signal; above it, ceiling effects compress the variance below the chi-squared test resolution. Detectable cells separate into two mechanism types: monotone A-to-D decrease (processing_load, in low-tier models) and non-monotone D-drop (content_ambiguity, in a narrow capability band). Standard MMLU places every frontier-tier model above the detection band, so absence of signal there should be read as not measurable, not unbiased. Together with the ceiling-effect characterisation in arXiv:2606.26185, this work brackets the detectable region of position-bias measurement and makes the field central question askable in a verifiable form. Package, data, preregistration under MIT.
Forced-choice probes with counterbalanced orientations are a standard tool for measuring language-model "value dispositions," and a concentration/extremity index over repeated draws is read as how sharply a model commits. We show this estimator is not identifiable at its low end: counterbalancing, meant to remove position bias, instead maps a position-lock (a model returning the same answer letter regardless of content) onto the same near-0.5 signature as genuine neutrality, so "softness" and "non-engagement" cannot be distinguished from a content-independent letter-bias. Across nine models the fraction of position-locked items tracks the index almost perfectly (r=-0.986) - a structural consequence, not a finding: the informative quantity is the residual from that bound, the commitment a model shows on the items it does engage. The three models the index reads as soft are the most position-locked, and the lock resolves under access configurations that permit reasoning (as-deployed CLI -> raw API -> reasoning-enabled), moving the index 0.06->0.63 and 0.34->0.66->0.64 while lock collapses (Opus: 0.61->0.39->0.22); this establishes the low readings as an identifiability failure, not a disposition. The reasoning-enabled reading is not a "true" value either; the point is that the index alone cannot identify commitment at its low end. Access configuration (deployment client, reasoning on/off) is one generator of this failure, shown on two access paths (an Anthropic subscription CLI and a DeepSeek client), and in a frozen benchmark it is confounded with provider. We contribute a position-lock diagnostic that must accompany any concentration reading, and show that a concentration-blind audit risks reporting neutrality where a reasoning-permitting condition yields concentrated choices. The direction-flip component, largely robust to the artifact, still identifies genuine cross-model disagreement.
Qiong Tang, Xiangkun Hu, Xiangyang Liu +2cs.CL cs.AI
Long-context language models suffer from position bias, where information in middle positions is underutilized. Attention Sorting addresses this by iteratively reordering documents based on attention patterns, but its multiple sort-and-generate cycles increase deployment cost. We hypothesize that position bias is the primary bottleneck and propose Debiased One-Pass Attention Sorting, which estimates a per-prompt position-bias curve from the low-attention majority of documents and uses it to correct raw attention scores (via subtraction or division) to enable single-pass sorting. Our experiments on two models refute this hypothesis in the tested setting: on LLaMA-2-7B-32K-Instruct, debiasing produces identical results to uncalibrated single-pass sorting (94.83\% containment accuracy), while on YaRN-Llama-2-7b-64k, debiasing improves accuracy by 8.67 percentage points but remains 14.84pp behind iterative sorting, closing only 37\% of the gap. These results suggest that position-bias correction is insufficient to match iterative sorting, and that repeated reordering provides additional benefits beyond bias correction.
Large Language Models (LLMs) still struggle with the ``lost-in-the-middle'' problem, where critical information located in the middle of long-context inputs is often underrepresented or lost. While existing methods attempt to address this by combining multi-scale rotary position embeddings (RoPE), they typically suffer from high latency or rely on suboptimal hand-crafted scaling strategies. To overcome these limitations, we introduce a layer-specific positional embedding scaling~(LPES) method that assigns distinct scaling factors to each layer. LPES achieves a more balanced attention distribution without fine-tuning model parameters or increasing inference delay. A specially designed genetic algorithm is employed to efficiently select the optimal scaling factors for each layer by incorporating Bézier curves to significantly reduce the search space. Extensive experiments demonstrate that LPES effectively mitigates positional attention bias and delivers consistent improvements across multiple long-context benchmarks, yielding up to an $11.2$\% accuracy gain on the key-value retrieval dataset.