A quantitative account of the failure of majority voting over multiple LLM samples to raise answer accuracy, yet its gain varies erratically: on hard questions it can even backfire.
Abstract
Majority voting over multiple LLM samples is widely used to raise answer accuracy, yet its gain varies erratically: on hard questions it can even backfire. This paper gives a quantitative account of this failure. A pluralistic agreement index Gamma is defined as the expected fraction of the samples of a wrong run that agree with the consensus, normalized by a reference scale d=(1-p)/(C-1), and is decomposed into a mechanical component (what a vote delivers given only a per-case answer preference) and a preference-unexplained residual. The mechanical null is difficulty-matched and leak-free: each case is resimulated at its own accuracy and option preference, estimated from the case's other runs, so no run predicts its own agreement. On GPT-4.1 the decomposition shows benchmark-associated direction (an observational ordering over n=4 cells per benchmark, not a significance claim). On multiple-choice GPQA-Diamond, the per-case answer preference explains 81-93% of the held-out test-run agreement index: the shared-bias-dominates account over-claims here, because a wrong but attractive option the whole cohort latches onto is captured by the per-case preference channel (whether that preference is induced by shared training bias is not identified). On open-domain AIME, the mechanical preference explains only 59-78% (21-29% if shrunk to pure noise), and a preference-unexplained residual of 1.56-2.80 Gamma units survives, which a run-level preference-heterogeneity reference more than absorbs (1.4-2.1). A self-consistency backfire on hard questions is reproduced (binned voting gap down to -0.09, coupled CI [-0.12,-0.07]), and the highest-agreement bin reaches an accuracy of only 0.42-0.83, a 1.2-3.6x lift over base rate: agreement is graded evidence, not certification. No new voting method is proposed; code and evidence are committed and reproducible.
Agreement among repeated samples of a language model is routinely read as evidence about answer reliability, yet wrong answers can agree just as strongly as right ones. This paper asks what information wrong-consensus agreement actually contains, and answers with a quantitative decomposition. A pluralistic agreement index Gamma, normalized by the reference scale d=(1-p)/(C-1), is split into a mechanical component (agreement delivered by a per-case answer preference alone) and a preference-unexplained residual. The mechanical reference is leak-free: each case's preference and accuracy are estimated from its other runs only. On public GPT-4.1 per-run data, coverage phi (the mechanical/empirical ratio) shows a benchmark-associated direction: 0.81-0.93 on multiple-choice GPQA-Diamond against 0.59-0.78 on open-domain AIME, where a residual of 1.54-2.80 Gamma units survives, more than absorbed by a calibrated run-level preference-heterogeneity reference. A controlled replication under one fixed protocol (four runs per question, K=32 votes) on five open-weights checkpoints (Qwen3.5-9B/122B, Qwen3.8-27B, Gemma4-26B/31B) finds near-complete mechanical coverage in all ten cells (phi approximately 1, with a small overshoot consistent with a quantified finite-donor plug-in bias), robust to a two-run design; the largest cell (qwen3.5-122b, p=0.222) sits inside the GPT-4.1 AIME accuracy range and still saturates (phi=1.041). A cross-system contrast at comparable aggregate accuracy contrasts near-complete mechanical agreement in the open-weights models against a larger preference-unexplained residual in the frontier family. This contrast is confounded with sampling protocol by design. Agreement is graded evidence, not certification. No new voting method is proposed; code and evidence are committed.
Lizhuo Zhang, Mengmeng Tang, Chenfeng Long et al.· 1 citation
Emergent misalignment (EM) --- the broad misbehaviour a language model acquires after fine-tuning on narrow harmful data --- is mediated in Qwen2.5 models by a latent persona direction, and that direction is causal in open weights. Transplanting it into a model that shares only pretraining with its source induces broad EM (2.83 $\pm$ 0.26\% misaligned against a random-direction floor of $\sim$1.1\%), and ablating a model's own direction roughly halves an overt inducer's broadcast (21\% to 10\%). The transplant doubles as a measurement method, causally assaying directions that a source model represents but cannot itself express. Whether a fine-tune recruits this persona depends on method and capacity, and since low-rank PEFT is the cheaper regime at scale, the recruiting method is also the economical one. On Qwen2.5-32B, LoRA at low ranks on insecure code recruits it (3.4\% misaligned) while full SFT on identical data does not (0.3\%) and moves against the persona axis (drift--persona cosine $+0.17$ at rank 1 to $-0.10$), the far-inducer, high-capacity exception consistent with a representational-distance $\times$ capacity account. The persona's causal role is itself conditional. Steering a bad-medical SFT run away from the direction during training raises the broadcast from ${\sim}24\%$ to ${\sim}50\%$ while matched random controls stay at or below baseline, replicated across three training seeds, so removing the direction is no blanket recipe. Because recruitment is a loss-reducing shortcut that capacity renders redundant, it can be screened for and prevented in the tested instances. Persona loss-relevance at the SFT solution orders four inducers'broadcasts rank-perfectly within Qwen2.5, inoculation removes recruitment selectively (4.75\% to 0.0\%, code coherence 65\% to 87\%), and fine-tuning orthogonal to the single behaviour-derived axis reduces it persona-specifically.
Benevolence bias is identified and measure, a small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions, and is easy to diagnose and straightforward to fix.
Yuanzi Li, Jun-Hao Wang, Minghui Liu et al.· 0 citations
In this paper we present a 30, 000-variant India-context bias audit to compare LLaMA3.2-1B and LLaMA3.1-8B across categories of Gender, Religion, Profession and Region that proposes the India Context Sensitivity Index (ICSI) as a category-weighted fairness metric. The larger model shows an improvement of the aggregate bias score that is statistically significant (Mann-Whitney p < 0.001) a biased-response rate that is significantly lower (χ² = 51.92, p < 0.001, 94.04% unbiased responses) at approximately double the inference latency. The breakdown of results by categories also shows that this overall gain is unevenly split: the 8B remains more sensitive on Gender-based prompts even as it improves on Religion, Profession and Region, underscoring the merit of reporting disaggregated, category-imbued fairness over a single number bias score for India deployed LLMs [9], [17], [23]. The future work will allow the scaling of the benchmark to be done to additional model sizes within the LLaMA family (for instance 3B, 70B) to instead fit a bias-versus-scale curve rather than a two-point comparison (14). Then, the extension of the category set to ‘caste-adjacent’ and intersectional categories (for instance gender × region) that are under-represented in this effort (18), (9). The next direction will be replacing the lexicon-based bias scorer with the LLM-as-judge scorer validated against human annotation, to measure scorer-induced bias in the evaluation pipeline itself (28). Finally, deploying the mitigation variants (few-shot, prompt-engineering) evaluated here as live runtime guardrails and measuring their effect on the ICSI in a closed deployment loop (29), (30).
Neha Neha, Amandeep Noliya· INTERNATIONAL JOURNAL OF SCI...· 0 citations
A new method, called CW-Net, translates the reasoning process of an autonomous vehicle’s AI system into understandable concepts that explain its behavior.
