Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models

• URL: http://arxiv.org/abs/2606.26079v1
• Date: Wed, 24 Jun 2026 17:53:26 GMT
• Title: Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models
• Abstract summary: We introduce Facet-Probe, a five-facet audit of multimodal large language models.<n>We find that none of the 18 MLLMs we audit are order-invariant.<n>We propose cross-ordering flip rate as a standard reporting axis for MLLMs.
• Score: 29.673427534201654
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Standard benchmarks for multimodal large language models (MLLMs) score each item on one canonical ordering and miss whether order-irrelevant shuffling changes the answer, a baseline reliability property called for by emerging AI evaluation guidelines. We introduce Facet-Probe, a five-facet audit (option, evidence-chunk, document-rank, image-set, and mixed-modality ordering) of 18 frontier and open-weight MLLMs. A Bayesian item-response model separates ordering noise from per-facet bias, and a same-ordering control estimates the decoder-stochastic floor for observed flips. We find that none of the 18 MLLMs we audit are order-invariant: screened per-facet panel-mean flip rates span 24-50%. A Gemini same-ordering control at temperature 0 estimates a substantial ordering excess over a same-input decoder-noise floor in verified cells. Capability predicts but does not eliminate flips; the best model still flips on 13.4% of trials. In our Gemini mitigation tests, training-free prompt changes are modality-conditional and do not transfer from text to visual reasoning. These results suggest that prompt-level mitigation alone is unlikely to provide general order robustness, motivating future work on training-time and architectural approaches. We propose cross-ordering flip rate as a standard reporting axis for MLLMs.

Related papers

• Verification-Aware Training for Speculative Decoding [57.84976863792784]
  We introduce Verification-Aware Training (VAT), a plug-in framework that simulates verification at every training step and turns the resulting accept and reject patterns into supervision.<n>VAT modifies only the training objective, so it can be layered on top of existing methods without changing the draft architecture, the target model, or the inference procedure.<n>VAT improves average acceptance length by up to 11.4% and wall-clock speedup by up to 8.7%, with consistent gains across math, code, and chat benchmarks.
  arXiv  Detail & Related papers  (2026-08-31T01:42:10Z)
• Position Bias is Hidden Behind Ceiling Effects: A Permutation Diagnostic for LLM Benchmarks [0.0]
  I introduce inspect_permute, an open-source extension to inspect_ai- evaluation framework.<n>It runs exhaustive answer_order permutations per question.<n>It reports the chi-squared / Cramer V signature of position bias with bootstrap confidence intervals.
  arXiv  Detail & Related papers  (2026-07-23T02:45:02Z)
• AgentFairBench: Do LLM Agents Discriminate When They Act? [2.3004655342211078]
  AgentFairBench is a cheap, reproducible, multi-domain benchmark for demographic disparity in the actions of LLM agents.<n>It spans three regulator-anchored domains: hiring, lending, and medical triage.<n>A NumPy-only harness computes counterfactual flip rate, mean absolute score difference (MASD), action-rate disparity, and tool-invocation disparity.
  arXiv  Detail & Related papers  (2026-06-15T13:50:26Z)
• Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models [59.51249894128724]
  Masked diffusion language models can reduce inference steps by revealing multiple tokens per denoising iteration.<n>We propose ADAS, a training-free reranking rule for parallel masked diffusion decoding.
  arXiv  Detail & Related papers  (2026-06-09T13:17:27Z)
• Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior [16.249731576262167]
  We study whether the same family label predicts behavior in interactive multi-LLM systems.<n>We find that hedging, a reasoning-distilled Llama checkpoint shifts by 18% depending on which same-base partner it replies to.<n>Results identify post-training recipe as a first-class axis for multi-LLM panel composition.
  arXiv  Detail & Related papers  (2026-05-30T07:26:26Z)
• Predicting Decisions of AI Agents from Limited Interaction through Text-Tabular Modeling [20.47884338846419]
  AI agents negotiate and transact in natural language with unfamiliar counterparts.<n>In such interactions, the counterpart's LLM, prompts, control logic, and rule-based fallbacks are hidden.<n>We ask whether an agent can predict an unfamiliar counterpart's next decision from a few interactions.
  arXiv  Detail & Related papers  (2026-05-12T17:09:32Z)
• BiAxisAudit: A Novel Framework to Evaluate LLM Bias Across Prompt Sensitivity and Response-Layer Divergence [22.315546054051143]
  Bias audits of large language models now operate within governance frameworks such as the EU AI Act.<n>We introduce textscBiAxisAudit, a protocol that reports each bias score together with a reliability estimate on two axes.
  arXiv  Detail & Related papers  (2026-05-09T16:26:49Z)
• DPRM: A Plug-in Doob h transform-induced Token-Ordering Module for Diffusion Language Models [76.12556589212666]
  We introduce DPRM (Doob h-transform Process Reward Model), a plug-in token-ordering module for diffusion language models.<n>DPRM starts from confidence-driven progressive ordering and gradually shifts to Doob h transform Process Reward guided ordering through online estimates.<n>Under tractable optimization assumptions, DPRM also yields a sample-complexity advantage over random and confidence-only ordering.
  arXiv  Detail & Related papers  (2026-04-27T11:50:26Z)
• ABCD: All Biases Come Disguised [4.603755953026689]
  Multiple-choice question (MCQ) benchmarks have been a standard evaluation practice.<n>We propose a simple bias-reduced evaluation protocol that replaces the labels of each question with uniform, unordered labels.<n>We show that this protocol substantially improves the robustness to answer permutations, reducing mean accuracy variance $3times$ with only a minimal decrease in the mean model's performance.
  arXiv  Detail & Related papers  (2026-02-19T15:12:33Z)
• CompassVerifier: A Unified and Robust Verifier for LLMs Evaluation and Outcome Reward [50.97588334916863]
  We develop CompassVerifier, an accurate and robust lightweight verifier model for evaluation and outcome reward.<n>It demonstrates multi-domain competency spanning math, knowledge, and diverse reasoning tasks, with the capability to process various answer types.<n>We introduce VerifierBench benchmark comprising model outputs collected from multiple data sources, augmented through manual analysis of metaerror patterns to enhance CompassVerifier.
  arXiv  Detail & Related papers  (2025-08-05T17:55:24Z)
• Cautious Next Token Prediction [62.74127603725369]
  We propose a new training-free decoding strategy, dubbed as Cautious Next Token Prediction (CNTP)<n>In the decoding process, if the model has comparatively high prediction entropy at a certain step, we sample multiple trials starting from the step independently and stop when encountering any punctuation.<n>We show that our proposed CNTP approach outperforms existing standard decoding strategies consistently by a clear margin.
  arXiv  Detail & Related papers  (2025-07-03T05:49:18Z)
• Confidence Diagram of Nonparametric Ranking for Uncertainty Assessment in Large Language Models Evaluation [20.022623972491733]
  Ranking large language models (LLMs) has proven to be an effective tool to improve alignment based on the best-of-$N$ policy.<n>We propose a new inferential framework for hypothesis testing among the ranking for language models.
  arXiv  Detail & Related papers  (2024-12-07T02:34:30Z)
• Are You Sure? Challenging LLMs Leads to Performance Drops in The FlipFlop Experiment [82.60594940370919]
  We propose the FlipFlop experiment to study the multi-turn behavior of Large Language Models (LLMs) We show that models flip their answers on average 46% of the time and that all models see a deterioration of accuracy between their first and final prediction, with an average drop of 17% (the FlipFlop effect) We conduct finetuning experiments on an open-source LLM and find that finetuning on synthetically created data can mitigate - reducing performance deterioration by 60% - but not resolve sycophantic behavior entirely.
  arXiv  Detail & Related papers  (2023-11-14T23:40:22Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.