Mechanistic Diagnostics of Spatial Lexical Bias in Multimodal Large Language Model Spatial Reasoning

• URL: http://arxiv.org/abs/2606.01914v1
• Date: Mon, 01 Jun 2026 08:49:47 GMT
• Title: Mechanistic Diagnostics of Spatial Lexical Bias in Multimodal Large Language Model Spatial Reasoning
• Abstract summary: Adding a spatial relation word to the answer options can attract the model's decision and make the newly added option likely to be selected.<n>We show that models can answer a binary spatial question correctly, yet consistently select an incorrect third spatial option once it is added to the answer set.<n>We show that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias.
• Score: 37.5846828763184
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Multimodal large language models (MLLMs) remain unreliable on spatial multiple-choice questions, and their failures are often attributed to poorly attended visual information. In this work, we identify a complementary failure mode, spatial lexical bias: adding a spatial relation word to the answer options can attract the model's decision and make the newly added option likely to be selected. Using nine open-weight MLLMs, we show that this phenomenon is widely observed. In particular, models can answer a binary spatial question correctly, yet consistently select an incorrect third spatial option once it is added to the answer set. We isolate such binary-stable but ternary-fragile cases as diagnostic examples and leverage mechanistic interpretability tools, revealing that a substantial part of the failure instead originates on the language side rather than the visual side: visual attention analyses and residual-stream probes show the correct spatial relation remains internally available on these failures, while irrelevant-option controls, activation patching, and sparse component interventions trace the bias to specific LLM-side channels and neurons. Based on this finding, we show that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias, lifting four-way robust accuracy by up to 100 points on synthetic data, and by 68.0, 32.6, and 20.1 points on broader evaluation datasets WhatsUp, SpatialMQA-Direct, and VSR.

Related papers

• Tracking the Truth: Object-Centric Spatio-Temporal Monitoring for Video Large Language Models [154.39583176906893]
  Large language models (MLs) have advanced understanding, they remain highly prone to hallucinations in scenes.<n>We argue this stems from a failure intemporal monitoring, the ability to persistently track object dynamic identities, states, and relations over time.<n>Existing benchmarks obscure this deficit relying on single final-answer queries that can often be resolved via local visual cues or statistical priors.
  arXiv  Detail & Related papers  (2026-05-09T14:32:36Z)
• SEM: Sparse Embedding Modulation for Post-Hoc Debiasing of Vision-Language Models [36.40093869747498]
  We propose Sparse Embedding Modulation (SEM), a post-hoc, zero-shot debiasing framework that operates in a Sparse Autoencoder latent space.<n>SEM identifies and modulates bias-relevant neurons while preserving query-relevant ones.<n>Our results demonstrate that sparse latent representations provide an effective foundation for post-hoc debiasing of vision-language models.
  arXiv  Detail & Related papers  (2026-03-19T15:28:08Z)
• From Indoor to Open World: Revealing the Spatial Reasoning Gap in MLLMs [65.04549036809557]
  We introduce a benchmark built from pedestrian-perspective videos captured with synchronized stereo cameras, LiDAR, and IMU/GPS sensors.<n>This dataset provides metrically precise 3D information, enabling the automatic generation of spatial reasoning questions.<n> Evaluations reveal that the performance gains observed in structured indoor benchmarks vanish in open-world settings.
  arXiv  Detail & Related papers  (2025-12-22T18:58:12Z)
• Distance Is All You Need: Radial Dispersion for Uncertainty Estimation in Large Language Models [13.41454380481593]
  We introduce bfRadial Dispersion Score (RDS), a simple, parameter-free, fully model-agnostic uncertainty metric.<n>RDS naturally extends to per-sample scoring, enabling applications such as best-of-$N$ selection and confidence-based filtering.
  arXiv  Detail & Related papers  (2025-12-04T00:53:49Z)
• HyCoVAD: A Hybrid SSL-LLM Model for Complex Video Anomaly Detection [0.0]
  Video anomaly detection (VAD) is crucial for intelligent surveillance, but a significant challenge lies in identifying complex anomalies.<n>We introduce HyVAD, Hybrid Complex Video Anomaly Detection, a hybrid SSL-LLM model that combines a multi-task SSL temporal analyzer with LLM validator.
  arXiv  Detail & Related papers  (2025-09-26T16:20:06Z)
• From Distributional to Overton Pluralism: Investigating Large Language Model Alignment [82.99849359892112]
  We re-examine previously reported reductions in response diversity post-alignment.<n>Our analysis suggests that an apparent drop in the diversity of responses is largely explained by quality control and information aggregation.<n>Findings indicate that current alignment techniques capture but do not extend the useful subset of assistant-like base LLM behavior.
  arXiv  Detail & Related papers  (2024-06-25T16:32:33Z)
• MM-SpuBench: Towards Better Understanding of Spurious Biases in Multimodal LLMs [38.93090238335506]
  Spurious bias, a tendency to use spurious correlations between non-essential input attributes and target variables for predictions, has revealed a severe pitfall in deep learning models trained on single modality data. We introduce MM-SpuBench, a comprehensive visual question-answering (VQA) benchmark designed to evaluate MLLMs' reliance on nine distinct categories of spurious correlations. Our findings illuminate the persistence of the reliance on spurious correlations from these models and underscore the urge for new methodologies to mitigate spurious biases.
  arXiv  Detail & Related papers  (2024-06-24T20:29:16Z)
• Anomaly Detection of Tabular Data Using LLMs [54.470648484612866]
  We show that pre-trained large language models (LLMs) are zero-shot batch-level anomaly detectors. We propose an end-to-end fine-tuning strategy to bring out the potential of LLMs in detecting real anomalies.
  arXiv  Detail & Related papers  (2024-06-24T04:17:03Z)
• MLAD: A Unified Model for Multi-system Log Anomaly Detection [35.68387377240593]
  We propose MLAD, a novel anomaly detection model that incorporates semantic relational reasoning across multiple systems. Specifically, we employ Sentence-bert to capture the similarities between log sequences and convert them into highly-dimensional learnable semantic vectors. We revamp the formulas of the Attention layer to discern the significance of each keyword in the sequence and model the overall distribution of the multi-system dataset.
  arXiv  Detail & Related papers  (2024-01-15T12:51:13Z)
• Exposing Attention Glitches with Flip-Flop Language Modeling [55.0688535574859]
  This work identifies and analyzes the phenomenon of attention glitches in large language models. We introduce flip-flop language modeling (FFLM), a family of synthetic benchmarks designed to probe the extrapolative behavior of neural language models. We find that Transformer FFLMs suffer from a long tail of sporadic reasoning errors, some of which we can eliminate using various regularization techniques.
  arXiv  Detail & Related papers  (2023-06-01T17:44:35Z)

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.