Sparse Visual Thought Circuits in Vision-Language Models

• URL: http://arxiv.org/abs/2603.25075v1
• Date: Thu, 26 Mar 2026 06:24:36 GMT
• Title: Sparse Visual Thought Circuits in Vision-Language Models
• Abstract summary: We develop a causal pipeline to localize and test sparse visual thought circuits in Qwen3-VL-8B.<n>We train SAEs at this layer, construct task-selective sets via an explicit rule, and perform inference time scaling and ablation while quantifying accuracy and drift.
• Score: 2.5754366051855837
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Sparse autoencoders (SAEs) improve interpretability in multimodal models, but it remains unclear whether SAE features form modular, composable units for reasoning-an assumption underlying many intervention-based steering methods. We test this modularity hypothesis and find it often fails: intervening on a task-selective feature set can modestly improve reasoning accuracy, while intervening on the union of two such sets reliably induces output drift (large unintended changes in predictions) and degrades accuracy, even under norm-matched perturbations. This non modular circuit interference is consistent with shared internal pathways where feature unions amplify activation shifts. We develop a reproducible causal pipeline to localize and test these sparse visual thought circuits in Qwen3-VL-8B. On a controlled synthetic benchmark with seven task types and three difficulty levels, linear probes identify a mid decoder locus for task type information. We train SAEs at this layer, construct task-selective sets via an explicit rule, and perform inference time scaling and ablation while quantifying accuracy and drift. Our findings-validated with bootstrapped subsamples and permutation controls, and replicated across multiple VLM families and five diverse datasets clarify the boundaries of SAE feature composability and provide a rigorous diagnostic framework for more reliable VLM control.

Related papers

• Multimodal Model Diffing for Feature Discovery and Control [25.349396112139214]
  We introduce MMDiff, a framework that turns sparse autoencoders into feature-level interfaces for discovering and controlling multimodal behavior.<n>We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR.
  arXiv  Detail & Related papers  (2026-08-10T17:59:30Z)
• Steering Vision-Language Models with Joint Sparse Autoencoders [2.367986397868328]
  We introduce the Joint Sparse Autoencoder (JSAE) to factorize sequence-pooled vision and language activations into shared, interpretable image/caption-level features.<n>Applying JSAE to vision-language models (VLMs) yields representations that are difficult to use as controllable cross-modal steering directions.
  arXiv  Detail & Related papers  (2026-06-24T10:11:04Z)
• Demystifying Variance in Circuit Discovery of LLMs [51.97489679448601]
  We develop a new circuit discovery method that substantially lessens resampling variance.<n>We show that rephrasing variance arises because prompts with different templates tend to activate different circuits in the model.<n>We argue that sparsity, which has been claimed to form more compact and interpretable task circuits, fails to solve this problem.
  arXiv  Detail & Related papers  (2026-06-15T16:25:07Z)
• What Semantics Survive the Connector? Diagnosing VLM-to-DiT Alignment in Video Editing [59.40855219373921]
  Flow matching based video generative models rely on prepended Vision-Language Models to handle complex, instruction-based video editing.<n>The prevailing assumption is that a connector module can seamlessly align the VLM's rich multi-modal reasoning with the original text embedding space of DiTs.<n>We propose a controlled data processing pipeline based on video composition that results in TRACE-Edit, a diagnostic dataset focusing on relation-based editing.
  arXiv  Detail & Related papers  (2026-05-20T06:42:15Z)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models [35.34635490741956]
  Visual latent reasoning lets a multimodal large language model (MLLM) create intermediate visual evidence as continuous tokens.<n>We propose textbfGAP, a textbfGranular textbfAlignment textbfParadigm for visual latent modeling.<n>On Qwen2.5-VL 7B, the resulting model achieves the best mean aggregate perception and reasoning performance among our supervised variants.
  arXiv  Detail & Related papers  (2026-05-12T16:41:09Z)
• Steering and Rectifying Latent Representation Manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection [52.5174167737992]
  Video anomaly detection (VAD) aims to identify abnormal events in videos.<n>We propose SteerVAD, which advances MLLM-based VAD by shifting from passively reading to actively steering and rectifying internal representations.<n>Our method achieves state-of-the-art performance among tuning-free approaches requiring only 1% of training data.
  arXiv  Detail & Related papers  (2026-02-27T13:48:50Z)
• PromptCD: Test-Time Behavior Enhancement via Polarity-Prompt Contrastive Decoding [85.22047087898311]
  We introduce Polarity-Prompt Contrastive Decoding (PromptCD), a test-time behavior control method that generalizes contrastive decoding to broader enhancement settings.<n>PromptCD constructs paired positive and negative guiding prompts for a target behavior and contrasts model responses to reinforce desirable outcomes.<n>Experiments on the "3H" alignment objectives demonstrate consistent and substantial improvements, indicating that post-trained models can achieve meaningful self-enhancement purely at test time.
  arXiv  Detail & Related papers  (2026-02-24T08:56:52Z)
• Same Answer, Different Representations: Hidden instability in VLMs [65.36933543377346]
  We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness.<n>We apply this framework to modern Vision Language Models (VLMs) across the SEEDBench, MMMU, and POPE datasets.
  arXiv  Detail & Related papers  (2026-02-06T12:24:26Z)
• VSCOUT: A Hybrid Variational Autoencoder Approach to Outlier Detection in High-Dimensional Retrospective Monitoring [0.0]
  VSCOUT is a distribution-free framework designed for retrospective (Phase I) monitoring in high-dimensional settings.<n>VSCOUT achieves superior sensitivity to special-cause structure while maintaining controlled false alarms.<n>Its scalability, distributional flexibility, and resilience position VSCOUT as a practical and effective method for retrospective modeling and anomaly detection in AI-enabled environments.
  arXiv  Detail & Related papers  (2026-01-28T18:30:48Z)
• When Domain Pretraining Interferes with Instruction Alignment: An Empirical Study of Adapter Merging in Medical LLMs [0.6345523830122167]
  Large language models can exhibit surprising adapter interference when combining domain adaptation and instruction alignment.<n>We study a two-stage LoRA pipeline for medical LLMs, where domain-oriented pre-training (PT) and supervised fine-tuning (SFT) are trained separately and later merged.
  arXiv  Detail & Related papers  (2026-01-26T10:54:06Z)
• SALVE: Sparse Autoencoder-Latent Vector Editing for Mechanistic Control of Neural Networks [0.0]
  We present SALVE, a framework that bridges mechanistic interpretability and model editing.<n>We learn a sparse, model-native feature basis without supervision.<n>We validate these features with Grad-FAM, a feature-level saliency mapping method.
  arXiv  Detail & Related papers  (2025-12-17T20:06:03Z)
• Unsupervised Conformal Inference: Bootstrapping and Alignment to Control LLM Uncertainty [49.19257648205146]
  We propose an unsupervised conformal inference framework for generation.<n>Our gates achieve close-to-nominal coverage and provide tighter, more stable thresholds than split UCP.<n>The result is a label-free, API-compatible gate for test-time filtering.
  arXiv  Detail & Related papers  (2025-09-26T23:40:47Z)
• CKAA: Cross-subspace Knowledge Alignment and Aggregation for Robust Continual Learning [80.18781219542016]
  Continual Learning (CL) empowers AI models to continuously learn from sequential task streams.<n>Recent parameter-efficient fine-tuning (PEFT)-based CL methods have garnered increasing attention due to their superior performance.<n>We propose Cross-subspace Knowledge Alignment and Aggregation (CKAA) to enhance robustness against misleading task-ids.
  arXiv  Detail & Related papers  (2025-07-13T03:11:35Z)
• Steering Large Language Model Activations in Sparse Spaces [21.55545768931058]
  A key challenge in AI alignment is guiding large language models (LLMs) to follow desired behaviors at test time.<n>We introduce sparse activation steering (SAS), a method that leverages sparse autoencoders (SAEs) to steer behavior in sparse spaces.
  arXiv  Detail & Related papers  (2025-02-28T20:43:45Z)
• LF-Steering: Latent Feature Activation Steering for Enhancing Semantic Consistency in Large Language Models [16.37602070339033]
  Large Language Models (LLMs) often generate inconsistent responses when prompted with semantically equivalent paraphrased inputs.<n>We propose LF-Steering, a novel activation steering approach to precisely identify latent feature representations responsible for semantic inconsistency.<n>Our method maps the hidden states of the relevant transformer layer into a sparsely activated, high-dimensional feature space based on a sparse autoencoder.
  arXiv  Detail & Related papers  (2025-01-19T13:06:51Z)
• Unsupervised Controllable Generation with Self-Training [90.04287577605723]
  controllable generation with GANs remains a challenging research problem. We propose an unsupervised framework to learn a distribution of latent codes that control the generator through self-training. Our framework exhibits better disentanglement compared to other variants such as the variational autoencoder.
  arXiv  Detail & Related papers  (2020-07-17T21:50:35Z)

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.