Diagnosing and Fixing Perception-Decision Misalignment in Omni-LLMs

Hongbo Jiang, Jie Li, Yunhang Shen, Tianyu Xie, Pingyang Dai· August 18, 2026 View original

Key takeaways

  • Omni-LLMs can suffer from Perceptual-Decision Misalignment (PDM), where decisions don't align with perceptions.
  • Causal Modality Sensitivity (CMS) and CausalMSBench provide tools to diagnose PDM.
  • Modality Subspace Activation (MSA) is a training-free method to restore CMS.
  • MSA dynamically balances modal projections to ensure decisions are genuinely multi-modal.

Who benefits

Autonomous VehiclesRoboticsHealthcare (diagnostic AI)DefenseVirtual Assistants

Summary

Researchers identified and formalized Perceptual-Decision Misalignment (PDM) in Omni-Large Language Models (Omni-LLMs), where decisions don't faithfully reflect multi-modal perceptions. They propose Modality Subspace Activation (MSA), a training-free inference-time framework that uses SVD to dynamically balance modal projections, effectively restoring causal modality sensitivity.

Omni-Large Language Models (Omni-LLMs) are critical for advanced multi-modal reasoning in applications like autonomous agents, yet their strong performance can mask a fundamental flaw: Perceptual-Decision Misalignment (PDM). This occurs when an Omni-LLM's decisions do not genuinely reflect its multi-modal perceptions. To diagnose this, researchers formalized Causal Modality Sensitivity (CMS) using a dual-lens framework, measuring Answer Retention Rate (ARR) for macro behavior and Logit Angular Discrepancy (LAD) for microscopic distribution shifts. They also created CausalMSBench, a diagnostic dataset designed to isolate language priors. Benchmarking revealed that popular Omni-LLMs exhibit alarmingly low CMS, showing minimal changes in decision distribution even when crucial modalities are removed. To address this, the study proposes Modality Subspace Activation (MSA), a training-free, inference-time framework. MSA utilizes Singular Value Decomposition (SVD) to estimate the activation strengths of different modalities. By dynamically balancing these modal projections within the last hidden state, MSA effectively restores CMS across various benchmarks, ensuring that decisions are more genuinely informed by all available sensory inputs.

Why it matters

This research is vital for building trustworthy and robust multi-modal AI systems, ensuring that their decisions are truly based on their perceptions rather than spurious correlations or language priors.

How to implement this in your domain

  1. 1Apply the Causal Modality Sensitivity (CMS) framework to diagnose PDM in existing multi-modal AI systems.
  2. 2Integrate Modality Subspace Activation (MSA) as an inference-time technique to improve decision fidelity in Omni-LLMs.
  3. 3Utilize CausalMSBench to benchmark and validate the perceptual alignment of multi-modal models.
  4. 4Develop internal guidelines for evaluating and mitigating PDM in AI agents and world models.

Original post by Hongbo Jiang, Jie Li, Yunhang Shen, Tianyu Xie, Pingyang Dai

"arXiv:2608.14655v1 Announce Type: new Abstract: Omni-Large Language Models (Omni-LLMs) power complex multi-modal reasoning in applications like World Action Models and autonomous agents. However, their strong performance often masks a profound Perceptual-Decision Misalignment (PD…"

View on X

Originally posted by Hongbo Jiang, Jie Li, Yunhang Shen, Tianyu Xie, Pingyang Dai on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses