MIND: Decoupling Model-Induced Label Noise via Latent Manifold Disentanglement

The paradigm of learning from automatic annotations—driven by pre-trained experts and Foundation Models—dominates data-hungry applications. However, it introduces a critical challenge: model-induced label noise. Unlike stochastic noise in classical robust learning, this noise stems from annotator inductive biases, manifesting as systematic errors tightly coupled with local feature manifolds. Existing methods relying on global transition matrices underfit these structural patterns, while learning instance-specific matrices remains mathematically intractable. We propose Model-Induced Noise Decoupling (MIND), a theoretically grounded framework addressing this dilemma. We demonstrate that the high-dimensional noise manifold can be decoupled into tractable, subspace-dependent components via Latent Manifold Disentanglement. Specifically, our Latent Decoupling Estimator (LDE) dynamically projects samples into latent structural clusters with consistent error modes, facilitating noise identifiability without ground-truth anchor points. To rigorously evaluate robustness, we adopt a hierarchical protocol: moving from controlled noise on CIFAR-100 to a structural stress test on large-scale real-world 3D datasets (S3DIS, ScanNet), where error patterns explicitly couple with geometric manifolds. Empirically, MIND significantly outperforms state-of-the-art methods on these complex benchmarks and effectively corrects zero-shot hallucinations from Vision-Language Models (e.g., OpenSeg), highlighting its potential as a robust distillation framework for Foundation Models.

Today's most powerful AI models—such as the vision systems that power self-driving cars or smart assistants—are increasingly used to automatically label massive amounts of data, replacing the slow and expensive process of human annotation. But there is a fundamental problem: these AI "teachers" are far from perfect. When asked to label a 3D indoor scene, they might look at a curtain and call it a wall. They might see a thin table leg and miss it entirely. Worse, these are not random slip-ups—the AI makes the same kinds of mistakes over and over, in predictable patterns tied to how objects look and are shaped. A model that confuses flat vertical surfaces will mislabel walls, boards, and doors alike. If we train new AI systems on data filled with these systematic errors, we end up amplifying the very mistakes we hoped to avoid. This paper introduces a method called MIND that tackles this problem head-on. The key idea is surprisingly intuitive: different types of mistakes happen in different geometric situations. Errors on flat surfaces have one pattern; errors along sharp edges have another. Instead of treating all mistakes as the same, MIND automatically separates the data into groups based on underlying geometric structure—like putting all flat-surface samples in one bucket and all thin-structure samples in another—then figures out what kinds of errors are happening in each group. Once those error patterns are understood, MIND can correct the faulty labels without ever seeing a single human-verified example. Think of it as a proofreader that learns the AI teacher's bad habits and fixes them accordingly. We tested MIND across a wide range of scenarios. On standard image recognition tasks, it cut the error rate nearly in half compared to existing correction methods, and it worked equally well regardless of whether the mistakes were structured or random. The real test came on complex 3D indoor environments—spaces filled with walls, beams, furniture, and clutter—where we deliberately used flawed AI teachers whose errors are tightly linked to physical geometry. Here too, MIND consistently outperformed all competing approaches. Most importantly, we showed that MIND can serve as a "safety net" for state-of-the-art vision-language models like OpenSeg and LSeg—systems that understand images and text together but often hallucinate when applied to unfamiliar 3D environments. By plugging MIND onto these models, we improved their accuracy by over 8 percentage points with zero extra human effort. This points toward a future where powerful but imperfect AI models can be deployed more safely: their predictable mistakes can be automatically identified and corrected before they cause harm.