Learnijoy
ResearchAI ResearchAI Engineering & DevTools

DMDIntel Interprets LLM Predictions Using Dynamic Mode Decomposition.

Amogh Joshi, Animesh Mukherjee, Sergey Utyuzhnikov· August 14, 2026 View original

Key takeaways

  • DMDIntel offers a new, more effective method for interpreting LLM classification predictions.
  • It decomposes LLM hidden states into "modes" to rank input token importance.
  • The technique consistently outperforms current state-of-the-art interpretability methods.
  • Improved interpretability is vital for building trust and debugging AI systems.

Who benefits

HealthcareBFSILegalAI Development

Summary

DMDIntel is a new method that uses dynamic mode decomposition (DMD) to make large language model (LLM) predictions interpretable in classification tasks. It decomposes hidden states into patterns and ranks input tokens based on their projection onto these patterns.

Researchers have introduced DMDIntel, a novel technique designed to enhance the interpretability of large language models, particularly in classification scenarios. This method leverages dynamic mode decomposition (DMD) to break down the complex internal states of an LLM into fundamental, prominent patterns, referred to as modes. By analyzing how input tokens project onto these identified modes, DMDIntel assigns relevance ranks to each token, effectively highlighting their contribution to the model's final prediction. Extensive testing across various datasets and LLM architectures demonstrates that DMDIntel significantly outperforms existing state-of-the-art interpretability methods. Techniques like principal component analysis (PCA), integrated gradients, and SHAP were consistently surpassed by DMDIntel's ability to accurately attribute input tokens. This robust performance suggests a more reliable and insightful approach to understanding the black-box nature of LLMs.

Why it matters

Understanding why an LLM makes a particular decision is crucial for trust, debugging, and compliance in professional applications. This research offers a more effective way to interpret LLM outputs, leading to more reliable and explainable AI systems.

How to implement this in your domain

  1. 1Evaluate: Assess the current interpretability methods used for your LLM applications.
  2. 2Research: Investigate the open-source availability and integration potential of DMDIntel.
  3. 3Pilot: Implement DMDIntel on a small-scale classification task to compare its interpretability insights against existing tools.
  4. 4Refine: Use the enhanced interpretability to debug model failures or biases more effectively.

Original post by Amogh Joshi, Animesh Mukherjee, Sergey Utyuzhnikov

"arXiv:2608.13048v1 Announce Type: new Abstract: In this work, we introduce DMDIntel which uses dynamic mode decomposition (DMD) to make the predictions made by LLMs in a classification task interpretable. It develops an input attribution pipeline, that first decomposes the hidden…"

View on X

Originally posted by Amogh Joshi, Animesh Mukherjee, Sergey Utyuzhnikov on X · view source

Want to go deeper?

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

Explore courses

More in AI Research

AI Engineering & DevToolsAI ResearchAI Investing

FlowLOB Generates Realistic, Controllable Limit Order Books Efficiently

This paper introduces FlowLOB, a conditional flow-matching generator for Limit Order Book (LOB) trajectories that offers realistic market dynamics, efficient sampling, and controllable scenario generation, outperforming existing agent-based and deep generative simulators. FlowLOB achieves high fidelity with significantly fewer computational steps than diffusion models and transfers effectively to unseen instruments.

Zhuohan Wang, Andreea Bacalum, Ollie Olby, Carmine Ventre, Namid StillmanAug 14, 2026
AI Engineering & DevToolsAI Research

Auditing Reveals Bias in Neural Combinatorial Optimization Benchmarks

This paper audits test-time budget allocation in Neural Combinatorial Optimization (NCO) solvers, revealing that reported gains from non-uniform sampling often stem from "sampling luck" rather than true allocation benefits on in-distribution data. It proposes a correction procedure and demonstrates real gains under distribution shift, emphasizing the need for rigorous evaluation.

Jinhyung BaeAug 14, 2026
AI Engineering & DevToolsAI Research

Diffusion Models Solve Mixed-Integer Optimization Problems Faster

This paper introduces Constrained Graph Diffusion (CGD), a novel learning-based approach that uses a graph-based generative diffusion model to approximately solve mixed-integer optimization problems (MIPs). CGD integrates a training-free feasibility projection operator into the diffusion process, significantly improving solution quality and feasibility while achieving substantial speedups over traditional numerical solvers.

Vincenzo Di Vito, Mehdi Taghizadeh, Deepjyoti Deka, Kaarthik Sundar, Ferdinando FiorettoAug 14, 2026