Explainability tools usually have to choose: be fast, or be faithful to the actual model. Surrogate-model methods (train a simple model to mimic the complex one, then explain the surrogate) are fast but only approximate. Perturbation-based methods are more faithful but painfully slow at scale, and neither approach is architecture-agnostic by default.
DLBacktrace skips the tradeoff. It computes relevance directly from a model's own weights and activations and propagates it layer by layer from output back to input, using consistency- and faithfulness-preserving scores rather than a surrogate. Because it works straight from the model's internals, it generalizes across architecture families: CNNs, RNNs, Transformers, LLMs, and Mixture-of-Experts models, on tabular, vision, and text tasks (classification, detection, segmentation, generation).
The speedup from that approach is not incremental. DLB v2 achieves a 20x to 1400x speedup over v1, cutting explainability time on LLaMA-3.2-3B from a range of 250 to 30,000 seconds down to 12 to 18 seconds. That's the difference between explainability being a research artifact and explainability being something you can run in a production pipeline.
What it covers:
- Layer-wise relevance propagation across a wide range of layer types: linear, Conv2D, pooling, LSTM, embedding, dropout, layer/batch normalization, and tensor manipulation (transpose, permute, squeeze), including negative indexing.
- Broad model coverage, tested against ResNet, VGG, DenseNet, EfficientNet, MobileNet, and ViT for vision; BERT, ALBERT, RoBERTa, DistilBERT, ELECTRA, XLNet, LLaMA-3.2, and Qwen for text; and JetMoE, OLMoE, Qwen3-MoE, and GPT-OSS for Mixture-of-Experts, with expert-level relevance analysis built in.
- A high-level pipeline interface for text and image tasks, with automatic model loading, so you're not wiring up relevance propagation by hand for every new architecture.
- DLB Auto Sampler, for text generation with greedy, beam-search, temperature, top-k, and top-p decoding, with token-level relevance tracked as it generates.
- CPU and GPU compatible, with CUDA acceleration and deterministic execution for production use.
Who it's for
- ML teams who need explainability at production scale and can't afford surrogate-model approximations or minutes-per-inference explanation cost.
- Teams working across multiple architecture families (vision, text, MoE) who want one explainability engine instead of one per architecture.
- Researchers and auditors who need relevance scores traceable straight to model weights, not to a proxy model.
Links:
- Docs: https://dlbacktrace.lexsi.ai/
- Authors: Vinay Kumar Sankarapu, Chintan Chitroda, Yashwardhan Rathore, Neeraj Kumar Singh, Pratinav Seth