Cracking open
AI's black box

The NSF National Deep Inference Fabric (NDIF) provides free remote access to large-scale AI models, democratizing access to model internals for transparent and reproducible experimentation.

Get Started

Learn About NDIF

Supported by

Research Papers Using our Tools

Ecosystem Repos Built on NNsight

GitHub Stars Across NDIF-Team Repos

GPUs for Remote NDIF Execution

Data sourced from NDIF internal metrics and public GitHub repository statistics.

What is NDIF?

The National Deep Inference Fabric is a unique nationwide research computing fabric that enables scientists to perform transparent and reproducible experiments on the largest-scale open AI systems. NDIF has three parts:

HPC Fabric

A nationwide high-performance computing fabric powered by NCSA's Delta — utilizing one 8xH200 node and six 4xA40 nodes — providing free remote access to run experiments on large-scale AI models.

NNsight Library

An open-source PyTorch-based toolkit (1100+ GitHub stars) that lets researchers inspect, modify, and customize internal computations of AI models, complete with remote access to large scale models.

Training Program

A nationwide training program developed with PIT-UN, a consortium of 63 universities and colleges, providing workshops, tutorials, and resources to build a broad community of AI researchers.

Learn more about NDIF

Remote Model Access

Access and inspect open-source model internals remotely on NDIF with the NNsight API. Deploy models on-demand with full transparency into their computations.

All models are free for research use — no GPU required on your end.

View All Models NNsight Docs ↗

Hot70B

Llama 3.1 70B

meta-llama/Llama-3.1-70B

Hot8B

Llama 3.1 8B

meta-llama/Llama-3.1-8B

Warm405B

Llama 3.1 405B

meta-llama/Llama-3.1-405B

Warm120B

GPT-OSS 120B

openai/gpt-oss-120b

Featured Research

Researchers worldwide use NDIF and NNsight to uncover how large-scale AI models work, with 140+ published papers at top venues including ICLR, NeurIPS, ICML, and EMNLP.

ArXiv 2026

A Pre-Specified Construction-Confirmation Test of Operation-Level Causal Transfer Across Finite Isomorphic Symbolic Domains

Xinyi Shan

The paper investigates whether a language model can transfer operational knowledge between symbolic domains by applying input-specific adjustments to its internal states. It tests a precise intervention: estimating how the model’s hidden state changes when performing an operation in one domain (like successor in integers mod 16), then applying that same change to a corresponding input in another domain (like letters) to see if it produces the correct output (predecessor). Using careful controls and a two-stage validation process — first identifying a promising candidate, then confirming it on independent data — the authors find that one specific route (integer mod 16 to letters, successor to predecessor) consistently works across two different intervention tools on a single model and layer range. While this suggests the model encodes some operation-level structure that can be transferred in this narrow case, the result is highly specific and does not imply broader capabilities like generalizing across models, domains, or operations. The work advances understanding of how symbolic reasoning might be localized and manipulated in neural

ArXiv 2026

A Unifying Perspective on Language Model Representations: From Filler-Role Structure to Mechanistic Interpretability

Zhang Enyan, R. Thomas McCoy

This work introduces Tensor Product Representations (TPRs) as a unifying framework for understanding how language models encode compositional structure in their internal representations. By modeling meaning as bindings between "fillers" (content) and "roles" (structural positions), TPRs provide a mathematical foundation that explains why diverse interpretability techniques — such as analogical reasoning, probing, sparse autoencoders, and activation patching — often reveal related patterns. The authors demonstrate both theoretically and through experiments across model scales that these methods can be seen as special cases of TPR-based analysis, suggesting a deeper, shared structure underlying model behavior. This approach moves interpretability beyond isolated observations toward a coherent account of how neural networks represent and manipulate structured knowledge.

TAIGR workshop, ICML 2026

Activation Probes Surface Code-Security Signals that the Model's Output Misses

Ivan Wiryadi

Researchers investigated whether the internal activations of open-weight AI models could detect security flaws in code more effectively than simply asking the model to judge code safety through prompts. They trained a lightweight probe on pairs of vulnerable and fixed Python functions, then tested it on real-world vulnerabilities the model had never seen during training. Across multiple models, the probe consistently identified the vulnerable version of a function as higher risk than its fixed counterpart, outperforming random guessing and surpassing the model’s own prompted judgments — even when those prompts included chain-of-thought reasoning. This suggests that the model’s internal representations contain usable security signals that are not accessible through standard prompting, offering a promising path for scalable code review in AI-assisted development.

ICLR 2026

Activation Steering via Generative Causal Mediation

Aruna Sankaranarayanan, Amir Zur, Atticus Geiger, Dylan Hadfield-Menell

Researchers tackle the challenge of pinpointing and controlling specific behaviors in language models, which can be dispersed throughout lengthy responses. By developing a new approach, they aim to provide more precise control over these models, enabling more effective intervention and modification of their outputs. This work has implications for improving the reliability and trustworthiness of language models in real-world applications.

ArXiv 2026

Attention Calibration for Position-Fair Dense Information Retrieval

Andrianos Michail, Elias Schuhmacher, Juri Opitz, Simon Clematide, Rico Sennrich

Researchers investigate a common issue in dense retrieval models, where their effectiveness decreases when relevant information appears later in a passage. To address this positional bias, they develop a method to adjust the model's attention during inference, without requiring retraining or sacrificing overall performance. This approach shows promise in improving fairness and effectiveness across various models and languages.

ArXiv 2026

Behind the Scenes: Mechanistic Interpretability of LoRA-adapted Whisper for Speech Emotion Recognition

Yujian Ma, Jinqiu Sang, Ruizhe Li

Conducts the first systematic mechanistic interpretability study of LoRA within Whisper encoder for speech emotion recognition. Using layer contribution probing, logit-lens inspection, SVD, and CKA analysis, the authors reveal two key mechanisms: delayed specialization that preserves general features before consolidating task-specific information, and forward alignment with backward differentiation between LoRA matrices. Findings clarify how LoRA reshapes encoder hierarchies, providing insights for efficient adaptation strategies.

View all research

Featured Projects

Open-source repositories using NDIF and NNsight — production-grade interpretability tools, libraries, and research code from the community.

PopularActiveCourse

ARENA-education/ARENA_materials

No description

Jupyter Notebook1,349887

Active 5 days ago

PopularActive

hijohnnylin/neuronpedia

open source interpretability platform 🧠

TypeScript1,162157

Active 2 days ago · Apache-2.0

PopularActiveCourse

callummcdougall/ARENA_3.0

No description

Jupyter Notebook1,124728

Active 4 months ago

PaperPopular

saprmarks/dictionary_learning

No description

Python421100

Active last year · MIT

PaperPopularActive

EleutherAI/delphi

Delphi was the home of a temple to Phoebus Apollo, which famously had the inscription, 'Know Thyself.' This library lets language models know themselves through automated interpretability.

Python27962

Active 2 days ago · Apache-2.0

Popular