Conference week can feel like drinking from a firehose—hundreds of posters, overlapping sessions, and a calendar that refuses to stay tidy. At AI Tech Inspire, a promising open-source tool crossed our radar that aims to calm the chaos for NeurIPS attendees: AgenticNAV, a co‑pilot for building personalized schedules and exploring papers more deeply.
Quick facts
- AgenticNAV is designed to help attendees create personalized schedules and explore NeurIPS papers in more detail.
- It’s an academic open-source initiative from researchers at the University of Exeter and the Technical University of Munich.
- Hosted on Hugging Face Spaces and available at: https://huggingface.co/spaces/CORE-AIx/AgenticNav.
- Free to use, no login required, and no intent to commercialize.
- Configurable to work with your preferred LLM and inference provider; default uses
GPT-OSS 120Bon Ollama Cloud. - Source code: https://github.com/core-aix/agentic-nav, ready for local deployment.
- It’s a prototype; the team welcomes feedback and pull requests.
Why a conference co‑pilot matters
NeurIPS schedules are notorious for FOMO. You skim abstracts, star items in the app, then miss half of them anyway because the poster hall is a maze and the clock is unforgiving. A focused tool that asks, “What do you care about?” and then composes a plan around that is more than a convenience—it’s a productivity boost. If you care about diffusion models, structured prediction, multimodal agents, or scaling laws, you don’t want to waste cycles hunting session IDs and time slots.
AgenticNAV frames itself as a one-stop workflow for two facts of conference life: 1) create a schedule you’ll actually follow, and 2) dive deeper into the papers you shortlist. It’s not another generic schedule viewer. Its stated goal is to help you navigate by intent: give it themes or keywords and get a focused route through the noise.
Key takeaway: turn interests into a day plan, and turn a shortlist of papers into actionable reading.
Hosted, free, and no login friction
The zero-friction angle stands out. The tool is live on Hugging Face Spaces and usable without accounts or paywalls. That matters in the week before a conference, when attention is scarce and setup time is the enemy. Spin it up in your browser, try a few topics, and see if the recommendations and schedule are helpful.
For developers, the simplicity also lowers the barrier to team adoption: you can paste the Space link in a group chat and everyone can try it. No shared credentials. No per-seat drama. Just the tool and your queries.
Bring-your-own model: an LLM-agnostic stance
Under the hood, AgenticNAV leans into configurability. It supports plugging in your preferred LLM and inference provider, which is a practical nod to the reality that teams have opinions—on latency, cost, data governance, and model behavior. By default, it runs GPT-OSS 120B via Ollama Cloud, but the point is flexibility. If your org mandates specific gateways or you’re optimizing for throughput with your own endpoints, you can adapt the stack.
This model-agnostic approach also means experimentation is part of the value. Want to compare a larger GPT-style model against a smaller, faster one on heuristic tasks like “pick the most relevant posters on control theory + RL”? You can. Curious whether your local quantized model on a laptop GPU is good enough? Configure it and see. If you’re already running models with PyTorch, TensorFlow, or tooling accelerated by CUDA, the local deployment angle will feel familiar.
Open-source and the sovereign AI angle
AgenticNAV is published as an academic open-source initiative by researchers from the University of Exeter and TUM, with source available at GitHub. The maintainers explicitly encourage local deployment, aligning with “sovereign AI” preferences where data locality and control matter. For teams that prefer to keep usage data off third-party platforms, running locally is a strong option.
Local control also creates room for deeper customization: adjusting prompt templates, inserting your own ranking heuristics, or integrating a custom vector index. If your group maintains a curated reading list or lab-specific taxonomy, you can impose it on the planning flow—something a closed, one-size-fits-all conference app rarely enables.
What using it might look like
The pitch is straightforward:
- Input your interests in plain language (e.g., “model-based RL for robotics,” “causal representation learning,” “efficient
Stable Diffusionpipelines”). - Get a prioritized plan: a shortlist of posters/sessions and a time-aware schedule optimized for your stated focus.
- Explore papers in more detail: prompts to summarize, compare, or tease out differences across related works.
Think of it as an assistant that turns keywords into an itinerary and turns a list of papers into context. Instead of scanning PDFs and using Cmd+F to hunt for methods, you can query for “contrastive vs. predictive objectives in self-supervised speech” and let the tool assemble a thread across relevant abstracts.
Importantly, the team labels this as a prototype. That suggests expectations should be calibrated: treat it as a smart helper, not an oracle. If it surfaces a must-see poster, trust but verify against the official schedule. And if something is missing, that’s precisely the feedback the maintainers are inviting.
Why it matters for developers and engineers
The engineering lens on this is less about novelty and more about leverage:
- Time-to-signal: Reduce the overhead between question and answer. Less clicking around, more “Here’s the subset that matters.”
- Composability: Because it’s open-source, you can extend the tool—e.g., plug into your lab’s knowledge base or your own ranking criteria.
- Portability: If local matters (compliance, offline usage, queue times), you can deploy it on your own machines.
- Cost control: Choose your inference path. Use a hosted model for convenience or a local model for predictable cost.
For teams sending multiple people to NeurIPS, there’s also a coordination angle: use a shared configuration to avoid overlap and cover more ground. One person tracks safety and alignment; another tracks optimization and systems; a third follows diffusion pipelines. A tool like this can help reduce redundant coverage and produce a consolidated reading plan after the conference.
Trying it today
Curious? You can use the hosted version here: AgenticNAV on Hugging Face Spaces. It’s free, no login, and configured out of the box. For those who want control and customization, the repository at core-aix/agentic-nav includes what you need to deploy locally and tweak behavior.
“This is a prototype” + “We welcome feedback and PRs.” That’s an invitation to shape a tool you might use every conference season.
Caveats and what to watch
- Prototype status means features and performance can evolve quickly; expect occasional rough edges.
- Relevance depends on your prompts and model choice; a smaller model might be faster but less precise.
- Always cross-check schedule details with the official NeurIPS program to avoid outdated slots or room changes.
Those caveats are normal for open tooling—and they’re also where contributors can have outsized impact. If you’re the type who builds internal aids for your lab or company, this is fertile ground to generalize those scripts into public utilities.
Where this could go next
AgenticNAV is pointed at NeurIPS 2025, but the core idea is portable: any content-heavy event where attendees want personalized discovery. Imagine plugins for co-authorship graphs, lab lineage overlays, or quick comparisons to prior conference editions. Or broader capabilities like “build me a reading list to ramp up a new teammate in 48 hours” based on their background and your tech stack (PyTorch vs. TensorFlow, GPU constraints, etc.).
Questions worth exploring:
- Can ranked recommendations be combined with your calendar to avoid hallway sprints between distant halls?
- How well do smaller, faster models perform for this planning task compared to larger ones?
- Could local deployments cache embeddings to support offline browsing during spotty venue Wi‑Fi?
AgenticNAV won’t eliminate the delightful chaos of NeurIPS—but it nudges it toward a plan you can execute. For developers and engineers who value autonomy and tinkering, the open-source, LLM-agnostic posture is the real story. If you’re heading to NeurIPS or simply want a blueprint for smarter technical event workflows, it’s worth a look—and maybe a pull request.
Recommended Resources
As an Amazon Associate, I earn from qualifying purchases.
