Bay Area Frontier Research Club #14 | Adapting and Unlearning (dinner + paper discussion)
Training is just the beginning. How do you keep changing a trained model — adapting what it knows, and unlearning what it shouldn't?
FRC #14 takes two routes to that question. Ruotong Liao makes the case for adaptation — models that keep training on themselves at test time, improving after deployment without a full retrain. David Khachaturov makes the case for removal — unlearning with formal guarantees, so a model can be made to provably forget rather than promised to.
The Frontier Research Club convenes researchers from the frontier labs, faculty and PhDs from Stanford and Berkeley, and engineers running these systems in production, around concrete work: papers, methods, and results. Discussion focuses on the questions that determine whether a result holds — the assumptions, the evaluation, the failure modes, and what would constitute convincing evidence.
Each session features two to three brief presentations, with most of the evening reserved for questions and critique. Papers are circulated in advance.
Agenda
5:30pm: Doors open
5:30pm – 6:30pm: Networking + light dinner
6:30pm – 8:00pm: Research presentations + discussion
8:00pm – 8:30pm: Networking
Presenters & topics
Talk 1: Test-Time Training and Self-Evolving Models
Presented by Ruotong Liao — PhD Candidate, LMU Munich · Visiting Researcher, Stanford University
What if a model kept training after it shipped — adapting to each new problem at inference, instead of waiting for the next full retrain? Ruotong's work on test-time training and self-evolution treats a deployed model as something that continues to change in the field, updating itself on the data in front of it.
Ruotong is a PhD candidate at LMU Munich and a visiting researcher at Stanford, advised by Serena Yeung-Levy, with work published at NeurIPS, EMNLP, and ECCV, and an invited talk at the UN's "AI for Good." Her talk covers where self-adaptation earns its keep, where it breaks, and how to tell the two apart.
📄 Pre-read: arxiv.org/abs/2603.02203
Talk 2: Unlearning with Guarantees
Presented by David Khachaturov — Founder, Egoist Machines (YC S26) · PhD, University of Cambridge
Deleting a record from a database is one line of SQL; deleting it from a trained model is an open problem — and most "unlearning" merely asserts that the model has forgotten, an assertion that state-of-the-art methods fail under poisoning audits and that can even be forged. David presents a yet-unannounced architecture — Mixture of Remote Experts — where revoking a data source becomes a checkable equality, and closes on the open problems: portable per-owner capability latents that survive base-model upgrades, and revocation certificates an outsider can verify.
David Khachaturov is the founder of Egoist Machines (YC S26), where he builds memory infrastructure that makes user data removable and portable by design. He holds a PhD in machine learning security from the University of Cambridge, supervised by the late Ross Anderson and Robert Mullins, and is an ICML and AAAI author.
📄 Pre-reads:
Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025) — arxiv.org/abs/2406.17216
On the Necessity of Auditable Algorithmic Definitions for Machine Unlearning (USENIX Security 2022) — arxiv.org/abs/2110.11891
Towards Reliable Forgetting: A Survey on Machine Unlearning Verification — arxiv.org/abs/2506.15115
LoRA-X: Cross-Model Low-Rank Adaptation — arxiv.org/abs/2501.16559
Ramp Labs — Research — labs.ramp.com/research
Want to present your work?
If you have a research paper you’d like to discuss at one of our next sessions, please submit it for consideration.
SUBMIT YOUR PAPER HERE.
Who should attend
Researchers and PhDs from frontier labs, leading institutions, research-backed startups and beyond
Anyone working on post-training, adaptation, continual learning, or machine unlearning
Research engineers and scientists building and evaluating models that keep changing after they ship
People who come for the methods, the evaluation, and the argument
If you've ever wished research moved faster, and was more open and collaborative, you belong here.
Capacity is limited.
We will take photos and short video clips for event recap and promotion. By attending, you consent to being photographed and recorded, and to the use of those images and clips by the organizers on social media and other event marketing channels.
🎥 Last session recap — AGI House SF, July 8
FRC #13 examined a single question from two directions: how do you get more capability out of a model at inference, without retraining it or paying for a bigger one? Two talks, one argument — a room of frontier-lab researchers, founders, and investors pressed on both.
Aditya Kusupati (Google DeepMind) — Matryoshka Principles for Adaptive Intelligence: the case for adaptive models. His Matryoshka Representation Learning now underpins the production embedding models at OpenAI (text-embedding-3) and Google (Gemini); he drew the line from MRL to MatFormer (behind Gemma 3n) into a broader set of principles for building compute-elastic intelligence — a single network that dials its own size up or down at inference.
Varun Sunkaraneni (Cobalt · MIT/Poggio group) — Agentic Systems as Boosting Weak Reasoning Models: the case for orchestration. New work reframing classical boosting theory for the test-time compute era — verifier-backed committee search, with weak-model committees reaching 76.4% on SWE-bench Verified, outperforming much stronger single models on code repair, theorem proving, and program synthesis.
Find presentation recordings on our YouTube channel, @FrontierResearchClub.
🌐 Connect with Frontier Research Club
Luma Calendar: luma.com/frontiersyndicate
YouTube: FrontierResearchClub
LinkedIn: frontier-research-club
Instagram: @frontierresearchclub
Email: kristopher@frontiersyndicate.vc
Hosted by
Frontier Syndicate is a venture community connecting frontier tech researchers, builders, and investors through curated convenings and early-stage capital. Across the Bay Area, we host a recurring series of research forums, builder nights, and intimate investor dinners — and back exceptional companies emerging from the labs, communities, and technical networks we convene.
Hexo Labs is a neolab for recursive self-improving AI. Their open-source SIA framework is the first to update both the harness AND the model weights of a task-specific agent in the same self-improvement loop — clearing state-of-the-art results across multiple domain benchmarks. Hexo backs the broader research community through grants and direct collaboration on hard problems in science and engineering.
Ashurst Perkins Coie is a top global law firm with a strategic focus on technology, energy and infrastructure, and financial services. Building on Perkins Coie's long history as counsel to the technology and venture community, the firm advises innovators from first financing through global scale — and hosts tonight's session at its Palo Alto office.