LLM Scaling Laws
PreprintPins down when LLM scaling laws actually extrapolate — from multi-node GPU pretraining on national NAIRR / NSF ACCESS supercomputers. Under review at NeurIPS 2026.
My bet: scale alone won’t make AI trustworthy — it has to verify itself
Models that learn, logic that verifies.
I build NeuroSymbolic AI — systems that fuse statistical learning with rigorous logic so they can detect, explain, and recover from their own errors. The work spans national-scale LLM pretraining to robotics, appears at AAAI, ICLR, and CIKM (NeurIPS under review), and grew out of shipping production systems at Dell and Intel.
- 18 fellowships & awards
- AAAI · ICLR · CIKM top venues
- NAIRR · NSF ACCESS national HPC
- 9 papers & preprints
- Harvard Galileo
- USC ICT
- Syracuse
- Arizona State
- Dell
- Intel
- AAAI 2026 reviewer
Now
& where it’s going
I’m pushing on the part of AI I think matters most: trust. Right now that means proving when LLM scaling laws actually extrapolate, from multi-node pretraining on national supercomputers (NeurIPS 2026, under review), and building perception that catches and corrects its own mistakes in novel environments. The throughline is simple: AI whose answers you can verify. Reviewing for AAAI 2026.
Interactive · scaling laws
When do neural scaling laws extrapolate?
You fit L(N, D) on small training runs — can you trust it on bigger ones? It hinges on tokens‑per‑parameter coverage: how diverse D/N is across the runs you fit. This widget refits the paper’s laws live in your browser, pinned to the original Python fits on the published runs — drag the coverage window and watch extrapolation hold, or fall apart.
The paper trains the same model family two ways: a collinear fan
(tokens locked to parameters, D = k · N,
twelve k levels between 1 and 5) and a non‑collinear grid
(N and D varied independently) — same corpora, same budget. Fitting
Chinchilla‑style laws such as
L(N, D) = E + A/N^α + B/D^β on each design and predicting
held‑out larger runs, the grid wins 97.3% of 1,500 seed‑paired comparisons
(95% CI [96.4%, 98.0%]) — and an interval‑variance diagnostic
VK ≥ τK over the fitted
tokens‑per‑parameter window predicts, before fitting, when a design has enough
coverage to extrapolate. (Animation paused to respect your reduced‑motion setting.)
The runs. Loading the paper’s measured training runs…
Interactive · neurosymbolic AI
A network that learns the shape of its own mistakes.
The core idea of my CIKM ’24 paper, running live below. A model labels each sample twice — fine class and coarse family. f‑EDR then learns logical rules for when those predictions are wrong, from the model’s own labeled mistakes — no hand-written hierarchy — and uses them to flag fresh errors and recover the constraints of a taxonomy the model was never given. Three datasets to try: the paper’s military-vehicle benchmark, CIFAR‑100, and 20 Newsgroups — all from the open PyEDCR example suite.
A classifier predicts fine + coarse labels → f‑EDR learns rules like
error(Tank) ← assign(Tank) ∧ assign(2S19‑MSTA) from its
observed mistakes → the rules flag errors on new images (no ground truth
needed) and double as recovered hierarchy constraints. (Animation paused to
respect your reduced-motion setting.)
Predict. A fine-tuned vision model assigns every image a fine label and a coarse family — most are right, a few are quietly wrong.
00 — About
Engineer turned computer scientist, working where logic meets learning.
Last updated: Loading…
Hi there!
I am a 3rd-year Computer Science Ph.D. student and a Research Associate at the Leibniz NeuroSymbolic Lab @ Syracuse University (SU), New York, USA under the direction of Prof. Paulo Shakarian.
Until recently, I served as a Visiting Researcher at the Learning Sciences group at the Institute for Creative Technologies (ICT) under the University of Southern California (USC).
Until 2023, I served as an AI/ML Researcher at Ben Gurion University of the Negev (BGU), Israel. I was a part of Prof. Gera Weiss and Prof. Shai Arogeti’s Intelligent Robotics Lab (IRL) at BGU, which I co-founded.
I hold a Master of Science (M.Sc.) in Computer Science and a Bachelor of Science (B.Sc.) in Mechanical Engineering, both from BGU.
My research interests are in the fields of Artificial Intelligence (AI), Robotics, Machine Learning (ML), and Modern Control Theory.
Feel free to take a look around to learn more about my work.
News
- Loading news...
01 — Research focus
Three threads, one agenda.
I want learning systems we can trust because they can check themselves. Three lines of work run toward that goal.
-
01
NeuroSymbolic AI
Infusing neural models with rigorous logic so learned systems can reason, justify, and recover from their own errors — without ground-truth labels. The core of my Ph.D.
-
02
Machine Learning & LLMs
From hierarchical multi-label classification to LLM scaling-law extrapolation and agentic frameworks for education — applied ML on real problems, with honest baselines.
-
03
Robotics & Control
Differential games and modern control for autonomous systems with competing objectives — bridging the mechanical engineering that started my path with the AI that continues it.
02 — Publications
Selected peer-reviewed work.
Across NeuroSymbolic AI, LLM scaling laws, hierarchical multi-label classification, vision-language sensor fusion, and differential games for control. Author name in bold.
Loading publications…
03 — Projects
Things I’ve built & shipped.
Open-source systems you can run today — each one a capability, with a paper behind it for the proof. Tooling named so you know exactly what’s under the hood.
Agentic Education
ResearchAn LLM agent framework that generates and grades personalized learning material at scale — built at USC ICT on Microsoft AutoGen. FLAIRS 2026.
VLC Fusion
Paper + codeUses language to tell a perception stack how to fuse its sensors — keeping object detection robust when conditions fall apart. ICLR 2026 (UCRL workshop).
Abductive Perception
ResearchLets multiple pretrained models reconcile their disagreements into one coherent reading of a novel scene — no ground truth required. AAAI 2026.
PyEDCR
Open sourceAutomatically catches and corrects a classifier’s mistakes — learning the logical rules that keep its predictions consistent, with no retraining and no hand-labeled data.
PyDiffGame
Open sourceA differential-game engine that solves for Nash-equilibrium controllers when multiple control objectives compete — open-source, and the most-starred project I maintain.
Years in research & industry
Papers & preprints
Talks & essays
Fellowships & awards
04 — Experience
Research and engineering across labs and continents.
Roles at Syracuse, USC ICT, ASU, BGU, Frenn, Dell, and Intel — always at the boundary of research and shipped systems.
Loading experience…
05 — Education
From mechanical engineering to a Ph.D. in AI.
Each step a deliberate move toward the questions I find most worth answering.
Loading education…
06 — Teaching & talks
Sharing the work.
Courses I’ve taught, talks I’ve given, and essays I’ve written — the public-facing side of the research.
Teaching
Loading teaching…
Talks & essays
Loading talks…
07 — Recognition
Selected fellowships, scholarships & awards.
Funding and recognition supporting my path from Israel to the United States — across mechanical engineering, AI, and security-focused research.
Loading awards…
08 — News
Lately.
A live feed from my research life — new papers, talks, affiliations, and milestones.
- Loading news…
09 — Contact
Get in touch.
Always glad to talk research, collaborations, and hard problems where learning meets logic. Email is the fastest way to reach me.