We built a frontier forecasting model from scratch in 15 days. Introducing Simons.
From a minimal baseline, automated research reached a checkpoint ranked ahead of TimesFM-3 on the GIFT-Eval leaderboard in about two weeks.
Full post: abel.ai/research/introducing…
We release the weights, the complete training code, the synthetic data generators and our public training data. Models and code are free for business use.
Check out
Code: github.com/Abel-ai-lab/Simon…
Models and data: huggingface.co/collections/a…
Demo: huggingface.co/spaces/abel-l…
39
Most exploratory runs stayed small. Within the first day, forecasting error dropped noticeably for a few dozen dollars of GPU time.
About three quarters of the total gain came from data and how a series enters the model; the backbone mattered very little.
1
21
Simons is open in every way: we release weights, the complete training code, the synthetic data generators and our public training data. Models and code are permissive for business use.
Code: github.com/Abel-ai-lab/Simon…
Models and data: huggingface.co/collections/a…
31
Abel retweeted
2/ For a long time, scale has held causal discovery back. This makes experimentation time-consuming and causal discovery on datasets with thousands of variables impractical.
That's why Andrey is built, for faster, accurate and agentic-native causal discovery.
Abel AI Lab is releasing Andrey, an open-source causal discovery package for Python, built for speed, accuracy, and agentic workflows.
This is just the first release 👶. We'd love to hear what you think, what's missing, and what you'd like to see next!
📦 Try Andrey: uv add andrey-core
🌐 Launch Post: andrey.abel.ai/blog/introduc…
💻 Code: github.com/abel-ai-lab/andre…
#CausalInference #CausalDiscovery #Python #OpenSource #AbelAI #AbelAILab
1/ I wanted to share something I’ve been building.
Andrey is a very fast causal discovery package. It's over 100x faster on PC and faster on most other supported methods, compared with other popular packages.
1
4
126
Abel retweeted
1/ I wanted to share something I’ve been building.
Andrey is a very fast causal discovery package. It's over 100x faster on PC and faster on most other supported methods, compared with other popular packages.
Good generation should depend on the right evidence.
SagE-QG, accepted to EMNLP and co-authored by Abel AI Lab, asks a simple question: if the evidence supporting an output is removed, should the output change?
The work aligns question generation with both the skill being tested and the evidence it should actually depend on — using counterfactual intervention as a training signal.
That shift matters beyond question generation:
from generating plausible outputs to learning the structure and dependencies behind them.
In language, that means evidence and reasoning.
In the measurable world, it means signals, state, dynamics, and interventions.
This is part of the direction behind Temporal Superintelligence.
Model behavior is output. The mechanism behind it is the harder question.
Uncovering Hidden Triggers, accepted to ICML 2026 and co-authored by Abel AI Lab , studies the internal mechanisms behind LLM backdoors.
The work traces backdoor behavior from hidden representations → sparse attention heads → controllable vectors, then validates those mechanisms through intervention.
Ablating a small subset of attributed heads can sharply reduce attack success, while manipulating the resulting Backdoor Vector can activate or suppress the behavior.
By moving from observing outputs to understanding underlying mechanisms, this work contributes to building more transparent and reliable AI systems.
At Abel, we believe the next generation of AI must do more than generate outputs — it must understand the hidden structures that shape the world.
Understanding those hidden structures may be key to defining the next foundational architecture beyond transformers.
paper: openreview.net/forum?id=wG7w…
We’re hiring: Full-Stack Engineer
We’re building the infrastructure that brings large-scale temporal foundation models into real-world use.
You’ll work across backend systems, data infrastructure, model inference, and cloud operations — turning research models into reliable production systems that users can actually build with.
We’re looking for someone who is strong in Python + backend engineering, comfortable around ML systems, and excited to work end-to-end across data, inference, APIs, and deployment.
Experience with ML platforms, model serving, GPU infrastructure, or time-series / structured data is a strong plus.
Remote · US / Global · Full-time
Apply: abel.ai/careers/full-stack-e…
1
5
229
A changing world cannot be modeled with a static state.
Two recent works co-authored by Abel AI Lab explore two parts of the same loop.
TIME-RA reasons over time series, context, and visual signals to identify what changed, where, and why — moving anomaly detection toward diagnosis and system understanding.
HyMEM lets internal memory evolve as new observations arrive, refreshing working state as the environment changes.
Together:
observe → infer state → reason → act → observe again
At Abel, we think this continuous state-update loop is fundamental to foundation models for dynamic systems.
The model should not just predict what comes next. It should maintain an evolving representation of the system itself.
TIME-RA:aclanthology.org/2026.findin…
HyMEM :aclanthology.org/2026.findin…
5
105
For time series, the bitter lesson may run backwards.
Language models can scale with more text. But history cannot be lengthened. There is only one history of the S&P 500, one operating history of a power grid, one realized trajectory of any system that matters.
At KDD MileTS 2026, Abel CEO @stephen_wzg argued that this changes what a time-series foundation model should learn.
Forecasting from history alone is not enough. Real systems are open, multivariate, and shaped by the actions we take. A world model should infer the state, structure, dynamics, and uncertainty of the system itself — then answer not only:
What happens next?
but also:
Why? What changes if we act differently?
Abel’s direction brings forecasting foundation models, temporal causal discovery, and spectral dynamics into one object: amortized system identification.
Evaluation has to evolve with the model. If benchmarks score only forecasting, a better world model and a slightly better forecaster can look identical. A meaningful benchmark should test prediction, explanation, intervention — and abstention when the evidence is weak.
This is the direction behind Abel:
Temporal Superintelligence — making the measurable world programmable.
kdd-milets.github.io/milets2…
We took SIMONS out of finance and into experimental biology and it recovered all 58 as directed causal source.
Then it surfaced 16 new candidates.
🤖 Made with AI
2
184
The ranking also tracked experimental influence.
The graph did not only recover names.
It recovered part of the experimental ordering.
1
29
Most AI models predict text.
SIMONS predicts state transitions.
SIMONS is Abel’s Temporal World Model.
Proving ground: finance.
Since live paper start:
Abel portfolio: +44.7%
QQQ:+24.9%
Excess:19.9%
This is why prediction needs a different substrate from language.
Text can be compressed.
Time has to be continued.
For reality-facing agents,the target is not the sequence itself.
It is the state transition that generates it.
I built the deep nets the whole field now uses as time-series baselines (FCN/ResNet). Then a bag of random convolutional kernels out-ran them — at a fraction of the compute.
I've sat with that for years. It isn't an engineering bug. It's an epistemic one.
Representation learning conquered vision and language because both rest on a compositional substrate: edges compose into objects, tokens into meaning. Stack layers, and depth pays off. So we assumed time worked the same way — and tokenized trajectories like sentences.
But a time series isn't compositional. It's spectral.
An image samples a static scene. A series samples a process. What you're representing was never the data — it's the evolution operator that emits it. Its features aren't stacked patches; they're the operator's modes: eigen-frequencies, decay rates, long memory.
Random kernels win because they never pretend there's a hierarchy to climb. They just blanket pattern-space — and accidentally respect that the structure is shallow-and-spectral, not deep-and-compositional.
Across markets, industry, and robotics, what's shared isn't a pattern. Maybe it's the branch cut.
One way out is - stop tokenizing trajectories. Time isn't read — it's continued.
2
132
Abel retweeted
I built the deep nets the whole field now uses as time-series baselines (FCN/ResNet). Then a bag of random convolutional kernels out-ran them — at a fraction of the compute.
I've sat with that for years. It isn't an engineering bug. It's an epistemic one.
Representation learning conquered vision and language because both rest on a compositional substrate: edges compose into objects, tokens into meaning. Stack layers, and depth pays off. So we assumed time worked the same way — and tokenized trajectories like sentences.
But a time series isn't compositional. It's spectral.
An image samples a static scene. A series samples a process. What you're representing was never the data — it's the evolution operator that emits it. Its features aren't stacked patches; they're the operator's modes: eigen-frequencies, decay rates, long memory.
Random kernels win because they never pretend there's a hierarchy to climb. They just blanket pattern-space — and accidentally respect that the structure is shallow-and-spectral, not deep-and-compositional.
Across markets, industry, and robotics, what's shared isn't a pattern. Maybe it's the branch cut.
One way out is - stop tokenizing trajectories. Time isn't read — it's continued.















