Agentic Fine-Tuning of Frontier Models

We adapt the strongest open-weight models available today with training recipes that improve agentic behavior while preserving the reasoning that made the base model worth choosing. Getting a frontier model to behave like a great agent is a training problem, not a prompting problem: our recipes combine reinforcement learning, supervised fine-tuning, and preference optimization — to teach the habits that separate good models from great ones: which tool to reach for, when to verify, when to stop.

Most fine-tuning quietly overwrites a model's reasoning in exchange for surface-level task compliance — you gain a format and lose a mind. Our approach protects the model's native thinking by construction, and spends the training budget on data that actually improves the model: demonstrations are screened to remove examples that game the task rather than solve it. The same recipes run on our own GPU clusters and on hosted training APIs alike, and the strongest results go back to the community on Hugging Face.

LiveBench and Evaluation

Benchmark scores mean little if the test set leaked into the training data. That is why we built LiveBench, our contamination-limited benchmark published as an ICLR Spotlight and regarded industry-wide as a go-to benchmark for evaluating LLMs for real-world use. New questions are released every month — drawn from recent math competitions such as AMC12, AIME, and IMO, LeetCode and AtCoder problems, Zebra Puzzles, Connections word puzzles, fresh arXiv abstracts, news articles, and Kaggle datasets — so models are always measured on questions they cannot have memorized.

Every answer is scored against verifiable, objective ground truth rather than by an LLM judge, across 17 tasks spanning six categories: reasoning, coding, math, data analysis, language comprehension, and instruction following. LiveBench evaluates prominent closed-source models alongside dozens of open-weight models, and it keeps our own fine-tuning honest — every recipe we ship is validated on it.

Past Work: Open Models and Benchmarks

Much of the work above rests on research we published in earlier years. We introduced DPO-Positive, which fixed a failure mode of preference optimization and produced the original Smaug models, and Giraffe, our work on expanding context lengths in open LLMs. We studied data contamination in code generation, showing how model performance degrades on problems released after a training cutoff — the class of finding that LiveBench is designed to guard against. And in tabular deep learning, our NeurIPS study mapped when neural networks actually outperform gradient-boosted trees across 176 datasets.

Time-Series Forecasting

Time-series forecasting is a ubiquitous problem in the industry, used for problems such as forecasting inventory demand and predicting stock prices. While statistical approaches such as ARIMA and Prophet are widely used, deep learning models such as transformers are getting increasingly more popular. However, many high-impact forecasting problems are either low-data, or low signal-to-noise ratio, and many forecastable datasets have standard attributes that carry over to other datasets.

With this intuition in mind, Abacus.AI has developed a radically new approach to time-series forecasting. Inspired by recent innovations in meta-learning and Bayesian inference on tabular data, we have designed the world's first foundation model for time-series forecasting. To achieve this, we pretrain a transformer on a mix of real-world and synthetic datasets, across a variety of different forecasting tasks, resulting in a state-of-the-art forecasting model that can run inference on a new dataset in less than a second. This work was published at NeurIPS as ForecastPFN, the first zero-shot forecasting model trained purely on synthetic data.

Deep recommender systems

Which recommender algorithm should you use? In the paper below, we ran one of the largest evaluations of recommender systems to date — dozens of algorithms across dozens of datasets and metrics — and found that the best algorithm depends heavily on the dataset and metric at hand, with well-tuned classical baselines often beating modern deep learning approaches. Building on this study, we developed a meta-learning approach that predicts the best algorithm and hyperparameters for a new dataset.

Training with Less Data

Real-world datasets are often small, noisy, and imbalanced, so we develop techniques that let models learn from less data. One of these is generative dataset augmentation: we build on Data Augmentation Generative Adversarial Networks to create synthetic examples that are combined with the original training data, producing more robust models on sparse or incomplete datasets.

Another is learning from expert knowledge, not just examples. Enterprises encode years of domain expertise as rules, and models trained on sparse data often struggle to beat those rule-based systems. In the paper below, we introduce differentiable loss functions that fold first-order logic constraints directly into deep learning training via weighted proof tracing, so a model can learn from its training data and from general knowledge rules at the same time.

AI-Assisted ML

Deep learning has seen great success across a wide variety of domains. The best neural architectures are often carefully constructed by seasoned deep learning experts in each domain. For example, years of experimentation have shown how to arrange bidirectional transformers to work well for language tasks and dilated separable convolutions for image tasks. A relatively new sub-field of deep-learning deals with automated machine learning, or as we prefer to call it: AI-assisted machine learning. The fundamental idea is that AI will create a first pass of the deep-learning model given a use-case or a dataset. Developers/data scientists can then either use that model directly or fine-tune. We are conducting cutting-edge research in the main pillars of AI-Assisted ML: hyperparameter optimization (HPO) and neural architecture search (NAS).

When developing a deep learning model, there are many knobs and dials to tune that depend on the specific task and dataset at hand. For example, setting the learning rate too high can prevent the algorithm from converging. Setting the learning rate too low can cause the algorithm to get stuck at a local minimum. There are countless other hyperparameters such as the number of epochs, batch size, momentum, regularization, shape, and size of the neural network. These hyperparameters are all dependent on each other and interact in intricate ways, so finding the best hyperparameters for a given dataset is an extremely difficult and highly nonconvex optimization problem.

Randomly testing different sets of hyperparameters may eventually find a decent solution but could take years of computation time. Efficiently tuning deep learning hyperparameters is an active area of research. Five years ago, the best algorithms weren't much better than random search. Now algorithms are capable of orders of magnitude speedups. At Abacus.AI, we use state-of-the-art HPO while training all our models.

Neural architecture search (NAS) is a rapidly developing area of research in which the process of choosing the best architecture is automated.

At Abacus.AI, we are using NAS to both fine-tune proven deep network paradigms, and learn novel architectures for new domains. Our goal is to empower data scientists and developers to create custom, production-grade models in days, not months. See this blog post to read about our method, BANANAS, which combines Bayesian optimization with neural predictors to achieve state-of-the-art performance. Since making our code open-source, dozens of developers have forked our repository, and two independent research groups have confirmed that it achieves state-of-the-art performance on NAS-Bench-101. BANANAS has even been cited in survey papers on NAS.

We are also actively conducting fundamental research on the theory of NAS. Recently, we studied local search for NAS - a simple yet effective approach. We showed experimentally that local search gives state-of-the-art performance on smaller benchmark NAS search spaces, but performs worse than random search on extremely large search spaces. Motivated by this stark contrast, we gave a complete theoretical characterization of local search. Our theoretical results confirm that local search performs well on smaller search spaces and when the search space exhibits locality.

Finally, we are conducting formal studies on the building blocks of NAS, including the architecture encoding. In most NAS algorithms, the neural architectures must be passed as input to the algorithm using some encoding. For example, we might encode the neural architectures using an adjacency matrix. Our recent work shows that this encoding can have a substantial impact on the final result of the NAS algorithm. We conduct a set of experiments with eight different encodings with various NAS algorithms. Our results lay out recommendations for the best encodings to use in different settings within NAS.

Deep learning models deployed in high-stakes settings often exhibit bias, and retraining a fairer model from scratch is not always practical — training is expensive, and the original training data may no longer be available. The two papers below debias pretrained models directly. In "Intra-Processing Methods for Debiasing Neural Networks" (NeurIPS), we introduce three techniques that take a trained model and a small validation dataset and remove bias through fine-tuning or post-processing, with strong results on tabular and structured data. In "DECO", we frame debiasing as a compositional optimization problem, jointly trading off accuracy against fairness metrics, with a focus on models used in financial services.

An AI Super Assistant For Enterprises And Professionals