We present AlphaLab, an autonomous research harness that leverages frontier LLM agentic capabilities to automate the full experimental cycle in quantitative, computation-intensive domains. Given only a dataset and a natural-language objective, AlphaLab proceeds through three phases without human intervention: (1) it adapts to the domain and explores the data, writing analysis code and producing a research report; (2) it constructs and adversarially validates its own evaluation framework; and (3) it runs large-scale GPU experiments via a Strategist/Worker loop, accumulating domain knowledge in a persistent playbook that functions as a form of online prompt optimization. All domain-specific behavior is factored into adapters generated by the model itself, so the same pipeline handles qualitatively different tasks without modification. We evaluate AlphaLab with two frontier LLMs (GPT-5.2 and Claude Opus 4.6) on three domains: CUDA kernel optimization, where it writes GPU kernels that run 4.4x faster than torch.compile on average (up to 91x); LLM pretraining, where the full system achieves 22

Diffusion language models (DLMs) enable parallel, order-agnostic generation with iterative refinement, offering a flexible alternative to autoregressive large language models (LLMs). However, adapting reinforcement learning (RL) fine-tuning to DLMs remains an open challenge because of the intractable likelihood. Pioneering work such as diffu-GRPO estimated token-level likelihoods via one-step unmasking. While computationally efficient, this approach is severely biased. A more principled foundation lies in sequence-level likelihoods, where the evidence lower bound (ELBO) serves as a surrogate. Yet, despite this clean mathematical connection, ELBO-based methods have seen limited adoption due to the prohibitive cost of likelihood evaluation. In this work, we revisit ELBO estimation and disentangle its sources of variance. This decomposition motivates reducing variance through fast, deterministic integral approximations along a few pivotal dimensions. Building on this insight, we introduce Group Diffusion Policy Optimization (GDPO), a new RL algorithm tailored for DLMs. GDPO leverages simple yet effective Semi-deterministic Monte Carlo schemes to mitigate the variance explosion of ELBO estimators under vanilla double Monte Carlo sampling, yielding a provably lower-variance estimator under tight evaluation budgets. Empirically, GDPO achieves consistent gains over pretrained checkpoints and outperforms diffu-GRPO, one of the state-of-the-art baselines, on the majority of math, reasoning, and coding benchmarks.

A bstract Genomic foundation models have emerged alongside the rapid progress of large language models, offering a promising framework for learning general-purpose sequence priors for DNA understanding, generation, and design. This connection to LLMs creates a major opportunity: modern architectures, scaling infrastructure, autoregressive training, and token-based modeling provide powerful tools for genomic sequence modeling. At the same time, DNA differs fundamentally from natural language. Genomic sequences are noisy, redundant, sparsely constrained, unevenly annotated, and shaped by evolutionary rather than communicative pressures. As a result, key components of the standard LLM recipe, including data construction, tokenization, and training objectives, must be reconsidered in the biological sequence setting. A central challenge in DNA modeling is reconciling single-nucleotide resolution with long-context reasoning. Single-nucleotide resolution is essential for variant effect prediction, splice-site analysis, and codon-level reasoning. Long-context modeling is equally important, as many genomic mechanisms depend on distal regulatory elements, gene neighborhoods, and long-range evolutionary constraints. However, the most direct path to nucleotide-level reasoning, single-nucleotide tokenization, makes genomic sequences extremely long and imposes substantial computational cost on Transformer models. We present C arbon , a family of efficient generative DNA language models designed as a practical reference point for this setting. C arbon includes 3B- and 8B-parameter decoder-only autoregressive models using non-overlapping 6-mer tokenization. C arbon -3B supports a maximum context length of 65,536 tokens, corresponding to approximately 393 kbp of DNA; C arbon -8B supports up to 131,072 tokens, roughly 786 kbp. This simple and controlled setup helps isolate a central question for DNA language modeling: whether current progress is limited primarily by model architecture and nominal context length, or by more basic alignment between data, tokenization, objectives, evaluation, and the biological structure of genomic sequence. In our training-free evaluation suite, C arbon -3B is competitive with Evo2-7B despite having less than half the parameters. C arbon -8B improves on C arbon -3B on every training-free task, with the largest gain on long-context retrieval. Both models deliver tens-fold faster inference under comparable settings. The C arbon recipe combines annotation-aware data curation, deterministic 6-mer tokenization, and a staged CE-to-FNS objective schedule, adapting the LLM recipe to the statistical and biological properties of DNA rather than directly transplanting it. We release the models, data, training code, and evaluation suite, including new training-free probes for sequence-level perturbation and DNA long-context retrieval. C arbon is intended as an open recipe for efficient generative DNA modeling rather than an argument for any specific architecture, tokenization strategy, or objective design as the optimal solution. Its strong performance provides grounded evidence that substantial room remains for domain-aware model design carefully aligned with the genomic sequence itself.

We present a novel framework for online learning in Stackelberg general-sum games, where two agents, the leader and follower, engage in sequential turn-based interactions. At the core of this approach is a learned diffeomorphism that maps the joint action space to a smooth spherical Riemannian manifold, referred to as the Stackelberg manifold. This mapping, facilitated by neural normalizing flows, ensures the formation of tractable isoplanar subspaces, enabling efficient techniques for online learning. Leveraging the linearity of the agents' reward functions on the Stackelberg manifold, our construct allows the application of linear bandit algorithms. We then provide a rigorous theoretical basis for regret minimization on the learned manifold and establish bounds on the simple regret for learning Stackelberg equilibrium. This integration of manifold learning into game theory uncovers a previously unrecognized potential for neural normalizing flows as an effective tool for multi-agent learning. We present empirical results demonstrating the effectiveness of our approach compared to standard baselines, with applications spanning domains such as cybersecurity and economic supply chain optimization.

Time-series foundation models (TSFMs) are increasingly explored as predictive experts within emerging agentic time-series systems. However, TSFMs exhibit heterogeneous inductive biases, and no single model consistently dominates across forecasting regimes, making expert selection a critical challenge. Existing systems often delegate this decision to LLM-based controllers, incurring substantial inference overhead. We present TimeRouter, an efficient routing framework that leverages empirical complementarity across a pool of pretrained TSFMs through lightweight discriminative routing, selective gating, and ensemble fallback. Concretely, TimeRouter combines a learned routing head, a selective gate, and an ensemble fallback, enabling adaptive expert selection without invoking an LLM at inference time. TimeRouter achieves state-of-the-art performance on the GIFT-EVAL leaderboard, with an LB MASE of 0.6765. Beyond benchmark performance, our ablation studies provide empirical insights into TSFM routing design, highlighting the importance of pool composition and selective gating. Taken together, these results position TimeRouter as a modular and lightweight routing layer for future agentic time-series systems built upon foundation-model pools. Our code is available at https://github.com/UConn-DSIS/TimeRouter.

Mixture-of-Experts (MoE) has become a cornerstone in recent state-of-the-art large language models (LLMs). Traditionally, MoE relies on $\mathrm{Softmax}$ as the router score function to aggregate expert output, a designed choice that has persisted from the earliest MoE models to modern LLMs, and is now widely regarded as standard practice. However, the necessity of using $\mathrm{Softmax}$ to project router weights into a probability simplex remains an unchallenged assumption rather than a principled design choice. In this work, we first revisit the classical Nadaraya–Watson regression and observe that MoE shares the same mathematical formulation as Nadaraya–Watson regression. Furthermore, we show that both feed-forward neural network (FFN) and Mixture-of-Experts (MoE) can be interpreted as a special case of Nadaraya–Watson regression, where the kernel function corresponds to the input neurons of the output layer. Motivated by these insights, we propose the **zero-additional-cost** Kernel Inspired Router with Normalization ($\mathrm{KERN}$), an FFN-style router function, as an alternative to $\mathrm{Softmax}$. We demonstrate that this router generalizes both $\mathrm{Sigmoid}$- and $\mathrm{Softmax}$-based routers. **Based on empirical observations and established practices in FFN implementation, we recommend the use of $\mathrm{ReLU}$ activation and $\ell_2$-normalization in $\mathrm{KERN}$ router function.** Comprehensive experiments in MoE and LLM validate the effectiveness of the proposed FFN-style router function $\mathrm{KERN}$.

Modern preference alignment methods, such as DPO, rely on divergence regularization to a reference model for training stability—but this creates a fundamental problem we call "reference mismatch." In this paper, we investigate the negative impacts of reference mismatch in aligning text-to-image (T2I) diffusion models, showing that larger reference mismatch hinders effective adaptation given the same amount of data, e.g., as when learning new artistic styles, or personalizing to specific objects. We demonstrate this phenomenon across text-to-image (T2I) diffusion models and introduce margin-aware preference optimization (MaPO), a reference-agnostic approach that breaks free from this constraint. By directly optimizing the likelihood margin between preferred and dispreferred outputs under the Bradley-Terry model without anchoring to a reference, MaPO transforms diverse T2I tasks into unified pairwise preference optimization. We validate MaPO's versatility across five challenging domains: (1) safe generation, (2) style adaptation, (3) cultural representation, (4) personalization, and (5) general preference alignment. Our results reveal that MaPO's advantage grows dramatically with reference mismatch severity, outperforming both DPO and specialized methods like DreamBooth while reducing training time by 15%. MaPO thus emerges as a versatile and memory-efficient method for generic T2I adaptation tasks.

Traditional Reinforcement Learning from Human Feedback (RLHF) often relies on reward models, frequently assuming preference structures like the Bradley–Terry model, which may not accurately capture the complexities of real human preferences (e.g., intransitivity). Nash Learning from Human Feedback (NLHF) offers a more direct alternative by framing the problem as finding a Nash equilibrium of a game defined by these preferences. While many works study the Nash learning problem directly in the policy space, we instead consider it under a more realistic policy parametrization setting. We first analyze a simple self-play policy gradient method, which is equivalent to Online IPO. We establish high-probability last-iterate convergence guarantees for this method, but our analysis also reveals a possible stability limitation of the underlying dynamics. Motivated by this, we embed the self-play updates into a proximal point framework, yielding a stabilized algorithm. For this combined method, we prove high-probability last-iterate convergence and discuss its more practical version, which we call Nash Prox. Finally, we apply this method to post-training of large language models and validate its empirical performance.

The capabilities of Large Vision-Language Models (LVLMs) have reached state-of-the-art on many visual reasoning tasks, including chart reasoning, yet they still falter on out-of-distribution (OOD) data, and degrade further when asked to produce their chain-of-thought (CoT) rationales, limiting explainability. We present Chart-RVR, a general framework that fine-tunes LVLMs to be more robust and explainable for chart reasoning by coupling Group Relative Policy Optimization (GRPO) with automatically verifiable rewards. Our framework comprises of three rewards that maximize: (i) correct chart-type classification, (ii) faithful chart table reconstruction, and (iii) process conformity. Applied to 3-billion-parameter LVLMs, Chart-RVR consistently outperforms standard supervised fine-tuning (SFT) on both in-distribution and out-of-distribution datasets, closing the OOD performance gap while improving rationale fidelity. The resulting models, the Chart-RVR-3B series, achieve state-of-the-art results on six chart-reasoning benchmarks spanning in-domain and OOD settings, surpassing all existing models of comparable size. Beyond accuracy, Chart-RVR yields more interpretable CoT rationales, strengthening trust and reliability - showcasing the power of verifiable rewards with GRPO for training reliable, interpretable chart-reasoning models.

Tokenization and transfer learning are two critical components in building state of the art time series foundation models for forecasting. In this work, we systematically study the effect of tokenizer design, specifically scaling and quantization strategies, on model performance, alongside the impact of pretraining versus random initialization. We show that tokenizer configuration primarily governs the representational capacity and stability of the model, while transfer learning influences optimization efficiency and alignment. Using a combination of empirical training experiments and theoretical analyses, we demonstrate that pretrained models consistently leverage well-designed tokenizers more effectively, particularly at smaller vocabulary sizes. Conversely, misaligned tokenization can diminish or even invert the benefits of pretraining. These findings highlight the importance of careful tokenization in time series modeling and suggest that combining small, efficient vocabularies with pretrained weights is especially advantageous in multi-modal forecasting settings, where the overall vocabulary must be shared across modalities. Our results provide concrete guidance for designing tokenizers and leveraging transfer learning in discrete representation learning for continuous signals.

Reinforcement Learning with Human Feedback (RLHF) enhances the alignment of Large Language Models (LLMs). However, its limitations have led to the development of Direct Preference Optimization (DPO), an RL-free approach designed to overcome these shortcomings. While studies have shown that DPO improves instruction-following capabilities, it negatively impacts the reasoning ability of LLMs. Additionally, DPO is highly sensitive to judgment noise in preference datasets and the size of the training set. Although several modifications to DPO have been proposed, they still fail to fully resolve these issues. To address these limitations, we propose Triple Preference Optimization (TPO), a new preference learning method designed to enhance both reasoning and instruction-following abilities through one-step optimization. We compare TPO against DPO and its recent variants using state-of-the-art training setups, including both base and instruction-tuned models such as Mistral and Llama 3. Our evaluation covers a comprehensive range of chat-based and reasoning benchmarks. The results demonstrate that TPO achieves significant improvements over existing methods without substantially increasing response length across different dataset sizes. Specifically, TPO outperforms DPO and SimPO by up to 7.0 10.4 respectively. Furthermore, TPO achieves these improvements while requiring less data than DPO.

Recent works have demonstrated the effectiveness of adapting pre-trained language models (LMs) for forecasting time series in the low-data regime. We build upon these findings by analyzing the effective transfer from language models to time series forecasting under various design choices including upstream post-training, time series tokenizer and language backbone size. In the low-data regime, these design choices have a significant impact on the validation loss, with clear-cut choices that outperform others. Contrary to Hernandez et al. (2021), we observe that the validation loss of the LMs continues to smoothly decrease long after the validation loss of the randomly initialized models has converged, leading to a non-vanishing transfer gap that holds across design choices. These findings not only help shed light on the effective use of compute-efficient training for time series, but also open the way for the study of modality-agnostic properties of data distributions leveraged by these models.

Large Language Models (LLMs) and Foundation Models (FMs) have recently become prevalent for time series forecasting tasks. While fine-tuning LLMs enables domain adaptation, they often struggle to generalize across diverse and unseen datasets. Moreover, existing Time Series Foundation Models (TSFMs) still face challenges in handling non-stationary dynamics and distribution shifts, largely due to the lack of effective mechanisms for adaptation. To this end, we present TS-RAG, a retrieval-augmented generation framework for time series forecasting that enhances the generalization and interpretability of TSFMs. Specifically, TS-RAG leverages pre-trained time series encoders to retrieve semantically relevant segments from a dedicated knowledge base, enriching the contextual representation of the input query. Furthermore, we propose an Adaptive Retrieval Mixer (ARM) module that dynamically fuses the retrieved patterns with the TSFM's internal representation, improving forecasting accuracy without requiring task-specific fine-tuning. Thorough empirical studies on seven public benchmark datasets demonstrate that TS-RAG achieves state-of-the-art zero-shot forecasting performance, outperforming the existing TSFMs by up to 6.84\% across diverse domains while also providing desirable interpretability. Our code and data are available at: https://github.com/UConn-DSIS/TS-RAG.

Recent studies in multivariate time series (MTS) forecasting reveal that explicitly modeling the hidden dependencies among different time series can yield promising forecasting performance and reliable explanations. However, modeling variable dependencies remains underexplored when MTS is continuously accumulated under different regimes (stages). Due to the potential distribution and dependency disparities, the underlying model may encounter the catastrophic forgetting problem, i.e., it is challenging to memorize and infer different types of variable dependencies across different regimes while maintaining forecasting performance. To address this issue, we propose a novel Structural Knowledge Informed Continual Learning (SKI-CL) framework to perform MTS forecasting within a continual learning paradigm, which leverages structural knowledge to steer the forecasting model toward identifying and adapting to different regimes, and selects representative MTS samples from each regime for memory replay. Specifically, we develop a forecasting model based on graph structure learning, where a consistency regularization scheme is imposed between the learned variable dependencies and the structural knowledge while optimizing the forecasting objective over the MTS data. As such, MTS representations learned in each regime are associated with distinct structural knowledge, which helps the model memorize a variety of conceivable scenarios and results in accurate forecasts in the continual learning context. Meanwhile, we develop a representation-matching memory replay scheme that maximizes the temporal coverage of MTS data to efficiently preserve the underlying temporal dynamics and dependency structures of each regime. Thorough empirical studies on synthetic and real-world benchmarks validate SKI-CL's efficacy and advantages over the state-of-the-art for continual MTS forecasting tasks.

Probabilistic time series forecasting is a challenging problem due to the long sequences involved, the large number of samples needed for accurate probabilistic inference, and the need for real-time inference in many applications. These challenges necessitate methods that are not only accurate but computationally efficient. Unfortunately, most current state-of-the-art methods for time series forecasting are based on Transformers, which scale poorly due to quadratic complexity in sequence length, and are therefore needlessly computationally inefficient. Moreover, with a few exceptions, these methods have only been evaluated for non-probabilistic point estimation. In this work, we address these two shortcomings. For the first, we introduce VQ-TR, which maps large sequences to a discrete set of latent representations as part of the Attention module. This not only allows us to attend over larger context windows with linear complexity in sequence length but also allows for effective regularization to avoid overfitting. For the second, we provide what is to the best of our knowledge the first systematic comparison of modern Transformer-based time series forecasting methods for probabilistic forecasting. In this comparison, we find that VQ-TR performs better or comparably to all other methods while being computationally efficient.

We aim to produce a smaller language model that is aligned to user intent. Previous research has shown that applying distilled supervised fine-tuning (dSFT) on larger models significantly improves task accuracy; however, these models are unaligned, i.e. they do not respond well to natural prompts. To distill this property, we experiment with the use of preference data from AI Feedback (AIF). Starting from a dataset of outputs ranked by a teacher model, we apply distilled direct preference optimization (dDPO) to learn a chat model with significantly improved intent alignment. The approach requires only a few hours of training without any additional sampling during fine-tuning. The final result, Zephyr-7B, sets the state-of-the-art on chat benchmarks for 7B parameter models, and requires no human annotation. In particular, results on MT-Bench show that Zephyr-7B surpasses Llama2-Chat-70B, the best open-access RLHF-based model. Code, models, data, and tutorials for the system are available at https://github.com/huggingface/alignment-handbook.

We introduce Hyper-Trees as a novel framework for modeling time series data using gradient boosted trees. Unlike conventional tree-based approaches that forecast time series directly, Hyper-Trees learn the parameters of a target time series model, such as ARIMA or Exponential Smoothing, as functions of features. These parameters are then used by the target model to generate the final forecasts. Our framework combines the effectiveness of decision trees on tabular data with classical forecasting models, thereby inducing a time series inductive bias into tree-based models. To resolve the scaling limitations of boosted trees when estimating a high-dimensional set of target model parameters, we combine decision trees and neural networks within a unified framework. In this hybrid approach, the trees generate informative representations from the input features, which a shallow network then uses as input to learn the parameters of a time series model. With our research, we explore the effectiveness of Hyper-Trees across a range of forecasting tasks and extend tree-based modeling beyond its conventional use in time series analysis.

Sequential models such as recurrent neural networks or transformer-based models became de facto tools for multivariate time series forecasting in a probabilistic fashion, with applications to a wide range of datasets, such as finance, biology, medicine, etc. Despite their adeptness in capturing dependencies, assessing prediction uncertainty, and efficiency in training, challenges emerge in modeling high-dimensional complex distributions and cross-feature dependencies. To tackle these issues, recent works delve into generative modeling by employing diffusion or flow-based models. Notably, the integration of stochastic differential equations or probability flow successfully extends these methods to probabilistic time series imputation and forecasting. However, scalability issues necessitate a computational-friendly framework for large-scale generative model-based predictions. This work proposes a novel approach by blending the computational efficiency of recurrent neural networks with the high-quality probabilistic modeling of the diffusion model, which addresses challenges and advances generative models' application in time series forecasting. Our method relies on the foundation of stochastic interpolants and the extension to a broader conditional generation framework with additional control features, offering insights for future developments in this dynamic field.

This work is the first to openly reproduce the Reinforcement Learning from Human Feedback (RLHF) scaling behaviors reported in OpenAI's seminal TL;DR summarization work. We create an RLHF pipeline from scratch, enumerate over 20 key implementation details, and share key insights during the reproduction. Our RLHF-trained Pythia models demonstrate significant gains in response quality that scale with model size, with our 2.8B, 6.9B models outperforming OpenAI's released 1.3B checkpoint. Our results highlight best practices in data, training, and evaluation for RLHF. We publicly release the trained model checkpoints and code to facilitate further research and accelerate progress in the field at https://github.com/vwxyzjn/summarize_from_feedback_details

Building on the remarkable achievements in generative sampling of natural images, we propose an innovative challenge, potentially overly ambitious, which involves generating samples of entire multivariate time series that resemble images. However, the statistical challenge lies in the small sample size, sometimes consisting of a few hundred subjects. This issue is especially problematic for deep generative models that follow the conventional approach of generating samples from a canonical distribution and then decoding or denoising them to match the true data distribution. In contrast, our method is grounded in information theory and aims to implicitly characterize the distribution of images, particularly the (global and local) dependency structure between pixels. We achieve this by empirically estimating its KL-divergence in the dual form with respect to the respective marginal distribution. This enables us to perform generative sampling directly in the optimized 1-D dual divergence space. Specifically, in the dual space, training samples representing the data distribution are embedded in the form of various clusters between two end points. In theory, any sample embedded between those two end points is in-distribution w.r.t. the data distribution. Our key idea for generating novel samples of images is to interpolate between the clusters via a walk as per gradients of the dual function w.r.t. the data dimensions. In addition to the data efficiency gained from direct sampling, we propose an algorithm that offers a significant reduction in sample complexity for estimating the divergence of the data distribution with respect to the marginal distribution. We provide strong theoretical guarantees along with an extensive empirical evaluation using many real-world datasets from diverse domains, establishing the superiority of our approach w.r.t. state-of-the-art deep learning methods.