Plan Before You Trade: Inference-Time Optimization for RL Trading Agents

Reinforcement learning agents for portfolio management are typically trained and deployed as static policies, with no mechanism for using price forecasts at inference time. We propose $\text{FPILOT}$ (**Fin**ancial **P**lugin **I**nference-time **L**earning for **O**ptimal **T**rading), a plugin inference-time optimization framework inspired by Model Predictive Control (MPC). Our key structural insight is that future prices mostly do not depend on one agent's portfolio allocation, so a suitable predictive model can produce a multi-step price trajectory without iterative action-conditioned rollouts as in typical reinforcement learning. At each decision step, we use the forecaster's predicted price trajectory to construct an allocation-based imagined return objective, and optimize the policy at inference-time before executing one step of the trade. Our framework is compatible with any pre-trained agent and adapts the policy to the forecaster's predictions without any retraining. Evaluated across five policy learning algorithms on the TradeMaster DJ30 benchmark, $\text{FPILOT}$ produces consistent improvements in total return and return-based risk-adjusted metrics (Sharpe, Sortino, Calmar), with stochastic policies benefiting more than deterministic ones. Further, using synthetic forecasts at calibrated quality levels, we show that gains consistently improve with forecaster quality, suggesting that our performance will improve based on advances in financial forecasting.

Replicable Bandits with UCB based Exploration

We study replicable algorithms for stochastic multi-armed bandits (MAB) and linear bandits with UCB (Upper Confidence Bound) based exploration. A bandit algorithm is $ρ$-replicable if two executions using shared internal randomness but independent reward realizations, produce the same action sequence with probability at least $1-ρ$. Prior work is primarily elimination-based and, in linear bandits with infinitely many actions, relies on discretization, leading to suboptimal dependence on the dimension $d$ and $ρ$. We develop optimistic alternatives for both settings. For stochastic multi-armed bandits, we propose RepUCB, a replicable batched UCB algorithm and show that it attains a regret $O\!\left(\frac{K^2\log^2 T}{ρ^2}\sum_{a:Δ_a>0}\left(Δ_a+\frac{\log(KT\log T)}{Δ_a}\right)\right)$. For stochastic linear bandits, we first introduce RepRidge, a replicable ridge regression estimator that satisfies both a confidence guarantee and a $ρ$-replicability guarantee. Beyond its role in our bandit algorithm, this estimator and its guarantees may also be of independent interest in other statistical estimation settings. We then use RepRidge to design RepLinUCB, a replicable optimistic algorithm for stochastic linear bandits, and show that its regret is bounded by $\widetilde{O}\!\big(\big(d+\frac{d^3}ρ\big)\sqrt{T}\big)$. This improves the best prior regret guarantee by a factor of $O(d/ρ)$, showing that our optimistic algorithm can substantially reduce the price of replicability.

Model Predictive Control with Differentiable World Models for Offline Reinforcement Learning

Offline Reinforcement Learning (RL) aims to learn optimal policies from fixed offline datasets, without further interactions with the environment. Such methods train an offline policy (or value function), and apply it at inference time without further refinement. We introduce an inference time adaptation framework inspired by model predictive control (MPC) that utilizes a pretrained policy along with a learned world model of state transitions and rewards. While existing world model and diffusion-planning methods use learned dynamics to generate imagined trajectories during training, or to sample candidate plans at inference time, they do not use inference-time information to optimize the policy parameters on the fly. In contrast, our design is a Differentiable World Model (DWM) pipeline that enables endto-end gradient computation through imagined rollouts for policy optimization at inference time based on MPC. We evaluate our algorithm on D4RL continuous-control benchmarks (MuJoCo locomotion tasks and AntMaze), and show that exploiting inference-time information to optimize the policy parameters yields consistent gains over strong offline RL baselines.

Gradual Fine-Tuning for Flow Matching Models

Fine-tuning flow matching models is a central challenge in settings with limited data, evolving distributions, or strict efficiency demands, where unconstrained fine-tuning can erode the accuracy and efficiency gains learned during pretraining. Prior work has produced theoretical guarantees and empirical advances for reward-based fine-tuning formulations, but these methods often impose restrictions on permissible drift structure or training techniques. In this work, we propose Gradual Fine-Tuning (GFT), a principled framework for fine-tuning flow-based generative models when samples from the target distribution are available. For stochastic flows, GFT defines a temperature-controlled sequence of intermediate objectives that smoothly interpolate between the pretrained and target drifts, approaching the true target as the temperature approaches zero. We prove convergence results for both marginal and conditional GFT objectives, enabling the use of suitable (e.g., optimal transport) couplings during GFT while preserving correctness. Empirically, GFT improves convergence stability and shortens probability paths, resulting in faster inference, while maintaining generation quality comparable to standard fine-tuning. Our results position GFT as a theoretically grounded and practically effective alternative for scalable adaptation of flow matching models under distribution shift.

Sketched Gaussian Mechanism for Private Federated Learning

Communication cost and privacy are two major considerations in federated learning (FL). For communication cost, gradient compression by sketching the clients' transmitted model updates is often used for reducing per-round communication. For privacy, the Gaussian mechanism (GM), which consists of clipping updates and adding Gaussian noise, is commonly used to guarantee client-level differential privacy. Existing literature on private FL analyzes privacy of sketching and GM in an isolated manner, illustrating that sketching provides privacy determined by the sketching dimension and that GM has to supply any additional desired privacy. In this paper, we introduce the Sketched Gaussian Mechanism (SGM), which directly combines sketching and the Gaussian mechanism for privacy. Using R\'enyi-DP tools, we present a joint analysis of SGM's overall privacy guarantee, which is significantly more flexible and sharper compared to isolated analysis of sketching and GM privacy. In particular, we prove that the privacy level of SGM for a fixed noise magnitude is proportional to $1/\sqrt{b}$, where $b$ is the sketching dimension, indicating that (for moderate $b$) SGM can provide much stronger privacy guarantees than the original GM under the same noise budget. We demonstrate the application of SGM to FL with either gradient descent or adaptive server optimizers, and establish theoretical results on optimization convergence, which exhibits only a logarithmic dependence on the number of parameters $d$. Experimental results confirm that at the same privacy level, SGM based FL is at least competitive with non-sketching private FL variants and outperforms them in some settings. Moreover, using adaptive optimization at the server improves empirical performance while maintaining the privacy guarantees.

Conservative Contextual Bandits: Beyond Linear Representations

Conservative Contextual Bandits (CCBs) address safety in sequential decision making by requiring that an agent's policy, along with minimizing regret, also satisfies a safety constraint: the performance is not worse than a baseline policy (e.g., the policy that the company has in production) by more than $(1+\alpha)$ factor. Prior work developed UCB-style algorithms in the multi-armed [Wu et al., 2016] and contextual linear [Kazerouni et al., 2017] settings. However, in practice the cost of the arms is often a non-linear function, and therefore existing UCB algorithms are ineffective in such settings. In this paper, we consider CCBs beyond the linear case and develop two algorithms $\mathtt{C-SquareCB}$ and $\mathtt{C-FastCB}$, using Inverse Gap Weighting (IGW) based exploration and an online regression oracle. We show that the safety constraint is satisfied with high probability and that the regret of $\mathtt{C-SquareCB}$ is sub-linear in horizon $T$, while the regret of $\mathtt{C-FastCB}$ is first-order and is sub-linear in $L^*$, the cumulative loss of the optimal policy. Subsequently, we use a neural network for function approximation and online gradient descent as the regression oracle to provide $\tilde{O}(\sqrt{KT} + K/\alpha) $ and $\tilde{O}(\sqrt{KL^*} + K (1 + 1/\alpha))$ regret bounds, respectively. Finally, we demonstrate the efficacy of our algorithms on real-world data and show that they significantly outperform the existing baseline while maintaining the performance guarantee.

Sketched Adaptive Federated Deep Learning: A Sharp Convergence Analysis

Combining gradient compression methods (e.g., CountSketch, quantization) and adaptive optimizers (e.g., Adam, AMSGrad) is a desirable goal in federated learning (FL), with potential benefits on both fewer communication rounds and less per-round communication. In spite of the preliminary empirical success of sketched adaptive methods, existing convergence analyses show the communication cost to have a linear dependence on the ambient dimension, i.e., number of parameters, which is prohibitively high for modern deep learning models. In this work, we introduce specific sketched adaptive federated learning (SAFL) algorithms and, as our main contribution, provide theoretical convergence analyses in different FL settings with guarantees on communication cost depending only logarithmically (instead of linearly) on the ambient dimension. Unlike existing analyses, we show that the entry-wise sketching noise existent in the preconditioners and the first moments of SAFL can be implicitly addressed by leveraging the recently-popularized anisotropic curvatures in deep learning losses, e.g., fast decaying loss Hessian eigen-values. In the i.i.d. client setting of FL, we show that SAFL achieves asymptotic $O(1/\sqrt{T})$ convergence, and converges faster in the initial epochs. In the non-i.i.d. client setting, where non-adaptive methods lack convergence guarantees, we show that SACFL (SAFL with clipping) algorithms can provably converge in spite of the additional heavy-tailed noise. Our theoretical claims are supported by empirical studies on vision and language tasks, and in both fine-tuning and training-from-scratch regimes. Surprisingly, as a by-product of our analysis, the proposed SAFL methods are competitive with the state-of-the-art communication-efficient federated learning algorithms based on error feedback.

On the Generalizability of Foundation Models for Crop Type Mapping

Foundation models pre-trained using self-supervised and weakly-supervised learning have shown powerful transfer learning capabilities on various downstream tasks, including language understanding, text generation, and image recognition. Recently, the Earth observation (EO) field has produced several foundation models pre-trained directly on multispectral satellite imagery (e.g., Sentinel-2) for applications like precision agriculture, wildfire and drought monitoring, and natural disaster response. However, few studies have investigated the ability of these models to generalize to new geographic locations, and potential concerns of geospatial bias -- models trained on data-rich developed countries not transferring well to data-scarce developing countries -- remain. We investigate the ability of popular EO foundation models to transfer to new geographic regions in the agricultural domain, where differences in farming practices and class imbalance make transfer learning particularly challenging. We first select six crop classification datasets across five continents, normalizing for dataset size and harmonizing classes to focus on four major cereal grains: maize, soybean, rice, and wheat. We then compare three popular foundation models, pre-trained on SSL4EO-S12, SatlasPretrain, and ImageNet, using in-distribution (ID) and out-of-distribution (OOD) evaluation. Experiments show that pre-trained weights designed explicitly for Sentinel-2, such as SSL4EO-S12, outperform general pre-trained weights like ImageNet. Furthermore, the benefits of pre-training on OOD data are the most significant when only 10--100 ID training samples are used. Transfer learning and pre-training with OOD and limited ID data show promising applications, as many developing regions have scarce crop type labels. All harmonized datasets and experimental code are open-source and available for download.

Optimization and Generalization Guarantees for Weight Normalization

Weight normalization (WeightNorm) is widely used in practice for the training of deep neural networks and modern deep learning libraries have built-in implementations of it. In this paper, we provide the first theoretical characterizations of both optimization and generalization of deep WeightNorm models with smooth activation functions. For optimization, from the form of the Hessian of the loss, we note that a small Hessian of the predictor leads to a tractable analysis. Thus, we bound the spectral norm of the Hessian of WeightNorm networks and show its dependence on the network width and weight normalization terms--the latter being unique to networks without WeightNorm. Then, we use this bound to establish training convergence guarantees under suitable assumptions for gradient decent. For generalization, we use WeightNorm to get a uniform convergence based generalization bound, which is independent from the width and depends sublinearly on the depth. Finally, we present experimental results which illustrate how the normalization terms and other quantities of theoretical interest relate to the training of WeightNorm networks.

Loss Gradient Gaussian Width based Generalization and Optimization Guarantees

Generalization and optimization guarantees on the population loss in machine learning often rely on uniform convergence based analysis, typically based on the Rademacher complexity of the predictors. The rich representation power of modern models has led to concerns about this approach. In this paper, we present generalization and optimization guarantees in terms of the complexity of the gradients, as measured by the Loss Gradient Gaussian Width (LGGW). First, we introduce generalization guarantees directly in terms of the LGGW under a flexible gradient domination condition, which we demonstrate to hold empirically for deep models. Second, we show that sample reuse in finite sum (stochastic) optimization does not make the empirical gradient deviate from the population gradient as long as the LGGW is small. Third, focusing on deep networks, we present results showing how to bound their LGGW under mild assumptions. In particular, we show that their LGGW can be bounded (a) by the $L_2$-norm of the loss Hessian eigenvalues, which has been empirically shown to be $\tilde{O}(1)$ for commonly used deep models; and (b) in terms of the Gaussian width of the featurizer, i.e., the output of the last-but-one layer. To our knowledge, our generalization and optimization guarantees in terms of LGGW are the first results of its kind, avoid the pitfalls of predictor Rademacher complexity based analysis, and hold considerable promise towards quantitatively tight bounds for deep models.