Recent progress in Language Models (LMs) has dramatically advanced the field of natural language processing (NLP), excelling at tasks like text generation, summarization, and question answering. However, their inference remains computationally expensive and energy intensive, especially in settings with limited hardware, power, or bandwidth. This makes it difficult to deploy LMs in mobile, edge, or cost sensitive environments. To address these challenges, recent approaches have introduced multi LLM intelligent model selection strategies that dynamically allocate computational resources based on query complexity -- using lightweight models for simpler queries and escalating to larger models only when necessary. This survey explores two complementary strategies for efficient LLM inference: (i) routing, which selects the most suitable model based on the query, and (ii) cascading or hierarchical inference (HI), which escalates queries through a sequence of models until a confident response is found. Both approaches aim to reduce computation by using lightweight models for simpler tasks while offloading only when needed. We provide a comparative analysis of these techniques across key performance metrics, discuss benchmarking efforts, and outline open challenges. Finally, we outline future research directions to enable faster response times, adaptive model selection based on task complexity, and scalable deployment across heterogeneous environments, making LLM based systems more efficient and accessible for real world applications.

Multi-agent reinforcement learning (MARL) methods have achieved state-of-the-art results on a range of multi-agent tasks. Yet, MARL algorithms typically require significantly more environment interactions than their single-agent counterparts to converge, a problem exacerbated by the difficulty in exploring over a large joint action space and the high variance intrinsic to MARL environments. To tackle these issues, we propose a novel algorithm that combines a decomposed centralized critic with decentralized ensemble learning, incorporating several key contributions. The main component in our scheme is a selective exploration method that leverages ensemble kurtosis. We extend the global decomposed critic with a diversity-regularized ensemble of individual critics and utilize its excess kurtosis to guide exploration toward high-uncertainty states and actions. To improve sample efficiency, we train the centralized critic with a novel truncated variation of the TD($λ$) algorithm, enabling efficient off-policy learning with reduced variance. On the actor side, our suggested algorithm adapts the mixed samples approach to MARL, mixing on-policy and off-policy loss functions for training the actors. This approach balances between stability and efficiency and outperforms purely off-policy learning. The evaluation shows our method outperforms state-of-the-art baselines on standard MARL benchmarks, including a variety of SMAC II maps.

Large language models (LLMs)-empowered web agents enables automating complex, real-time web navigation tasks in enterprise environments. However, existing web agents relying on supervised fine-tuning (SFT) often struggle with generalization and robustness due to insufficient reasoning capabilities when handling the inherently dynamic nature of web interactions. In this study, we introduce WorkForceAgent-R1, an LLM-based web agent trained using a rule-based R1-style reinforcement learning framework designed explicitly to enhance single-step reasoning and planning for business-oriented web navigation tasks. We employ a structured reward function that evaluates both adherence to output formats and correctness of actions, enabling WorkForceAgent-R1 to implicitly learn robust intermediate reasoning without explicit annotations or extensive expert demonstrations. Extensive experiments on the WorkArena benchmark demonstrate that WorkForceAgent-R1 substantially outperforms SFT baselines by 10.26-16.59%, achieving competitive performance relative to proprietary LLM-based agents (gpt-4o) in workplace-oriented web navigation tasks.

We consider the problem of sampling from a probability distribution $π$. It is well known that this can be written as an optimisation problem over the space of probability distribution in which we aim to minimise the Kullback--Leibler divergence from $π$. We consider several partial differential equations (PDEs) whose solution is a minimiser of the Kullback--Leibler divergence from $π$ and connect them to well-known Monte Carlo algorithms. We focus in particular on PDEs obtained by considering the Wasserstein--Fisher--Rao geometry over the space of probabilities and show that these lead to a natural implementation using importance sampling and sequential Monte Carlo. We propose a novel algorithm to approximate the Wasserstein--Fisher--Rao flow of the Kullback--Leibler divergence which empirically outperforms the current state-of-the-art. We study tempered versions of these PDEs obtained by replacing the target distribution with a geometric mixture of initial and target distribution and show that these do not lead to a convergence speed up.

Self-attention has emerged as a core component of modern neural architectures, yet its theoretical underpinnings remain elusive. In this paper, we study self-attention through the lens of interacting entities, ranging from agents in multi-agent reinforcement learning to alleles in genetic sequences, and show that a single layer linear self-attention can efficiently represent, learn, and generalize functions capturing pairwise interactions, including out-of-distribution scenarios. Our analysis reveals that self-attention acts as a mutual interaction learner under minimal assumptions on the diversity of interaction patterns observed during training, thereby encompassing a wide variety of real-world domains. In addition, we validate our theoretical insights through experiments demonstrating that self-attention learns interaction functions and generalizes across both population distributions and out-of-distribution scenarios. Building on our theories, we introduce HyperFeatureAttention, a novel neural network module designed to learn couplings of different feature-level interactions between entities. Furthermore, we propose HyperAttention, a new module that extends beyond pairwise interactions to capture multi-entity dependencies, such as three-way, four-way, or general n-way interactions.

Offline reinforcement learning (RL) is crucial when online exploration is costly or unsafe but often struggles with high epistemic uncertainty due to limited data. Existing methods rely on fixed conservative policies, restricting adaptivity and generalization. To address this, we propose Reflect-then-Plan (RefPlan), a novel doubly Bayesian offline model-based (MB) planning approach. RefPlan unifies uncertainty modeling and MB planning by recasting planning as Bayesian posterior estimation. At deployment, it updates a belief over environment dynamics using real-time observations, incorporating uncertainty into MB planning via marginalization. Empirical results on standard benchmarks show that RefPlan significantly improves the performance of conservative offline RL policies. In particular, RefPlan maintains robust performance under high epistemic uncertainty and limited data, while demonstrating resilience to changing environment dynamics, improving the flexibility, generalizability, and robustness of offline-learned policies.

Constrained Reinforcement Learning (CRL) addresses sequential decision-making problems where agents are required to achieve goals by maximizing the expected return while meeting domain-specific constraints. In this setting, policy-based methods are widely used thanks to their advantages when dealing with continuous-control problems. These methods search in the policy space with an action-based or a parameter-based exploration strategy, depending on whether they learn the parameters of a stochastic policy or those of a stochastic hyperpolicy. We introduce an exploration-agnostic algorithm, called C-PG, which enjoys global last-iterate convergence guarantees under gradient domination assumptions. Furthermore, under specific noise models where the (hyper)policy is expressed as a stochastic perturbation of the actions or of the parameters of an underlying deterministic policy, we additionally establish global last-iterate convergence guarantees of C-PG to the optimal deterministic policy . This holds when learning a stochastic (hyper)policy and subsequently switching off the stochasticity at the end of training, thereby deploying a deterministic policy. Finally, we empirically validate both the action-based ( C-PGAE) and parameter-based ( C-PGPE) variants of C-PG on constrained control tasks, and compare them against state-of-the-art baselines, demonstrating their effectiveness, in particular when deploying deterministic policies after training.

We study the problem of peeling a Velcro strap from a surface using a robotic manipulator. The surface geometry is arbitrary and unknown. The robot has access to only the force feedback and its end-effector position. This problem is challenging due to the partial observability of the environment and the incompleteness of the sensor feedback. To solve it, we first model the system with simple analytic state and action models based on quasi-static dynamics assumptions. We then study the fully-observable case where the state of both the Velcro and the robot are given. For this case, we obtain the optimal solution in closed-form which minimizes the total energy cost. Next, for the partially-observable case, we design a state estimator which estimates the underlying state using only force and position feedback. Then, we present a heuristics-based controller that balances exploratory and exploitative behaviors in order to peel the velcro efficiently. Finally, we evaluate our proposed method in environments with complex geometric uncertainties and sensor noises, achieving 100% success rate with less than 80% increase in energy cost compared to the optimal solution when the environment is fully-observable, outperforming the baselines by a large margin.

Constrained decoding enables Language Models (LMs) to produce samples that provably satisfy hard constraints. However, existing constrained-decoding approaches often distort the underlying model distribution, a limitation that is especially problematic in applications like program fuzzing, where one wants to generate diverse and valid program inputs for testing purposes. We propose a new constrained sampling framework based on Markov Chain Monte Carlo (MCMC) that simultaneously satisfies three core desiderata: constraint satisfying (every sample satisfies the constraint), monotonically converging (the sampling process converges to the true conditional distribution), and efficient (high-quality samples emerge in few steps). Our method constructs a proposal distribution over valid outputs and applies a Metropolis-Hastings acceptance criterion based on the LM's likelihood, ensuring principled and efficient exploration of the constrained space. Empirically, our sampler outperforms existing methods on both synthetic benchmarks and real-world program fuzzing tasks.

Biological and psychological concepts have inspired reinforcement learning algorithms to create new complex behaviors that expand agents' capacity. These behaviors can be seen in the rise of techniques like goal decomposition, curriculum, and intrinsic rewards, which have paved the way for these complex behaviors. One limitation in evaluating these methods is the requirement for engineered extrinsic for realistic environments. A central challenge in engineering the necessary reward function(s) comes from these environments containing states that carry high negative rewards, but provide no feedback to the agent. Death is one such stimuli that fails to provide direct feedback to the agent. In this work, we introduce an intrinsic reward function inspired by early amygdala development and produce this intrinsic reward through a novel memory-augmented neural network (MANN) architecture. We show how this intrinsic motivation serves to deter exploration of terminal states and results in avoidance behavior similar to fear conditioning observed in animals. Furthermore, we demonstrate how modifying a threshold where the fear response is active produces a range of behaviors that are described under the paradigm of general anxiety disorders (GADs). We demonstrate this behavior in the Miniworld Sidewalk environment, which provides a partially observable Markov decision process (POMDP) and a sparse reward with a non-descriptive terminal condition, i.e., death. In effect, this study results in a biologically-inspired neural architecture and framework for fear conditioning paradigms; we empirically demonstrate avoidance behavior in a constructed agent that is able to solve environments with non-descriptive terminal conditions.