arXiv subjects

Gauri Joshi

Publications and source records attributed to Gauri Joshi.

At least 19 recordsLinked to original sources

arXiv2026

Creating Power Flexibility in Datacenters via Query Routing

The growing energy requirements of large language model inference highlight the need for datacenters to support safe grid operation by providing power demand flexibility, while maintaining service for interactive workloads. This work investigates query routing as a mechanism to enable such capability by dynamically allocating queries across different language model instances without deferring requests. We develop a general analytical framework that connects query arrival rates and sequence length distributions to token iteration time, request residence time, serving capacity, and power consumption of model instances. Leveraging this framework, we characterize the feasible query routing allocations and the corresponding power demand flexibilities, enabling the selection of routing strategies to achieve desired power consumption behavior. We validate our framework across four GPU platforms on language models ranging from 1B to 70B parameters, obtaining 2.8% mean absolute percentage error in estimating achievable power flexibility. We evaluate our approach on a real trace-driven simulation of a 10,000 H100 GPUs datacenter, exhibiting flexibility up to 16% of total power consumption, with 90% of the desired power reduction achieved within approximately 10 seconds. These results establish query routing as an effective mechanism for datacenter power demand flexibility and demonstrate the broader applicability of our proposed framework to datacenter level planning and orchestration.

cs.AI↗

arXiv2026

Democratizing LLM Training with the Edge: Challenges, Systems, and the Road Ahead

The growing computational power of modern GPU data centers has let a few industrial giants train and deploy large language models (LLMs) on massive amounts of data. Yet the sheer scale of these models and their training data makes centralized development increasingly hard to sustain, and at the same time shuts out smaller institutions that hold relevant private data but lack comparable infrastructure. In this paper, we organize our positions on distributed LLM training around the pretraining and fine-tuning paradigms, covering federated learning, split learning, and decentralized approaches, and we assess their readiness for geo-distributed settings with heterogeneous hardware and unreliable connectivity. We then present a {\em unified} system architecture built around a cloud-hosted parameter server that coordinates heterogeneous clients over outbound gRPC connections, reusing one communication, compression, and fault-tolerance stack for both pretraining and a family of fine-tuning methods. We validate the design with real geo-distributed experiments on GPT-2 Medium and Llama3-1B pretraining across multiple university sites, showing stable convergence under realistic conditions. Finally, we chart the road ahead, identifying open problems that span frontier model support, elastic fault tolerance, adaptive compression, federated alignment, privacy and trust, incentive design, and system--algorithm co-design, all of which must be addressed to make democratized LLM training a reality.

cs.DC↗

arXiv2026

Mind the Delay: Latency-aware Query Routing for Heterogeneous LLM pools

Large language model (LLM) query routers typically assign each query to a suitable model that balances response accuracy and monetary cost. Subsequently, separate system-level load balancers based on policies such as join-the-shortest-queue or round-robin distribute the queries across replicas of each model to reduce latency. However, decoupling accuracy-cost and latency considerations can lead to suboptimal routing decisions. We introduce the Batch Evolution Aware (BEA) estimator to provide a fine-grained estimate of the latency, that is, the time-to-first-token (TTFT) experienced by an arriving query. Unlike existing load balancers that rely on coarse signals such as queue lengths, per-model processing rates, or historical average latencies, BEA replays the serving framework's token-batch formation using a snapshot of an instance's current workload, stopping when the incoming query's first token is generated. Taking into account the compute and memory requirements of the prefill and decode phases of language queries as well as the contention between multiple queries that are served in the same batch, BEA predicts TTFT within an error of approximately 5%. Each prediction takes about 0.1 ms, which is fast enough to run on the critical routing path. We also design an analytical framework for a model instance under a limited processor-sharing scheme that calculates expected TTFT in closed form and captures the qualitative trend of latencies as the load increases. The TTFT estimates produced by BEA are then used in our latency-aware router, BEAR, which simultaneously balances accuracy, cost, and latency to select the best model instance from a heterogeneous pool of LLMs. Against seven baselines, BEAR improves the utility of responses delivered within their targets by 63% on average over existing latency-aware routers, and by 30.5% over the best baseline at each arrival rate.

cs.AI↗

arXiv2026

Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning

Federated learning (FL) enables collaborative model training across distributed clients without sharing raw data, yet its scalability is limited by synchronization overhead. Asynchronous federated learning (AFL) alleviates this issue by allowing clients to communicate independently, thereby improving wall-clock efficiency in large-scale, hardware-heterogeneous environments. However, asynchrony introduces updates computed on outdated global models (staleness) that can destabilize optimization and hinder convergence. We propose FedRevive, an AFL framework that revives stale updates through data-free knowledge distillation (DFKD). FedRevive integrates parameter-space aggregation with a server-side DFKD process that transfers knowledge from stale client updates to the current global model without access to data. A meta-learned generator synthesizes pseudo-samples for multi-teacher distillation. A hybrid aggregation scheme combining raw with DFKD updates effectively mitigates staleness while retaining AFL scalability. Experiments on vision and text benchmarks show that FedRevive achieves faster training by up to 29.7% and higher final accuracy by up to 11.4 percentage points than baselines.

cs.LG↗

arXiv2026

Federate the Router: Learning Language Model Routers with Sparse and Decentralized Evaluations

Large language models (LLMs) are increasingly accessed as remotely hosted services by edge and enterprise clients that cannot run frontier models locally. Since models vary widely in capability and price, routing queries to models that balance quality and inference cost is essential. Existing router approaches assume access to centralized query-model evaluation data. However, these data are often fragmented across clients, such as end users and organizations, and are privacy-sensitive, which makes centralizing data infeasible. Additionally, per-client router training is ineffective since local evaluation data is limited and covers only a restricted query distribution and a biased subset of model evaluations. We introduce the first federated framework for LLM routing, enabling clients to learn a shared routing policy from local offline query-model evaluation data. Our framework supports both parametric multilayer perceptron router and nonparametric K-means router under heterogeneous client query distributions and non-uniform model coverage. Across three benchmarks, federated collaboration improves the accuracy-cost frontier over client-local routers, both via increased effective model coverage and better query generalization. Our theoretical results also validate that federated training reduces routing suboptimality.

cs.LG↗

arXiv2026

FedGuide: Diffusion Prior Alignment and Value Baseline Guidance for Heterogeneous Federated Reinforcement Learning

Federated Reinforcement Learning (FRL) enables collaborative policy learning across distributed agents with heterogeneous environments. While recent methods based on variance reduction, divergence penalization, and momentum optimization improve FRL under heterogeneous settings, they still primarily synchronize policy or value-network parameters and do not explicitly address distributional mismatch among heterogeneous clients. Therefore, we propose \textbf{FedGuide}, a FRL framework that uses diffusion priors as behavior models to provide personalized data supported distributions for heterogeneous local policy learning. Instead of directly averaging local policies, FedGuide aggregates those diffusion priors through Optimal-Transport Mixture-of-Experts (OT-MoE), preserving heterogeneous behavior modes in distribution space. It further develops a Distribution Correction Estimation (DICE) value baseline to provide low-variance, return-aware guidance for local policy improvement. Experiments across heterogeneous environments show that FedGuide outperforms representative FRL methods in client-average returns, final-round performance, and worst-round robustness, while maintaining stable learning under stronger heterogeneity.

cs.LG↗

arXiv2026

On-Policy Attention Linearization

Hybrid transformer architectures that replace most softmax attention layers with linear attention offer transformer-level quality at a fraction of the memory cost. Rather than pretraining such models, a growing body of work distills them from already trained full-attention transformers. However, these distilled models often collapse on long-context retrieval and reasoning tasks, particularly when operating in thinking mode, where the efficiency gains of hybrid architectures matter most. Since linear attention layers must compress context into a fixed-size state, their errors compound over long sequences. As off-policy distillation never teaches the student model to recover from this drift, tasks that necessitate longer sequence lengths become especially challenging. We introduce On-Policy Attention Linearization (OPAL) in which the hybrid attention student samples its own long-context trajectories and receives dense supervision from the frozen full-attention teacher. Applying OPAL to Qwen3-4B and MiMo-7B-RL-0530, we recover $87$--$94\%$ of full-attention performance on commonsense reasoning, $100\%$ on needle-in-a-haystack (NIAH) retrieval, and $83$--$93\%$ on mathematical reasoning with only 3B training tokens. We achieve these results without supervised fine-tuning (SFT) or reinforcement learning with verifiable rewards (RLVR). Compared with the strongest prior linearization method, which recovers $68\%$ of its teacher's retrieval performance and $21.6\%$ absolute average mathematical reasoning accuracy, OPAL fully recovers retrieval and achieves $67.6$--$72.2\%$ on math reasoning.

cs.LG↗

arXiv2026

Not All Turns Are Equally Hard: Adaptive Thinking Budgets For Efficient Multi-Turn Reasoning in Agents

As LLM reasoning performance plateaus, improving inference-time compute efficiency is crucial to mitigate overthinking and long thinking traces even for simple queries. Prior approaches including length regularization, adaptive routing, and difficulty-based budget allocation primarily focus on single-turn settings and fail to address the sequential dependencies inherent in multi-turn reasoning. In this work, we formulate multi-turn reasoning as a sequential compute allocation problem and model it as a multi-objective Markov Decision Process. We propose TAB: Turn-Adaptive Budgets, a budget allocation policy trained via Group Relative Policy Optimization (GRPO) that learns to maximize task accuracy while respecting global per-problem token constraints. Consequently, TAB takes as input the conversation history and learns to adaptively allocate smaller budgets to easier turns and save appropriate number of tokens for the crucial harder reasoning steps. Our experiments on diverse agentic benchmarks demonstrate that TAB achieves a superior accuracy-cost tradeoff saving up to 35% tokens and reducing latency up to 30% while maintaining accuracy over static and off-the-shelf LLM budget baselines. Further, for systems where a plan of all turns is available apriori, we propose TAB All-SubQ, a budget allocation policy that budgets tokens based on the conversation history and entire plan saving up to 40% tokens over baselines.

cs.LG↗

arXiv2026

Emergent and Subliminal Misalignment Through the Lens of Data-Mediated Transfer

Fine-tuning LLMs on narrow harmful datasets can induce Emergent Misalignment (EM), where models exhibit misaligned behavior far beyond the fine-tuning distribution. We argue that emergent misalignment can be better understood as a data-mediated transfer phenomenon: harmful fine-tuning examples do not induce uniform behavioral spillover, but interact with the structural properties of the dataset and the difficulty of the tasks relative to the model. Across our experiments, we find that misalignment appears more readily when fine-tuning and evaluation prompts share similar underlying functional structure, when prompts leave more room for coherent harmful completions, and when the target behavior has been more reliably learned by the model. The training pipeline itself also matters: pretraining composition shapes later misalignment. We further study Subliminal Learning (SL), where misalignment is transmitted by fine-tuning on seemingly benign data generated by a harmful teacher. Moving beyond the standard SFT setting, we for the first time compare this transfer under off-policy and on-policy distillation as well, allowing us to separate the roles of the teacher guidance and the training data distribution in transmitting misalignment. Together, these results argue for a data-centric view: Emergent/subliminal misalignment should not be treated as a simple consequence of isolated harmful fine-tuning examples, but as the result of interactions between fine-tuning data structure, pretraining distributions, and training channels.

cs.LG↗

arXiv2026

Navigating the Accuracy-Size Trade-Off with Flexible Model Merging

Model merging has emerged as an efficient method to combine multiple single-task fine-tuned models. The merged model can enjoy multi-task capabilities without expensive training. While promising, merging into a single model often suffers from an accuracy gap with respect to the fine-tuned models. On the other hand, deploying all individual fine-tuned models incurs high storage costs. We propose FlexMerge, a novel data-free model merging framework that: (a) flexibly generates merged models of varying sizes, spanning the full spectrum from a single merged model to retaining all fine-tuned models; and (b) supports multiple merging algorithms in a unified framework. Using FlexMerge, we systematically characterize the accuracy-size trade-off of different algorithms. Our study reveals two key findings: first, even modestly larger merged models can yield steep accuracy gains (up to 13.5% when just doubling the size); second, algorithm rankings are not consistent as size increases, with some methods overtaking others beyond the one-model regime. These results uncover a new design dimension for model merging: developing and comparing algorithms across the full spectrum of sizes rather than only at the single-model limit. Extensive experiments on vision and NLP benchmarks, with up to 30 tasks, confirm the generality and practicality of FlexMerge.

cs.CV↗

arXiv2026

PubSwap: Public-Data Off-Policy Coordination for Federated RLVR

Reasoning post-training with reinforcement learning from verifiable rewards (RLVR) is typically studied in centralized settings, yet many realistic applications involve decentralized private data distributed across organizations. Federated training is a natural solution, but scaling RLVR in this regime is challenging: full-model synchronization is expensive, and performing many local steps can cause severe client drift under heterogeneous data. We propose a federated RLVR framework that combines LoRA-based local adaptation with public-data-based off-policy steps to improve both communication efficiency and cross-client coordination. In particular, a small shared public dataset is used to periodically exchange and reuse response-level training signals across organizations, providing a lightweight anchor toward a more globally aligned objective without exposing private data. Our method selectively replaces locally incorrect responses with globally correct ones during public-data steps, thereby keeping training closer to the local policy while still benefiting from cross-client coordination. Across mathematical and medical reasoning benchmarks and models, our method consistently improves over standard baselines. Our results highlight a simple and effective recipe for federated reasoning post-training: combining low-rank communication with limited public-data coordination.

cs.LG↗

arXiv2026

Improving the Convergence of Private Shuffled Gradient Methods with Public Data

We consider the problem of differentially private (DP) convex empirical risk minimization (ERM). While the standard DP-SGD algorithm is theoretically well-established, practical implementations often rely on shuffled gradient methods that traverse the training data sequentially rather than sampling with replacement in each iteration. Despite their widespread use, the theoretical privacy-accuracy trade-offs of private shuffled gradient methods (\textit{DP-ShuffleG}) remain poorly understood, leading to a gap between theory and practice. In this work, we leverage privacy amplification by iteration (PABI) and a novel application of Stein's lemma to provide the first empirical excess risk bound of \textit{DP-ShuffleG}. Our result shows that data shuffling results in worse empirical excess risk for \textit{DP-ShuffleG} compared to DP-SGD. To address this limitation, we propose \textit{Interleaved-ShuffleG}, a hybrid approach that integrates public data samples in private optimization. By alternating optimization steps that use private and public samples, \textit{Interleaved-ShuffleG} effectively reduces empirical excess risk. Our analysis introduces a new optimization framework with surrogate objectives, varying levels of noise injection, and a dissimilarity metric, which can be of independent interest. Our experiments on diverse datasets and tasks demonstrate the superiority of \textit{Interleaved-ShuffleG} over several baselines.

cs.LG↗

arXiv2026

Internal Planning in Language Models: Characterizing Horizon and Branch Awareness

The extent to which decoder-only language models (LMs) engage in planning, that is, organizing intermediate computations to support coherent long-range generation, remains an important question, with implications for interpretability, reliability, and principled model design. Planning involves structuring computations over long horizons, and considering multiple possible continuations, but how far transformer-based LMs exhibit them without external scaffolds, e.g., chain-of-thought prompting, is unclear. We address these questions by analyzing the hidden states at the core of transformer computations, which capture intermediate results and act as carriers of information. Since these hidden representations are redundant and encumbered with fine-grained details, we develop a pipeline based on vector-quantized variational autoencoders that compresses them into compact summary codes. These codes enable measuring mutual information and analyzing the computational structure of the underlying model behavior. Using this framework, we study planning in LMs across synthetic grammar, path-finding tasks, and natural language datasets, focusing on two planning properties: (i) the planning horizon of pre-output computations, and (ii) the extent to which the model considers alternative valid continuations. As a separate downstream use of the same pipeline, we also analyze how decision-relevant information is distributed across layers and earlier prefix blocks when producing next-token predictions. Together, these analyses advance our understanding of planning in LMs and provide a general-purpose pipeline for inspecting internal model dynamics. Our results reveal that the effective planning horizon is task-dependent, that models implicitly preserve information about unused correct continuations, and that predictions draw most on recent computations, though earlier blocks remain informative.

cs.AI↗

arXiv2026

Achieving Logarithmic Regret in KL-Regularized Zero-Sum Markov Games

Reverse Kullback-Leibler (KL) divergence-based regularization with respect to a fixed reference policy is widely used in modern reinforcement learning to preserve the desired traits of the reference policy and sometimes to promote exploration (using uniform reference policy, known as entropy regularization). Beyond serving as a mere anchor, the reference policy can also be interpreted as encoding prior knowledge about good actions in the environment. In the context of alignment, recent game-theoretic approaches have leveraged KL regularization with pretrained language models as reference policies, achieving notable empirical success in self-play methods. Despite these advances, the theoretical benefits of KL regularization in game-theoretic settings remain poorly understood. In this work, we develop and analyze algorithms that provably achieve improved sample efficiency under KL regularization. We study both two-player zero-sum matrix games and Markov games: for matrix games, we propose OMG, an algorithm based on best response sampling with optimistic bonuses, and extend this idea to Markov games through the algorithm SOMG, which also uses best response sampling and a novel concept of superoptimistic bonuses. Both algorithms achieve a logarithmic regret in $T$ that scales inversely with the KL regularization strength $β$ in addition to the traditional $\widetilde{\mathcal{O}}(\sqrt{T})$ regret without the $β^{-1}$ dependence.

cs.LG↗

arXiv2026

Ravan: Multi-Head Low-Rank Adaptation for Federated Fine-Tuning

Large language models (LLMs) have not yet effectively leveraged the vast amounts of edge-device data, and federated learning (FL) offers a promising paradigm to collaboratively fine-tune LLMs without transferring private edge data to the cloud. To operate within the computation and communication constraints of edge devices, recent literature on federated fine-tuning of LLMs proposes the use of low-rank adaptation (LoRA) and similar parameter-efficient methods. However, LoRA-based methods suffer from accuracy degradation in FL settings, primarily because of data and computational heterogeneity across clients. We propose Ravan, an adaptive multi-head LoRA method that balances parameter efficiency and model expressivity by reparameterizing the weight updates as the sum of multiple LoRA heads $s_i\textbf{B}_i\textbf{H}_i\textbf{A}_i$ in which only the core matrices $\textbf{H}_i$ and their lightweight scaling factors $s_i$ are trained. These trainable scaling factors let the optimization focus on the most useful heads, recovering a higher-rank approximation of the full update without increasing the number of communicated parameters since clients upload $s_i\textbf{H}_i$ directly. Experiments on vision and language benchmarks show that Ravan improves test accuracy by $2-8\%$ over prior parameter-efficient baselines, making it a robust and scalable solution for federated fine-tuning of LLMs.

cs.LG↗

arXiv2026

MELINOE: Fine-Tuning Enables Memory-Efficient Inference for Mixture-of-Experts Models

Mixture-of-Experts (MoE) model architectures can significantly reduce the number of activated parameters per token, enabling computationally efficient training and inference. However, their large overall parameter counts and model sizes have precluded their widespread usage in resource-constrained settings as all of the parameters must still be loaded into GPU memory. Prior works aim to address this memory bottleneck by offloading certain experts into CPU memory and porting them to GPU memory only when they are activated. In practice, these methods suffer from the significant I/O latency incurred by expert transfer. We present MELINOE, a method that fine-tunes an MoE model to more strongly prefer activating a smaller number of experts per sequence. Caching these preferred experts in GPU memory reduces expert churn and CPU-GPU transfer overhead. MELINOE increases throughput by $1.2-3\times$ over efficient baselines and up to $14.7\times$ over transfer-heavy baselines while retaining or even improving the performance of the model on a downstream task, making it a reliable method for improving MoE inference efficiency.

cs.LG↗

arXiv2026

LOCUS: Low-Dimensional Model Embeddings for Efficient Model Exploration, Comparison, and Selection

The rapidly growing ecosystem of Large Language Models (LLMs) makes it increasingly challenging to manage and utilize the vast and dynamic pool of models effectively. We propose LOCUS, a method that produces low-dimensional vector embeddings that compactly represent a language model's capabilities across queries. LOCUS is an attention-based approach that generates embeddings by a deterministic forward pass over query encodings and evaluation scores via an encoder model, enabling seamless incorporation of new models to the pool and refinement of existing model embeddings without having to perform any retraining. We additionally train a correctness predictor that uses model embeddings and query encodings to achieve state-of-the-art routing accuracy on unseen queries. Experiments show that LOCUS needs up to 4.8x fewer query evaluation samples than baselines to produce informative and robust embeddings. Moreover, the learned embedding space is geometrically meaningful: proximity reflects model similarity, enabling a range of downstream applications including model comparison and clustering, model portfolio selection, and resilient proxies of unavailable models.

cs.LG↗

arXiv2026

Sample Complexity of Average-Reward Q-Learning: From Single-agent to Federated Reinforcement Learning

Average-reward reinforcement learning offers a principled framework for long-term decision-making by maximizing the mean reward per time step. Although Q-learning is a widely used model-free algorithm with established sample complexity in discounted and finite-horizon Markov decision processes (MDPs), its theoretical guarantees for average-reward settings remain limited. This work studies a simple but effective Q-learning algorithm for average-reward MDPs with finite state and action spaces under the weakly communicating assumption, covering both single-agent and federated scenarios. For the single-agent case, we show that Q-learning with carefully chosen parameters achieves sample complexity $\widetilde{O}\left(\frac{|\mathcal{S}||\mathcal{A}|\|h^{\star}\|_{\mathsf{sp}}^3}{\varepsilon^3}\right)$, where $\|h^{\star}\|_{\mathsf{sp}}$ is the span norm of the bias function, improving previous results by at least a factor of $\frac{\|h^{\star}\|_{\mathsf{sp}}^2}{\varepsilon^2}$. In the federated setting with $M$ agents, we prove that collaboration reduces the per-agent sample complexity to $\widetilde{O}\left(\frac{|\mathcal{S}||\mathcal{A}|\|h^{\star}\|_{\mathsf{sp}}^3}{M\varepsilon^3}\right)$, with only $\widetilde{O}\left(\frac{\|h^{\star}\|_{\mathsf{sp}}}{\varepsilon}\right)$ communication rounds required. These results establish the first federated Q-learning algorithm for average-reward MDPs, with provable efficiency in both sample and communication complexity.

stat.ML↗