Trade-Off

Efficient routing across multiple LLMs enables cost-quality tradeoffs by directing queries to the cheapest capable model. Prior work attributes routing headroom to an "unsolvability ceiling", queries no model in the pool can solve. We present a large-scale study of multi-tier LLM routing with 206,000 query-model pairs across six benchmarks (MMLU, MedQA, HumanEval, MBPP, Alpaca, ShareGPT) using the Gemma 4 and Llama 3.1 families. Evaluating with both LLM-as-a-judge and exact-match metrics, we show that a substantial portion of reported unsolvability stems from evaluation artifacts: (i) systematic judge biases favoring verbosity over correctness, (ii) truncation under fixed generation budgets, and (iii) output format mismatches. Through dual-judge validation and exact-match grounding, we reduce measured unsolvability across tasks. We introduce a decomposition framework attributing failures to these artifacts, revealing consistent patterns across domains and model families. These artifacts also distort router training signals: standard routers collapse to majority-class prediction (~79% smallest-tier optimal), confirmed via random-feature and shuffled-label controls, incurring a 13-17 percentage point opportunity cost. We provide actionable recommendations including dual-judge validation, exact-match anchoring, and cost-sensitive objectives. Our findings suggest existing routing headroom estimates are substantially inflated, underscoring the need for reliable evaluation protocols in multi-LLM systems.

V2X can warn an autonomous vehicle about hazards beyond line-of-sight, but it also brings uncertainty: messages may be delayed, dropped, or even forged. Meanwhile, map knowledge may change during a trip, forcing the vehicle to replan under tight real-time budgets. This paper studies how to make motion planning and low-level control robust to such uncertain, event-driven updates. We present MORPH-U, a CARLA-based closed-loop stack that fuses LiDAR/radar/camera with V2X (CAM/DENM) into a Local Dynamic Map (LDM) and triggers Hybrid-A* replanning when validated hazards or map changes affect the planned route. We expose the planning/control trade-offs via a multi-objective formulation over tracking error, safety margin (minimum TTC), responsiveness, and smoothness, and select operating points using Pareto-frontier analysis. To avoid unsafe replanning from faulty V2X triggers, MORPH-U adds a lightweight Byzantine-inspired acceptance gate that combines a quorum rule with an on-board sensor veto. Experiments in dynamic CARLA scenarios show that V2X-augmented LDM improves downstream safety, Pareto tuning provides controllable accuracy-comfort trade-offs, and the gate prevents replanning under saturated false-DENM injection ($p_{\text{attack}}=1.0$).

Preserving model fidelity is essential for stealthy text-to-image (T2I) backdoor attacks. Existing methods such as Learning without Forgetting (LwF) rely on output-based distillation, which provides limited regularization. We introduce Elastic Weight Consolidation (EWC) as a parameter-based alternative for preserving fidelity in backdoor learning. While stronger in principle, we show that standard static EWC with a fixed regularization weight lambda and mean-squared utility loss creates an artificial trade-off between attack success rate (ASR) and fidelity, particularly degrading performance on weak triggers. To address this, we propose Cosine-Aware Adaptive EWC, which dynamically adjusts EWC regularization using a cosine-based semantic utility and adaptive scheduling. This approach transforms EWC from a fixed penalty into a context-sensitive constraint, maintaining high ASR while preserving model fidelity. Experiments demonstrate improved ASR-fidelity balance and enhanced robustness on out-of-domain (OOD) datasets compared to existing baselines.

Large language models (LLMs) are trained and evaluated as though perfect reliability is achievable for any task given sufficient scale. We show that this assumption is information-theoretically unjustified. Every generative task has a reliability ceiling that no model can exceed, determined by how much output uncertainty is resolvable from observable context. The gap decomposes into a resolvable component closable with additional context and a subjective component inherent to task ambiguity. Autoregressive generation further degrades this ceiling at a rate governed by the task's dependency kernel, which quantifies inter-token correlations in the output. From these two primitives, we derive a first-principles scaling law where LLM performance is bottlenecked by the scarcer resource: training data or model capacity. This law recovers the Chinchilla scaling law as a special case and provides a structural account of when scaling improves reliability. Beyond scaling, our framework unifies diverse practical phenomena, such as the benefits of retrieval-augmentation and the spectral mechanics of catastrophic forgetting. Our work formalizes the resource-complexity tradeoffs that govern model performance across domains, offering a unified theory of performance limits in generative language models.

Object detection on heterogeneous edge devices must satisfy strict energy, latency, and memory constraints while still providing reliable perception for downstream autonomy. Existing energy-aware NAS methods often target limited deployment settings, while real energy remains difficult to optimize because it is highly device-dependent and costly to measure. We address these challenges with an energy-adaptive framework that combines an energy-aware XiResOFA search space, a two-stage energy estimator, and iterative search to identify a single energy-efficient base architecture. We then apply compound scaling to transform this base design into the XiYOLO family across deployment budgets, enabling interpretable accuracy-energy tradeoffs under sparse hardware measurements. Experiments on PascalVOC, COCO, and real-device deployment show that XiYOLO achieves a stronger energy-accuracy tradeoff than YOLO baselines. On PascalVOC, the medium XiYOLO model reaches 86.15 mAP50 while reducing energy relative to YOLOv12m by 20.6% on GPU and 35.9% on NPU. On COCO, XiYOLO reduces energy relative to YOLOv12 by up to 53.7% on GPU and 51.6% on NPU at the small scale. The proposed two-stage estimator also improves sample efficiency over a joint predictor under few-shot adaptation with only 2-20 target-device samples.

As large language models (LLMs) increasingly mediate both content generation and moderation, linguistic evasion strategies known as Algospeak have intensified the coevolution between evaders and detectors. This research formalizes the underlying dynamics grounded in a joint action model: when Algospeak increases, detectability and understandability decrease. Further, the concept of Majority Understandable Modulation (MUM) is introduced and defined as the modulation level at which additional evasive alteration increases detector evasion but loses comprehension for the majority of recipients. To empirically probe this trade-off, we introduce a reproducible framework that can be used to create meaning-preserving, Algospeak-style variants, based on an existing taxonomy and with tunable modulation levels. Using COVID-19 disinformation as a first proof-by-example setting, we construct a reference dataset of 700 modulated items, drawn from twenty base sentences across five modulation levels and seven strategies. We then run two linked evaluations with seven different language models: one testing for interpretation through meaning recovery and one for disinformation detection through classification. Curve fitting over modulation levels yields an estimate of the Majority Understandable Modulation threshold and enables sensitivity analyses across strategies and models, see Figure 1. Results reveal the characteristic relationships between understandability and modulation. This study lays the groundwork for understanding the dynamics behind Algospeak and provides the framework, dataset, and experimental setups described.

Model cascades, in which a cheap LLM defers to an expensive one on low-confidence queries, are widely used to navigate the cost-quality tradeoff at deployment. Existing approaches largely treat the deferral threshold as an empirical hyperparameter, with limited guidance on the geometry of the resulting cost-quality frontier over a model pool. We develop a decision-theoretic framework grounded in constrained optimization and duality. For a two-model cascade, we establish piecewise concavity of the cost-quality frontier on decreasing-benefit regions of the confidence support, with reciprocal shadow prices linking the budget- and quality-constrained formulations. Given a pool of $k$ models, we characterize the frontier achievable by deterministic two-model threshold cascades as the pointwise envelope over $\binom{k}{2}$ pairwise cascades, with switching points where the optimal pair changes. For $k$-model cascades, we derive first-order conditions in which a single shadow price equalizes marginal quality-per-cost across stage boundaries. We validate the framework on five benchmarks (MATH, MMLU, TriviaQA, SimpleQA, LiveCodeBench) across eight models from five providers. Within the deterministic threshold-cascade class, full fixed chains underperform the pairwise envelope, and optimized subsequence cascades do not deliver practically meaningful held-out gains over it. A lightweight pre-generation router exceeds the best cascade policy on four of five datasets, mainly because it avoids the cheap model's generation cost on queries sent directly to a larger model rather than because of a stronger routing signal. These results suggest that cascade performance is limited primarily by structural cost, since cascades pay the cheap model before any escalation decision, rather than by a shortage of intermediate stages.

We derive a tight analysis of the trade-off function for Differentially Private Stochastic Gradient Descent (DP-SGD) with subsampling based on random shuffling within the $f$-DP framework. Our analysis covers the regime $σ\geq \sqrt{3/\ln M}$, where $σ$ is the noise multiplier and $M$ is the number of rounds within a single epoch. Unlike $f$-DP analyses for Poisson subsampling, which yield non-closed implicit formulas that can be machine computed but are non-transparent, random shuffling admits a tight analysis yielding transparent and interpretable closed-form bounds. Our concrete bounds, derived via the Berry-Esseen theorem, are tight up to constant factors within the proof framework. We demonstrate worked parameter settings for a single epoch ($E=1$) with a corresponding trade-off function $\geq 1-a-δ$, that is, only $δ$ below the ideal random guessing diagonal $1-a$: For $δ= 1/100$ and $σ= 1$, roughly $M \approx 1.14\times 10^6$ rounds and $N \approx 1.14\times 10^7$ training samples suffice to achieve meaningful differential privacy. This is in contrast to recent negative results for the regime $σ\leq 1/\sqrt{2 \ln M}$. Our concrete bounds can be composed over multiple epochs leading to $δ$ having a linear in $E$ dependency, which restricts $E=O(\sqrt{M})$. To go beyond Berry--Esseen, we introduce a new proof technique based on a generalization of the law of large numbers that yields an asymptotic random guessing diagonal-limit result: if $E=c_M^2M$ with $c_M\to 0$, then the $E$-fold composed trade-off function satisfies $f^{\otimes E}(a)\to 1-a$ uniformly in $a\in[0,1]$ with $δ$ having only an $O(\sqrt{E})$ dependency. We compare this asymptotic regime with the corresponding Poisson subsampling asymptotic, and highlight the characterization of explicit convergence rates as an open question.

Semantic communication (SemCom) with learned encoder-decoder architectures enables end-to-end learning of compact task-oriented representations optimized for the wireless channel, reducing channel resources needed to convey task-relevant information and improving spectrum efficiency. This paper studies semantic image transmission over block Rayleigh fading with AWGN using a multi-task semantic autoencoder that jointly reconstructs images and predicts labels from the received waveform. The latent dimension (complex channel uses per source sample) serves as a cross-layer control variable governing semantic fidelity and channel resource usage. We characterize the resulting latency-task fidelity tradeoff: larger latent representations improve inference accuracy but increase service time, channel uses, and queueing delay. Building on this insight, we develop online semantic-rate controllers that adapt the latent dimension per update under a long-term semantic error constraint. A queue-aware drift-plus-penalty policy minimizes delay subject to an average semantic error cap, while a complementary age-aware policy minimizes time-average Age of Information (AoI). By adapting the semantic rate to congestion and fidelity requirements, the proposed framework improves spectrum utilization and enables timely semantic updates with significantly lower delay and AoI than fixed-rate baselines.

We identify and prove a fundamental trade-off governing long-sequence models: no model can simultaneously achieve (i) per-step computation independent of sequence length (Efficiency), (ii) state size independent of sequence length (Compactness), and (iii) the ability to recall a number of historical facts proportional to sequence length (Recall). We formalize this trade-off within an Online Sequence Processor abstraction that unifies Transformers, state space models, linear recurrent networks, and their hybrids. Using the Data Processing Inequality and Fano's Inequality, we prove that any model satisfying Efficiency and Compactness can recall at most O(poly(d)/log V) key-value pairs from a sequence of arbitrary length, where d is the model dimension and V is the vocabulary size. We classify 52 architectures published before March 2026 into the triangle, showing that each achieves at most two of the three properties and that hybrid architectures trace continuous trajectories in the interior. Experiments on synthetic associative recall tasks with five representative architectures validate the theoretical bound: empirical recall capacity lies strictly below the information-theoretic limit, and no architecture escapes the triangle.

Per-instance algorithm selection (PIAS) takes advantage of complementarity between a set of algorithms by deciding which algorithm to run on a given instance. This decision is based on features of the instances, which, in the context of black-box optimization (BBO), require a part of the optimization budget to be computed. This raises two questions: (a) from which fraction of the budget spent on feature computation does PIAS become worth it for BBO, and (b) which fraction of the budget optimizes the tradeoff between feature accuracy and PIAS performance. To this end, we perform a broad study where PIAS with varying sampling budgets for feature computation is compared to the single best algorithm on a broad range of algorithm selection scenarios. These scenarios consist of two portfolio sizes, three problem sets, 4 dimensionalities, and 10 target budgets. We find that PIAS is viable for the majority of tested scenarios, even when as much as a quarter of the total budget is spent on feature computation. The tradeoff for the fraction of the budget spent on feature computation to maximize the benefit of PIAS is highly dependent on the specific AS scenario. Further, on average 20 percent of PIAS loss to the virtual best solver is explained by the budget spent on feature computation, highlighting the importance of properly accounting for the feature budget.

We present an experimental study of energy-to-solution (ETS) of hybrid quantum-classical applications, enabled by direct instrumentation of power consumption of a Forte Enterprise trapped-ion quantum processor. We apply this methodology to a hybrid quantum-classical pipeline for quantum fine-tuning of foundational AI models, and validate the approach end-to-end on quantum hardware. Despite noise and limited qubit counts, the resulting models achieve accuracy competitive with and exceeding classical baselines such as logistic regression and support vector classifiers. Our results show that QPU energy consumption scales approximately linearly with qubit number for shallow circuits, while classical simulation exhibits exponential scaling, indicating a break-even for ETS around 34 qubits. The classification error improvement of the best quantum fine-tuned model over the best classical fine-tuned model considered in this study is around 24%. We further contextualize these findings with comparisons to tensor network methods. This work establishes energy-to-solution as a measurable and scalable metric for evaluating quantum applications and provides experimental evidence of favorable energy-accuracy trade-offs.

Place recognition is essential for long-term autonomous navigation, enabling loop closure and consistent mapping. Although deep learning has improved performance, deploying such models on resource-constrained platforms remains challenging. This work explores efficient LiDAR-based place recognition for EdgeAI by leveraging Bird's Eye View representations to enable lightweight image-based networks. We benchmark representative architectures without aggregation heads using a unified descriptor scheme based on global pooling and linear projection, and evaluate performance under FP32, FP16, and INT8 quantization. Experiments reveal trade-offs between accuracy, robustness, and efficiency: FP16 matches FP32 with lower cost, while INT8 introduces architecture-dependent degradation. Overall, the presented results are a strong basis for future research on 'use-case'-aware quantisation of Neural Networks for Edge deployment.

The growing relevance of multi-agent systems has drawn increasing focus on communication-efficient filters for collaborative perception to alleviate the system's communication burden. While the event-triggered (ET) mechanism can improve communication efficiency in collaborative state estimation, an inevitable trade-off exists between estimation accuracy and communication cost in ET filters. This paper proposes a fast and accurate ET diffusion-based filter for real-time multi-agent collaborative target tracking, aiming to reduce the system's data transmission without compromise in tracking performance. The proposed filter achieves improved tracking accuracy, reduced data transmission, and accelerated convergence using an error-minimized ET cubature information filter (CIF) for local estimation, and a correlation-aware diffusion strategy for global fusion. The experimental results confirm the scalability of the proposed EDC-CIF algorithm and demonstrate its efficacy in simultaneously reducing estimation error and computation time while significantly enhancing communication efficiency.

In multi-armed bandits, the most-explored arms are the most informative, while reward maximization typically pulls only the best arm. We study the tradeoff between identifying arm means accurately and accumulating reward, and present an algorithm with regret guarantees that interpolates between the two objectives. We provide both upper and lower bounds and validate empirically.

Mixture-of-experts (MoE) architectures have turned LLM serving into a cluster-scale workload in which communication consumes a considerable portion of LLM serving runtime. This has prompted industry to invest heavily in expensive high-bandwidth scale-up networks. We question whether such costly infrastructure is strictly necessary. We present the first systematic cross-layer analysis of network cost-effectiveness for MoE LLM serving, comparing four representative XPU (e.g., GPU/TPU) topologies (scale-up, scale-out, 3D torus, and 3D full-mesh). We find that lower-cost switchless topologies are more cost-effective than the scale-up topology across all serving scenarios explored, improving cost-effectiveness by 20.6-56.2%. In particular, the 3D full-mesh topology is Pareto-optimal in terms of the performance-cost tradeoff. We also find that current scale-up link bandwidths are over-provisioned: reducing the link bandwidth improves throughput per cost by up to 27%. A forward-looking analysis of upcoming GPU generations indicates that the cost-performance advantage of switchless networks will likely persist.

Industrial systems increasingly depend on Machine Learning (ML), and operate on heterogeneous nodes that must satisfy tight latency, energy, and memory constraints. Dynamic ML models, which reconfigure their computational footprint at runtime, promise high energy efficiency and lower average latency for modest accuracy tradeoffs; however, their deployment is complex due to the additional hyperparameters they rely on. These hyperparameters, controlling the accuracy versus average latency tradeoff, are often tuned on a calibration dataset that must match the test time distribution, an assumption that rarely holds in real-world scenarios, leading to suboptimal operational conditions, possibly below static models. We propose a two-tier adaptive architecture that co-optimizes model and system decisions. At the global level, a scheduler configures and deploys, for each edge node, a cascade of classifiers composed of lightweight specialized models and a generalist fallback, satisfying latency and memory constraints. At the node level, a local controller tracks data drifts and hardware resources, enabling or disabling specialized predictors (SP) to preserve high energy efficiency and avoid latency-constraint violations under varying conditions. This design allows longer operating times without forcing a global redeployment step, and enables efficient execution in case of an unreachable remote global controller. We evaluate the approach on two datasets under controlled distribution mismatch scenarios, showing average per-inference reductions of latency up to 2.45x and energy up to 2.86x, with less than 4% accuracy drop compared to static baselines. Our contributions are:(1) a budgeted SP-cascade formulation that preserves worst-case latency constraints;(2) a hierarchical controller that maintains efficiency under data and resource changes; and (3) an experimental evaluation on embedded hardware.

Early detection of plant diseases is critical for improving crop productivity, while it also facilitates the foundations of precision agriculture. Recent advances in distributed deep learning have enabled plant disease classification models to be trained across geographically distributed agricultural sensing infrastructures. However, deploying such systems in large-scale Internet of Things (IoT) environments, introduces significant challenges related to computational cost, energy consumption, and system efficiency. In this paper, we present a design-space exploration of hierarchical federated learning architectures for plant disease classification, with a particular focus on the trade-offs between predictive performance and energy efficiency. We further introduce a power- and energy-aware optimization framework that enables the systematic evaluation and selection of model-aggregator configurations under varying deployment constraints. The hierarchical federated architecture organizes distributed clients through intermediate aggregation layers, reducing communication and computational overhead. We evaluate multiple convolutional neural network architectures, including EfficientNet-B0, ResNet-50, and MobileNetV3-Large, in combination with different federated aggregation strategies such as FedAvg, FedProx, and FedAvgM. Experimental results demonstrate that different model-aggregator combinations exhibit distinct performance-energy trade-offs. Consequently, we highlight configurations that achieve competitive diagnostic accuracy and significantly reduce system resource requirements.

Frontier artificial intelligence is reshaping all aspects of society, from economic output or military capability to democratic institutions. The EU is entering this transformation from a position of structural dependence: frontier models originate almost exclusively from the United States or China, the US holds approximately sixteen times the EU's AI supercomputing capacity, and only 15% of global hyperscale data centre capacity resides within EU borders. Although the European Commission has accelerated its policy response, existing initiatives remain fragmented and lack a cohesive vision for securing strategic autonomy across the full frontier AI value chain. Here we propose a unified framework connecting five sovereignty pillars (economic competitiveness, resilience, security and defence, European values, and foreign relations) to a decomposition of the frontier AI stack comprising five layers, 26 components, and 29 sub-components. This framework allows the identification of critical gaps, redundancies, and inter-pillar trade-offs that current EU policy leaves implicit. Our analysis of the AI Gigafactory Initiative illustrates how a sovereignty-centred lens reveals conflicts that narrowly economic framings obscure. Moreover, this framework offers policymakers a structured basis for designing, evaluating, and prioritising frontier AI interventions across multiple dimensions of European strategic autonomy across the 92 initiatives from four major Commission communications we. identify, and beyond.

Edge computing environments impose strict constraints on energy consumption and latency, making the deployment of deep neural networks a significant challenge. Therefore, smart and adaptive inference strategies that dynamically balance computational cost or latency with predictive accuracy are critical in edge computing scenarios. In this work, we build on Adaptive Deep Neural Networks (ADNNs) that employ the Multi-Armed Bandit (MAB) framework. Current literature leverages the first version of the Upper Confidence Bound (UCB1) strategy to dynamically select the optimal confidence threshold, enabling efficient early exits without sacrificing accuracy. However, we introduce four additional Upper Confidence Bound strategies in ADNNs, namely UCB-V, UCB-Tuned, UCB-Bayes, and UCB-BwK, and perform, for the first time, a comparative study of these strategies with respect to trade-offs between accuracy, energy consumption, and latency. The proposed UCB strategies are employed on the ResNet and MobileViT neural networks, and are evaluated on the benchmark datasets of CIFAR-10, CIFAR-10.1, and CIFAR-100. Experimental results demonstrate that all strategies achieve sub-linear cumulative regret, with UCB-Bayes converging the fastest, followed by UCB-Tuned and UCB-V. Finally, UCB-V and UCB-Tuned dominate the Pareto Frontiers of accuracy-latency and accuracy-energy trade-offs. The implementation code is available here: https://github.com/gr3gor1/MAB_UCB

We present JigsawRL, a cost-efficient framework that explores Pipeline Multiplexing as a new dimension of RL parallelism. JigsawRL decomposes each pipeline into a Sub-Stage Graph that exposes the intra-stage and inter-worker imbalance hidden by stage-level systems. On this abstraction, JigsawRL resolves multiplexing interference through dynamic resource allocation, eliminates fragmented utilization by migrating long-tail rollouts across workers, and formulates their coordination as a graph scheduling problem solved with a look-ahead heuristic. On 4-64 H100/A100 GPUs across different agentic RL pipelines and models, JigsawRL achieves up to 1.85x throughput over Verl on synchronous RL, 1.54x over StreamRL and AReaL on asynchronous RL, and supports heterogeneous pipelines with moderate latency trade-off.

We study learned memory tokens as a computational scratchpad for a single-block Universal Transformer with Adaptive Computation Time (ACT) on Sudoku-Extreme, a combinatorial reasoning benchmark. Memory tokens are empirically necessary: no configuration without them reaches non-trivial performance. The optimal count has a sharp lower threshold (T=0 always fails, T=8 reliably succeeds) followed by a stable plateau (T=8-32, 57.4% +/- 0.7% exact-match) and a dilution boundary at T=64. Under halt-side pressure (lambda warmup), mean halt drops monotonically with memory size across the plateau (from 11.6 at T=8 to 8.3 at T=64), showing that memory tokens and ponder depth substitute as resources at fixed accuracy. We also identify a router initialization trap that causes the majority of training runs to fail: both default zero-bias and Graves' recommended positive bias settle into a shallow halt equilibrium the model cannot escape. Inverting the bias to -3 ("deep start") eliminates the failure mode, and ablation shows the trap is inherent to ACT initialization rather than an artifact of our architecture. With reliable training, ACT yields an order of magnitude lower seed variance than fixed-depth processing (+/-0.7 vs +/-9.3 pp); lambda warmup recovers 34% of compute at matched accuracy; and attention heads specialize into memory readers, constraint propagators, and integrators across recursive depth. Code: https://github.com/che-shr-cat/utm-jax.

Modern AI systems increasingly rely on workflows composed of multiple interacting agents, some powered by large language models (LLMs) and others by conventional computational modules. This paper analyzes the fundamental tradeoffs between latency, reliability, and cost in LLM-enabled agentic workflows. We introduce performance models for both LLM and non-LLM agents that capture the relationship between computational effort and output quality, incorporating the impact of reasoning and output tokens for LLM agents using a parametric exponential reliability function. Then, we study the design of sequential workflows under latency and cost constraints. Main results include a water-filling token allocation policy and characterizations of optimal workflow reliability in terms of shadow prices.

Despite the impressive capabilities of large language models, their substantial computational costs, latency, and privacy risks hinder their widespread deployment in real-world applications. Small Language Models (SLMs) with fewer than 10 billion parameters present a promising alternative; however, their inherent limitations in knowledge and reasoning curtail their effectiveness. Existing research primarily focuses on enhancing SLMs through scaling laws or fine-tuning strategies while overlooking the potential of using agent paradigms, such as tool use and multi-agent collaboration, to systematically compensate for the inherent weaknesses of small models. To address this gap, this paper presents the first large-scale, comprehensive study of <10B open-source models under three paradigms: (1) the base model, (2) a single agent equipped with tools, and (3) a multi-agent system with collaborative capabilities. Our results show that single-agent systems achieve the best balance between performance and cost, while multi-agent setups add overhead with limited gains. Our findings highlight the importance of agent-centric design for efficient and trustworthy deployment in resource-constrained settings.

The key-value (KV) cache is the dominant memory bottleneck during Transformer inference, yet little is known theoretically about how aggressively it can be compressed before multi-step reasoning degrades. We study this through $k$-hop pointer chasing on $n$ tokens under a shared KV cache of size $s$, attention dimension $m$, $H$ heads, $p$-bit precision, and a locality-respecting cache controller (satisfied by all standard KV-compression methods). We give three results. (1) Product depth lower bound (conjectured). We conjecture that any such Transformer ($n \geq 4k$, $s \leq \sqrt{n}/4$) requires depth $L = Ω(\lceil k/s \rceil \cdot \lceil \log_2 n/(Hmp) \rceil)$, and isolate the sole remaining gap as a probabilistic step on the joint distribution of cache trace and pointer chain. Unconditionally, we prove a matching upper bound $L = O(\min(k, \lceil k/s \rceil \log s) \cdot \log n/(mp))$ via windowed pointer doubling, and a max-bound $L = Ω(\max(\lceil k/s \rceil, \log n/(Hmp)))$. Closing the conjecture amounts to upgrading max to product. (2) Bandwidth barrier. The product bound binds only when $Hmp \lesssim \log n$. Any lower bound provable via per-window distinguishability counting -- including reachability, bandwidth, and combinations -- cannot exceed $\lceil k/s \rceil$ once $Hmp \geq \log_2 n$. Breaking this requires lifting unconditional communication-complexity bounds for pointer chasing to Cache-Transformer depth. (3) Adaptive vs oblivious error scaling. Under random cache over $T = \lceil \log_2 k \rceil$ doubling stages, oblivious caches give $\Pr[\mathcal{E}] \leq (s/(n-T))^T + 2T^3/n$ (exponential in $T$), while adaptive locality-respecting caches achieve $\Pr[\mathcal{E}] = s/n$ exactly, independent of $T$. The $Ω((n/s)^{T-1})$ separation explains why heavy-hitter eviction empirically dominates random eviction for multi-hop reasoning.

Open-weight Small Language Models(SLMs) can provide faster local inference at lower financial cost, but may not achieve the same performance level as commercial Large Language Models (LLMs) that are orders of magnitudes larger. Consequently, many of the latest applications of LLMs, such as software engineering agents, tend to be evaluated on larger models only, leaving the issue of improving the cost-benefit trade-off of such applications neglected. This paper proposes Atropos, a predictive early-termination analysis and hotswap technique that aims to improve the cost-benefit trade-off for LLM-based agents that use self-consistency. The core component of ATROPOS is a predictive model based on structural properties of LLM inferences: after merging multiple agentic inference paths into a graph representation, ATROPOS uses Graph Convolutional Network (GCN) to predict whether an ongoing inference will eventually succeed or not. If an agentic task instance running on the source LLM is predicted to fail, ATROPOS subsequently performs hotswapping, i.e., migrating the on-going inference context onto the more capable target LLM: this is feasible because LLM contexts are stateless. An empirical evaluation of ATROPOS using three recent LLM-based agents shows that ATROPOS can predict early termination of eventually failing inferences with the accuracy of 0.85 at the midpoint of the inference. Hotswapping LLMs for such inferences can convert up to 27.57% of them to be successful. Consequently, ATROPOS achieves 74.35% of the performance of closed LLMs with as low as only 23.9% of the cost.

Large-scale renewable energy deployment introduces pronounced volatility into the electricity system, turning grid operation into a complex stochastic optimization problem. Accurate electricity price forecasting (EPF) is essential not only to support operational decisions, such as optimal bidding strategies and balancing power preparation, but also to reduce economic risk and improve market efficiency. Probabilistic forecasts are particularly valuable because they quantify uncertainty stemming from renewable intermittency, market coupling, and regulatory changes, enabling market participants to make informed decisions that minimize losses and optimize expected revenues. However, it remains an open question which models to employ to produce accurate forecasts. Should these be task-specific machine learning (ML) models or Time Series Foundation Models (TSFMs)? In this work, we compare four models for day-ahead probabilistic EPF (PEPF) in European bidding zones: a deterministic NHITS backbone with Quantile-Regression Averaging (NHITS+QRA) and a conditional Normalizing-Flow forecaster (NF) are compared with two TSFMs, namely Moirai and ChronosX. On the one hand, we find that TSFMs outperform task-specific deep learning models trained from scratch in terms of CRPS, Energy Score, and predictive interval calibration across market conditions. On the other hand, we find that well-configured task-specific models, particularly NHITS combined with QRA, achieve performance very close to TSFMs, and in some scenarios, such as when supplied with additional informative feature groups or adapted via few-shot learning from other European markets, they can even surpass TSFMs. Overall, our findings show that while TSFMs offer expressive modeling capabilities, conventional models remain highly competitive, emphasizing the need to weigh computational expense against marginal performance improvements in PEPF.

The fundamental limit of natural signal compression has traditionally been characterized by classical rate-distortion (RD) theory through the tradeoff between coding rate and reconstruction distortion, while the rate-distortion-perception (RDP) framework introduces a divergence-based measure of perceptual quality as a modeling principle, leaving its theoretical origin unclear. In this paper, motivated by a synonymity-based semantic information perspective, we reformulate perceptual reconstruction as recovering any admissible sample within an ideal synonymous set (synset) associated with the source, rather than the source sample itself, and establish a synonymous source coding architecture. On this basis, we develop a synonymous variational inference (SVI) analysis framework with a synonymous variational lower bound (SVLBO) for tractable analysis of synset-oriented compression. Within this framework, we establish a synonymity-perception consistency principle, showing that optimal identification of semantic information is theoretically consistent with perceptual optimization. Based on this result, we further derive a tight-bound synonymous source coding rate characterization and show that its Jensen-limit relaxation leads to a synonymous rate-distortion-perception form for practical optimization. These analytical results show that the distributional divergence term arises naturally from the synset-based reconstruction objective, clarify its compatibility with existing RDP formulations and classical RD theory, and suggest the potential advantages of synonymous source coding.

Experimental evidence suggests that AI tools raise worker productivity, but also that sustained offloading can erode expertise. This creates a tradeoff when expertise is a complement to AI. To explore the consequences of this tradeoff, we develop a dynamic model in which a decision-maker chooses the intensity of practice-displacing AI offloading for a worker over time, trading immediate productivity against the erosion of worker skill. We decompose the tool's productivity effect into two components, one independent of worker expertise and one that scales with it. The model produces three main results. First, a decision-maker who fully anticipates skill erosion still rationally adopts AI when front-loaded gains outweigh long-run skill costs, lowering long-run productivity. The decomposition sorts deployments into five regions by their long-run effect, separating beneficial from harmful adoption. Second, the tradeoff introduces the potential for misaligned incentives to harm workers. When managers are short-termist or worker skill has external value, AI use can leave the worker worse off than with no AI, the outcome we call the augmentation trap. Third, when AI substitutes strongly enough for expertise, offloading can generate permanent divergence, with high-skill workers realizing their potential and low-skill workers deskilling. Small differences in managerial incentives can determine which path a worker takes.

Visual odometry (VO) frontends are typically tuned offline by domain experts on pre-recorded datasets and then deployed with fixed hyperparameters. Yet a configuration that performs best on a benchmark is not guaranteed to remain best when texture, illumination, motion blur, sensor noise, or computational conditions change at deployment. We propose a frontend that instead adapts its parameters automatically and continuously. We formulate frontend tuning as a sequential decision-making problem and introduce an image-conditioned reinforcement-learning policy that combines a lightweight embedding of the current image with a compact set of frontend statistics. At each decision step, the policy selects the FAST detection threshold, KLT patch size, and RANSAC rejection threshold; a privileged critic provides additional context only during training. Trained on synthetic TartanAirV2 data, the policy transfers zero-shot to a monocular-inertial OpenVINS pipeline on EuRoC, TUM-VI, and UZH-FPV. On synthetic test sequences, the learned policy improves the tracking-computation trade-off over an optimized static parameter configuration. On the three real-world benchmarks, it reduces mean ATE by up to 8% and runtime by up to 57% relative to static parameters baseline. These results show that image-conditioned online adaptation can improve the accuracy-computation trade-off beyond a single configuration selected offline.

As AI usage becomes more prevalent in social contexts, understanding agent-user interaction is critical to designing systems that imp rove both individual and group outcomes. We present an online behavioral experiment (N=243) in which participants play three multi-tu rn bargaining games in groups of three. Each game, presented in randomized order, grants access to a single LLM assistance modality: proactive recommendations from an Advisor, reactive feedback from a Coach, or autonomous execution by a Delegate. All three modalitie s are powered by an LLM with super-human performance within this negotiation setting. On each turn, participants privately decide whe ther to act manually or use the AI modality available in that game. We document a preference-performance misalignment: participants s trongly prefer the higher-control Advisor (44%) over the Delegate (19%), yet groups only significantly increase collective surplus un der Delegate access. Adjusting for voluntary non-compliance, delegating to the AI yields suggestive individual welfare gains, roughly 1.5x the intent-to-treat estimate. A mechanism analysis traces this gap to a human filter: AI-generated proposals create more joint surplus than manual proposals across all conditions, but in the Advisor and Coach modes users modify, override, or ignore the AI's su ggestions, reverting toward human-baseline trade patterns. The Delegate advantage arises not from a different AI capability but from bypassing this filtering step altogether. Realizing these welfare gains depends not only on model capability, but on the interaction structure through which that capability is delivered. We argue that assistance modalities should be designed as mechanisms with endog enous participation; adoption-compatible interaction rules are a prerequisite to improving welfare with automated assistance.

Low-Rank Adaptation (LoRA) methods have emerged as crucial techniques for adapting large pre-trained models to downstream tasks under computational and memory constraints. However, they face a fundamental challenge in balancing task-specific performance gains against catastrophic forgetting of pre-trained knowledge, where existing methods provide inconsistent recommendations. This paper presents a comprehensive analysis of the performance-forgetting trade-offs inherent in low-rank adaptation using principal components of weight matrices as initialization. Our investigation reveals that fine-tuning intermediate components leads to better balance and robustness to high learning rates than first (PiSSA) and last (MiLoRA) components in existing work. Building on these findings, we provide practical guidelines for initialization of LoRA methods to balance the performance-forgetting trade-off. In a thorough empirical study on a variety of computer vision and NLP tasks we confirm that these guidelines achieve high accuracy and reduced forgetting.

In speech machine learning, neural network models are typically designed by choosing an architecture with fixed layer sizes and structure. These models are then trained to maximize performance on metrics aligned with the task's objective. While the overall architecture is usually guided by prior knowledge of the task, the sizes of individual layers are often chosen heuristically. However, this approach does not guarantee an optimal trade-off between performance and computational complexity; consequently, post hoc methods such as weight quantization or model pruning are typically employed to reduce computational cost. This occurs because stochastic gradient descent (SGD) methods can only optimize differentiable functions, while factors influencing computational complexity, such as layer sizes and floating-point operations per second (FLOP/s), are non-differentiable and require modifying the model structure during training. We propose a reparameterization technique based on feature noise injection that enables joint optimization of performance and computational complexity during training using SGD-based methods. Unlike traditional pruning methods, our approach allows the model size to be dynamically optimized for a target performance-complexity trade-off, without relying on heuristic criteria to select which weights or structures to remove. We demonstrate the effectiveness of our method through three case studies, including a synthetic example and two practical real-world applications: voice activity detection and audio anti-spoofing. The code related to our work is publicly available to encourage further research.

LLM services need to offer a family of models spanning different capability--cost trade-offs to accommodate diverse user preferences. Model merging offers a practical way to construct such a model family by combining a reasoning-enhanced model with an instruction-tuned model. Compared with model-level merging, layer-wise merging offers finer control over the capability--cost trade-off by assigning different merge weights to individual layers. However, it introduces two practical challenges: the layer-wise search space is high-dimensional, and existing methods rarely exploit informative signals from the source models; moreover, the highly variable runtime of LLM evaluations makes synchronous batch optimization inefficient by leaving GPU resources idle while waiting for slow evaluations. To address these challenges, we propose Asynchronous Prior-Guided Bayesian Model Merging (AP-BMM), which formulates layer-wise merging as a multi-objective optimization problem to approximate a Pareto set of merged LLMs, yielding a family of Pareto-optimal merged models with diverse capability--cost trade-offs. AP-BMM leverages parameter and activation discrepancies between the source models to guide the early layer-wise search, employs asynchronous pending-aware Bayesian optimization to maximize GPU utilization through asynchronous evaluations and to select high-quality candidates via pending awareness, and applies lightweight ranking over an oversized candidate pool to improve Pareto-front coverage. Under fixed evaluation budgets, AP-BMM achieves higher hypervolume (HV) and broader Pareto-front coverage than synchronous layer-wise and representative model-level merging methods, while reducing wall-clock time through higher GPU utilization.

The structure of naming systems in natural languages hinges on a trade-off between high informativeness and low complexity. Focusing on the domain of kinship naming, we analyze such trade-off while addressing simplifying assumptions of prior work, namely: (i) universal communicative need across languages, and (ii) optimal listeners. To that aim, we collect data from four different languages, and analyze how different communicative needs and variations in the listener model influence the informativeness--complexity trade-off. Adopting a referential game setup from emergent communication, we further show that trade-off optimality is not only theoretically achievable but also emerges empirically in learned communication systems.

The "alignment tax" of post-training is typically framed as a drop in task accuracy. We show it also involves a severe loss of calibration, making models overconfident, less reliable, and model outputs less diverse. We show that this trade-off can be navigated effectively via a simple post-hoc intervention: interpolating between a model's weights before and after alignment. Crucially, this is not a strict trade-off. We find that the process consistently reveals Pareto-optimal interpolations - models that improve accuracy beyond both parents while substantially recovering the calibration lost during alignment. Our work demonstrates that simple model merging provides a computationally efficient method for mitigating the full scope of the alignment tax, yielding models that are more capable and more reliable.

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose ParallelBench, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We release our benchmark to help accelerate the development of truly efficient dLLMs.

Transformers have achieved remarkable successes across a wide range of applications, yet the theoretical foundation of their model efficiency remains underexplored. In this work, we investigate how the model parameters -- mainly attention heads and head dimensions -- should be allocated across layers to balance expressivity and efficiency. We first provide mathematical analysis on the role of early layers in information extraction from an approximation perspective, with a theoretical characterization on the trade-off between the number of heads and head dimension under a fixed parameter budget. In addition, we uncover and prove the \emph{saturation} behavior of softmax activations: Continuously increasing head dimensions can lead to diminishing returns in learning errors, particularly for long sequences. Supported by both theory and experiments, this saturation pattern suggests that later layers can operate more efficiently with reduced parameters. Combining these insights, we propose principled strategies for allocating attention heads and dimensions across Transformers' layers, shedding light on theoretically-grounded model efficiency of Transformer-based architectures.

The factors driving the performance of in-context learning (ICL) in large language models (LLMs) remain poorly understood despite ICL's surprising effectiveness, enabling models to adapt to new tasks from only a handful of examples. To clarify and improve these capabilities, we characterize how the statistical properties of the pretraining distribution (e.g., tail behavior, coverage) shape ICL. We develop a theoretical framework that encompasses generalization and task selection and show how distributional properties govern sample efficiency, task retrieval, and robustness. To this end, we generalize existing concentration results to heavy-tailed priors and dependent sequences, better reflecting the structure of LLM pretraining data. Our framework reveals a fundamental design trade-off: heavy-tailed pretraining distributions facilitate robust task selection under distribution shifts but are detrimental to generalization, especially in low-data regimes. We then empirically evaluate our predictions by studying how ICL performance varies with the pretraining distribution on challenging tasks such as stochastic differential equations and stochastic processes with memory. Together, these findings suggest that controlling key statistical properties of the pretraining distribution is essential for building ICL-capable and reliable LLMs.

The retrieval stage of retrieval-augmented generation (RAG) for scientific question answering depends on how documents are segmented and how chunks are represented in embedding space. This dependence is especially relevant to chemistry texts, which contain dense terminology, symbolic notation, quantitative evidence, and context associated with document structure. However, benchmark-based evidence on the interaction between chunking strategy and embedding model remains limited for chemistry-specific retrieval. Using ChemQuests, a corpus of 952 question-answer pairs from 151 ChemRxiv papers across 17 chemistry subfields, we construct chunk-level, Massive Text Embedding Benchmark (MTEB)-compatible retrieval benchmarks for controlled evaluation. We first screen 41 embedding models on the external chemistry retrieval benchmarks ChemNQRetrieval and ChemHotpotQARetrieval using a geometric-mean metric at rank 10 (Geom@10), which we validate against the full retrieval-metric profile. We then evaluate shortlisted models on ChemQuests-derived tasks across five chunking strategies, seven chunk sizes, and multiple overlap settings. Embedding choice is associated with the largest observed differences in evidence retrieval, with retrieval-tuned E5, Beijing Academy of Artificial Intelligence General Embedding (BGE), and Nomic models among the strongest overall. Within the evaluated grid, medium-to-large chunks combined with fixed-token, recursive-token, or hierarchical-section chunking provide a practical starting point for the retrieval stage of chemistry-aware RAG. Low overlap was generally favored where overlap variation was evaluated.

When asked to explain their decisions, LLMs can often give explanations which sound plausible to humans. But are these explanations faithful, i.e. do they convey the factors actually responsible for the decision? In this work, we analyse counterfactual faithfulness across 75 models from 13 families. We analyze the tradeoff between conciseness and comprehensiveness, how correlational faithfulness metrics assess this tradeoff, and the extent to which metrics can be gamed. This analysis motivates two new metrics: the phi-CCT, a simplified variant of the Correlational Counterfactual Test (CCT) which avoids the need for token probabilities while explaining most of the variance of the original test; and F-AUROC, which eliminates sensitivity to imbalanced intervention distributions and captures a model's ability to produce explanations with different levels of detail. Our findings reveal a clear scaling trend: larger and more capable models are consistently more faithful on all metrics we consider. Our code is available at https://github.com/google-deepmind/corr_faith.

Reliability measures associated with the prediction of the machine learning models are critical to strengthening user confidence in artificial intelligence. Therefore, those models that provide not only predictions, but also reliability, enjoy greater popularity. In the field of recommender systems, reliability is crucial, since users tend to prefer those recommendations that are sure to interest them, that is, high predictions with high reliabilities. In this paper, we propose Restricted Bernoulli Matrix Factorization (ResBeMF), a new algorithm aimed at enhancing the performance of classification-based collaborative filtering. This model is based on a collection of restricted matrix factorizations that jointly generate, for each user-item pair, a full probability distribution over the possible rating scores. To prove its effectiveness, the proposed model has been compared to other existing solutions in the literature in terms of prediction quality (Mean Absolute Error and accuracy scores), prediction quantity (coverage score) and recommendation quality (Mean Average Precision score). The experimental results demonstrate that the proposed model provides a good balance in terms of the quality measures used compared to other recommendation models.

We provide a theoretical analysis of Adam under non-stationary stochastic objectives, separating two regimes: Euclidean tracking under adaptive strong monotonicity of the Adam-preconditioned mean-gradient operator, and high-probability projected stationarity guarantees under general $L$-smooth objectives. In the tracking regime, we derive finite-time expected and high-probability bounds that decompose sharply into four components: initialization, objective drift, a first-moment tracking error governed by $\beta_1$, and a preconditioner perturbation governed by $\beta_2$. We characterize the burn-in time required for the transient terms to decay to the asymptotic tracking bound under constant and step-decay schedules. We also prove a high-probability bound on the average projected stationarity gap for Adam under distribution shift. Across both analyses, our bounds reveal a noise--drift tradeoff: in noise-dominated regimes, first-moment averaging and adaptive preconditioning can yield favorable upper guarantees, whereas in drift-dominated regimes, stale first-moment information and preconditioner perturbations can enlarge Adam's tracking guarantee, potentially allowing vanilla SGD to attain a smaller tracking error. Our explicit $(\beta_1,\beta_2,\epsilon)$-dependent bounds identify mechanisms through which adaptive step-sizing can help or hurt under nonstationarity and provide theoretical explanations consistent with Adam's empirical instability and stabilization under distribution shift.