Overview

Efficient inference studies how to get useful model outputs with less computation, memory, time, or energy, which matters for large-scale serving and devices with limited resources. Approaches include quantization and distillation, caching and reuse, sparse or low-rank computation, early exits, amortized inference, and adaptive allocation of test-time compute. Systems work complements these methods with batching, scheduling, compilation, and hardware-aware execution, while probabilistic methods seek faster posterior or marginal estimates without losing calibration or exactness. The central tension is reducing cost without undermining accuracy, uncertainty estimates, or robustness, and results can depend strongly on the workload and hardware used for evaluation.

Latest papers

118 papers14 papers in the last four weeks, down 22% on the four weeks before. 0.1% of all new papers.

Nov 13, 2026

Speed-Variation Distillation for Generative Modeling

While diffusion models excel at image generation, their denoising trajectories pass through corrupted intermediate states far from the image manifold. We ask whether generation can instead proceed through recognizable images throughout the entire trajectory. We introduce a generative framework built around a variation-reducing degradation process that contracts diverse images toward a shared prototype while preserving recognizable intermediate states. Because such manifold traversal lacks a simple analytical form, we construct the trajectories using generative-prior-based image morphing. We then distill the reverse process into a diversification operator that transforms the prototype into increasingly varied samples. To resolve the one-to-many ambiguity of the reverse process, we associate each trajectory with its initial noise latent and apply perturbations to improve robustness against accumulated errors. At inference, every step yields a recognizable image, allowing early termination and enabling a controllable speed--variation trade-off.
Nov 8, 2026

How Much Training-Inference Mismatch Can On-Policy Distillation Tolerate?

On-policy distillation (OPD) trains a student language model on its own responses, guided by a fixed teacher. Generating these responses is expensive, so systems use quantized inference and asynchronous rollouts to reduce the cost. Both, however, introduce training-inference mismatch between the rollout engine and the model being trained. Such mismatch can destabilize reinforcement learning (RL), but its effects on OPD remain unclear. This paper investigates how much mismatch OPD tolerates, where its effects arise, and how training systems can use this tolerance. Measuring all mismatch sources on one scale, we examine quantization across three student-teacher pairs, with additional tests of weight staleness and injected noise on one pair. OPD tolerates substantial mismatch without meaningful loss of training quality but degrades at higher levels, revealing an empirical mismatch budget. Both the quality cost of mismatch and the mismatch introduced by quantization depend on the model, with some layers particularly sensitive to reduced precision. In controlled comparisons, OPD remains stable under rollout configurations that destabilize RL methods. Under weight noise, mismatch also shifts the prefixes reached during generation, introducing errors that per-token corrections cannot address. To put these findings into practice, we introduce a lightweight admission check that screens rollout configurations against a calibrated budget without requiring a full training run for each candidate. Model-specific FP8 recipes preserve training quality while providing capacity for at least 1.93×1.93\times as many KV-cache tokens as BF16. Combining FP8 with asynchronous rollouts reaches the same training quality 1.62×1.62\times sooner than synchronous BF16 training. Code is available at https://github.com/benyucong/train-inference-mismatch-opd
Nov 6, 2026

Half-precision inference of AI weather prediction models: negligible loss of skill at a fraction of the cost

AI weather prediction models are released with 32-bit (fp32) weights and run in fp32 by default, although current GPUs compute much faster in 16-bit formats. We run five publicly available models -- Aurora, Pangu-Weather, FourCastNet, SFNO and FastNet -- on a single 16 GB GPU from 52 initial dates in 2024, in fp32 and with their networks in half precision (fp16), and verify the forecasts against ERA5. The physical state must stay in fp32: converting the whole model rounds it in physical units, which in fp16 overflows and in bf16 discards meaningful information. With the state, its normalisation and denormalisation in fp32, running the network in fp16 changes the global RMSE by at most 1.4%, and by about 0.5% or less for four of the five models, equivalent to less than one hour of lead time up to day 8 and at most two hours at day 10. The differences show no systematic pattern by region, scale or for tropical and extratropical storms, apart from a small tropical penalty in FourCastNet, and fp16 leaves the CRPS, spread and calibration of Aurora's ensemble almost unchanged. No value overflowed fp16's numeric range (+/-65,504); the largest stay 38 to 544 times below it. A 10-day forecast in fp16 is 1.3-3.5 times faster than in fp32 and uses 1.4-3.7 times less energy. Sensitivity experiments show that rules of thumb from other applications of reduced precision carry over only partly: the first and last layers dominate for some models but not for others, and no type of layer is sensitive on its own. For running AI weather models on a single GPU, fp16 with the state in fp32 is practically lossless at a fraction of the cost. bf16, often preferred for its wider range, is as fast but less accurate than fp16 for every model, and simple 8-bit quantisation costs far more skill.
Oct 31, 2026

BitLLM: Efficient Low-Bit, Conversion-Hiding LLM Decoding on Mobile GPUs

The rapid scaling of large language models (LLMs) has unlocked remarkable capabilities across various domains. However, deploying LLMs on mobile devices remains challenging because of their high computational and memory requirements. Low-bit quantization (e.g., 4-bit weights, 8-bit activations) can reduce model size while preserving accuracy, but its practical efficacy on mobile GPUs has been limited so far. In this paper, we identify that the performance gap stems primarily from frequent data type conversions between quantized and floating-point values, and dynamic and inefficient memory access patterns caused by the growing Key-Value (KV) cache during autoregressive decoding. We introduce BitLLM, a mobile inference engine that addresses type conversion overhead and suboptimal memory accesses in quantized LLMs. Our approach features an adaptive layout and packing strategy for the KV cache that ensures efficient memory access throughout LLM decoding. This is coupled with an integer-native execution pipeline that defers and overlaps type conversion outside the compute hot path, keeping the critical MatMul stages in the integer domain. Implemented on off-the-shelf mobile GPUs, BitLLM outperforms leading engines, i.e., LLAMA.CPP, MNN-LLM, LiteRT-LM, and MLC-LLM, achieving average speedups of 1.78×\times, 1.73×\times, 1.58×\times, and 1.47×\times, respectively, across various models and quantization levels. Under long-context generation, BitLLM reduces average energy consumption (per token) by up to 88.9%.
Oct 29, 2026

Gaussian Distillation of Flow Policies

Flow policies are among the strongest robot controllers and are well suited to modelling multimodal action distributions. Our analysis shows that the continuous-action distributions of the studied flow policies are well approximated by Gaussians. We propose the Gaussian Action Readout (GAR), which replaces a flow policy's final layer with a Gaussian readout trained on just 0.02% of eligible states, while keeping the rest of the network frozen. It samples with one action-network pass, gives closed-form likelihoods, and achieves near-teacher task success at up to twice the inference speed.
Oct 27, 2026

Frugal Handwritten Text Recognition for Sub-Saharan Civil-Status Registers: A Smartphone-to-Records Pipeline with CPU-Only Inference

Handwritten civil-status registers across Sub-Saharan Africa are being photographed, but rarely transcribed, which leaves their content unsearchable. Local administrations that hold these records typically lack GPUs, reliable connectivity, and the budget for external digitization services. We study whether a frugal pipeline, with consumer-grade capture and CPU-only inference, can make such registers machine-readable. We introduce a corpus of 1,568 smartphone images of handwritten birth, marriage, and death records from a municipal registry in Chad, with line-level transcriptions for 185 records. Our pipeline combines a fully classical preprocessing stage, running at about 2.6 images per second on a laptop CPU, with a Kraken recognizer fine-tuned from a pretrained model in about four hours on a free-tier GPU. On manually corrected line segmentation and a single-registry test set whose writers also appear in training, fine-tuning reduces the character error rate from 14.53% to 0.50% and the word error rate from 56.93% to 3.37%. On a six-core laptop CPU, the recognizer weighs 61 MiB, peaks below 2.6 GiB of RAM, and processes an image in under 6 s on average, excluding line segmentation. Finally, a downstream information-extraction task shows that residual recognition errors cost about 9 F1 points on role resolution, compared with a 36.7-point gap between the fine-tuned and pretrained recognizers. These results are evidence of feasibility within one registry, not of performance on unseen writers or sites; we discuss these limitations in detail.
Oct 26, 2026

Training-Free Inference Acceleration for Video Large Language Models: Methods, Trade-offs, and Evaluation

Video Large Language Models (VLMs) have demonstrated strong capabilities in video understanding and reasoning, but processing long visual sequences incurs substantial computational, latency, and memory overhead. This survey examines training-free inference acceleration, focusing on reducing inference costs while preserving task-relevant spatiotemporal evidence. We organize existing methods by their intervention stages in the inference pipeline, covering input, visual encoding, projector, prefill, and decoding. Post-training quantization (PTQ) is discussed separately as a cross-stage extension with varying calibration requirements. We analyze how intervention position affects evidence preservation and computational savings, and why local efficiency improvements may not translate into end-to-end acceleration. We complement this analysis with quantitative evidence from prior studies and our own experiments under a shared evaluation setting to assess accuracy-latency trade-offs.
Oct 26, 2026

WorldPager: Zero-Stall Tiered Memory for Long-Horizon World Models on a Single Consumer GPU

Interactive world models keep the KV of past chunks as retrieval memory so that revisited places stay consistent. On a single consumer machine (one RTX 4090, 32 GiB of host memory, one NVMe SSD), long sessions with such memory hit two walls. The first is capacity: the KV of Matrix-Game 2.0 exhausts host memory after 45 seconds of video in bf16 and after 2.8 minutes in int4. The second is the model's positional horizon of 4.3 minutes. We show that this memory is capacity-bound rather than bandwidth-bound and that its access set is known before compute, because the camera pose is an input. An analytical model built on these observations predicts spill points to within one chunk and classifies zero versus non-zero stalls correctly on held-out configurations. Guided by it, WorldPager stores int4 KV in a GPU/DRAM/NVMe hierarchy instead of recomputing it, prefetches each chunk's memory from the extrapolated pose, never waits on a miss, and writes back in the background; to our knowledge, it is the first to place the KV memory of an interactive world model on an NVMe tier with zero-stall prefetching. With training-free position handling, it generates 37.5 minutes of video continuously with zero stalls and constant host memory, at about 9.5 frames per second. On a sealed test set, int4 storage shows no measurable quality loss; on long-gap revisits it lowers image quality by about one MUSIQ point.
Oct 26, 2026

Inference Latency-Aware Tsetlin Machine Training

The Tsetlin Machine (TM) is a logic-based machine learning model well suited to low-latency inference on resource-constrained edge devices. Because each clause is an AND over Boolean literals, both the literals and their associated Tsetlin Automaton (TA) actions can be represented at the bit level, enabling bitwise evaluation and an early exit that is triggered as soon as any included literal evaluates to zero, forcing the clause output to zero. In this work, we study how the two TM training hyperparameters, T and s, govern both classification accuracy and inference latency. Building on the inversely proportional T-s relationship along which accuracy is nearly constant, we find that inference latency is reduced primarily by increasing s; rather than exhaustively searching the (T, s) grid at high training cost, we propose a directed search strategy that treats accuracy as a constraint and advances toward larger s to reduce inference latency within a fixed evaluation budget. Experimental results on the gem5 simulator with an ARM processor show that, for certain datasets, the search is able to increase s while preserving the baseline accuracy, thereby reducing inference latency at negligible accuracy cost: across six datasets, the method reduces inference latency by up to 16.34%, while some datasets exhibit an accuracy reduction of only 0.06%. The proposed strategy requires far fewer configuration evaluations than an exhaustive grid search.
Oct 26, 2026

GUIFlow: Temporal Redundancy-Aware Incremental Inference for GUI Agents

GUI agents built on vision-language models interact with continuously evolving interfaces, yet conventional inference repeatedly recomputes representations of largely overlapping observations and histories. This mismatch makes multi-step interaction costly, while changes in context, interface content, and action arguments prevent straightforward cache reuse. We introduce GUIFlow, a framework for full-pipeline incremental inference in multi-image GUI interaction. Our central insight is that advancing a decision need not require rebuilding all supporting representations: each interaction step can instead selectively update computational states maintained across the trajectory. To support new decisions as the context evolves, GUIFlow adapts how past computation is reused: historical KV states are position-aligned and injected into the current context; change signals guide selective updates of current-observation states in the vision encoder and language model; and earlier action responses serve as speculative drafts verified by the target model. Together, these mechanisms enable cross-step reuse throughout visual encoding, prefill, and decoding without additional training, a separate draft model, or changes to the model-visible inputs and action interface. Experiments demonstrate up to 3.39x geometric-mean speedup across offline multi-image benchmarks and up to 3.49x in online, closed-loop interaction on OSWorld, while maintaining task performance broadly comparable to the dense baseline.
Oct 25, 2026

HIFI: A Highly Efficient Framework for Statistical Inference with Foundation Models

Foundation models provide powerful representations for unstructured data, such as text and images, and parameter-efficient fine-tuning (PEFT) makes their adaptation to downstream tasks increasingly practical. For neural estimators that lack a tractable sampling distribution, resampling-based procedures offer a flexible route to statistical inference, but they often require hundreds of model fits, making them computationally prohibitive at foundation-model scale. We introduce HIFI, a two-stage fine-tuning framework for statistical inference on parameters such as conditional means. Stage 1 adapts a foundation model once on a larger sample, capturing the complex structure of the data and controlling estimation bias through a single fit. Stage 2 applies a resampling inference procedure to a smaller independent sample, concurrently fitting multiple lightweight PEFT estimators to the residual to quantify estimation variability without relearning the full task. HIFI accommodates a broad class of resampling procedures, including ensemble subsampling and the nonparametric bootstrap. Across vision and text experiments, HIFI scales to pretrained models ranging from tens of millions to more than ten billion parameters while providing accurate point estimates and well-calibrated confidence intervals across diverse architectures, and reduces training compute by up to 91.9% relative to direct resampling. An application to gene-expression inference illustrates the practical value of the framework for uncertainty-aware scientific inference.
Oct 21, 2026

Understanding Target Probability Reuse in Speculative Decoding

Speculative decoding uses a target model to verify a tree of draft tokens in parallel. We study whether the resulting next-token distributions can help select the next tree. For unrestricted target models, we show that guaranteed exact reuse requires a distribution computed at the next round's complete context. We characterize when retaining this distribution requires sacrificing accepted tokens and establish conditions under which this loss persists across decoding rounds. Across six targets, we find that reranking first-token branches with a separately computed distribution at the current context improves expected acceptance over draft ranking, especially with small tree budgets. Drawing the first token from this distribution before allocating the tree further improves expected acceptance over pruning a preselected tree, even at equal mean tree size. In our experiments, both selection rules also decode faster than draft ranking when the separate target pass overlaps standalone drafting. Our findings show that making target probabilities available before tree allocation can improve the efficiency of speculative decoding. Our code is available at https://github.com/js-lee-AI/TargetRows.
Oct 19, 2026

Learning Goal-Conditioned Subgoals through Frozen Hierarchical World Models

Hierarchical world models plan over long horizons by passing predicted states from a coarse model to a lower-level controller. Choosing these states can be expensive: the upper-level planner repeatedly differentiates through candidate macro-action trajectories. We replace this search with a lightweight goal-conditioned network trained through the frozen high-level predictor. The network supplies a conditioning vector, while the predictor generates the next latent waypoint. Training uses hindsight goals and state-prediction losses on offline trajectories, without search-solution targets or macro-action labels. Inference retains an autoregressive high-level rollout, but requires no high-level candidate optimization. We evaluate this approach on the shared-latent HWM checkpoints distributed with H-JEPA, using three backbone seeds and three head initializations per backbone. With a common evaluation protocol, mean success is 70.4% on Cube, 48.7% on Push-T, 100.0% on FourRoom, and 71.1% on Visual AntMaze. The method improves over the corresponding HWM search baseline on three environments and has the same mean on Push-T. Dedicated measurements on an NVIDIA L40S give 1.28-4.39x end-to-end speedups for batched evaluation. Ablations identify state-adaptive conditioning as a consistent contributor, while predictor-mediated subgoals and unrestricted conditions have environment-dependent benefits.
Oct 16, 2026

Replay Pruning and Few-Bit Integer Inference for Hardware-Efficient Speed-Sign Recognition

Speed-limit sign recognition on an automotive camera system-on-chip (SoC) must fit a detector and a classifier (the reader) into the memory, arithmetic, and energy budgets of the device. In our two-stage recognizer, the detector, a YOLO-family network with ReLU and hard-sigmoid operators, is compressed by channel replay: an iterative pruning cascade identifies the channels to remove, the recorded deletions are replayed on the original task-trained network, the parent, and one continuous knowledge-distillation recovery follows. On a YOLOv5n, replay removes 60.3% of the parameters at the accuracy of the parent and, in validation mAP50-95 over five matched seeds, outperforms iterative pruning and training from scratch, and its per-step variant (per-step replay) is the most accurate pruning method. On a YOLOv8n, replay and per-step replay outperform every other method and the unpruned parent. The reader, a five-layer CNN with 18k parameters, and the post-processing of the YOLOv5n and YOLOv8n heads use integer arithmetic only. With 4-bit weights and 8-bit activations (W4A8), the reader matches FP32 accuracy in 12.1 KB of on-chip memory, and an executable streaming model reproduces the INT8 class predictions of its runtime on all 9,850 test crops. On a Jetson AGX Orin, INT8 is faster than FP16 only in the implicit formulation of TensorRT, and the replay detector runs fastest on the deep-learning accelerator (DLA). Trained on public data only, the combined system of 0.72 M parameters recognizes 95.2% of the signs of a held-out automotive camera at 90% precision and 94 frames per second.
Oct 13, 2026

FastEEP: Fast Intra-Instance Elastic Scaling for Expert-Parallel MoE Serving

Large-scale expert parallelism (EP) has become the standard deployment configuration for Mixture-of-Experts (MoE) models in production. Meanwhile, the highly dynamic nature of LLM workloads requires providers to continuously adjust serving capacity to match fluctuating demand. Scaling MoE serving cannot be efficiently achieved by simply adding or removing serving instances; instead, it often requires resizing the EP group within an existing instance, which we refer to as intra-instance scaling.Existing inference engines primarily rely on instance-level scaling, which destroys and reconstructs entire serving instances, incurring substantial scaling overhead. We present FastEEP (Fast Elastic Expert Parallel), a fast intra-instance scaling system for MoE serving based on the principle of Partial Reuse. FastEEP reuses existing runtime state across worker execution, GPU memory, and communication by separating reusable execution structures from the resources activated by the current EP configuration. This separation enables coordinated resource reclamation and reconfiguration while preserving costly-to-reconstruct state, allowing fast EP scaling without full instance teardown and reconstruction. Our evaluation shows that FastEEP achieves 4.5x--11.7x faster scaling than the existing vLLM-based approach across different models, enabling efficient, fine-grained scaling of MoE serving.
Oct 8, 2026

LiteNWM: Efficient Latent World Models for Onboard Visual Navigation in the Wild

Direct visual navigation policies generate trajectories efficiently but do not explicitly evaluate their future consequences. Generative navigation world models provide this foresight through visual rollouts, which are costly when evaluating multiple candidates. We present LiteNWM, a latent navigation world model that shares visual encoding across candidates and jointly predicts their action-conditioned future representations at multiple horizons, while a learned scorer uses these predictions to select trajectories. In offline evaluations on RECON, SCAND, and SACSoN, LiteNWM reduces macro-averaged trajectory error by 17.56% relative to NoMaD+NWM-XL and achieves a 128.00-fold end-to-end speedup on an RTX 5090. The same evaluator transfers from NoMaD to MBRA without proposer-specific retraining, reducing MBRA's macro-averaged trajectory error by 16.2%. In real-robot experiments in unseen indoor and outdoor environments, LiteNWM improves navigation success from 43.3% to 83.3% relative to NoMaD. These results demonstrate that LiteNWM can be deployed for future-aware planning and closed-loop navigation on a physical robot.
Oct 7, 2026

RealtimeWAM: How Fast Can I Run My World Action Model?

World Action Models (WAMs) combine visual dynamics modeling with action generation, but their high inference latency limits responsive robot control. Recent efforts accelerate inference by removing explicit future-video generation at test time, as in FastWAM, an approach that requires a specially tailored architectural design. More general caching strategies exploit feature redundancy, but redundancy alone does not capture the changing computational demands of closed-loop control. To address these challenges, we present RealtimeWAM, a general, training-free framework that coordinates parallel execution with adaptive computation for low-latency inference across diverse WAM architectures. We exploit layerwise dependencies to overlap observation processing with prediction. However, concurrent branches still compete for GPU resources, limiting the benefit of parallel execution. We therefore adapt computation throughout the pipeline through selective reuse, caching observation features in visually stable regions and reusing Transformer residuals while reserving additional refinement for small predicted adjustments. We evaluate RealtimeWAM on FastWAM and OpenWAM across RoboTwin, LIBERO, and LIBERO-Plus. On an RTX 4090, measured mean inference latencies are 24.09 and 63.09 ms, corresponding to average speedups of 8.90×\times and 10.67×\times. Average success rates are 82.75% and 87.41%, respectively, within 0.02 and 0.53 percentage points of native inference. Across five real-world tasks, RealtimeWAM improves average success rates over native inference by 17.2 and 37.2 percentage points on FastWAM and OpenWAM, respectively.
Oct 5, 2026

Finite-Sample Distribution Theory and Efficient Large-Scale Inference for Online Quantile Regression

This paper studies online quantile regression for large-scale and streaming data using Stochastic SubGradient Descent (SSGD) with constant learning rates. Classical offline inference for quantile regression is computationally and memory intensive. Existing works of online inference for quantile regression provide only asymptotic guarantees and typically require sub-exponential tail conditions for distribution theory. To bridge these gaps, we introduce new techniques to prove a quenched central limit theorem (CLT) and finite-sample Gaussian approximation for SSGD under a finite-moment assumption. We further show that Ruppert-Polyak averaging with a constant learning rate has a non-vanishing bias and fails to satisfy CLT centering at the population target. Hence we propose suffix averaging to address this issue and establish its finite-sample Gaussian approximation. Based on these results, we provide an efficient online inference method for quantile regression that avoids covariance estimation. Numerical experiments show that our method achieves desirable empirical coverage rates and competitive performance compared to other inference methods. We also apply our approach to U.S. wage data to demonstrate its practical effectiveness.
Oct 4, 2026

SFlexRCA: Lightweight, Scalable, and Flexible Root Cause Analysis for IIoT Edge Systems

Industrial Internet of Things (IIoT) systems generate high-dimensional sensor telemetry from interconnected components, where faults can propagate across the system. To address these challenges, we propose SFlexRCA (Scalable and Flexible Root Cause Analysis), a topology-free RCA framework designed for resource-constrained IIoT environments. SFlexRCA transforms multivariate telemetry into compact orthogonal representations and applies shared lightweight linear modeling, avoiding explicit graph construction, message passing, and per-variable or lag-specific parameter growth. We evaluate SFlexRCA on three publicly available IIoT datasets, BATADAL, SWaT, and WADI, spanning different numbers of monitored variables, temporal characteristics, and training-data regimes. SFlexRCA is compared with 10 statistical, causal, and non-causal baselines in terms of RCA accuracy, training efficiency, inference latency, and memory consumption. In addition, inference efficiency and memory consumption are evaluated on Raspberry Pi 3 and Raspberry Pi 5, while energy consumption is additionally measured on Raspberry Pi 5. We further investigate temporalcontext sensitivity, architectural and loss components, and alternative representations. Notably, SFlexRCA maintains strong localization performance on BATADAL despite its limited normal-operation training data, while its compact shared architecture avoids the parameter growth associated with causal and graph-based approaches. Its lightweight shared architecture further enables efficient deployment on resource-constrained IIoT edge devices. The SFlexRCA code is available at https://github.com/theamrzaki/RootCause- Analysis-Correlation-Attentive-Modeling.
Oct 2, 2026

TreeWalker: Partial Evaluation for Grouped Tree-Ensemble Inference

Many prediction tasks evaluate a trained tree ensemble on groups of rows that share feature values: discrete-time survival models expand each patient into GG time steps, click-through-rate models score every item in a search session, and scenario analyses vary a few inputs while holding the rest fixed. Standard inference treats each row independently and ignores this input-side correlation. We present TreeWalker, which instantiates partial evaluation for grouped tree-ensemble inference: constant features are static, varying features dynamic. TreeWalker walks each tree once per group, partitions a row bitmask at varying splits, and skips empty subtrees via an unsplit shortcut. It reads standard LightGBM and XGBoost models; training is unchanged. We prove a structural work decomposition: per-tree work splits into the constant-projected subtree size ∣Tc∣|T_c|, GG leaf writes, and a predicate-mask provisioning cost QQ. For the trace evaluator, per-row work approaches (dv+1)/(d+1)(d_v+1)/(d+1) as G→∞G \to \infty, giving asymptotic speedup 1/(1−ρˉG)1/(1-\barρ_G). On Intel, TreeWalker is ∼3×{\sim}3\times faster than a row-independent traversal at the reference model size (T=500T{=}500, L=8L{=}8; G=16G{=}16 for SUPPORT/FLCHAIN and empirical groups for Expedia) and 7.77.7--9.1×9.1\times faster at G=128G{=}128 on the survival datasets; on Arm, 2.52.5--2.9×2.9\times and 5.55.5--6.0×6.0\times. On a scenario-analysis benchmark it is faster in all 1616 configurations on both machines. For f64 models, outputs match treelite GTIL up to tree-ordering roundoff; for f32 models, TreeWalker's f64 leaf accumulation is closer than native f32 to an exactly rounded reference on every audited row.
Oct 2, 2026

Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation

Flow Matching enables high-quality visual generation via continuous-time dynamics, but inference remains costly due to multiple sequential function evaluations. Existing acceleration methods reduce the number of function evaluations but often introduce additional training overhead, degrade quality, or fail to account for input-dependent variability. We propose COFLOW, an inference-time method that adaptively selects the step counts each generation based on the prompt features. Our context-aware COFLOW is trained online with an unsupervised reward that balances inference efficiency and generation fidelity. Our method is plug-and-play, requiring no retraining of the underlying generative model. It generalizes to image and video generation, achieving over 2.5x speedup while preserving perceptual and semantic quality. We further provide a theoretical analysis establishing an O(1/K) forward-Euler discretization error bound under standard regularity conditions.
Oct 1, 2026

STEER: Reducing Inference Cost in Relational Foundation Models through Semantically Informed Sampling

Relational foundation models (RFMs) are pretrained once on a collection of relational databases and prediction tasks, and then applied zero-shot to previously unseen databases and tasks. To make a prediction for a target row, an RFM samples a neighborhood of rows linked to that row through foreign keys and uses this neighborhood as its inference context. Lowering inference cost is an important goal for any foundation model, and for RFMs this cost grows with the size of the context. The simplest ways to shrink the context is to drop some of the sampled rows, but this ignores the semantics of the database schema, so it is as likely to discard informative rows as uninformative ones. We propose STEER, a sampling approach that shrinks the inference context by concentrating it on the tables most relevant to the prediction task at hand. STEER obtains relevance information by prompting a large language model to rank the foreign-key edges of the database schema into relevance tiers for the given task, and then maps each tier to a probability of following that edge during traversal. Because the ranking uses only the schema, it is computed once per task and reused across all subsequent predictions, amortizing its cost. We evaluate STEER on three state-of-the-art RFMs (RT, RT-J, and Griffin) and show that it reduces inference context size by about 40% on average while maintaining, and in some cases improving, prediction accuracy.
Sep 29, 2026

High-Dimensional Simulation-Based Inference in Latent Spaces

Neural simulation-based inference (SBI) has been widely successful in inferring a relatively small number of interpretable parameters from potentially high-dimensional observations, such as images or time series. Accordingly, representation learning in SBI has focused almost exclusively on compressing the observations used to condition the posterior. More recently, however, SBI has begun to target increasingly high-dimensional parameter spaces, raising the complementary question of whether the inference target itself should be compressed. Our answer is a practical merger of SBI and latent generative modeling, which learns a low-dimensional representation of the simulator parameters, performs posterior inference directly in this latent space, and maps posterior samples back to the original parameter space. We characterize the conditions under which latent-space inference recovers the desired target posterior and systematically study its empirical trade-offs. Across four case studies and three generative families, we compare latent and standard estimators while controlling for network capacity, regularization, optimization, and training compute. At matched training compute, latent-space inference achieves accuracy and marginal calibration comparable to direct target-space inference while sampling up to more than an order of magnitude faster.
Sep 28, 2026

Accelerator Choice Is Not Enough: AlphaFold2 Inference on Cloud TPUs

AlphaFold2 is written in JAX, so the same inference code compiles and runs unchanged on CPUs, GPUs and Google Cloud TPUs. That portability makes the accelerator look like the main decision a user has to make. We show that it is not. Running one AlphaFold2 inference workload across a Colab CPU runtime, an NVIDIA T4 GPU and a dedicated eight-chip Cloud TPU v5e slice, we find a large hardware advantage for the TPU, 0.47 s per call in steady state on a single chip against 13.1 s on the T4 in the same measurement campaign, and three ways in which the software layer decides how much of it a user actually gets. The default execution path uses one chip of the eight, and at list prices the idle capacity makes the slice cost about as much per prediction as the GPU. Batching with JAX's vmap never exceeds single-query throughput, while mapping queries across chips with JAX's pmap gives eight chips 6.5-7.9x the throughput of one on a matched grid; automatic sharding leaves the per-chip footprint unchanged, consistent with replication, most plausibly because AlphaFold2 carries no sharding annotations. Our retained trace analysis of a first call at a new input shape reports about three quarters of the traced span in JAX tracing and compilation rather than execution. Reruns five weeks later reproduced neither cloud baseline, the GPU one off by roughly a factor of two, so the hardware ratio above is specific to one campaign.
Sep 28, 2026

AD-E2E-JEPA: A Joint-Embedding Predictive Architecture For End-to-End Autonomous Driving

Autonomous driving requires world models that can understand the physical world, reason and plan, and operate safely. In this paper, we first systematically evaluate existing action-conditioned joint-embedding predictive architecture (JEPA) world models, including LeWM, DINO-WM, and JEPA-WM for end-to-end autonomous driving (E2EAD). To isolate world-model quality from policy learning, we employ a goal-conditioned zero-shot planning setting that evaluates these models using ground-truth future observations as goals, without training any driving policy. We find that existing JEPA-based world models are either accurate for driving but computationally expensive, or computationally efficient but insufficient for planning. To address this trade-off, we propose AD-E2E-JEPA, which introduces a SIGReg-regularized learnable projector applied to projected patch embeddings. The projector reduces the number of planning patches by 16×16\times and the embedding dimension by 4×4\times, achieving a 100×100\times inference speedup while retaining planning performance, with a 0.8-second runtime for an 8-frame rollout over 256 candidate trajectories. Without training any driving policy, the world model itself reaches the goals located 20 meters away on average within the displacement of respectively 4.0/2.8 meters, using world-model rollouts over trajectory vocabularies of respectively 256/8,192 candidates. On the NAVSIMv2 benchmark, it achieves 67.3/72.9 EPDMS with multiplicative safety metrics and 84.1/86.5 EPDMS†^{\dagger} without them in goal-conditioned zero-shot planning. Experiments further show that the self-supervised pretrained projector improves downstream imitation learning performance from 80.2 to 85.4 EPDMS. The source code is available at https://github.com/HaoranZhuExplorer/AD-E2E-JEPA
Sep 27, 2026

Can Tabular Foundation Models Amortize Statistical Inference?

For decades, statistical inference has largely been developed one problem at a time. Given a scientific target, such as a treatment effect or a regression function, statisticians design a problem-specific estimator together with a procedure for quantifying its uncertainty. This paper proposes a different paradigm. We focus on a classical problem in statistical inference, confidence interval construction, and develop TabCon, an amortized inference system built on a tabular foundation model that produces confidence intervals for new datasets through a simple forward pass. The key methodological ingredients of TabCon are a sparse mixture-of-experts architecture and reinforcement-learning-based post-training that calibrate the resulting confidence intervals to a desired coverage level. Across a wide range of benchmark datasets, TabCon attains near-nominal coverage while producing short confidence intervals. At inference time, it also offers considerably greater computational efficiency, running 50 times faster than the classical bootstrap procedure, even when the latter uses only 50 bootstrap samples.
Sep 26, 2026

Superposed Inference for Hyperdimensional Computing

Hyperdimensional computing (HDC) is attractive for efficient and robust learning, but conventional inference still encodes every query independently, repeatedly paying the cost of high-dimensional projection. We introduce SupHDC, a new inference paradigm that processes multiple queries through a shared encoding computation. SupHDC assigns lightweight random slot keys, superposes the keyed queries before encoding, and uses slot-specific classifiers to recover their individual predictions. A random-feature kernel view explains why exact recovery of each hypervector is unnecessary: inference only needs to preserve the class evidence that determines the prediction. Across ten datasets, SupHDC achieves 1.39x analytical speedup with no average accuracy loss, and up to 2.08x speedup with only a 2.67 percentage-point mean accuracy loss. On a Raspberry Pi~5, it delivers 2.01x measured wall-clock speedup with a 2.26 percentage-point loss in mean prediction accuracy. SupHDC shows that high-dimensional redundancy can be used not only for robustness, but also as capacity for shared inference.
Sep 24, 2026

Rolling-WAM: World Action Models with Rolling Imagination

World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action updates and limiting closed-loop responsiveness. We present Rolling-WAM, a formulation that distributes joint denoising across successive replanning cycles. Our method maintains a sliding window of video-action chunks at staggered noise levels. At each step, a rolling noise schedule fully denoises the imminent action chunk for execution, while partially refining farther-future chunks. As the window advances with new camera observations, the retained future chunks continue their denoising process. This distributes the computational cost over time while carrying an evolving visual-action context across chunk boundaries. Evaluations on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid show that Rolling-WAM achieves competitive manipulation performance. By removing the need to denoise the entire prediction horizon from scratch, it delivers a 4.5x steady-state replanning speedup over standard joint WAMs.
Sep 24, 2026

Direct Message Approximation (DMA): A Consistency-Based Framework for Tractable Approximate Inference on Factor Graphs

Approximate message passing on factor graphs underlies two dominant families of probabilistic inference algorithms: expectation propagation (EP) and variational message passing (VMP). Both methods approximate the marginal at each factor edge, forcing an iterative round-robin schedule, risking negative-precision messages, and, for VMP, collapsing to point estimates at Dirac-delta factors. We introduce Direct Message Approximation (DMA), which approximates factor-to-variable messages directly rather than the marginal. For normalisable factors, we define a consistency condition (requiring exactness when all other incoming messages are Dirac deltas) to guide message construction. We prove a master theorem (proper messages, any graph) bounding marginal KL from message KL, with three structural corollaries: Dirac-input consistency, no EP-style inner-loop iteration, and no negative-precision messages. Further, we prove a complementary O(1/r2)O(1/r^2) guarantee for the inherently improper backward message of the product factor, whose closed-form treatment has resisted prior work. As a concrete instantiation, we derive explicit DMA messages for the product and leaky-ReLU factors and assemble a Bayesian neural network (BNN) inference algorithm with one forward/backward sweep per training example and no gradient learning-rate hyperparameter, validating that the structural guarantees translate to predictive uncertainty that widens in data-sparse regions, including under model mismatch.
Sep 22, 2026

Visual Jev: Accurate and Efficient Decisions from Shared Visual Context

Many vision applications ask several independent, forced-choice questions about the same image. Visual Jev encodes the image and public context once, executes isolated question suffixes as a batch, and reads candidate probabilities from the backbone's language-model head. Across four benchmarks, answer-supervised post-training raises equal-weight macro accuracy from 70.6% to 76.1%, with the gain concentrated on the two task families represented in training. At N=32 questions per image, shared batched execution is 8.9x faster in warm amortized time than independent serial execution and remains 3.4x faster than an already-batched baseline that recomputes the prefix, at the cost of higher peak memory. A matched typed-head control offers no consistent accuracy advantage over the language-model-head readout. The supported design is therefore simple: adapt the backbone for quality, retain the existing readout, and share execution for efficiency.