Adaptive Computation

Follow this topic Ask about this topic

Momentum

14 papers in the last four weeks, up 180% on the four weeks before. 0.1% of all new papers.

Jul 13Week of Sep 28

Latest papers 59

Oct 8, 2026cs.CL

Easy to anticipate, hard to compute: boundary dependence finds the computed outputs that entropy patching misses

Byte-level language models such as the Byte Latent Transformer (BLT) group bytes into patches and run their large global model once per patch. BLT starts a patch where a small model's next-byte entropy is high, so global compute goes where the next byte is hard to predict. We show that this rule has a systematic blind spot: positions whose type is predictable but whose value must be computed, such as the number after "=" in a worked math solution. Under tight patch budgets, entropy-triggered layouts skip these positions, and accuracy on them collapses. In Meta's BLT-1B with patch starts on 10% of bytes, the entropy rule puts a patch start at 16% of the computed results in GSM8K solutions and gets 19.0% of them exactly right; a boundary after each "=" at the same patch count gets 51.8%, and entropy combined with a label-free boundary-dependence signal gets 67.1% (default layout at 26% of bytes: 76.8%). The gap survives adapting BLT-1B to the budget with low-rank fine-tuning (32.9% vs 72.7%, three runs per rule, paired p < 1e-200) and grows with model size in byte models trained from scratch at a 10% budget: at 1M, 12M and 50M parameters, boundary dependence beats entropy on final answers by -1.6, +10.1 and +19.8 points, and at 50M it gets 35.9% of computed results against 13.9% (3 seeds each). BLT's entropy-jump rule helps neither target at 50M. The entropy trigger of Scratchpad Patching is likewise indistinguishable from random scratchpads on final answers (5.6% vs 5.9%, 5 seeds), while answer-start scratchpads give 38.1%. The effect is specific to computed values: copies and lookups gain little, and values the model cannot compute gain nothing. Boundary dependence, the rise in the model's own loss when a patch start is removed, measured per two-byte context, finds these positions without labels: combined with entropy it beats the hand-written rule on computed results.

Nicolás Vera Zúñiga

Independent Researcher, Chile

Byte-Level Language Model Efficient Language Model Inference Adaptive Computation

Oct 7, 2026cs.LG

Input-Blind Controls Produce Substantial Oracle Headroom for Layer Programs in Multiple-Choice Evaluation

Adaptive computation aims to improve language-model inference by tailoring execution to each input. For layer programs, oracle evaluations use known answers to estimate the potential gain from this flexibility, before a practical selector is available. However, a gain from selection does not by itself explain why the chosen programs help. This study examines this distinction using 32 layer-skipping and repetition programs on two models and 4,413 multiple-choice items. The analysis compares their gains over a fixed action selected without the evaluation prompt with those of input-blind perturbations at the same sites, re-evaluating selections on another prompt. With shared option order, the controls give 10.2-11.8 and 15.6-19.4 percentage points of headroom on Qwen3-4B-Base and Llama-3.1-8B, exceeding the real programs' 9.0 and 10.1 in all three random-direction draws per model. They match answer-change rate only, and the ordering depends on the menu: in post hoc comparisons, real programs lead on Llama's repeat-only menu in every draw. A smaller KL-calibrated comparison, including an input-dependent control, favours real programs in point estimate, with inconclusive corrected tests. Fixed letter offsets produce headroom of similar scale. Rotating options sharply reduces both families' headroom, while leaving positive real-minus-control differences of 1.4-2.3 and 3.7-4.5 points; their magnitudes and statistical support depend on further adjustments and the reference. A supplementary generated-answer test finds that search-selected programs keep a 26.0-point advantage over programs selected for other problems after rewording, without a placebo comparison. These results show that substantial headroom can persist across prompts with shared option order without establishing a benefit specific to the selected layer computation; neither ordering against these controls identifies that benefit.

Yibei Guo, Rui Liu

College of Aeronautics and Engineering, Kent State University

Adaptive Computation LLM Evaluation

Oct 6, 2026cs.AI

Enabling Dynamic Computation in Looped LMs

Looped LMs are parameter efficient and promise dynamic computation (saving memory and FLOPs on easy tokens). However, state-of-the-art open Looped LMs trained with this dynamic computation capability (Ouro models) do not realize it in practice as each loop iteration (depth) requires its own level of KV-cache, necessitating all loop computations. Moreover, Ouro's early-exit prior is enforced on each token equally, which results in static lower-depth like processing of all tokens regardless of difficulty. In this work, we propose a simple "best-available" KV caching strategy that works out-of-the-box, creating a new frontier in the performance vs depth space. Our approach enables up to 30% reduction in FLOPs and KV memory while retaining full-depth performance, showing the true flexibility of Looped LMs. Furthermore, training looped LMs with awareness about this KV caching strategy improves performance and efficiency. Finally, we apply a small but effective fix to the early-exit prior enforcement objective that makes tokens exit at truly heterogeneous depths based on effort. Our findings are validated on Ouro models as well as smaller looped LMs pre-trained from scratch.

Aayush Mishra, Arnau Padrés Masdemont, Victor Conchello Vendrell +3

Qualcomm AI Research · Johns Hopkins University

LLM Inference Acceleration Adaptive Computation KV-Cache Management Recursive Language Models

Oct 1, 2026cs.LG

BranchIP: Learning Adaptive Equivariant Computation for Interatomic Potentials

Equivariant machine learning interatomic potentials (MLIPs) have revolutionized atomistic modeling, but accurate treatment of complex materials and molecular systems demands expensive models. This limits simulation length- and time-scales, with tensor products a key computational bottleneck. The recent emergence of foundation-scale MLIPs further exacerbates this challenge. We present Branch Interatomic Potential (BranchIP), a single-model framework for learned adaptive tensor product computation, trained with a novel distillation loss. In our experiments on two systems of physical interest, a heterogeneous catalysis system and a proton-conducting solid acid electrolyte, BranchIP accelerates MLIPs across model sizes by up to

2.4\times

while reducing memory usage by up to

2.6\times

. This is achieved while maintaining physical fidelity. Furthermore, the learned adaptive computation provides model interpretability by revealing which interactions demand deeper computation and showing how computational depth relates to chemical complexity and dynamics.

Laura Zichi, Gil Harari, Chuin Wei Tan +6

Harvard University · Robert Bosch LLC Research and Technology Center

Machine Learning Interatomic Potentials Equivariant Neural Networks Adaptive Computation Computational Materials Science

Sep 30, 2026cs.LG

A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models

Looped models reason by applying the same block of weights many times, so compressing that block saves memory traffic on every loop. Compressed looped models, however, often collapse, and the collapse is usually blamed on rounding error that accumulates from loop to loop. In this work we test that account on more than 30 models from five families and find, to our surprise, that it holds only for loops that never settle. When a loop settles, a fixed rounding error does not accumulate. It moves the point where the loop settles, much as tilting a bowl moves where a ball comes to rest, and the answer is lost only when the shift is larger than the readout tolerates. This picture lets us predict which models fail from a single label-free measurement, and it tells us why failed models recover: their loops still settle, so a few final loops with 8-bit weights bring the answer back. Motivated by these findings, we build a controller that stops when the model's halting head fires and then finishes with 8-bit loops. On Sudoku-Extreme and Maze-Hard it beats fixed-depth inference by up to 15 points under a third of the weight traffic.

Steven Kolawole, Pearse Jim, Opegbemi M. Busoye +2

Carnegie Mellon University · ML Collective

Efficient Neural Network Inference Recurrent Neural Networks Neural Network Quantization Neural Network Compression Adaptive Computation

Sep 30, 2026cs.RO

DeepJEPA: Scaling World Models from Within

World-model planners typically scale outward by rolling farther, sampling more trajectories, or optimizing longer, while assigning the same computation to every imagined transition. We show that making every transition uniformly deeper wastes computation and can degrade planning because useful refinement is concentrated at a small set of decision-critical events. We introduce DeepJEPA, a weight-tied joint-embedding predictive world model that treats transition depth as an inner test-time scaling axis and learns when another recurrent update is worth computing for each candidate and rollout step. Across five visual-control settings, DeepJEPA improves or matches the strongest fixed-depth planner while averaging only 1.00-1.26 updates per transition. Its additional computation concentrates at contact onset and sustained object interaction, where latent corrections can change which candidates enter the planner's elite set and which action is selected. Representation probes further show that improved planning does not require uniformly better object-state decodability. DeepJEPA therefore reframes world-model scaling as a problem of allocating internal computation where it can change the planner's decision: think deeper at decision-critical transitions instead of making every rollout uniformly deeper or longer.

Zijian Jin, Yunbei Zhang, Yuanzhe Liu +5

New York University · Tulane University · UIUC +4

World Model Learning World Model-Based Planning Joint-Embedding Predictive Architecture Test-Time Scaling Model-Based Planning Adaptive Computation

Sep 30, 2026cs.CV

AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks

Embodied tasks demand accurate, flexible, and semantically rich 3D scene representations. 3D semantic occupancy is well suited to this requirement, as it can model holistic 3D spaces by encoding geometric occupancy along with semantic categories. However, existing occupancy prediction methods struggle to meet practical deployment requirements, such as adapting to varying computing budgets, sensor setups, and observation views. In this paper, we propose a point-based Adaptive 3D Occupancy Prediction method, called AdaOcc, tailored for embodied scenarios. To accommodate heterogeneous sensor inputs, AdaOcc uses an adaptive geometry-guided dual-branch encoder that can support RGB images in various numbers of views with (estimated) depth maps or LiDAR scans. AdaOcc represents occupied regions via sparse semantic points trained with a progressive query learning strategy, allowing the prediction computational budget to be flexibly adjusted through query point numbers and decoder layers. To facilitate high-fidelity geometric modeling for lightweight point-based occupancy learning, we further propose a novel containment loss that regularizes predicted points to reside within valid occupied regions. Extensive experiments show that our method achieves a new state-of-the-art on Occ-ScanNet with considerable performance improvements over previous methods. Moreover, our framework demonstrates strong practical applicability as an adaptive 3D perception module in real-world embodied systems.

Jinglong Wang, Yunjie Wang, Zhiyang Zhang +4

Beihang University · Beijing Academy of Artificial Intelligence · XYZ Embodied AI +3

3D Occupancy Prediction Adaptive Computation

Sep 28, 2026cs.LG

LoopICL: Looping a single transformer block to solve tabular tasks

Tabular foundation models using in-context learning have recently surpassed gradient-boosted trees on predictive tabular tasks. However, recent mechanistic insights suggest that parameters in these models are largely redundant. We introduce LoopICL, a looped transformer whose core design decouples parameter count from computational depth. LoopICL consists of a single block, processing data through two coupled streams: a cell stream capturing per-cell feature representations and a row stream capturing in-context example representations, jointly refined through within-column and cross-column attention. During pre-training, we vary loop counts, allowing the block to be unrolled for a varying number of iterations at test-time and use a learned exit-gate to automatically exit. In its standard setting, LoopICL performs competitively with TabICLv2 on TabArena and TALENT at the same computational cost (FLOPs), while using nearly 90% fewer parameters. Furthermore, its recurrent design enables users to also trade off inference cost and performance, providing a resource-aware TFM.

Amir Rezaei Balef, Katharina Eggensperger

TU Dortmund University, Dortmund, Germany · Lamarr Institute for Machine Learning and Artificial Intelligence, Dortmund, Germany · University of Tübingen, Tübingen, Germany

Tabular Foundation Models Cost-Aware Inference Recurrent Transformers Tabular ICL Adaptive Computation

Sep 28, 2026cs.CL

Improving Test-Time Scaling with Adaptive Looped Transformers

Looped transformers have demonstrated promising parameter efficiency by reusing layers for latent computation. Prior studies compare looped and non-looped models at matched parameters or per-token FLOPs. However, to the best of our knowledge, whether looping improves test-time scaling as outputs grow longer remains underexplored. Through post-training looped transformers, we study the accuracy-compute slope, measured as the accuracy gain per doubling of test-time decoding FLOPs. We find that existing looped transformers often yield steeper slopes than their non-looped baseline, yet underperform it at matched compute. While fixed-depth looping spends extra iterations on every token, our analysis shows that many tokens do not benefit from extra iterations. We therefore propose TaH2, which enables the model to focus extra iterations on the tokens that benefit from looping. It jointly post-trains the backbone and an iteration decider through lookahead depth supervision, which uses online labels indicating whether further iteration improves the prediction. TaH2 improves both the efficiency and attainable accuracy of test-time scaling. On challenging AIME benchmarks, TaH2 improves the accuracy-compute slope by 53% (2.74 vs. 1.79) over the non-looped baseline, exceeding the baseline's peak accuracy by about 3.4 points at matched test-time compute. As the maximum iteration depth increases, existing looped models largely plateau, while TaH2's gain over the non-looped baseline continues to grow from +2.8 points at depth 2 to +3.9 points at depth 8. Our code is available at https://github.com/thu-nics/TaH.

Yichen You, Tianyu Fu, Aosong Feng +4

Tsinghua University · Yale University

Test-Time Scaling Recurrent Transformers LLM Inference Acceleration Adaptive Computation

Sep 28, 2026cs.AI

GenoMorph: Pathway-Grounded Genomic Disease Reasoning via Adaptive Latent Computation

Large language models (LLMs) have demonstrated strong capabilities in biological reasoning; however, genomic disease inference remains largely dependent on memorized gene-disease associations rather than understanding biological pathways. This shortcut learning undermines robustness and generalization, and breaks down when molecular identifiers are unavailable. We present GenoMorph, a multimodal genomic reasoning framework that shifts disease prediction from associative gene-disease mapping toward pathway-grounded reasoning. GenoMorph couples a frozen DNA foundation model with question-conditioned cross-attention fusion, self-adaptive latent reasoning (LatentSp), a residual reasoning gate for iterative genomic evidence reinjection, and rejection sampling fine-tuning regularized by hierarchical optimal transport (OT). Rather than learning direct gene-disease mappings, GenoMorph aligns genomic sequence representations with latent pathway dynamics, enabling reasoning trajectories that follow molecular interactions before producing disease predictions. LatentSp dynamically allocates computation according to reasoning confidence, reducing unnecessary reasoning steps and improving inference efficiency. We further construct an anonymized benchmark from the Kyoto Encyclopedia of Genes and Genomes (KEGG), replacing every gene and molecular identifier with anonymous symbols while preserving sequences and pathway topology, thereby removing memorization shortcuts. GenoMorph raises the weighted F1 from 0.7863 (BioReason) to 0.9412, and rejection sampling fine-tuning with self-adaptive latent reasoning pushes it to 0.9725 while cutting latency nearly 60%. On the anonymized benchmark it reaches 0.9465 F1, substantially outperforming prior systems and confirming that accurate disease prediction can arise from pathway reasoning rather than memorized gene-disease associations.

Tanmoy Kanti Halder, Akash Ghosh, Arijit Roy +1

Department of Computer Science and Engineering, Indian Institute of Technology Patna, Bihta, 801106, Bihar, India · Prasannadeb Women’s College, India

LLM Reasoning with Graphs Genomics Adaptive Computation

Sep 24, 2026cs.AI

CounterRoute: Self-Routed Reasoning via Hierarchical Counterfactual Credit Assignment

Reasoning-capable language models often produce long chains of thought when direct answers suffice, wasting inference compute. Many dual-mode models leave this choice to users. Automating it is challenging because routing targets evolve with the policy, initial mode preferences destabilize exploration, and sequence-level objectives entangle routing with response learning. We introduce CounterRoute, an online reinforcement-learning framework that jointly learns routing and modeconditioned responses in one shared policy directly from a native dual-mode checkpoint, without method-specific SFT warm-up. Paired current-policy counterfactual rollouts assign cross-mode credit only to the routing token, while within-mode GRPO trains response tokens. A paired-to-self-routed curriculum stabilizes early training with forced rollouts from both modes, then increases self-routed updates to improve autonomous routing. Across nine benchmarks, CounterRoute better balances accuracy and efficiency than heuristic and learned adaptive-routing methods. Relative to always-thinking checkpoints, it improves macro-average accuracy while reducing mean generated tokens by 51% for Qwen3-8B and 41% for Qwen3-14B. On instruction-following and commonsense benchmarks where direct answering is strong, think rates fall as low as 1% while response quality improves. Despite training only on math and instruction following, its routing behavior and response quality generalize to held-out coding, science, knowledge, and commonsense benchmarks.

Ruochen Jiao, Besnik Fetahu, Zhenyu Shi +1

RL for Language Model Reasoning Efficient Language Model Reasoning LLM Routing Credit Assignment Efficient Language Model Inference Counterfactual Credit Assignment+1

Sep 23, 2026cs.CV

Learning Spectral Allocation: A Fractional Diffusion Framework for Adaptive Volumetric Segmentation

We address adaptive computation in 3D medical image segmentation: instead of designing another backbone, we ask how much spectral mixing each network stage needs and let optimization answer. We derive FHEAT, a two-parameter operator family, from the discrete cosine transform (DCT) solution of a fractional heat equation. A fractional order alpha and a diffusion strength D govern the operator, and at D=0 it is exactly the identity. Reparametrized by the semigroup time tau = D*alpha, same-resolution instances compose exactly, so any distribution of diffusion across same-resolution stages amounts to a single Sobolev-type regularizer of learned strength. This identity limit lets the optimizer of each layer, not the designer, decide whether global mixing is needed and how sharp it should be. We instantiate FHEAT in a lightweight U-shaped architecture (Light-UNETR) paired with a Kolmogorov-Arnold mixer (KAN3D) with adaptive rational activations, yielding FHEAT-Seg. At 5% to 20% label rates on three public benchmarks, training produces gradient-driven spectral sparsification: seven of the eight stage-level operators drive D to zero, and the survivor saturates at the sharpest low-pass (alpha ~ 0.9) in the decoder layer feeding the semi-supervised attention map. The retired layers become exact identity shortcuts at inference, cutting FLOPs from 4.29G to 0.90G (a 79% drop) at 0.975M parameters. Under a standard semi-supervised protocol, FHEAT-Seg reaches Dice scores of 90.47% (left atrium), 78.79% (Pancreas-CT), and 81.90% (BraTS 2019), ahead of five semi-supervised methods and the Light-UNETR baseline. The large variant also surpasses Light-UNETR-L under full supervision (Dice 93.09%, 85.11%, and 87.19%) with 2.851M parameters and 55.75G FLOPs. These results suggest that the allocation of spectral computation is a learnable property of optimization dynamics, not a manual design commitment.

Yi-Hui Shen, Tie-Qiang Li

School of Arts and Sciences, Fujian Medical University, Fuzhou, China · School of Medical Imaging, Fujian Medical University, Fuzhou, China · Department of Medical Radiation Physics and Nuclear Medicine, Karolinska University Hospital, Stockholm, Sweden +2

Semi-Supervised Medical Image Segmentation 3D Medical Image Segmentation Adaptive Computation

Sep 20, 2026cs.CV

VGGT-Prime: Compute-Adaptive Mixture-of-Heads for Efficient Visual Geometry Transformers

Feed-forward visual geometry models such as the Visual Geometry Grounded Transformer (VGGT) have recently enabled direct 3D reconstruction from multi-view images. Despite their promising performance, these models scale quadratically with the number of input views due to their global attention mechanism, resulting in substantial latency for long sequence inputs. There have been some recent efforts to accelerate VGGT, but they primarily focus on reducing \emph{token redundancy} through token merging or key/value sparsification. Our work resolves this bottleneck from a different perspective by investigating \emph{architectural redundancy} in visual geometry transformers. We show that the multi-head attention modules in VGGT's global-attention layers contain substantial architectural redundancy, with only a subset of heads carrying critical geometric information. In light of this observation, we propose VGGT-Prime, a compute-adaptive mixture-of-heads model that resolves this redundancy to accelerate visual geometry transformers while maintaining competitive reconstruction quality. The key idea of VGGT-Prime is to estimate the appropriate computation level for each global-attention head using a lightweight router and then dynamically assign each head to different computation modes. Extensive experiments on multiple datasets demonstrate that VGGT-Prime can achieve an {

8\times

} inference speedup over VGGT while maintaining competitive performance on camera pose, depth, and point-cloud predictions. We further show that VGGT-Prime is complementary to existing acceleration methods, such as token merging, further improving inference speed by up to

14{\times}

over VGGT. An overview of our work is available on our project page.

Abteen Arab, Guile Wu, Chengjie Huang +1

Huawei Noah’s Ark Lab · University of British Columbia

3D ViTs Multi-View 3D Reconstruction Efficient ViTs Feed-Forward 3D Reconstruction Adaptive Computation

Sep 17, 2026cs.AI

When2Think: Learning When and How Much to Reason

Large Reasoning Models (LRMs) often overthink easy problems and underthink hard ones, leading to inefficient computation allocation. Existing methods regulate generated computation or select between direct answering and explicit reasoning, but do not jointly control whether}to reason and how much computation to allocate within reasoning. We call the resulting difficulty-dependent loss in accuracy under computation reduction the efficiency tax. We propose When2Think, an RLVR-based post-training framework for instance-adaptive computation allocation. Its core mechanism, Instance-level Difficulty-Aware Control (IDAC), uses cached reference statistics of success and token cost to modulate a correctness-gated efficiency bonus based on generated token count. Importance sampling supports exploration of Think and NoThink, while Batch-Wise Standardization constructs standardized advantages for critic-free optimization. The framework requires neither a learned reward model nor a learned critic, and offline reference caching avoids online reference-model queries during policy updates. On AIME24, When2Think improves Pass@3 by 10.0 percentage points while reducing token usage by 27.9% relative to the backbone.

Jaejun Shim, HyunJin Kim, Young Jin Kim +1

Sungkyunkwan University · Microsoft

Overthinking in Language Models RL for Language Model Reasoning Efficient Language Model Reasoning Efficient Language Model Inference Adaptive Computation

Sep 14, 2026cs.AI

T-LoopFormer: Token-Level Elastic-Depth Looped Transformers for Latent Reasoning With Dynamic Routing

Looped Transformers have recently demonstrated strong performance in both reasoning and language tasks by reusing a shared set of parameters across multiple iterations, achieving parameter efficiency without sacrificing representational power. Besides, looped Transformers perform inference directly in the latent space (latent reasoning) to reduce the number of tokens consumed during inference, thereby achieving improved sample efficiency. However, these models typically apply a fixed recursion depth uniformly to every token, leading to suboptimal compute allocation and leaving significant efficiency gains on the table. In this work, we propose \textbf{dynamic token-choice routing} for looped transformers, enabling each token to adaptively determine its own number of loop iterations based on its hidden state. We use a dynamic router to decide whether a token should continue recursing or exit early, allowing simple tokens to bypass unnecessary computation while hard tokens receive deeper processing. To ensure that this adaptive mechanism does not compromise decoding efficiency, we further introduce recursion-wise KV caching, which maintains an independent key-value cache for each recursion loop. This design ensures that tokens at different depths only attend to their corresponding cached states, effectively eliminating redundant computations for exited tokens and enabling fast autoregressive decoding. Extensive experiments show that T-LoopFormer reaches the sota performance under the same parameters on PPL and 10 zero-shot reasoning tasks, even surpassing the base model at 24x FLOPs and our model could reach the lowest inference latency, which validate the effectiveness of token-choice router and recursion-wise KV cache. Code: https://github.com/YuMingQian1234/T-LoopFormer.

Mingqian Yu, Wenpeng Zhang, Peilin Zhao

Institute of Automation Chinese Academy of Sciences Beijing, China · Independent Reasearcher · School of Artificial Intelligence Shanghai Jiao Tong University Shanghai, China

KV Caching Recurrent Transformers LLM Inference Acceleration Adaptive Computation Latent Reasoning

Sep 14, 2026cs.AR

Adaptive AI: Energy Efficient Multi-exit TinyML on Intelligent Vision Systems at the Edge

Traditional TinyML systems for edge devices achieve high accuracy by relying on fixed-depth models that require a constant number of multiply-accumulate (MAC) operations regardless of the input complexity. This approach wastes critical resources in battery-powered Internet-of-Things (IoT) devices and limits the real-time performance of edge cyber-physical systems. Multi-exit execution schemes mitigate these issues and are widely used on high-end devices such as GPUs, but are rarely exploited on edge IoT devices because they require substantial rethinking given their strict memory and computational constraints. We address these aspects by designing and deploying, on an ultra-low-power GWT GAP9 System-on-Chip (SoC), a novel multi-exit computational scheme, demonstrating it on a MobileNetV2 convolutional neural network (CNN) for the ImageNet-100 classification task. Our approach introduces multiple exits at different CNN depths, each with a confidence-based gating mechanism that dynamically and autonomously decides whether to continue or stop inference. Comparing our multi-exit strategy to the standard MobileNetV2 on a GAP9 SoC, we show a 41% reduction in the average computational cost (from 313 MMAC to 185 MMAC), a 29% lower inference time (from 49 to 35 ms), and an energy saving of 24% (from 2.1 to 1.6 mJ per frame). All these improvements come with a ~1% loss in accuracy compared to the full-depth MobileNetV2, which achieves 80.5%. Finally, comparing our adaptable multi-exit scheme with a third-party state-of-the-art adaptive CNN, also deployed on the GAP9, we achieve more than 2x its computational efficiency, increasing it from 8.1 to 17.2 MAC/cycle.

Luca Crupi, Lorenzo Lamberti, Alessandro Giusti +1

IDSIA, SUPSI, Lugano, Switzerland · IIS, ETH Zürich, Switzerland

Efficient Neural Network Inference Energy-Efficient ML Edge Inference Tiny ML Adaptive Computation Early-Exit Neural Networks

Sep 8, 2026cs.LG

MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models

Universal multimodal embedding (UME) increasingly demands encoder's capacity for handling a broad range of tasks and modalities with increased complexity. Prior scaling methods either increase the representation size, retrieval effort, or scales the encoder into a heavy multimodal LLM. Recent works, such as Think-Then-Embed (TTE), explore scaling via reasoning tokens. However, embedding models are hard to scale up: increasing parameters directly tradeoffs for the large training batch size that contrastive learning needs, and retrieval has to be served under tight latency. Moreover, UME tasks are diverse in complexity, where scaling up embedders can bring significant redundant computation. In this work, we propose MOEMB, which instead scales UME along the expert axis through mixture-of-experts (MoE), growing encoder capacity while preserving single-vector, non-autoregressive encoding. Through a systematic study of the design space and training recipes for MoE-based UME, MoEMB sets a new state of the art on both MMEB-V2 and MRMR among models trained on public MMEB-family data: with only 3B active parameters, MoEMB surpasses TTE-based methods with >4x active parameters, using significantly less computes. To further improve the scalability and efficiency, we conduct the first comprehensive study of adaptive computation for MoE-based embedding, spanning diverse strategies across training-based and inference-only methods. Together, these results support expert scaling as an effective and efficient direction for UME, with adaptive computation further improving efficiency for MLLM-based embedding models towards large-scale retrieval and recommendation systems.

Xuanming Cui, Shlok Kumar Mishra, Wentao Bao +6

University of Central Florida · AI at Meta

Efficient Multimodal Inference Multimodal Embedding Mixture-of-Experts Models Adaptive Computation

Sep 3, 2026cs.CV

Efficient Semantic Understanding from Digital Foveation

Dense semantic segmentation allocates computational resources uniformly across the entire image, regardless of scene complexity or task relevance. Inspired by biological vision, we investigate whether semantic understanding can be achieved more efficiently through digital foveated perception. We introduce a lightweight active-vision pipeline that combines saliency-driven fixation selection, high-resolution foveal observations, low-resolution contextual information, semantic accumulation, and adaptive computation. Beyond conventional dense prediction metrics, we use object-level evaluation to measure semantic understanding under sparse observations. On ADE20K-Object, a single foveated observation achieves 95.9% of the baseline Top-1 accuracy and 96.9% of the baseline Top-3 accuracy while requiring only 4.7% of the computational cost. At the scene level, semantic accumulation recovers 90.6% of the baseline object recall while using 58.6% of the computation. These results suggest that substantial semantic understanding can emerge from sparse observations when computation is allocated selectively, highlighting active vision as an efficient alternative to uniform dense processing and motivating evaluation protocols beyond conventional pixel-wise segmentation metrics.

Caterina Caccavella, Vittorio Fra, Andreas Ziegler +2

Zurich University of Applied Sciences (ZHAW), Wädenswil, Switzerland · ETH Zürich, Zürich, Switzerland · Politecnico di Torino, Turin, Italy +2

Active Visual Perception Semantic Segmentation Adaptive Computation

Sep 3, 2026cs.LG

RecurTrace: Adaptive Latent Reasoning with Loop-Time Memory

Repeating a small block of middle layers increases a language model's effective inference depth without adding parameters or generating extra tokens, and recent work shows that this latent recurrence improves reasoning. However, two design choices limit these gains. Each iteration sees only the previous output and cannot directly access earlier computations. Moreover, a fixed loop count wastes depth on easy inputs while leaving hard ones with too little computation. We introduce RecurTrace, which addresses both limitations using the loop's own trajectory. Specifically, Loop Memory Attention lets each looped layer attend to its own states from previous iterations along the loop-time axis, so the model can revisit earlier computations instead of relying on the latest state alone. A halting head then reads the loop state and predicts whether to continue, with supervision from an oracle that identifies when additional depth still reduces loss. In a controlled MathQA comparison on the same looped backbone, RecurTrace achieves 56.9% accuracy with an average of 2.0 loops, exceeding the best fixed loop depth by 2.2 points at matched compute. By comparison, ACT and PonderNet collapse to one loop, and CALM reaches only 54.1% with 5.6 loops, while the stronger LoopUS-Conf and TaH-Mismatch baselines reach 55.3% at 3.2 loops and 55.7% at 2.1 loops. Finally, RecurTrace improves generation accuracy over same-budget fine-tuned baselines at 0.6B, 1.7B, 4B, and 8B, with the gain growing with model size from 0.6 to 3.4 points.

Yuxiang Wang, Kunyu Feng, Yingda Shen +3

The Chinese University of Hong Kong, Shenzhen · The Chinese University of Hong Kong · Tianjin University +2

Memory-Augmented Language Models Recurrent Transformers LLM Inference Acceleration Adaptive Computation

Sep 1, 2026cs.AI

A Score Is Not a Policy: Measuring the Value of Adaptive Revision

As agentic systems become compound systems, increasingly important decisions move above task execution itself: when should a higher-level controller preserve the strategy guiding another process, and when should it revise it? We study this meta-level control problem in a hierarchical latent reasoner whose manager can retain or replace a commitment governing lower-level computation. Across three precommitted training seeds, learned revision timing produces qualitatively different policies, ranging from an almost deterministic early clock to substantially more state conditioned schedule distributions, yet none outperforms the best forced timing policy evaluated on the same frozen checkpoint. This separates state dependence from decision value: a controller can vary its actions with internal state without turning that variation into a reproducible task-performance benefit. A deeper intervention study on the original checkpoint shows that timing itself is consequential and order-sensitive, while exhaustive enumeration reveals that a strong fixed schedule captures most of the measurable value available from timing at this decision budget. Counterfactual PERSIST/REPLAN diagnostics further show why score-level evidence can be misleading when predictability is dominated by decision position rather than within-position discrimination. Together, these results argue that learned meta-level control should be evaluated along three separate axes: whether its score depends on state, whether that dependence changes realized behavior, and whether those changes capture outcome value beyond a strong non-adaptive policy.

Ayushi Chadha

Hierarchical RL Adaptive Computation

Aug 25, 2026cs.AI

Memory Is Not Always Needed: Characterizing Conditional Memory in Scientific Reasoning

Scientific reasoning requires language models to retrieve specialized knowledge and incorporate it reliably into multi-step computation. Conditional memory provides an explicit lookup pathway that complements dense neural representations, but its usefulness is inherently input- and computation-dependent: retrieved information may repair missing scientific associations, yet it may also introduce distracting shortcuts or interfere with reasoning that the base model can already perform correctly. In this work, we systematically investigate when, where, and to what extent conditional memory should participate in scientific reasoning. We characterize the scientific knowledge boundary and controlled interventions on memory-enabled knowledge-circuit nodes. Based on these analyses, we propose a Knowledge Boundary-Aware Router that uses task-specific input proxies available before generation to determine whether memory is activated, which layer-stage nodes receive memory signals, and how strongly these signals contribute. Experiments on biological and chemical reasoning benchmarks, covering two backbone families and six task types, show that memory effects vary substantially across inputs, tasks, and injection locations. Compared with static and activation-rate-matched random routing, our approach more consistently preserves beneficial memory contributions while suppressing memory-induced regressions, establishing selective memory allocation as an important principle for reliable scientific reasoning.

Zhen Bi, Xueshu Chen, Yan Wang +6

Huzhou Normal University · Alibaba Group · Bota Biosciences +1

Memory-Augmented Language Models Scientific Reasoning Adaptive Computation

Aug 13, 2026cs.LG

Reduced Matrix Multiplication: Input-Adaptive Matrix-Product Reduction for LLM Inference

Transformer-based language models achieve strong performance but incur substantial inference cost due to repeated high-dimensional matrix multiplications. We propose Reduced Matrix Multiplication (RMM), a training-free, input-adaptive inference method that reduces Transformer matrix products by selecting informative slices along their contraction dimensions, without modifying model weights. Under a simple retention-ratio control, RMM provides a smooth and predictable accuracy-efficiency trade-off. Across language models ranging from 1B to 70B parameters, we find that reduction tolerance depends on the model family, task, component, and retention ratio, although it often improves with model scale. Under moderate reduction, RMM remains robust across the evaluated discriminative, autoregressive generation, and long-context settings. We further show that the same principle extends to multimodal vision-language inference. Mechanistic ablations reveal a structural asymmetry within Transformers: attention-side computations are substantially more reducible than MLP components. Finally, wall-clock benchmarks with custom kernels on an NVIDIA A100 show that these computational savings can translate into practical runtime gains, especially at longer sequence lengths. Together, these results position RMM as a scalable direction for input-adaptive inference-time optimization.

Zixuan Lan, Yanhong Li, Jiawei Zhou

University of Chicago · Independent Researcher · Stony Brook University

Inference-Time Optimization LLM Inference Acceleration Adaptive Computation

Jul 27, 2026cs.AI

Exploring Budgeted Image Classification with Content-Sensitive Resource Allocation

The ever-growing adoption of Artificial Intelligence (AI) creates the need to deploy Deep Neural Networks in a variety of computational environments. We consider dynamic environments, where computational requirements are subject to change, and we pose the following question: How do we adjust the complexity of an AI classification system, in order to maximize its accuracy, while meeting changing computational constraints? We call this problem Budgeted Image Classification, and we formally formulate it as a resource allocation integer program. Given a computational budget, a batch of images, and a classification system that can make decisions with varying complexity (it has multiple decision points), we explore strategies to allocate images to decision points, in order to maximize accuracy within the available budget. The original integer program is NP-Hard, so, we propose a continuous relaxation, leading to a content-agnostic allocation strategy which assigns images to decision points without considering their particular content. We address this issue by proposing a content-sensitive strategy, that we experimentally show it leads to superior performance. We theoretically study the behavior of our strategies, deriving conditions that must be satisfied by decision points to be suitable for budgeted classification. We analyze fails cases, offering insights for future research directions.

Athanasios G. Papadopoulos

Tandon School of Engineering, New York University

Cost-Aware Inference Classification Inference-Time Compute Allocation Image Classification Adaptive Computation

Jul 23, 2026cs.LG

Learning to Access Computation: Accessibility Plasticity as a Principle of Adaptive Intelligence

Modern neural networks primarily adapt through parameter modification within predefined computational structures. While recent methods introduce modularity, conditional computation, and parameter-efficient adaptation, they generally do not distinguish computational capability from computational accessibility as separate adaptive variables. This work introduces Accessibility Plasticity, a principle of adaptive computation in which systems adapt not only by changing what computation exists, but also by reorganizing which existing computations can interact and participate. We formalize Accessibility Plasticity through a relationship-based operational realization and establish a reuse-first hierarchy of adaptation, where accessibility modification precedes more costly capability and structural changes. A proof-of-concept evaluation on sequential learning tasks shows that accessibility adaptation can reduce capability modification while maintaining comparable task performance. These results suggest accessibility as a distinct adaptive dimension and provide a foundation for future dynamic neural systems whose computational relationships evolve with changing environments.

Zhaowen Fan

Department of Computer Science, Sichuan University

Adaptive Computation

Jul 22, 2026cs.CV

Look Less, Think Faster: Joint Token-Compute Adaptation for Multimodal LLMs

Multimodal Large Language Models (MLLMs) have recently demonstrated strong performance across vision-language tasks. However, their high inference cost, arising from both the large number of input visual tokens and the heavy computation of the large language model (LLM), remains a key barrier to practical deployment. Recent work attempts to reduce the cost by adaptively optimizing individual dimensions, e.g., pruning redundant visual tokens or skipping LLM layers and heads. Nonetheless, prior approaches typically treat these dimensions independently and overlook a fundamental coupling: the available compute resources must be dynamically allocated across all dimensions based on the input content. To bridge the gap, we propose SmartVL, a unified adaptive inference framework that jointly controls vision token number and model compute capability in response to varying input contents and compute budgets. SmartVL introduces a vision-side token controller that dynamically selects informative visual tokens and an LLM-side compute controller that adaptively adjusts LLM computation. Importantly, these controllers are trained to coordinate with each other so that the overall inference cost satisfies a target budget. To allow this joint scheduling, we connect the controllers using a shared budget encoding and leverage a differentiable latency estimator for end-to-end training. This design enables SmartVL to learn cross-stage allocation strategies that adapt to both input complexity and runtime compute constraints. Experiments across multiple MLLM benchmarks demonstrate that, with joint scheduling, SmartVL consistently outperforms prior adaptive methods and achieves superior accuracy-efficiency Pareto frontiers. Project page: https://www.schaterji.io/publications/2026/jointtokencompute.

Pengcheng Wang, Zhiquan Wang, Jayoung Lee +5

Purdue University, West Lafayette, IN, USA · University of Wisconsin–Madison, Madison, WI, USA · NVIDIA, Santa Clara, CA, USA

Efficient Multimodal Inference Efficient VLM Inference Multimodal Large Language Models Multimodal Foundation Models Visual Token Pruning Adaptive Computation

Jul 22, 2026cs.CL

SLPO: Scaling Latent Reasoning via a Surrogate Policy

Reinforcement learning with verifiable rewards has become the predominant recipe for eliciting test-time scaling in explicit Chain-of-Thought reasoners. Yet this scaling path remains computationally costly, since every intermediate step must be decoded as a language token. Latent reasoning instead carries intermediate computation as continuous vectors and already matches or surpasses explicit CoT at far shorter horizons. Despite this promise, latent reasoners remain largely imitation-bound, while explicit CoT has already moved past imitation via outcome-reward RL. Latent trajectories lack a tractable per-step likelihood and an adaptive stopping interface under fixed thinking budgets, so outcome rewards cannot elicit latent test-time scaling. We introduce Surrogate Latent Policy Optimization (SLPO) to bring outcome-reward RL to autoregressive latent reasoners: an empirical surrogate policy density over latent transitions for trajectory-level credit assignment, and a correctness-supervised stopping head that outcome-reward optimization refines into a variable-horizon policy. Across continuous and soft thinking settings, SLPO improves Pass@

k

under parallel sampling and allocates longer latent computation to harder instances with higher deterministic accuracy.

Runyang You, Zhiyuan Liu, Yongqi Li +1

The Hong Kong Polytechnic University · Sichuan University · Work done as an intern at PolyU.

Reinforcement Learning RL for Language Model Reasoning Reinforcement Learning with Verifiable Rewards Continuous Latent Reasoning Adaptive Computation Latent Reasoning

Jul 19, 2026cs.LG

TAPAS: Throughput-adaptive Perception for Autonomous Systems

Autonomous systems rely on a perception module to navigate through dynamic environments. In real-world scenarios, the perception module's throughput requirements vary at runtime due to changes in scene complexity. However, existing perception strategies assume a fixed FPS and static model-to-cluster mapping, resulting in either over/under provision of throughput requirements or unnecessary energy consumption across diverse scenes. Addressing this challenge requires tightly coupled \textit{scene complexity awareness} to estimate an appropriate FPS target and \textit{dynamic model-to-cluster mapping} to deliver the required throughput at minimum energy. We propose a throughput-adaptive perception strategy for mobile/edge platforms, enabling intelligent runtime resource allocation based on varying FPS targets. We use Reinforcement Learning (RL) with RRM (Reward Reasoning Model) and a GRU (Gated Recurrent Unit) agent to orchestrate perception tasks across heterogeneous mobile/edge platforms. We evaluate TAPAS on Jetson Orin NX across KITTI and unseen nuScenes. On the \textit{KITTI} dataset's test sequences, TAPAS achieves 93-100% throughput met rate while saving energy by 76%. On the unseen \textit{nuScenes} dataset, TAPAS maintains 97% throughput met rate with 64% lower energy compared to \textit{SOTA} approaches, proving its robustness.

Aman Vyas, Vasista Kodumagulla, Zain Taufique +2

Department of Computing, University of Turku, Finland

Energy-Efficient ML Edge Computing Autonomous Driving Perception Adaptive Computation

Jul 15, 2026cs.AI

Per-Token Fixed-Point Convergence in Depth-Recurrent Transformers

A depth-recurrent transformer applies a weight-tied core a variable number of times, and prior work has shown that training with a randomized recursion count yields one checkpoint usable across a range of inference depths. We ask what such a model actually computes per token, and measure it directly. On a 135M-class model trained on FineWeb-Edu, the recurrent state converges to a per-token fixed point: mean successive-output KL divergence falls from 3.9e-1 at the second loop to 8.5e-6 by the sixteenth, and per-token state change decays in step. Crucially, this convergence is not uniform across tokens. The median token converges by loop six, while approximately 10 percent of tokens continue to update at the training-mean depth of eight, and mean convergence depth is ordered by token type (whitespace shallowest, content words deepest). This per-token variation is the central object of the paper. We show it is directly readable and that reading it outperforms learning to predict it: a training-free rule that halts each token once its output stabilizes attains uniform depth-8 quality at 4.94 average loops (a 38 percent reduction in average depth) and matches uniform depth across the average-depth range, whereas a linear router trained on convergence labels harvested from the same model requires nearly full depth and yields no reduction. The elasticity that makes this possible reproduces here as background (validation loss decreases monotonically from 3.80 at one loop to 3.20 at eight and remains stable to 32 loops). We report average depth as a FLOP proxy with a three-point wall-clock bracket rather than a realized speedup, make no FLOP-matched parity claim, and note that the allocation results are established at a single scale and seed. The complete study runs on a single RTX 4090 in approximately 100 GPU-hours.

Joe Logan

Independent researcher

Transformer Inference Recurrent Transformers Convergence Analysis Adaptive Computation

Jul 11, 2026cs.AI

CHASE: Cache-Hole-Adapted Skip Exit for Looped State-Space Language Models

Recent work on looped language models suggests that many reasoning problems benefit from greater computational depth rather than from additional independent parameters. Existing studies, however, focus almost exclusively on Transformer backbones, leaving open whether this principle also applies to state-space language models. We investigate Looped Mamba and Looped Hybrid Mamba-Transformer architectures, which repeatedly apply a shared Mamba (or hybrid) block to introduce explicit finite-depth recurrent computation. On two controlled reasoning tasks-Mano (modular-arithmetic manipulation) and p-hop induction-Looped Mamba consistently outperforms parameter-matched non-looped baselines and, in several settings, matches or exceeds non-looped models of equal effective depth. We then extend the study to language model pre-training under matched iso-parameter and iso-FLOPs protocols, which jointly disentangle the effects of parameter sharing and effective depth: looped models remain competitive on downstream benchmarks with substantially fewer distinct parameters, although deeper non-looped models retain an advantage in validation perplexity under strict iso-FLOPs comparisons. Finally, we adapt Ouro's two-stage exit gate to Looped Mamba for threshold-controlled selection among recurrent-step outputs. Executing such exits on a state-space backbone, however, leaves the recurrent state without its deeper updates, and validation perplexity then degrades severely. We therefore introduce a cache-hole adaptation that aligns continued training with skipped-state inference. At the scales studied, the adapted model keeps perplexity close to full computation and matches or exceeds full-compute exit-state selection on downstream benchmarks while executing roughly half of the recurrent steps, which translates into measured inference speedups once the prefill is compute-bound.

Zhenxuan Yu, Takeshi Kojima, Yutaka Matsuo +1

Rikkyo University · The University of Tokyo

Efficient Language Model Inference LLM Reasoning Adaptive Computation State Space Models

Jul 8, 2026cs.CC

Computing with Stochastic Oracles in AI-Augmented Computation

The Stochastic-Oracle Turing Machine (SOTM) framework models AI-augmented computation as the interaction of a probabilistic Turing machine with an oracle whose responses are drawn from context-dependent distributions. This paper studies what an SOTM can achieve under two oracle-response schemes: in a cached-response oracle, each distinct query receives one response that is reused on later calls to the same query, while in a fresh-response oracle, each call returns an independent response. In both schemes, the SOTM first computes from its input and internal random source to generate its first query, then proceeds adaptively, computing from its query-response transcript (the record of queries issued and responses received) to generate each subsequent query or produce a final output. Cached responses impose two transcript-based ceilings on achievable performance: a correct-identification ceiling governed by the total variation distance between the transcript distributions induced by the hidden states of the oracle, and an output quality ceiling equal to the expected score of the best output the SOTM can compute from the transcript. Fresh responses can raise these ceilings by allowing repeated calls to accumulate independent evidence toward correct or high-quality outputs. In the binary single-informative-query case, the error probability decreases exponentially in the number of calls to the same query at the Chernoff rate. For output quality, query-count bounds characterize threshold stopping when the score function is incorporated as part of the SOTM, and majority-based amplification bounds characterize the binary candidate-output model when it is not. Together, the results identify how response reuse, transcript information, and access to the score function determine what an SOTM can compute and at what token cost.

Jie Wang

Richard A. Miner School of Computing and Information Sciences, University of Massachusetts, Lowell, MA 01854, USA.

Oracle Complexity Adaptive Computation