Sparse
-
UniAVC: Unified Audio-Video Creation via Progressive Training and Asymmetric Scheduling
2026-10-20 07:19:16Kien T. Pham, Yuhao Liu, Chong Hou Choi, Zhenqi He, Yanghao Wang, Chenyang Lei, Rui Liu, Qifeng Chen, Long ChenAbstract
Unified audio-video creation requires a single model capable of text-conditioned joint synthesis, bidirectional cross-modal generation, and instruction-guided editing across both modalities. Existing methods rely heavily on paired corpora, which are scarce for joint generation and even rarer for editing. Consequently, they are spent learning single-branch generative and editing capabilities that abundant unimodal data could readily teach, rather than on the cross-modal coupling. Furthermore, prevailing systems apply a single shared noise schedule to both modalities, neglecting two critical limitations. First, due to the significant discrepancy in information density, video requires prolonged high noise to organize spatial and motion structure, whereas the sparser audio signal can settle earlier during the synthesis process, the dynamics a uniform schedule might compromise. Second, lockstep denoising leaves asymmetric workflows, such as conditioning on one modality to synthesize or edit another, untrained. We present UniAVC, which addresses both obstacles. It first trains each branch on unimodal dataset to full generative and editing competence, and subsequently dedicates paired data to exploit cross-modal correspondence. Modality-decoupled schedulers meanwhile warp the noise schedule independently from a shared timestep, while subtasks with one clean modality directly supervise asymmetric configurations. These settings respectively enable intent-aware editing, which preserves one modality intact while enlisting it to constrain the other, and post-inference self-refinement, which regenerates only an unsatisfactory branch against an accepted one. UniAVC thus spans multiple creation tasks from a single set of weights without task-specific adapters.
Abstract (translated)
URL
https://arxiv.org/abs/2610.24791
PDF
-
VEOcc++: Learning Persistent Voxel Feature Maps for Embodied Online Semantic Occupancy Mapping and Offline Refinement
2026-10-20 06:21:40Ruoyu Wang, Jiahan Li, Yong Liu, Kexin ZhangAbstract
Semantic occupancy mapping supports embodied perception during exploration and detailed scene understanding afterward, giving rise to two complementary requirements: online incremental mapping for maintaining an up-to-date global map under causal constraints, and offline global map refinement for improving the completed scene representation with additional computation. We present VEOcc++, a unified voxel-centric framework that connects both stages through a persistent sparse voxel feature map. During exploration, the map expands with newly observed regions and integrates aligned local features through reliability-guided updates, while context-aware decoding produces the current semantic occupancy map. After exploration, cached local features are consolidated as complementary evidence and combined with the completed online representation for multi-scale spatial refinement. On EmbodiedOcc-ScanNet, VEOcc++ achieves 67.12\% IoU and 58.40\% mIoU, establishing new state-of-the-art performance for online incremental mapping. With the local predictor fixed, our online mapping pipeline outperforms VEOcc fusion by 3.70 IoU and 3.74 mIoU points, demonstrating substantial gains from the proposed global mapping design. Offline global map refinement further improves the completed map to 67.79% IoU and 59.37% mIoU using only the observations acquired during exploration. Zero-shot deployment on a mobile robot further demonstrates transferability to unseen environments, onboard feasibility, and scalability to large-scale indoor mapping. Project page: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27704
PDF
-
E8-QAT: Efficient 2-Bit Quantization-Aware Training with Implicit Lattice Codebooks
2026-10-20 06:17:09Xinyu Lei, Hong Huang, Jiebin Zhang, Hongyu Li, Jinlong Song, Decheng Wu, Guanghua Yu, Jianchen Zhu, Sujian LiAbstract
Low-bit quantization reduces the memory footprint of large language models, but preserving accuracy with efficient training remains challenging. Scalar quantization encodes weights independently, while explicit-codebook vector quantization introduces costly search during training. We propose E8-QAT, a 2-bit quantization-aware training framework that uses the $E_8$ lattice as an implicit eight-dimensional codebook. It jointly quantizes weight groups through arithmetic projection without storing or searching an explicit codebook during the training forward pass. A 2-bit core encoding and sparse extension preserve the selected lattice points exactly, adding less than 0.1\% to core index storage. On Qwen3.6-35B-A3B, E8-QAT matches the performance of the full-precision model in terms of the average score across the four tasks. E8-QAT improves over the state-of-the-art baseline by 2.03\% on IFEval and by 1.51\% on LiveCodeBench, under the same data, schedule, and bit width. The 2-bit representation reduces peak memory by 80\% and supports a $4\times$ larger batch. At each representation's largest feasible batch, E8-QAT reaches $1.97\times$ the decoding throughput of BF16.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27699
PDF
-
HiQGS: Hierarchical Spatiotemporal Querying for Feed-Forward Gaussian Reconstruction
2026-10-20 04:10:57Qiaowei Miao, Qifeng Cheng, Yijing Guo, Yuru Xiao, Ziqian Ni, Sheng YangAbstract
Feed-forward Gaussian reconstruction from sparse posed monocular frames requires reasoning over observations that vary jointly in viewpoint and time. Existing methods either predict Gaussian primitives from observation-aligned features or aggregate multi-frame information through compact latent queries. However, the former tightly couples Gaussian allocation to individual frames, while the latter does not explicitly structure observed evidence according to its temporal relation to the requested timestamp. Consequently, evidence across spatial scales, viewpoints, and timestamps can be difficult to align and integrate. We present \textbf{HiQGS}, a query-based hierarchical framework for feed-forward Gaussian scene reconstruction from sparse monocular frames. HiQGS introduces a \textbf{Hierarchical Spatiotemporal Encoder} that organizes multi-resolution geometry-appearance visual features according to image location, observation time, and camera pose, thereby preserving complementary appearance and structural cues. Our \textbf{Time-Adaptive Decoder} use target timestamp queries the resulting hierarchical representation to retrieve relevant spatiotemporal evidence and predicts the corresponding Gaussian scene state. Experiments on static and dynamic scene reconstruction benchmarks demonstrate strong rendering quality and geometric accuracy, validating the effectiveness of hierarchical visual encoding and target-time-aware Gaussian querying.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27604
PDF
-
Meta Expert: Cross-Layer Expert Sharing for Storage- and Retrieval-Efficient MoE
2026-10-20 04:05:56Ermo Hua, Xiangyu Hong, Baiting Wu, Che Jiang, Haozheng Hou, Yicheng Gu, Hanjing Wang, Qirui Zeng, Ming Zhang, Ning Ding, Qipeng Guo, Youbang Sun, Biqing Qi, Bowen ZhouAbstract
As model parameters scale up, sparsely activated Mixture-of-Experts (MoE) approaches have become essential for model scaling. However, mainstream models currently adopt hierarchical architectures, where experts in different layers, although topologically independent, exhibit substantial similarity and redundancy in their stored content. Compared to the hierarchically arranged \textbf{Local Expert} in existing mainstream models, we hypothesize that inter-layer shared \textbf{Meta Expert} hold greater potential. To this end, we conduct a multi-dimensional empirical comparison between Local and Meta Experts, and find that models incorporating Meta Experts exhibit lower parameter redundancy, together with lower cross-entropy loss and higher downstream task accuracy. With the activated quantity and the expert granularity held fixed, \textbf{our model matches the performance of Vanilla MoE using only 70--80\% of the experts}, substantially improving the model's scaling potential. Beyond removing inter-layer redundancy among experts and improving information storage efficiency, our theoretical analysis further shows that, compared with Local Experts, Meta Experts provide hidden states with more flexible residual pathways, thereby improving information retrieval efficiency.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27598
PDF
-
Diffusion Subgoal Planning for Long-Horizon Offline Goal-Conditioned Reinforcement Learning
2026-10-20 00:56:00Hengrui Zhang, Yuhu Cheng, C. L. Philip Chen, Xuesong WangAbstract
Offline goal-conditioned reinforcement learning (GCRL) learns goal-directed policies from reward-free data, but in long-horizon tasks, goal-conditioned value functions often provide unstable guidance due to sparse rewards and discounting. Hierarchical methods partially mitigate this issue via subgoal decomposition; however, high-level decision-making still relies on noise-sensitive value estimates, leading to unstable behavior in complex environments. We address this limitation by proposing \textbf{D}iffusion \textbf{S}ubgoal \textbf{P}lanning (\textbf{DSP}), a diffusion-based framework for high-level subgoal generation. DSP casts high-level planning as guided generative inference over goal-conditioned subgoals and learns both conditional and unconditional flows, enabling classifier-free guidance to introduce a goal-directed bias at inference time. By removing explicit value-based guidance from high-level planning, DSP generates reachable and goal-directed subgoals through a generative model while retaining hierarchical execution. Experiments on offline GCRL benchmarks demonstrate that DSP outperforms prior methods on a range of navigation and manipulation tasks, with particularly strong performance in maze environments that require multi-step subgoal planning.
Abstract (translated)
URL
https://arxiv.org/abs/2609.34575
PDF
-
MLIPS: Multi-Layer Indexer Pool Sharing
2026-10-20 00:37:29Samkit Jain, Sai Kishan Pampana, Shlok AgarwalAbstract
Sparse attention limits the keys used by each query, but finding those keys can still require a full-prefix scan at every layer. We introduce \textbf{MLIPS (Multi-Layer Indexer Pool Sharing)}, a training-free method that shares candidate search while retaining each consumer's own scoring and selection. A producer publishes a top-$M$ pool larger than the attention budget $K$; subsequent layers select their own top-$K$ from that pool. Measurements across three sparse-attention model families show substantially higher coverage from broad pools than overlap between final selections. We implement layerwise sharing, persistent decode reuse, and a coarse producer in vLLM for DeepSeek-V4-Flash. Across four long-context suites, three consumer layers per producer give $1.14$--$1.40\times$ speedups and $14.8$--$27.4\%$ lower GPU energy, with observed accuracy changes of $-1.49$ to $+0.29$ percentage points in single runs. More extensive sharing exposes task-dependent quality losses. Sampled prefill traces also reveal recurring selected KV positions across query groups and layers. We characterize this shared working-set structure and use it to motivate Top-$G$, an exploratory bounded pool construction for prefill. Its deployment, task accuracy, and serving benefits remain open.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27438
PDF
-
The Router Already Knows
Abstract
Mixture-of-experts (MoE) models already make a sparse choice for every token: which experts should process it. We ask whether a sparse dictionary should reuse that choice. NR-SAE gives each routed expert a small dictionary and opens only those the frozen router selects. At block 15 of Gemma 4 26B-A4B, after ten million C4 training tokens and across three seeds, the expert-aligned autoencoder explains 0.518 of the variance in router-weighted expert outputs on a fixed evaluation capture. A rank-addressed control with the same parameters but sixteen times as many candidates reaches 0.481, a candidate-matched control 0.458, and an architecture-matched control with randomized expert identity 0.387. A transcoder variant also leads all three controls: 0.432, against 0.347, 0.371, and 0.118. Part of this advantage is architectural: each expert writes into a subspace of at most a quarter of the model width, so naming the expert names the subspace. How much it explains, we have not measured. At this layer and budget, the router already knows something a dictionary would otherwise have to learn. Our implementation did not train faster, and semantic interpretability, downstream loss, and causal faithfulness remain untested.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27421
PDF
-
A latent control model for realistic rodent motion
2026-10-19 20:54:21Aidan Sirbu, Charles Y. Zhang, Yuanjia Yang, Bence P. \"Olveczky, Talmo D. Pereira, Blake A. RichardsAbstract
The development of embodied models of animal behavior is a major goal of the neuro-AI community. Embodied models will enhance our ability to make direct comparisons between biological and artificial agents. To make such comparisons effective, though, we need embodied models that actually move like animals. Previous work has developed virtual animal bodies and techniques for training realistic motion through imitation. However, to date, there is no system for embodied animal modeling that produces natural movements both during trained goal-directed behavior, and during exploration driven by random generation. Here, we introduce Simulated Control via Augmented Motor Priors for Embodied Reinforcement learning (SCAMPER), a system for training an embodied virtual rat model, enabling natural movements during both goal-directed behavior and random exploration. SCAMPER uses a variational architecture with two distinct priors, a motion prior for control and a behavior prior for exploration. We show that combining both of these priors allows SCAMPER to generate realistic behavior during exploration and exploit the task structure effectively. SCAMPER can also learn from sparse rewards as a result of the diverse movements induced by the behavior prior, allowing the system to model realistic sparse-reward tasks from experimental neuroscience. Altogether, SCAMPER provides the neuro-AI community with an effective new system for training embodied models of rodents on realistic tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27279
PDF
-
Sparse RL Relaxation for LLM Reasoning
2026-10-19 19:50:46Jiabin Fan, Matthias Lindemann, Zita Marinho, Andr\'e FT Martins, Lili MouAbstract
Recent work has explored continuous representations in place of discrete tokens during Chain-of-Thought (CoT) reasoning, often referred to as soft CoT. This is motivated by the idea that a mixture of token embeddings can represent a superposition of multiple reasoning possibilities within a single step. However, it remains challenging to train large language models to realize this potential. Existing work has explored reinforcement learning (RL)-style training for soft CoT. However, these approaches treat soft tokens as actions in a Markov decision process rather than as differentiable intermediate computation steps. Thus, they do not propagate gradients across the soft tokens and miss out on the signal of how the choice of one soft token affects the next. In this paper, we propose a sparse relaxation method for RL (called SpaRLax), which leverages the differentiability of soft-CoT reasoning over time steps and makes training more stable and effective. We adopt an entmax sparsity method that adaptively determines the degree of softness, which truncates long-tail noise at confident steps while retaining multiple plausible candidates when uncertain. Across nine reasoning benchmarks, our method achieves the highest average performance compared with baseline methods. Further analyses validate the effective use of soft-token superposition in our approach, and also demonstrate the contribution of adaptive sparsity in our approach.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27220
PDF
-
Reconstruction-Optimized Expert Pruning for Mixture-of-Experts Language Models
2026-10-19 18:24:03Jelena Markovic-Voronov, Petros Prastakos, Kayhan Behdin, Jincheng Cao, Zhipeng Wang, Shayan Mohajer Hamidi, Yuanda Xu, Zhengze Zhou, Wenhui Zhu, Rahul MazumderAbstract
Mixture-of-Experts (MoE) models reduce computation through sparse expert activation, but deployment still requires storing a large number of expert parameters. We study post-training expert pruning using task-domain calibration data as a subset-selection problem: given a pretrained MoE layer, select a fixed number of experts that best reconstruct its original output. Exact subset search is combinatorial and becomes intractable for modern MoE models with many experts per layer. We develop OptEP, a scalable continuous relaxation that directly optimizes differentiable expert-selection variables, followed by a lightweight reconstruction-based update of the retained expert weights. Across multiple MoE architectures and tasks spanning mathematical reasoning and code generation, our approach improves over the strongest existing expert-pruning baselines, with absolute accuracy gains of up to 15.05% on Qwen3-30B-A3B, 14.55% on Qwen3-235B-A22B, 11.06% on GLM-4.5-Air, and 16.20% on Mixtral-8x7B. The improvements are particularly pronounced under aggressive pruning, demonstrating that reconstruction-driven expert selection and weight adaptation can preserve substantially more of the original model's performance while reducing the number of retained experts.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27123
PDF
-
Counting Contributions, Not Views: Occlusion-Aware Densification for 3D Gaussian Splatting
2026-10-19 18:09:10Marcel Rogge, Ren\'e Schuster, Didier StrickerAbstract
3D Gaussian Splatting grows a dense scene from a sparse initialization through adaptive density control (ADC), driven by per-Gaussian statistics gathered during rendering. These statistics pass a frustum-based visibility filter blind to occlusion: fully occluded Gaussians still count as observed, diluting the averaged gradient signal and suppressing densification precisely where geometry is occluded. We correct this with an occlusion-aware filter that restricts densification statistics to views in which a Gaussian actually contributes, and replace the criterion's raw view-count normalizer with an average-contribution one. Integrated into a StopThePop-based framework, the two together improve every image-quality metric over the occlusion-blind baseline on occlusion-heavy scenes (up to 6% lower LPIPS), at less than half the Gaussians a view-count normalizer requires. Under the longer schedules these large scenes require, the count stays bounded where most criteria diverge, and on smaller benchmarks our method stays competitive at a fraction of the competitors' count.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27108
PDF
-
Decomposed Relation Extraction from Aviation Accident Narratives: Evaluating Transformer-Based Safety Knowledge Discovery under Relational Sparsity
2026-10-19 12:25:20Aziida Nanyonga, Mustafa Tahsin Yilmaz, Saad Ezzini, Ugur Turhan, Graham WildAbstract
Relation extraction from aviation accident narratives is challenging because genuine safety relations are sparse among possible concept pairs. This study evaluates a decomposed framework separating relation classification from concept identification. A benchmark of 300 National Transportation Safety Board (NTSB) narratives containing 2,524 unique concept-text nodes was expanded into 21,038 ordered candidate pairs, with only 248 (1.18%) representing RESULTS_IN or CONTRIBUTES_TO relations. A TF-IDF Linear Support Vector Machine (SVM) and weighted RoBERTa classifier were developed using an independent graph-annotated resource with hard-negative sampling and frozen before sealed evaluation. Development Macro-F1 scores of 0.4738 (SVM) and 0.4809 (RoBERTa) declined to 0.2741 and 0.2420 under exhaustive evaluation, with Matthews correlation coefficients of 0.0128 and 0.0197. False positives dominated: the SVM recovered 96 reference relations but generated 7,018 false positives, while RoBERTa recovered 125 but generated 9,824. Post-test diagnostics revealed overlapping RoBERTa confidence distributions and false positives even within valid relation-category directions. Complementary end-to-end SmolLM2 evaluation identified missed and unsupported relations alongside malformed outputs. These findings demonstrate that known concept endpoints do not guarantee reliable relation extraction and that sampled-negative development performance may poorly reflect exhaustive evaluation. Controlling false positives and managing candidate-space sparsity remain essential for aviation safety knowledge extraction.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27013
PDF
-
FoundPDE: A generative foundation model for data-driven symbolic discovery of partial differential equations
2026-10-18 15:40:42Langchen Liu, Junlin Luo, Min Zhu, Zhuoran Yang, Lu LuAbstract
Discovering governing partial differential equations (PDEs) from observed data is a central inverse problem in computational science and engineering: the goal is not only to predict trajectories, but to identify interpretable symbolic laws that can be analyzed and reused. Existing PDE discovery methods often depend on hand-designed libraries, accurate numerical derivatives, or per-equation symbolic search, making them fragile under sparse observations, noisy data, broad equation families, and unseen symbolic compositions. Here, we introduce FoundPDE, a pretrained data-to-symbol foundation model for PDE discovery. FoundPDE maps discretized spatio-temporal trajectories to symbolic equations with a transformer, uses Fourier and first-order spatial-derivative augmentation to expose complementary trajectory information, refines numerical coefficients, and supports downstream extension with improved generalization through few-shot fine-tuning and hierarchical discovery. On a clean in-domain benchmark spanning 20 PDE families, our model reaches 97.50% type recognition. Under coefficient shifts beyond the training support, type recognition remains 90.53%; under multiplicative Gaussian noise on the model input, it is 94.43% at a 2% noise level but falls to 52.09% at 10%. For unseen composite equations, few-shot fine-tuning reaches over 90% exact recovery on four held-out families, and we further develop a hierarchical discovery algorithm that recovers two-term composites without fine-tuning FoundPDE. By amortizing symbolic discovery across equation families, FoundPDE shifts PDE discovery from per-dataset search toward a reusable pretrained model that extends to new physics with limited additional data, a step toward foundation models that uncover interpretable governing laws directly from scientific data.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26963
PDF
-
Expert-Guided Rubrics as Dense Reward Signals for Reinforcement Learning in Non-Verifiable Domains
2026-10-17 19:01:11Ashish Bansal, Hossam Amer, Keerat Kaur Guliani, Parinita Edke, Lesley Matheson, Leizer TeranAbstract
Large language models (LLMs) are increasingly deployed as autonomous agents in non-verifiable domains such as financial services, where they must interpret user intent, rewrite queries, and retrieve information under domain-specific constraints. Although reinforcement learning (RL) can train such agents from the consequences of their actions, recent successes largely rely on verifiable tasks with automatically computed binary rewards. In non-verifiable retrieval tasks, assessing query quality requires scarce subject-matter expert (SME) judgment that is inherently graded rather than binary. With Group Relative Policy Optimization (GRPO), sparse rewards yield no useful learning signal when all sampled outputs within a group receive the same outcome, causing optimization to stall. We introduce a method that converts a small, fixed investment of SME judgment into a dense reward and combines it with retrieval-based metrics to form a hybrid, GRPO-compatible training signal. Experiments show improvements of up to 6\% in Recall@3 and 6.7\% in MRR@3 compared with sparse-reward baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26932
PDF
-
Learning to Correct Geometry in Generated Videos
2026-10-16 17:59:55Weijie Lyu, Tiancheng Shen, Jason Li, Yujing Wang, Xiaoyuan Wang, Haobo Yuan, Yizhou Zhao, Ming-Hsuan YangAbstract
Generated videos often lose geometric consistency under camera motion: structures warp under viewpoint change, parallax becomes inconsistent, and smooth camera movement may collapse into shot-like transitions. We argue that this failure is largely a data problem. In web-scale video corpora, most visible motion comes from dynamic objects captured by static or weakly moving cameras, leaving only sparse supervision for viewpoint-induced appearance changes. To address this, we propose Video Geometry Corrector (VGC), a post-hoc video-to-video model trained on synthetic paired supervision. Starting from real videos with diverse camera motion and pose annotations, we construct VG-Pairs with a synthetic paired-supervision data generation pipeline. Each video clip is treated as a geometry-consistent target and paired with a distorted counterpart synthesized by a pretrained image-to-video inpainting model using trajectory-adaptive keyframes. VGC learns to map the distorted input to the geometry-consistent target without modifying the source generator, and uses optical flow as an auxiliary cue to help disentangle camera motion from object motion. Experiments on pose-conditioned and prompt-driven benchmarks show that VGC improves geometric consistency across multiple source generators, and scaling to a larger backbone further extends the framework toward more general video refinement.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23856
PDF
-
SparseMax-dSTL: Sound Robustness with Horizon-Independent Gradient Allocation for Temporal Logic Trajectory Optimization
2026-10-16 17:54:38Sai Sandeep Damera, Ryan Matheu, John S. Baras, Calin BeltaAbstract
A gradient-based solver that enforces a Signal Temporal Logic (STL) specification uses a smooth approximation of its quantitative semantics, the robustness. In safety-critical trajectory optimization, this approximation must have two properties. First, it must be sound, that is, nonnegative only on trajectories that satisfy the specification. Second, at each temporal operator, the weight of its gradient on the violating samples must stay bounded away from zero, however long the horizon. The approximations in use, including the log-sum-exp, its sound variant, and the generalized-mean robustness, do not have both. We introduce lower approximations of the minimum and the maximum whose gradients are sparsemax projections, and we prove that the robustness built from them has both. It lies below the exact robustness and is therefore sound, and its error has a bound that is independent of the number of samples and known before optimization. The gradient of each approximation is exactly zero on every sample beyond a data-dependent threshold, so the weight on violating samples a gap beyond the others has a positive lower bound at every horizon. On a planar unicycle and on a seven-joint manipulator beside a person in simulation, the log-sum-exp approximation is positive on a violating trajectory. As the horizon of the manipulator's task grows from 10 to 80 s, the gradient weight of the sound variant on the violating samples falls from 0.22 to 0.04, while that of sparsemax stays at 1. In a constrained trajectory optimization of the unicycle, the constraint on the log-sum-exp robustness returns a violating trajectory, whereas the constraint on the sparsemax robustness returns a trajectory that satisfies the specification with the required margin.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23837
PDF
-
Beyond SAEs: Consistent Concept Discovery with Sparse Coding
2026-10-16 17:25:11Jinho Choi, Hyesu Lim, Jaegul Choo, Steffen SchneiderAbstract
Sparse autoencoders (SAEs) explain foundation model representations by decomposing activations into sparse combinations of learned dictionary atoms. However, the learned dictionaries can vary across training runs, limiting the reliability and reproducibility of SAE-based analyses. We propose Consistent Sparse Dictionary Learning (ConSDL), which replaces the standard linear encoder with per-sample sparse coding. This is motivated by our synthetic analysis showing that coefficients from a single linear encoder pass are suboptimal compared with those obtained by per-sample optimization, suggesting that sparse coding is a reliable mechanism for dictionary learning. Across 16 synthetic settings, ConSDL achieves near-perfect dictionary and coefficient recovery. These gains translate to state-of-the-art cross-seed consistency across four vision and language settings, where ConSDL improves dictionary and coefficient consistency by 17.6% and 35.9%, respectively, over the average of six SAE baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23801
PDF
-
Trajectory-Triggered Sparse Collaborative Decoding
2026-10-16 15:24:05Amit Ranjan Trivedi, Sruthikesh Surineni, Kyong Ho Lee, Mohit Mittal, Steven Li, Rex HillAbstract
Dense assistance methods such as speculative decoding repeatedly verify draft tokens with a target model, keeping target computation coupled to generation. We study sparse collaborative decoding, where a draft model generates most tokens and a target model is invoked only for short corrective segments. The key problem is deciding when to call the target. We show that local uncertainty is a weak trigger: entropy, negative log-likelihood, logit margin, top-$k$ mass, and velocity features give near-chance failure prediction, with AUC 0.50-0.53. In contrast, aggregate trajectory-shape features over the draft confidence path reach AUC 0.86 using a lightweight logistic predictor trained only from draft correctness labels. This induces a trigger-timing tradeoff: early calls have more leverage but weaker evidence, while later calls have stronger evidence but less ability to redirect the trajectory. Used online, learned sparse collaborative decoding improves over the best open-loop schedule by +5.3 to +20.0 accuracy points across five GSM8K draft-target pairs. On MATH-500, it gains 7.0-13.5 points when early prediction and target competence align, but loses when either is weak. On StrategyQA, it improves over open-loop by 5.0-8.3 points, showing transfer beyond arithmetics.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23613
PDF
-
Metro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals
2026-10-16 15:23:50Royson Lee, Fady Rezk, Titouan Parcollet, Timothy Hospedales, Cristina CornelioAbstract
Model-predictive control with Joint-Embedding Predictive Architectures (JEPAs) provides a strong zero-shot goal-reaching planner, but it is only effective over short planning horizons. Hierarchical extensions attempt to bridge this gap by learning a macro planner to predict intermediate latent sub-goals to guide the micro planner. In this work, we demonstrate that unconstrained latent sub-goal prediction is fundamentally flawed. A rigorous evaluation reveals that a leading state-of-the-art macro planner routinely emits physically unrealisable sub-goals. To resolve this, we introduce Metro-WM, a hierarchical framework that issues sub-goals by retrieving genuine states from prior experience rather than generating ungrounded latent vectors. Specifically, Metro-WM constructs a graph whose vertices are observed frames from offline expert demonstrations or random-action trajectories, allowing frames from different episodes to be connected and stitched into routes to the goal. Planning over the full graph also makes the system highly robust to execution errors: if the micro planner drifts off course, Metro-WM instantly finds a new optimal path from the current state. Our experiments show that Metro-WM achieves superior long-horizon success rates of up to 37.33 percentage points over the next best hierarchical approach while being up to 10.9 times faster, requiring both 13-56 times less offline compute and fewer tuned hyperparameters. Additional analysis reveals that Metro-WM finds shorter paths than the offline demonstrations, outperforms an oracle relying on the query's own demonstration, and maintains robust performance under extremely sparse dataset conditions.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31868
PDF
