Diffusion
-
ForeTac: Predictive Contact Guidance for Generative Robot Policies
2026-10-20 10:22:41Zhiyuan Chen, Shuai Chen, Haoran Zhou, Yizhe Zhao, Kaiyuan Yang, Zhixiang Duan, Shaoqing Xu, Ran Cao, Diyun Xiang, Yangwei YouAbstract
In contact-rich manipulation, actions that appear equally plausible from vision can produce different contact outcomes, while reactive control based on current tactile feedback may respond too late to prevent failure. We present ForeTac, a predictive contact-guidance framework organized as a represent-predict-evaluate-guide loop. ForeTac encodes the current tactile state, predicts future tactile latent trajectories conditioned on candidate action sequences, and evaluates each predicted trajectory with a contact-quality objective learned from successful and failed physical interactions. At inference, gradients of this objective are backpropagated through the tactile predictor to guide action generation before execution while keeping all model parameters frozen. We evaluate ForeTac on four real-robot tasks (planar and curved-surface wiping, card swiping, and fragile-object grasping) with diffusion and flow-matching policies. Compared with vision-only Diffusion Policy, ForeTac improves the average task success rate (SR) from 36.3% to 71.3% and the average contact success rate (CSR) from 20.0% to 66.3% across these tasks, and yields gains in both metrics with a flow-matching policy. Component analyses support the learned tactile representation, action-conditioned foresight, and generation-time gradient guidance, while predicted tactile latent trajectories preserve contact-quality ordering. Project page: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18118
PDF
-
Decoupling Action from Egocentric Observation for World Simulation
2026-10-20 09:52:46Yue Ma, Pengjie Song, Xinyu Wang, Yi He, Zeqian Long, Fangneng Zhan, Kaichen Zhou, Hongyu Liu, Hongfa Wang, Peihao Li, Haoyang Huang, Nan Duan, Qifeng ChenAbstract
Egocentric action transfer aims to decouple the semantic action from egocentric observations and reproduce them in new visual contexts, enabling scalable world simulation, embodied policy learning, and robotic data generation. Existing approaches to transferring actions rely on condition-guided video generation, which converts the observation video into explicit geometric controls (e.g., hand poses or meshes) to drive synthesis. However, these methods introduce geometric estimation errors, making it difficult to preserve physically plausible interactions when transferred to new scenes. Alternatively, motion transfer methods directly extract motion patterns from the observation video, yet motion-level features alone cannot encode rich contact dynamics, frequently leading to severe hand structural collapse and implausible contact layout. To address both limitations, we present EgoACT, a test-time framework for egocentric action transfer that operates directly on the denoising process of video diffusion models without requiring additional geometric estimators. EgoACT uses Velocity-guided Structure Anchoring to stabilize reference-consistent hand-object structure in the early denoising stage, and Sparse Correspondence Calibration to refine reliable local correspondences in the mid-to-late denoising stages. Together, these two components preserve transferable action semantics while improving temporal coherence and interaction realism. We further establish EgoActionBench, a benchmark for evaluating action preservation, visual quality, and hand-object plausibility across diverse egocentric manipulation scenarios. Experiments show that EgoACT generates more coherent and physically plausible interaction videos than strong baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27946
PDF
-
What Novice Operators See Changes What the Policy Learns: A User Study of Keypoint Hand Visual Representation for Teleoperated Demonstration
2026-10-20 09:22:21Amane Yamaguchi, Shohei Kamiguchi, Masashi Hamaya, Shigeo Yoshida, Takuji Narumi, Mai NishimuraAbstract
Imitation learning increasingly relies on novice operators, whose demonstrations fall short of experts' in number and inter-trial consistency. Operators act on what they are shown, yet how the rendering of the robot affects their demonstrations, and the policies trained on them, has not been examined. We study this question with PointHand, a teleoperation interface that renders a multi-fingered robot hand as sparse keypoints at its fingertips and joints, the same keypoints a policy can observe. In a simulated mug-in-box task, 18 novices and 4 experts collected 2,200 trials under three renderings: the real shape of the hand, all keypoints emphasized, or the thumb and index finger alone emphasized. The novices shown the thumb and index finger emphasized collected 441 successful demonstrations from 600 trials, against 320 for those shown the real shape. We then trained diffusion policies on these demonstrations, replacing the scene point cloud and proprioceptive state they observe with the hand keypoints and two object keypoints. Within 400 steps these demonstrations went from the weakest of the three novice conditions under the point-cloud policy (46.3%) to the strongest under the keypoint policy (57.3%), a gain of +11.0 percentage points in all five seeds. Those collected with all keypoints emphasized changed by at most +0.7 percentage points, and in no more than three seeds. The value of a demonstration therefore depends on the rendering shown during collection and on what the policy observes together. A shared representation can pay off on both sides, in the number of successful demonstrations and in the success rate of the policy trained on them. Project page: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.27907
PDF
-
Gaze2HOI: Gaze-Aware Intention Disambiguation for Egocentric Hand-Object Interaction Forecasting
2026-10-20 08:59:54Chuanyang Ma, Chenye Wang, Yixuan Wang, Huiling Zhou, Xingqun Qi, Muyi Sun, Man Zhang, Caifeng Shan, Zhenan SunAbstract
Anticipating how users will interact with their surroundings is essential for proactive systems that provide timely assistance in AR, VR, and embodied applications. Prior work typically forecasts hand-object interactions directly from historical observations, leaving intention and contact geometry implicit. However, similar observation histories can precede semantically or geometrically distinct futures. Inspired by behavioral evidence that gaze can provide early cues for underlying intention, we propose Gaze2HOI, a gaze-aware framework that formulates hand-object interaction forecasting as a hierarchical problem comprising intention inference, contact grounding, and coupled hand-object motion generation. Specifically, Gaze2HOI first weights gaze cues by fixation strength and recency, and uses them to modulate the visual representations of a vision-language model (VLM) through Temporal Gaze Saliency Modulation (TGSM), thereby reducing semantic ambiguity in intention inference. Subsequently, an Anticipatory Contact Grounding (ACG) module maps these structured intentions into explicit contact, support, and placement fields, effectively mitigating geometric ambiguity. Driven by these explicit conditions, a diffusion-based model jointly generates the 6-DoF object trajectories and future articulated motion of both hands. To evaluate Gaze2HOI, we further introduce EgoAIM, a phase-aware benchmark that organizes egocentric hand-object interactions into four phases defined by evolving hand-object relations and provides fine-grained intention annotations. Extensive experiments demonstrate that Gaze2HOI achieves state-of-the-art performance in both intention inference and motion forecasting, even approaching oracle-level end-to-end performance in specific interaction phases. Project page: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27877
PDF
-
TamedFlow: Trajectory-Aware Distillation for Teacher-Assisted Diffusion Sampling
2026-10-20 08:02:36Shuaiqi Cheng, Jiahao Zhang, Boyuan Gu, Yuxi Chen, Haiying Xu, Zexin Lu, Huiyu Zhou, Xuming HuAbstract
Diffusion Transformers (DiTs) deliver high-fidelity image and video generation, but inference requires many sequential model evaluations. Two acceleration strategies are increasingly important: model compression makes each evaluation cheaper, while adaptive inference avoids selected teacher computation through reuse, forecasting, or speculative execution. Their combination, however, is brittle: directly applying the evaluated adaptive methods to a compressed student can degrade quality, an observed failure mode we call the \textbf{Combination Gap}. We link this gap to a mismatch between pointwise distillation and short-horizon execution: a student trained at teacher states need not remain reliable when its own predictions are chained. We propose \textbf{TamedFlow}, which trains a compressed DiT for this deployed role through cross-architecture feature alignment, a near-identity Velocity Alignment Head, and randomized rollout supervision. At inference, a lightweight Learnable Trajectory Estimator routes completed drafts: accepted drafts bypass the teacher, while rejected drafts roll back and receive one teacher rescue step. Distilling DiT-XL/2 into DiT-B/2, TamedFlow reaches FID $2.58$ at $2.25\times$ measured speedup over DDIM-50 on ImageNet $256\times256$. Ablations support trajectory supervision, and routing controls favor LTE over random decisions at nearly matched latency. Evaluations on ImageNet $512\times512$, MS-COCO, and Latte/FaceForensics extend the empirical scope.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26025
PDF
-
PixelSR: Pixel-Space Orthogonal Flow Decoupling for Real-World Image Super-Resolution
2026-10-20 07:21:29Xiang Lv, Kai Zhao, Qiang Wang, Qiao Mo, Chao Zhou, Mingwen ShaoAbstract
Latent diffusion models have recently achieved remarkable progress in Real-world Image Super-Resolution (Real-SR). However, existing methods consistently exhibit a persistent seesaw effect: enhancing perceptual realism compromises structural fidelity. This effect, we argue, is not intrinsic to diffusion models, but largely stems from the architectural constraints of prevailing two-stage frameworks. To address this challenge, we propose PixelSR, a novel single-stage Pixel Diffusion Transformer for Real-SR that operates directly in pixel space. Specifically, PixelSR consists of two core designs: Pixel-space Conditional Diffusion and Orthogonal Flow Decoupling. The former establishes an effective pixel-space restoration framework by integrating spatially aligned LR conditioning, visual semantic guidance, and lightweight texture refinement, thereby bypassing VAE compression and enabling end-to-end optimization directly in RGB space. Complementarily, the latter decomposes the asymmetric flow target into a low-rank semantic component and a noise-free orthogonal texture component. By routing these components to specialized branches, it achieves decoupled optimization with reduced mutual interference. Benefiting from these designs, our PixelSR restores perceptually realistic details while faithfully preserving structural content. Extensive experiments across multiple benchmarks demonstrate that PixelSR not only achieves competitive performance, but also effectively mitigates the fidelity--realism seesaw effect.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27759
PDF
-
AWLS: Advantage-Weighted Latent Steering for One-Step Generative Robotic Manipulation Policies
2026-10-20 06:58:12Jiahong Tang, Jiaju He, Zhiyuan Yang, Ruyang Han, Haiyue ZhuAbstract
Human-collected real-world robotic manipulation demonstrations often contain mixed-quality trajectories, including hesitant, indirect, or inefficient action segments. While diffusion- and flow-based policies can model such multimodal data, standard behavior cloning (BC) assigns equal likelihood to all demonstrated actions and may therefore reproduce suboptimal modes. Offline RL value estimates offer a natural signal for biasing policies toward higher-value behavioral modes, but directly fine-tuning the generative policy or using critic gradients for policy steering can amplify critic errors and induce prior drift. We propose Advantage-Weighted Latent Steering (AWLS), an offline post-training method for improving frozen one-step generative manipulation policies without updating their decoders. AWLS learns a residual noise shaper, an observation-conditioned residual correction to the policy's latent noise, and trains it with an advantage-weighted action reconstruction objective. The advantages are detached, scale-normalized, and used only as supervised weights on demonstrated actions, avoiding critic-gradient policy updates while preserving fast single-step inference. We further introduce a rollout-free decoder-sensitivity criterion for selecting pretrained checkpoints that remain responsive to latent-noise perturbations. Experiments on simulated and real-world robotic manipulation tasks with clean and mixed-quality demonstrations show that AWLS improves over the frozen base policy while limiting drift from the pretrained behavior prior. Project page: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.27736
PDF
-
Push-Pull Attentional Anchoring for Diffusion Concept Erasure
2026-10-20 06:38:51Nattanat Chatthee, Tagon Sompong, Ekapol Chuangsuwanich, Supasorn SuwajanakornAbstract
Rapid advances in diffusion models have raised concerns about privacy, harmful content, and copyright violations. Concept erasure methods aim to address these risks by removing undesirable concepts from pretrained models without full retraining. However, existing approaches often degrade non-target generation quality through heuristic strength scaling (e.g., classifier-free guidance or negative prompt scaling), leading to uncontrolled and excessive semantic drift. To address this, we propose Push-Pull Attentional Anchoring (PPAA), a mechanism in the cross-attention space that displaces target representations while anchoring the scale of erasure via bounded cosine-similarity constraints. By using a relative similarity ratio instead of an absolute difference, our adaptive threshold represents a meaningful percentage of semantic retention, facilitating a single value to be applied uniformly across layers. We conduct extensive experiments across multiple concept categories, including celebrities, artistic styles, nudity, and intellectual property. Our method achieves favorable trade-offs against prior approaches on Stable Diffusion v1.4 and demonstrates its applicability to modern architectures such as FLUX.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27719
PDF
-
BCT: Bayes-Consistent Tempering for Flow Language Models
2026-10-20 05:03:03Daisuke Oba, Hiroki Furuta, Naoaki OkazakiAbstract
Temperature scaling is typically applied as a post-hoc transformation of model probabilities. In posterior-induced flow language models, this creates two mismatches. First, posterior tempering implicitly changes both the endpoint distribution and the distribution of intermediate noisy states, while the denoiser remains trained on the original states. Second, the tempered posterior weights are combined with transport dynamics derived for the untempered flow. These mismatches are reflected empirically: lowering temperature improves GSM8K performance much less reliably in flow LMs than in discrete diffusion LMs. We introduce Bayes-Consistent Tempering (BCT), which makes temperature explicit in both the distributions and dynamics of the generative flow. Its learned realization trains the denoiser on tempered intermediate states and tempers the endpoint distribution using the corresponding tempered dynamics. We also derive training-free variants for conventionally trained checkpoints. Across hyperspherical flow LMs with two denoiser architectures and a von Mises--Fisher flow LM, BCT improves GSM8K generation across broad temperature ranges and achieves the strongest tempering performance. The gains also extend to Sudoku, demonstrating benefits on tightly constrained generation. At very low temperatures, the learned realization is also markedly more robust than the training-free variants, supporting temperature as a property of the generative flow across training and inference rather than a post-hoc probability adjustment. Code and model weights are available at this https URL .
Abstract (translated)
URL
https://arxiv.org/abs/2610.27641
PDF
-
DART: Scaling a Shape Representation Across Anatomical Classes and Imaging Modalities
2026-10-20 04:12:20Koushik Swarna, Pedro R. A. S. Bassi, Lam Ngo, Wenxuan Li, Jiancheng Yang, Alan Yuille, Zongwei ZhouAbstract
Current state-of-the-art segmentation models still produce anatomically implausible shapes, such as broken vessels and fragmented glands, on the structures that are hardest to segment. Learned shape representations capture what structures look like and can help correct these errors, but most cover a fixed set of anatomical classes. Extending this coverage across classes and imaging modalities is costly because it often requires new voxel-level annotations. We propose Diffusion for Anatomical Reshaping with Text (DART), which learns a shared shape representation from segmentation probabilities. It uses anatomical descriptions to correct these probabilities without receiving medical images. Across 13 public benchmarks, absolute Dice gains reach 31% for individual gallbladder cases in CT and 10% for individual prostate cases in MRI. Using existing segmentation predictions, we scale a single shape representation from 50 to 377 anatomical classes without new voxel-level annotations for correction training. The shared representation also corrects four lesion classes absent from training and transfers from CT to MRI without retraining. Corrected labels also improve segmentation models, raising one model's Dice from 57.1% to 67.5%. The same pipeline supports annotations for 45,000 CT scans. These results suggest that sharing shape information within and across classes can reduce the need for new voxel-level annotations when extending segmentation to more anatomical classes and imaging modalities. The model is available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.27605
PDF
-
Self-Decomposing Diffusion for Code Generation
2026-10-20 04:07:10Lize Shao, Zichen Xie, Zhizhen Chen, Zhe Zeng, Wenxi WangAbstract
Diffusion language models (DLMs) are emerging as a promising paradigm for code generation, offering the potential for fast, parallel decoding. However, generating complex programs remains challenging, as their interdependent components must maintain consistent interfaces, data flows, and shared assumptions. Aggressive parallel decoding can compromise these dependencies and undermine program correctness, while conservative decoding sacrifices efficiency. To address this trade-off, we propose planning program decomposition before implementation to coordinate interdependent units during parallel generation. We introduce Self-Decomposing Diffusion for Code (SDDC), a novel framework that decomposes complex programs and implements their units in parallel, entirely through diffusion. SDDC first generates a contract specifying the program's decomposition and inter-unit dependencies, then uses it to guide coordinated parallel implementation. Across five benchmarks and two 7B diffusion backbones, SDDC substantially improves the accuracy--latency trade-off, raising pass@1 by up to 31.1 percentage points while decoding 7.5--13.7$\times$ faster than the base models, and outperforming a fine-tuning baseline by up to 23.5 points with lower latency.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27600
PDF
-
Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?
2026-10-20 03:58:22Xingsong Ye, Yongkun Du, Jiaxin Zhang, Chong Sun, Zhineng ChenAbstract
Aiming to modify textual content while preserving the original style, Scene Text Editing (STE) has significant practical value in many applications. STE typically requires paired images before and after editing for training. However, recent approaches rely heavily on synthetic paired data and complex disentanglement strategies. They struggle to faithfully capture the diversity and complexity of natural scenes, often producing artifacts or hallucinations. In this work, we challenge the common practice and investigate whether STE can be driven by turning each real sample into its own paired supervision. Concretely, we explore multiple ways to disrupt the original image (e.g., Crop, Shuffle, Mask) to construct self-paired training data, and build a strong baseline (TextRIRO}) upon a conditional diffusion architecture. Our experiments show that combining character-column shuffling with image cropping (Shuffle&Crop) effectively destroys text semantics while preserving key style cues, enabling the model to learn editing behaviors directly from the real text image. With this concise yet real-oriented design, TextRIRO achieves promising performance in both accuracy and visual realism across real-world benchmarks, substantially mitigating the synthetic-to-real domain gap. Furthermore, we leverage TextRIRO to build TextRIRO-3M, a large-scale realistic dataset for long-text recognition by concatenating edited and original images. Experimental results show that training recognition models with TextRIRO-3M significantly improves recognition accuracy, demonstrating that TextRIRO provides highly useful training resources for downstream tasks. Overall, TextRIRO lifts STE to realistic, reliable, and widely applicable new levels. Code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27592
PDF
-
Beyond Model Weights: Sampling-Dependent Fairness in Diffusion Language Models
2026-10-20 03:49:20Koshiro Saito, Diasuke Oba, Danushka Bollegala, Naoaki OkazakiAbstract
Diffusion language models (DLMs) are emerging as a practical alternative to autoregressive language models (ARLMs), yet their fairness remains largely uncharacterized. DLMs introduce sampling controls with no direct autoregressive counterpart, such as block length and remasking strategy, that directly set their speed--quality operating point. This raises a basic question: does a DLM's measured social bias remain stable across these settings? Across LLaDA, LLaDA-1.5, and Dream on Open-BBQ, we find that all three models exhibit social bias, but both the magnitude of measured bias and the relative ordering of models by bias are highly sensitive to the sampling configuration. This sampling dependence also changes what successful debiasing means. While mitigation strategies established for ARLMs transfer effectively to DLMs and reduce bias without degrading downstream performance, sampling dependence persists even after mitigation. Our results show that fairness in DLMs cannot be characterized independently of the sampling configuration, even after debiasing. Project page is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27584
PDF
-
TeCoP: Accelerating Diffusion Large Language Model Decoding via Tentative Context Pipelining
2026-10-20 03:48:01Jui-Chien Tsou, Hsuan-Chi Liu, Wei-Chin Wang, Yu-Syuan Xu, Chun-Yi LeeAbstract
Diffusion large language models (dLLMs) predict multiple token positions per forward pass, yet block-wise decoders serialize generation and repeatedly recompute stable predictions. We attribute this inefficiency to coupling a prediction's readiness for irreversible commitment with its utility as context. We introduce Tentative Context Pipelining (TeCoP), a training-free decoder that decouples these roles through staggered block-parallel pipelining and draft-and-verify. TeCoP reuses stable but uncommitted predictions as provisional cross-block context, with commitment governed by confidence and stability checks. Across four LLaDA and Dream dLLMs on GSM8K, Minerva MATH, and MBPP, TeCoP reduces number of forward evaluations by up to 87% and achieves up to 7.50-fold throughput without KV caching and 11.77-fold with caching over the uncached original decoder. It outperforms Fast-dLLM and DSB in throughput across all 12 model-benchmark settings, with and without caching, while keeping average cached accuracy within one percentage point of the original decoder.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27583
PDF
-
Factual Access States in Diffusion Language Models: From Mechanism to Knowledge Editing and Retrieval
2026-10-20 03:45:45Yuchen Guo, Zhuoyuan Mao, Yiwen GuoAbstract
Factual recall is a lookup: some representation inside the network serves as the key, and what that key is made of decides which queries can retrieve the fact, what an edit must cover, and what a retriever can read. We call it the factual access state. In autoregressive (AR) models the subject position enriches attributes largely independently of the relation, and relation-specific selection happens later, at the readout position. In masked diffusion language models (DLMs) we find the state is assembled differently, and much earlier: in the first attention layer the wording of the relation is routed into the position of whichever entity is visible and the early feed-forward layers look the fact up on this fused representation. The state therefore varies with the relation's surface form and with the query direction, so the queries about one fact occupy a region of state space rather than a point, and an operation on the fact succeeds only for queries that enter it. We treat the state as an interface to a DLM's factual knowledge, and put it to two uses. Writing to the state gives a knowledge editor: from one declarative statement, Multi-View Edit fits one shared value per direction across that region and installs it in the early-MLP block, preserving source-prompt reliability while substantially improving paraphrase and reverse question answering, where AR editors gain near zero. Reading it suggests a complementary use: the state at each entity position can serve as a relation-specific retrieval vector with no training, returning both pieces of evidence for two-entity queries on which pooled, prompted and lexical retrievers almost never return both.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27580
PDF
-
CityPaint: A Free Lunch from 2D Diffusion for City-Scale 3D Texturing
Abstract
Generating high-fidelity, city-scale 3D environments plays a crucial role in applications such as virtual reality, gaming, and urban digital twins. However, the scarcity of real-world city-scale 3D training data forces existing methods to either stitch independently generated assets or learn from synthetic urban corpora, often sacrificing realism and global consistency. In contrast, 2D image data and their corresponding generative models are abundant and encode rich visual priors. To bridge this gap, we propose CityPaint, a novel framework that leverages 2D diffusion models to generate globally consistent city-scale 3D textures without relying on city-scale 3D training data. Specifically, CityPaint integrates geometric priors with the generative power of 2D diffusion models, introducing a robust top-$K$ camera baking strategy for iterative texture projection. To address the artifacts of 2D-to-3D projection, we further propose a decoupled refinement mechanism that separately optimizes terrain and building textures. Extensive experiments show that CityPaint produces globally coherent textures with fine local details and achieves favorable results against strong baselines in both visual quality and cross-view consistency.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27564
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
-
LoRASpace: A Pool-Wide Shared Substrate for Static and Dynamic Multi-LoRA Composition
2026-10-20 00:46:19Thamirawaran Sathiyalogeswaran, Navindu De Silva, Chamikara Siriwardane, W. K. Alwis, Jason Mars, Udaya Sampath K. Perera Miriya Thanthrige, Ranga RodrigoAbstract
Low-rank adaptation (LoRA) attaches small low-rank matrices to a frozen backbone, and public repositories host many LoRAs per backbone, each for a different concept or task. Combining several LoRAs in one generation forces a trade-off. Merging their weights into one update needs a single forward pass, but LoRAs trained for different concepts interfere and quality drops. Running each LoRA in its own forward pass (composite) preserves quality, but inference time grows linearly with the number of active LoRAs. We study a property of LoRA pools overlooked by prior multi-LoRA work. In real LoRA pools, the stacked low-rank factors need far fewer directions than their summed ranks. Using singular value decomposition, we find the minimal subspace that keeps 99% of the factor energy. On three diffusion pools, only 46% to 57% of the summed LoRA ranks is needed, and the subspace grows sub-linearly with pool size. A 48-LoRA sub-pool of LoRA-Hub on FLAN-T5 tasks shows comparable overlap (62%), while the diversity-curated LoRA-Retriever benchmark sits near the no-overlap baseline (96%), so pool composition rather than modality predicts where compression is large. We build on this overlap with LoRASpace, a training-free method that computes the subspace once per pool and reuses it in two inference modes, a static mode at merge cost and a dynamic mode for input-dependent routing. On the diffusion pools, static LoRASpace matches composite quality at merge cost, and dynamic LoRASpace is more than 6x faster than composite with six active LoRAs on a 17-LoRA pool, mainly because it runs one backbone pass instead of six. In an ablation, static variants at the same rank match LoRASpace, so the quality gain does not depend on the pool-wide basis. On the near-orthogonal LoRA-Retriever benchmark, compression is small and gains come from coupling with external retrieval.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27444
PDF
-
DStill: Trajectory Consistency Distillation for Accelerated One-Step Drafting in Block Diffusion Speculative Decoding
2026-10-19 23:05:16Shreyas Kulkarni, Uttam Gurram, Madhav Kanda, Krishna Teja Chitty Venkata, Shichu Chen, Kartik Balasubramaniam, Supriyo Chakraborty, Sambit SahuAbstract
Block-diffusion speculative decoding accelerates large language model inference by drafting multiple tokens in parallel. However, one-step drafting in discrete diffusion is limited by token factorization error: each token is predicted from the same masked block, without conditioning on the tokens selected at other positions in that step. This can produce inconsistent drafts and reduce acceptance length. Iterative refinement mitigates this limitation by conditioning later predictions on previously committed tokens, though additional passes increase drafting cost. To address this trade-off between draft fidelity and inference latency, we introduce DStill, a two-stage framework that distills a target-conditioned drafter's own multi-step refinement into one-step drafting. We first train the drafter to complete partially revealed blocks, then use its refinement trajectories to train a student to predict the final completion in one pass. Using a curriculum that gradually transitions from partial to fully masked inputs, we successfully transfer the multi-step refinement capabilities into one step. Across reasoning, math, code benchmarks on three model families DStill improves acceptance length by up to 28% and speedup by up to 25% over one-step DFlash, achieving up to 6.45x speedup over auto-regressive decoding.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27379
PDF
-
Learning to Be Different: Component Specialization for Compositional Diffusion Policies
2026-10-19 19:29:58Alec Laprevotte, Haonan Chen, Xiaoshen Han, Yilun DuAbstract
Diffusion policies with linearly composed scores have emerged as a promising approach to multitask robotic learning by combining separate factors to enable modular, task-wise adaptation, compositional generalization, and interpretable functional decomposition. However, current training objectives do not encourage factor specialization, often resulting in uninterpretable denoisers whose composed prediction is an average over identical copies, undermining their architectural premise. To address this, we propose a drop-in score-decorrelation auxiliary loss, motivated by projective composition theory. Our approach encourages specialization by penalizing overlap directly between component noise predictions, with little architectural change or computational overhead. We demonstrate across multiple benchmarks that our decorrelation loss consistently reduces measurable double-counting between components, improves task decomposition, and produces behaviorally meaningful specialization without compromising multitask performance. Website: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27194
PDF
