Transformer
-
UniRAP: Towards Unified Part-level Physical Affordance Reasoning and Actionable Perception
2026-10-21 14:02:25Linfei Li, Ruining Hu, Lin Zhang, Zhong Wang, Fengyi Zhang, Ying Shen, Binqiang Wang, Xin ZhangAbstract
Vision-language perception has achieved impressive progress in aligning natural language with visual observations, yet grounding high-level semantics into part-level physical interaction remains challenging. To address this gap, we propose UniRAP, a unified model for inferring part-level physical affordances and mapping language instructions to actionable geometric representations. UniRAP formulates this problem as a conditional multimodal generation task, integrating visual inputs, textual instructions, and optional prompts into a shared spatiotemporal representation space through a unified interface token mechanism. To predict executable contact geometry, we introduce a Unified Affordance Decoder (UAD), which jointly performs object detection, part-level affordance segmentation, and 4-DoF interaction pose estimation by leveraging intermediate segmentation features. In addition, we propose a curriculum-based transfer training strategy that progressively adapts the model from general visual parsing to interaction-aware perception, improving data efficiency under limited textual annotations. Experiments show that UniRAP achieves state-of-the-art performance on referring expression segmentation, affordance grounding, and interaction pose estimation, while maintaining strong spatiotemporal consistency in dynamic video scenarios. These results demonstrate the effectiveness of UniRAP as a unified perception framework for language-guided physical manipulation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29768
PDF
-
ALIAS: Evaluating Performance and Memorization in LLM-Written TLA+ Specifications
2026-10-21 14:02:02Arslan Bisharat, Mujtaba Nazari, Khushboo Bhadauria, Brian Ortiz, Eric Spencer, Beatriz Santos, TaiNing Wang, George K. Thiruvathukal, Konstantin Laufer, Mohammed AbuhamadAbstract
Memorization inflates the benchmark scores of large language models (LLMs), since a model can reproduce answers it saw in training. Researchers test for it with consistent identifier renaming (CIR), which gives every identifier in a task a new name, and treat a lower score on the renamed task as evidence of recall. We test this method on the Temporal Logic of Actions Plus (TLA+) language, whose model checker acts as an oracle by confirming whether a renamed task preserves the exact behaviors of the original. Across 111 TLA+-Bench tasks, renaming lowers the correctness of Claude Opus 4.8 from 28% to 18% and of GPT-5 from 9% to 7%. Most of this decrease comes from how the test grades, which requires exact names that most descriptions never state. Grading by behavior shrinks the decrease to about 3 percentage points for Opus 4.8 and to 0 for GPT-5. In two model families fine-tuned on disjoint halves of the tasks, renaming removes about half of the training gain in likelihood and costs 4.9 percentage points (Qwen) and 1.3 percentage points (Gemma) on unseen tasks. A lower score after renaming therefore mixes memorization, name-based grading, and the cost of new names.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29767
PDF
-
OMAR: Open-World Multi-Attribute Reasoning for Scene-Conditioned Shelf Organization
2026-10-21 13:59:13Benedikt Kreis, Rohit Menon, Benno Wingender, Muhammad Ehsan Ullah, Nils Dengler, Maren BennewitzAbstract
With the advent of foundation models, robots are increasingly capable of understanding and executing explicitly specified human tasks. Extending these capabilities to open-ended tasks, however, requires robots to infer appropriate actions from the observed scene. Household organization exemplifies this challenge, as robots must often act without detailed instructions, predefined object vocabularies, or known object-receptacle associations. While robots may follow instructions or learn organization preferences from demonstrations, inferring the latent organization of an occasionally changing scene and placing a novel object consistently with it remains challenging. We propose an open-world multi-attribute reasoning pipeline that grounds unknown objects across views and constructs an instance-level 3D scene graph for placement reasoning. First, a vision-language model proposes labels together with attributes such as content and packaging type, which are grounded through segmentation, depth, and multi-view association. Then, an LLM reasons over this representation to infer the latent shelf organization and rank scene-consistent object placements without explicitly user-provided preferences. The resulting placement rankings guide subsequent geometric and kinematic feasibility checks for execution. Our experimental evaluation shows that our approach leads to improved instance mapping and placement coherence over single-view language-grounding and geometry-only baselines, enabling autonomous placement of previously unseen objects without predefined object-receptacle mappings or class-specific retraining.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29761
PDF
-
Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies
2026-10-21 13:34:03Johannes Keller, Albrecht M. Blo{\ss}e, Adrian Lindenmeyer, Thomas Neumuth, Daniel SchneiderAbstract
Vision-language-action (VLA) policies need to recognize gaps in their training knowledge to request useful demonstrations or assistance in new situations, especially when failures are costly. Disagreement within an ensemble of independently trained policies can guide these decisions, but requires training and running multiple policies. We propose post-hoc kernel-variance heads that estimate epistemic uncertainty from the embeddings of a single policy. Each head computes the conditional variance of a Gaussian process conditioned on embeddings from the training demonstrations. Fitting requires no regression targets or gradient updates and leaves the policy unchanged. In experiments with SmolVLA on LIBERO-10, the kernel variance heads perform competitively with two-policy ensemble scores in task-level correlation with task success, in demonstration acquisition during active learning, and in referral to a stronger supervisor policy as a proxy for human assistance. For SmolVLA, the head on the vision-language backbone reduces uncertainty-scoring time by up to approximately a factor of 50 compared with ensemble scores, while also substantially reducing setup costs. We also show that the established evaluation protocol on LIBERO-10 primarily rewards recognition of coverage differences between tasks, while execution variability weakens assessment of within-task rankings. Thus, controlled perturbations provide a complementary evaluation: modality-specific heads identify unfamiliar camera, instruction, or state inputs, while fused heads can detect unfamiliar combinations of individually familiar inputs. With a supervisor proxy trained on the evaluated perturbation families, the decoder head improves within-task referral over ensemble disagreement. These findings show that useful uncertainty estimates for adaptive robot learning can be obtained from a single policy.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29730
PDF
-
Reconstruction-Augmented Visuomotor Pre-Training of 3D-Grounded Bimanual Dexterity from RGB-Only Human Videos
2026-10-21 13:28:19Xingyi He, Zhiwen Qiu, Adhitya Polavaram, Yunhao Cao, Cory Fan, Julius Samwer, Saleh Kalantari, Kuan FangAbstract
We present SpatialDex, a framework that learns 3D-grounded bimanual dexterity from RGB-only human videos through reconstruction-augmented visuomotor pre-training. The central idea is to learn human behavior together with its spatial context. From each video, we reconstruct the scene geometry and estimate the hand motion, align them at a shared video-level scale using the hand-size prior of 3D hand estimates. We retain the reconstructed geometry as both a policy input and a prediction target. The policy grounds the token grid of a pre-trained vision-language model in this geometry through gated residual fusion, is trained to predict future scene geometry alongside wrist and finger actions, and generates finger motion conditioned on the predicted trajectories of both wrists. Pre-trained on 1M human-video clips and fine-tuned with a small number of robot demonstrations, SpatialDex outperforms the state-of-the-art methods by 16.5 points on nine simulated dexterous tasks across two embodiments and by 38.3 points on six real-world bimanual dexterous tasks. Under a matched comparison, pre-training with reconstructed geometry improves success by 14.2 points over pre-training on hand motion alone.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29716
PDF
-
FlowVLA: Learning Flow-Aware Physical Completion for Quantitative Robotic Pouring
2026-10-21 13:07:39Liwen Yang, Yutian Zhang, Jiaqi Zhai, Li Yang, Tutian Tang, Qiwei Shan, Yancheng Bi, Qiankuan Shangguan, Yangkun Zhu, Ce Hao, Dong Wei, Qiaojun Yu, Dibo HouAbstract
During steady pouring, images and robot pose may remain nearly unchanged while transferred mass accumulates, making the return decision ambiguous for a whole-task vision-language-action (VLA) policy. We present FlowVLA, which separates motor generation from physical-completion estimation. A shared $\pi_{0.5}$ action expert learns approach, pour, and return with boundary-aware supervision. A target-independent estimator fuses visual, joint, wrist-wrench, and recurrent force-belief histories to maintain an estimate of transferred mass. The controller initiates return when this estimate plus predicted residual outflow reaches the requested mass. In coffee-pot/water experiments, FlowVLA achieves 13.1 g goal-balanced mean absolute error (MacroMAE), lower than all tested VLA baselines. Across all attempts, 53.7% both complete the task and finish within +/-10% of the requested mass. The same model interpolates to held-out 150/250-g targets without retraining. Independently trained models further evaluate the formulation in three additional container-fluid settings.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29692
PDF
-
Spatially Gated Demographic Steering for Debiasing Multi-Person Text-to-Image Generation
2026-10-21 12:32:16Venkatesh Thirugnana Sambandham, Torsten SchonAbstract
Text-to-Image models reproduce societal stereotypes, and existing mitigation applies one demographic correction to the whole scene. In multi-person scene generation this creates a trade-off: global steering improves aggregate balance only by homogenizing the people within an image. Steering-based methods like Embedding Arithmetic drive the intra-image diversity below the unsteered baseline. Text- and latent-projection approaches shift the aggregate skew with only marginal gains in intra-image diversity. We present Spatially Gated Steering (SGS), which steers the demographics of each person independently by extracting per-person masks from the Diffusion Transformer's (DiT) internal attention maps at a critical denoising timestep and applying a distinct demographic vector within each mask for the remaining steps. SGS needs no fine-tuning, inversion or layout specifications, and uses a lightweight segmentor over frozen DiT attention features to localize subjects at early timesteps. On skewed occupational prompts, SGS reaches high intra-image racial and gender diversity while keeping aggregate balance competitive and image quality near baseline, a result we validate on two architecturally distinct diffusion transformers, FLUX.1-Dev and SD3.5-Large. Codebase and model weights are available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.29651
PDF
-
MS2CAST: A Video Anomaly Detection Framework via Multi-Scale Spatial Cross-Attention and Efficient Spatiotemporal Networks
2026-10-21 12:29:24Ali Safaei, Mohammad KazemiAbstract
Video anomaly detection (VAD) is a challenging task for surveillance systems due to the scarcity and diversity of anomalous events. While semi-supervised VAD approaches have advanced the field, many rely on computationally intensive optical flow networks to capture motion dynamics. Convolutional neural networks (CNNs) are also widely used for spatial encoding, but their effective receptive field is highly local. To address these limitations, we introduce MS2CAST, an end-to-end semi-supervised predictive VAD framework. For spatial encoding, we employ a multi-scale cross-attention encoder that fuses the hierarchical feature representations from a vision transformer (MViTv2-S) and a CNN (ConvNeXt-T) to exploit their distinct receptive fields. To model spatiotemporal dynamics without optical flow estimation, we integrate two interchangeable spatiotemporal modules in the latent space. One module (SwinLSTM) incorporates a Swin Transformer, while the other (VMRNN) integrates Vision Mamba blocks. Our objective function combines L1 loss, Gradient Difference Loss (GDL), and Multi-Scale Structural Similarity Index Measure (MS-SSIM) to ensure pixel-wise accuracy, sharpness, and structural integrity. Evaluations across the UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets demonstrate frame-level Receiver Operating Characteristic (ROC) Area Under the Curve (AUC) scores that are competitive with top-tier techniques while maintaining a peak GPU memory usage of under 1 GB.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29649
PDF
-
FogTTA: Online Test-Time Adaptation for Robust Transformer-based Object Detection in Foggy Weather
2026-10-21 11:57:36Ali Abedi, Q. M. Jonathan Wu, Ning Zhang, Shiqi TianAbstract
Object detection models for autonomous driving commonly experience substantial performance drops when deployed under adverse weather due to the domain shift between training data and real-world operating conditions. This degradation is especially evident when models trained on clear-weather images encounter foggy environments with reduced visibility and contrast. To address this challenge, we introduce FogTTA, an online test-time adaptation framework designed to improve the robustness of Transformer-based object detectors in fog. Using RF-DETR as the underlying object detector, FogTTA enables real-time adaptation to the streaming target domain without requiring source data or retraining. The framework follows a teacher-student design, where the deployed model serves as the teacher and generates pseudo labels from weakly augmented target inputs. These predictions are subsequently refined through non-maximum suppression and confidence filtering. The student model then learns from strongly augmented target sample using the Varifocal loss to mitigate pseudo-label noise. The teacher is updated via exponential moving averaging to ensure stable and continuous adaptation. Experiments show that FogTTA outperforms prior baselines, delivering improved detection accuracy and stability while maintaining real-time performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29618
PDF
-
A Diagnostic Study of Faithfulness in Diffusion Language Models for Table-to-Text Generation
2026-10-21 11:42:19Zaynab Merimi, Noussayba El Marrakchi, Arthur SatoufAbstract
Diffusion language models generate by iteratively denoising a masked sequence under bidirectional attention, without the strict left-to-right commitment of autoregressive decoding. Whether this difference affects how reliably they remain grounded in structured inputs is largely unexplored. We study this question in table-to-text generation using ToTTo, where highlighted source cells make factual grounding directly testable. We compare the masked diffusion models MDLM and LLaDA with GPT-2 and GPT-Neo under an identical retrieval-based few-shot protocol, together with a QLoRA-adapted LLaDA, and we score reference similarity and source grounding as separate axes. The two pull apart. Within LLaDA, additional demonstrations raise BLEU from 11.3 to 22.0 while NLI entailment against the source table falls from 0.78 to 0.52; QLoRA reaches the best BLEU of the study, 24.15, and grounds no better for it. Manual annotation shows the failure modes differ in kind: the autoregressive baselines tend to wander off the table, while LLaDA-based models more often stay on topic and swap out individual entities. Reference-based metrics alone are insufficient to assess faithfulness in table-to-text generation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29602
PDF
-
CineEvolve: An Agentic Self-Evolution Framework for Cinematic Multi-Shot Video Generation
2026-10-21 11:30:09Jingzhe Zhang, Peixuan Zhang, Genying Li, Bing Ma, Shuchen Weng, Si Li, Boxin ShiAbstract
Cinematic multi-shot video generation requires both narrative planning and iterative refinement. Existing methods either lack narrative reasoning or rely on feed-forward workflows without policy updates. In this paper, we propose CineEvolve, an agentic self-evolution framework for cinematic multi-shot video generation that unifies a Director and a Critic within a single vision-language policy. Emulating real-world filmmaking that couples iterative direction with critical feedback, we train CineEvolve via alternating reinforcement learning, where the Director surfaces flaws to sharpen the Critic, and the updated Critic guides the Director's targeted revisions. To stabilize co-evolution, we introduce Evolutionary GRPO (E-GRPO), leveraging historical anchoring and revision safeguarding to mitigate advantage inflation. Additionally, we contribute CineTraj, a 40K cinematic trajectory dataset spanning creation, critique, and revision. Extensive experiments demonstrate that CineEvolve outperforms existing approaches. Our code will be available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29590
PDF
-
XGLASS: X-Ray, Gaze and Language Assisted Slice Sequencing for CT Volumes
2026-10-21 11:27:48Michael Gl\"ockler, Timo RopinskiAbstract
Vision-language models (VLMs) achieve remarkable performance on 2D medical images, such as X-rays or pathology images, while 3D VLMs that directly process CT/MRI volumes lag behind due to limited 3D training data and the predominance of 2D foundation architectures. To leverage powerful 2D VLMs for analyzing medical volumes, which contain hundreds of slices, it is therefore necessary to focus only on those few slices, which are diagnostically relevant. To enable such a prioritization, we propose XGLASS, a method that transfers medical expertise from 2D X-rays to 3D CT for task-aware slice prioritization. XGLASS uses disease-conditioned importance maps derived from language-aligned radiologist gaze on X-rays and spatially aligns them with coronal CT projections, bringing gaze priors and CT content into a shared 2D representation. The fused gaze-prior/CT projection is then used to predict a disease-specific ordering of axial slices, prioritizing diagnostically relevant slices. XGLASS significantly improves budgeted retrieval over 2D and 3D-native baselines, achieving a top-10 hit rate up to 75.9% for the best-performing disease. In practice, this means that after inspecting only the top 10 slices (4.17% of a 240-slice scan), XGLASS identifies a diagnostically relevant slice in more than 75% of cases. For downstream slice-wise finding classification, an XGLASS-selected 30-slice candidate set reduces model calls by 87.5% while improving F1 by 11.3 percentage points over all-slice inference, 4.2 points over the strongest 2D selector, and remains competitive with 3D-native selectors while requiring up to 130.8 times less time and 10.9 times less GPU memory per finding, supporting its use as a lightweight preselector for volumetric CT analysis. The code and pretrained checkpoint are available on this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26255
PDF
-
Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
2026-10-21 11:22:58Martin Thi{\ss}en, Elke Hergenr\"otherAbstract
Despite the success of foundation models in image synthesis, achieving fine-grained spatial control remains a significant challenge, typically requiring auxiliary structural adapters or complex inference-time steering. While modern vision-language (VLM) text encoders possess the latent capacity to represent spatial coordinates, generative models often fail to leverage this information. We demonstrate that the signal-to-noise ratio (SNR) distribution is a critical determinant of spatial grounding in flow-matching models. Using a novel synthetic benchmark of geometric primitives, we isolate structural mechanics from textural complexity inherent in natural imagery and introduce a rigorous color-aware layout metric to quantify spatial fidelity. We show that training on a centered SNR distribution results in poor layout adherence (25.9%). Instead, we identify high-noise regimes as a "critical learning phase" for spatial grounding. Shifting the training focus toward these stages increases layout adherence to 80.8%, effectively reaching the 83.6% reconstruction ceiling of the underlying variational autoencoder. We find that the necessity for this shift scales with resolution and aspect ratio variance, and that the conservative noise schedules employed by models like FLUX.1 are insufficient for precise layout adherence. Finally, we show that these findings translate to real-world imagery via the COCO dataset (1.0 vs. 57.3 Average Precision), where state-of-the-art layout adherence is unlocked through distribution tuning alone, without architectural modifications.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29581
PDF
-
From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation
2026-10-21 10:49:49Tingwen Du, Haoyu Guo, Hanxue Zhang, Xuanxu Wang, Junting Dong, Chao Gao, Mingfei Han, Tao Lu, Mulin Yu, Chunhua Shen, Xiaojun Chang, Zhihui LiAbstract
Vision-language-action (VLA) models achieve strong performance across embodied manipulation tasks by unifying visual understanding, language conditioning, and action generation. However, long-horizon tasks can involve non-Markovian dependencies: successful decisions may require retaining task progress and past interaction outcomes, while partial observability and execution errors compound over time. Existing work has begun to address these challenges by augmenting foundation-model policies with memory and by introducing hierarchical architectures that separate high-level planning from low-level execution. We organize the high-level VLM as a robotic agent through an editable harness to maintain task context and improve planning through execution experience. We instantiate this approach as SERA (Self-Evolving Robotic Agent), with a fixed VLA executor providing low-level control. The harness maintains and updates task context using observations and execution feedback to guide subsequent planning. Self-evolution uses environment rollouts to propose and evaluate harness revisions, with the selected configuration fixed before held-out testing. On RoboMME, SERA achieves 51.38% overall success with autonomous selection, compared with 44.38% without self-evolution. Our exploration suggests that agentic planning is a promising direction for history-dependent robotic manipulation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29560
PDF
-
SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model
2026-10-21 10:06:03Kai Tang, Peidong Jia, Zhong Chu, Jiajun Cao, Jixian Wu, Bohua Zhang, Rui Ma, Fangyuan Zhao, Sixiang Chen, Yichen Guo, Xiaowei Chi, Chun-Kai Fan, Kevin Zhang, Jinchang Xu, Fubing Yang, Weishi Mi, Xiaozhu Ju, Jian Tang, Shanghang ZhangAbstract
Safe control is a prerequisite for real-world embodied intelligence, for which safe reinforcement learning has emerged as a promising paradigm. However, existing safe reinforcement learning methods either require costly real-world exploration or depend on hand-crafted safety functions. Neither scales to vision-language-action models deployed in open-world physical environments. We propose SafeDojo, a model-based safe reinforcement learning framework that improves vision-language-action policies through world-model imagination. SafeDojo performs online reinforcement learning on top of an interactive video world model. A ResNet success classifier scores predicted future frames, while a lightweight safety head predicts per-step contact probabilities from the current latent context and proposed action chunk, jointly assessing task execution and trajectory safety. These decoupled signals are balanced through Lagrangian-constrained GRPO. On SafeLIBERO Level I, SafeDojo improves average safe-success rate by 8.25 percentage points over the strongest baseline; on the zero-shot Level II setting, it is competitive with the strongest model-free baseline. Real-world Franka deployment further shows the best average task and safe-success rates across five tasks. These results support world-model-based safe reinforcement learning as a promising path toward safer embodied intelligence.
Abstract (translated)
URL
https://arxiv.org/abs/2606.20698
PDF
-
Balancing Foundation Models and Trait-Specific Learning for Forehead-Crease Recognition
2026-10-21 09:50:08Geetanjali Sharma, Aryan Jain, Dev Nagaich, Aditya Nigam, Raghavendra RamachandraAbstract
Biometric recognition often degrades under partial facial vis- ibility, where conventional facial cues are occluded. The forehead region remains consistently visible and contains distinctive crease patterns that can serve as a complementary biometric trait. We present a unified eval- uation of forehead-crease recognition across explainable zero-shot rea- soning, domain-adaptive learning, and trait-specific modeling. FIRE en- ables interpretable zero-shot verification using multimodal large language models, while pretrained face-recognition models are adapted to the fore- head domain through margin-based learning. Under open-set evaluation, the best zero-shot model GPT-4o+FIRE achieves 18.22% EER, while Edge-Forehead achieves 8.30% EER and 54.01% TMR@0.1% FMR. The results show that although MLLMs offer interpretable zero-shot verifi- cation, trait-specific learning remains more effective for robust forehead- crease biometric recognition.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29495
PDF
-
Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers
2026-10-21 09:01:48Tianze Xia, Xiaoling Zhou, Ruya Lan, Yiqiao Liao, Chengru Song, Shiming XiangAbstract
Full 3D attention is costly in video Diffusion Transformers. Block-sparse attention reduces this cost by computing only selected attention blocks. However, existing methods either overlook the potential to optimize sparsity allocation or rely on heuristic rules or local metrics to guide it. In this work, we observe that sparsifying attention in different layers and at different denoising steps affects final-video fidelity to substantially different degrees. Yet the identities of the most sensitive layer-step pairs remain nearly unchanged across inputs. Building on these findings, we propose Allocating Layer-step Latency by Output Tests (ALLOT), a training-free sparsity allocation method. Treating each layer-step pair as a cell, ALLOT estimates each cell's contribution to final-video error through (i) cell probing and (ii) trajectory calibration, then jointly optimizes cell densities under a measured-latency budget. It preserves existing selection strategies and kernels, replacing only the allocation component. At equal or shorter wall-clock time, ALLOT outperforms every compared allocation strategy with XAttention, with an 11.8 dB PSNR gain on HunyuanVideo. Extensive experiments across four additional published pipelines on Wan2.1, Wan2.2, and HunyuanVideo show that ALLOT outperforms or ties the default in all nine settings.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29441
PDF
-
Don't Stop Long-Context Continual Pre-Training Too Early: An Empirical Scaling Law
2026-10-21 07:46:55Ziqing Qiao, Xiaoyue Xu, Sikui Zhang, Yinuo Xu, Bingxiang He, Chaojun Xiao, Xu HanAbstract
Long-context continual pre-training (LCCP) is widely used to improve long-context capability, but how its benefits scale with the number of tokens used for LCCP remains unclear. We investigate this question through modeling, monitoring, and validating the benefits of scaling LCCP. First, we establish an empirical additive power law relating long-context performance to model size and the respective token budgets of short-context pre-training and LCCP. Experiments across five model scales and more than one hundred short- and long-context budget configurations show that the fitted law reproduces the observed trends and extrapolates well to held-out model scales. Second, we introduce \textsc{Constant-Last4}, a training and observation protocol that averages the four most recent consecutive checkpoints from a constant-learning-rate run to enable stable monitoring along a single extensible training trajectory. Under this protocol, the same scaling-law form fits well and accurately predicts the LCCP performance trajectory of a 1.22B-parameter model trained on approximately 100B LCCP tokens. Finally, downstream evaluations along this trajectory show continued improvements in long-context performance, with no clear sign of saturation within the evaluated range. These gains persist after both 128K context extension and supervised fine-tuning, with benefits extending to long-context understanding and agentic tasks. Together, these results argue for continuing to scale LCCP rather than stopping early: its gains are stably observable along an extensible trajectory, quantitatively predictable, and transferable to downstream tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29352
PDF
-
Reader--Writer Coupling in Transformer MLPs
Abstract
A transformer MLP gives each hidden channel two roles: a reader that decides when the channel responds, and a writer that decides what it adds back to the residual stream. Read as a key--value memory, the layer can be measured one factor set at a time: the pairing of readers to writers is carried only by their shared index, invisible to any quantity computed from one factor set alone. Yet training couples the two factors through their joint update, so we ask what reading them together shows. The block supplies the joint instrument: the MLP writes back into the very stream it reads, so the exact Jacobian of its update is square, and its agreement with its own transpose resolves into reader--writer terms whose inner products give each channel, read as a couple, a signed pairing coordinate. Ranked by this coordinate, each block we measure has a small core of strongly coupled channels; where initialization gives either sign equally, the core is nearly all opposed: each writer points against the direction its reader responds to. This opposed core recurs from the smallest models trained from scratch to released checkpoints; its sign is in place within a few percent of training and holds while the coupling grows. Read as couples, most channels return their contribution against the reader that fired it, and the readers of later blocks take it up; few show in the logits. Training holds the couples in place: flipping the opposed writers moves the output toward common tokens, and a writer permuted onto another reader regrows to fit that reader while the reader stays.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29340
PDF
-
Plan First, Act Later: Streaming Reasoning for Executable Memory in Robot Control
2026-10-21 07:27:58Qinghang Liu, Conglin Wang, Zongsheng Li, Boyong He, Han Gao, Pengcheng Wang, Tao Sun, Lige Liu, Chang LiuAbstract
Long-horizon robot manipulation requires a robot to remember task-relevant information beyond its short observation window. Existing memory-augmented policies typically retrieve relevant information from history and reason over it at execution time, placing a heavy reasoning burden on the high-level model. To address this challenge, we propose PAL (\textbf{P}lan first, \textbf{A}ct \textbf{L}ater), a hierarchical memory framework for vision-language robot manipulation. It maintains a short-term visual buffer, a semantic-event memory, and an executable \texttt{subtask\_list}. A high-level vision-language model (VLM) converts streaming visual observations into the \texttt{subtask\_list}. At execution time, the VLM selects the currently relevant subtask from the \texttt{subtask\_list} and passes it to a low-level vision-language-action (VLA) policy. To teach the model what to write to memory and when to update it, we introduce a two-stage data generation procedure that first expands sparse update supervision and then uses self-labeling with task-grounded verification to generate training data that better match the state distribution encountered during execution. We evaluate our approach on memory-dependent benchmarks and real-world manipulation tasks. On RoboMME, PAL achieves a \textbf{56.05\%} success rate, outperforming the strongest baseline, FrameSamp-Modul, at 44.51\%. In real-world experiments across four manipulation tasks, \pal{} reaches \textbf{60.83\%} overall success, surpassing MemER (45.00\%) and FrameSamp-Modul (28.33\%). These results show that proactively writing future-relevant subtasks into hierarchical memory improves long-horizon visual reasoning and robot control. Code and videos can be found on our website: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.29329
PDF
