Attention
-
Judging the Judges: Confidence and Disagreement in Rubrics as Rewards
2026-10-15 03:35:00Michael Tian-Yue Liu, Anne Harrington, Trevor Darrell, Yutong BaiAbstract
Rubrics as Rewards (RaR) extends reinforcement learning to non-verifiable domains by using an LLM judge to score responses against a structured rubric. Prior work has focused on improving the rubric or strengthening the single judge, with comparatively little attention to how different single or multiple judges might be used during RaR training. We first show that judge identity alone produces a substantial spread in scores when training Qwen3-4B-Base on RaR-Medicine and RaR-Science across five frontier and open-weight judges. Building on this, we propose RaRE (Rubrics as Rewards with Ensembles), a novel judge ensemble strategy for RaR with two mechanisms: a soft-Likert reward that uses judge logprobs to weight more confident judges more heavily, and a drop-on-disagreement mask that excludes high-disagreement rollouts from the GRPO update. We find that this design is critical: while a naive average of judge scores fails to outperform the best single judge, RaRE outperforms every single-judge RaR variant on both HealthBench and GPQA Diamond, with per-theme analysis showing that it inherits the complementary strengths of its constituent judges. Our findings establish judge ensembling as a compelling axis for improving RaR.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21179
PDF
-
Language Models Share a Readout Bottleneck in Working Memory
Abstract
Large language models (LLMs) reproduce a human working-memory signature on the N-back task: performance declines as memory load N increases, and errors concentrate on lures, items that repeat the current stimulus at a neighboring but incorrect serial position. We ask whether this shared behavioral signature reflects a shared mechanism. The literature offers two candidate accounts. On a precision account, softmax attention cannot concentrate its weight on the item N positions back, so adjacent items are retrieved instead. On an interference account, attention selects the correct item, but the content of competing items corrupts the decision. We dissociate these accounts in 15 pretrained LLMs from three families (70 million to 14 billion parameters) and in minimal transformers trained from scratch. In all three families, the attention heads that carry the task attend to the target on most missed matches. Linear probes decode near-perfectly whether a matching letter occurred at the target or at a lure position, even on trials where the model false-alarms to the lure, although its own decisions barely distinguish the two. Sharpening attention yields negligible gains, whereas suppressing the letter identity of non-target items improves sensitivity (d') in nearly all model-by-load conditions. In larger models the task is spread over more heads, and removing interference helps more. In minimal transformers, precision limits a single attention layer, but a second layer removes this constraint. The working-memory limit of LLMs therefore lies in reading items out under interference rather than in locating them, consistent with interference accounts of human working memory.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21176
PDF
-
GSM: Efficient Language Modeling with Shared Global State
2026-10-15 03:28:02Yunao Zheng, Bin Wen, Xiaojie Wang, Kaiyu Jiang, Xuanyu Zheng, Changyi Liu, Hongyi Fu, Jianxiong Wang, Tianke Zhang, Haonan Fan, Yingxin Li, Jiankang Chen, Xu Wang, Tingting Gao, Han LiAbstract
Improving language model efficiency requires not only cheaper operations but also eliminating redundant computation. We argue that effective modeling requires information accessibility and computational depth, but not repeated historical retrieval across layers. Guided by this principle, we introduce the Global State Model (GSM), a causal encoder--decoder architecture that separates historical information acquisition from deep computation. The encoder progressively aggregates long-range information through multi-stage sparse retrieval into a fixed-size shared state, while decoder layers perform successive transformations over this state without repeated historical access. This design makes decoder attention and cache costs independent of history length, while dependency-aware pruning further reduces prefill computation. Experiments across multiple model scales show that GSM achieves downstream performance comparable to matched DeepSeek-V4.1-Flash-based baselines, with improvements at some scales, while preserving long-context retrieval capabilities. GSM also improves training and inference throughput and reduces cache overhead. These results demonstrate that eliminating architectural redundancy offers an effective path toward efficient language modeling without sacrificing information accessibility or computational depth.
Abstract (translated)
URL
https://arxiv.org/abs/2609.33465
PDF
-
Automatic Sentence Completion Using Bengali Audio Speech with a Transformer-Based Early Speech Recognition System
2026-10-15 03:24:24Md Asif Akram, Arif Faisal, Sakir Hossain Faruque, Sharun Akter KhushbuAbstract
Dementia severely impairs communication abilities and turns everyday conversations into challenging experiences. Individuals affected by dementia frequently struggle to express their thoughts or complete sentences, making it difficult for family members and caregivers to understand their needs. This study proposes a novel approach to assist dementia patients by developing an integrated system that detects signs of dementia through Bangla speech and text analysis and provides sentence auto-completion support during conversations. To achieve this, recorded voices from individuals with and without dementia were analysed, and key acoustic features were extracted to capture vocal patterns associated with cognitive decline. A transformer-based model was employed to transcribe voice signals into text for linguistic analysis. In the classification task, several Machine Learning (ML) and Deep Learning (DL) models were evaluated using both acoustic and textual representations, and the LSTM model architectures demonstrated strong capabilities for modelling contextual patterns in speech and language data. For the sentence auto-completion task, a sequential learning model based on a Bidirectional LSTM (Bi-LSTM) network with an attention mechanism was developed to sequentially complete partially expressed Bangla sentences during conversation. This research bridges the gap in dementia analysis in an underrepresented language and implements a sentence auto-completion framework for cognitive disorders.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21164
PDF
-
VoxTurn: Answering Several Turn-Taking Questions in One Pass with One Model
2026-10-15 03:06:58Yang XiaoAbstract
A cascaded voice agent must decide, throughout each silence of its user, whether the user has finished their turn: answering too early cuts the user off, answering too late makes the agent slow. It often needs related decisions too: will the user soon go on, is speech during its own turn an interruption or a backchannel, is a sound speech at all. With open detectors, each extra decision costs another model or another call, and on the 14-language eot-bench they cut users off in more than a quarter of mid-turn pauses. Because they are trained on segmented utterances, this mismatch is the largest single source of error. We present VoxTurn, a 0.8B speech LLM that casts these decisions as multiple-choice questions, scores the hidden state of each option, and answers any set of them in one forward pass with a block-isolating attention mask, exactly as if asked one at a time. Trained on real conversational pauses with labels derived from their timing, and re-queried while a silence lasts, VoxTurn reaches 11.5% false cut-offs in English at a 300 ms latency budget, comparable to two commercial detectors, and a 14-language mean of 19.9%, the lowest among open models, at 8.3 ms per decision. In Pipecat and LiveKit it cuts off 17% and 9% of turns, and it ranks fifth of 24 in TurnBench end-of-turn detection without any TurnBench training data. The model, code and framework plugins are released.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21150
PDF
-
Sparse Attention for Video Generation Acceleration via Growing Sparsity and Reduced Search
2026-10-15 02:37:51Yun Dai, Yuyun Chen, Jiarui Wen, Huiping Zhuang, Cen Chen, Ziqian ZengAbstract
Diffusion Transformers (DiTs) dominate video generation but suffer from high inference cost due to the quadratic complexity of self-attention. Sparse attention mitigates this by skipping unnecessary computation in the attention matrix, but existing methods apply a fixed sparsity ratio throughout denoising and struggle to accurately determine which computations to skip. To address these limitations, we analyze attention dynamics through attention entropy and mean attention distance, and find that attention becomes increasingly concentrated across denoising steps. We further observe that blocks selected early continue to capture a large fraction of the total attention weight at later steps. These findings suggest two acceleration opportunities: the sparsity ratio should grow as denoising progresses, and earlier search results can be reused to reduce the search space. Based on these findings, we propose Growing Sparsity and Reduced Search (GSRS), a training-free framework that retains more computation early and progressively reduces it later, while restricting subsequent searches to an early candidate set to reduce search overhead. GSRS adopts Top-p selection for adaptive sparsity control, supported by a precise evaluation scheme that computes exact attention weights within the FlashAttention computation at negligible overhead. Experiments on HunyuanVideo and Wan 2.1 validate that GSRS achieves 1.55$\times$--2.21$\times$ speedup while maintaining generation quality comparable to full attention.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21129
PDF
-
DST-Net: Decomposed Spatio-Temporal Network for Time Series Forecasting
2026-10-15 02:28:22Yaohua Zha, Hai Wu, Shu-Tao XiaAbstract
Series decomposition has emerged as a widely acknowledged paradigm in deep time series forecasting models. However, existing approaches predominantly employ a homogeneous modeling strategy for all subsequences, thereby overlooking the distinct inductive biases inherent in different components. We identify that the trend component exhibits dominant spatial coupling across channels but carries significant temporal redundancy, whereas the seasonal component is characterized by local temporal evolution but remains susceptible to attention distraction. To address this, we propose a novel disentangled architecture that tailors specific modeling strategies to align with the intrinsic properties of each component. For the trend branch, we design a Latent Graph Block, employing latent graph modeling to capture global topological dependencies and system-wide co-movements. Conversely, for the seasonal branch, we introduce a Gaussian Multi-scale Attention mechanism to mitigate attention distraction. By injecting learnable Gaussian priors, we impose structural constraints that compel the model to focus on multi-granular local contexts, effectively filtering out global noise and accurately capturing local evolutionary phases. Extensive experiments on multiple real-world benchmarks demonstrate the effectiveness of our proposed model.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21123
PDF
-
Recursive Reward Improvement via Adversarial Meta-Prompt Evolution
2026-10-15 00:31:18Wei Dai, Qianchu Liu, Guanghui Qin, Xing Han, Tristan Naumann, Paul Pu Liang, Sheng ZhangAbstract
Recursive Self-Improvement (RSI) has gained wide attention for its ability to work with limited data. In a proposer-solver loop, a proposer model writes training tasks and a solver model learns to solve them through reinforcement learning. Such loops have succeeded in mathematics and coding, where every answer can be checked automatically. Open-ended expert tasks, like clinical report generations, have no answer key, which makes simple question-answer generation not viable. Here the proposer must also write a rubric that lists the content an answer should include and the errors it should avoid. However, the unique challenge behind model-written rubric is the comprehensiveness of it. For naively generated rubric, solvers tend to exploit it, naming the expected topics without the substance behind them, which corrupts the learning process. To solve this problem, we introduce RRI, Recursive Reward Improvement, which uses a reward hacking adversary to supervise the reward. Before the solver trains on a generated task, an adversary writes an answer that games its rubric while being as useless as possible, the same model writes an honest answer without seeing the rubric, and up to two repair steps revise the rubric to make the honest answer win. The exploits found also rewrite how future tasks and rubrics are generated. On HealthBench Professional, RRI improves the adjusted score over the Generation Only self-improvement baseline by 10.9 points at 9B and 12.4 at 27B, and by 10.3 points without any external model. In particular, the gains are largest on difficult and writing tasks. We show that the 27B model trained on the HealthBench Professional description improves by 6.6 points on PRBench Hard and 8.0 on ProfBench without seeing a task from either. We further analyze the exploitable points of model-generated question-rubric pairs to facilitate future research on open-ended self-play.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21040
PDF
-
Decision-FLIP: Budget-Free Structured Pruning for Gaussian Occupancy Models
Abstract
Occupancy predictors rely on large persistent banks of Gaussians or learned queries whose size is fixed by hand. Common pruning proxies, such as opacity, activation frequency, or parameter magnitude, do not measure whether removing a row changes the model's voxel decisions. We introduce Decision-FLIP, a budget-free compilation framework whose scoring and structure selection need only unlabeled calibration inputs. Exact Mean-FLIP scores each persistent row by the number of voxel argmax decisions that change when the row is removed from the readout; a guarded Otsu split of the log-score distribution selects the survivor count without a target and abstains when the split is unsupported; and the survivors are physically compiled into a smaller checkpoint, so camera sampling, attention, and aggregation shrink with the bank. For streaming banks, scores are integrated over each row's lifecycle through warp, death, and refill. Across four Gaussian-bank model-dataset settings and one persistent-query transfer, the automatically selected cuts remove 12-79% of eligible rows and cut bank/query-dependent latency by 7-45% across settings with component profiles, with whole-model savings of up to 8.0% on the primary endpoints bounded by fixed-cost image, voxel, and sensor modules and peak-memory savings of up to 16.6%. Under matched one-epoch recovery controls, semantic mIoU changes by at most 0.18 point. Same-$K$ controls show that the ranking, not only the bank size, matters: random removal collapses GaussianFormer-1 and OccFormer, while label-free Mean-FLIP matches or outperforms supervised Taylor ranking on all five settings where that control was run. The method applies when removable primitives keep a stable checkpoint-row identity; it is budget-free in survivor count, not parameter-free.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20917
PDF
-
DefoliationCAM: An Edge-AI System for Real-Time Insect Defoliation Severity Assessment in Grapevines
2026-10-14 20:12:17Fuchen Li, Wei-Zhen Liang, Pin-Chu Lai, Xin Qiao, Jingqiu Chen, Haw Yen, Zonglin Yanga, Chi-En Chiang, Rudolph Avalos, Abbigail L. RathAbstract
Insect-induced defoliation threatens grapevine health and yield, yet severity assessment relies on subjective manual scouting. We developed DefoliationCAM (D-CAM), a solar-powered edge AI system that estimates canopy defoliation and transmits results via LoRa to a web dashboard. The pipeline uses YOLOX-Nano to localize grapevine cordons in 4K images and a MobileNet-based classifier with lightweight channel attention to classify 320 x 320 pixel patches. The 29,674 field-collected patches were labeled as background or one of four damage levels: no to negligible, mild, moderate, and severe. Merging moderate and severe classes reduced the class imbalance ratio from approximately 40:1 to 8.8:1. Synthetic augmentation increased macro-F1 from 0.768 to 0.838. The classifier outperformed ResNet-50 in macro-F1 with fewer parameters (1.08 M versus 23.52 M). Patch predictions were aggregated into a canopy-level Non-linear Global Defoliation Severity (NGDS) score and validated against an entomologist's ratings of 86 held-out canopy images. Quadratic discriminant analysis identified boundaries of 9% and 29% separating negligible damage, mild damage requiring monitoring, and action-level damage warranting on-site scouting and insecticide treatment consideration. The 29% boundary aligned with the 25%-30% action threshold range for Japanese beetle management in grapevine. Using TensorRT FP16 on an NVIDIA Jetson Orin Nano, the pipeline processed each 4K image in approximately 1.5 s, with a complete operating cycle of 57-60 s. A five-week deployment of two field units demonstrated the feasibility of autonomous solar-powered operation, LoRa communication, and remote defoliation monitoring. D-CAM supports frequent, objective canopy damage assessment for targeted scouting and timely pest-management decisions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20879
PDF
-
CogDrift-4D: Hierarchical Multimodal Modeling of Temporal Behavioral Decay for MCI Detection
2026-10-14 20:08:01Sankalp Vaish, Sumukh Veeramalla, Rishank Singh, Caroline Summerour, John A. Batsis, Xiaohui LiangAbstract
Early detection of mild cognitive impairment remains a critical bottleneck in dementia care, demanding scalable tools capable of capturing subtle behavioral decay. To capture the subtle temporal changes of cognitive decline, we introduce a behavioral dataset collected across seven unscripted multimodal interactions over an 18-month horizon. This dataset formalizes continuous, multi-session interactions into a dense 4D tensor (Batch * Session * Time * Feature) that explicitly preserves concurrent local dynamics alongside long-term changes. To effectively process these complex behavioral patterns, we propose CogDrift-4D: a novel, hierarchical 4D Transformer framework. To capture the complex trajectory of cognitive decline, CogDrift-4D implements a dual-level behavioral fatigue and drift modeling architecture. At the intra-session level, the model computes localized least-squares regression slopes to quantify the acute ``fading'' of frame-level facial action units and acoustic energy. At the inter-session level, a recursive cross-attention feedback encoder acts as a cumulative fatigue accumulator, mapping the patient's continuous behavioral drift across successive sessions. To mitigate environmental noise and track failures across sessions, a localized Quality Gate dynamically filters low-integrity features before final fusion, also an early fusion that integrates semantic insights with raw physical dynamics. Evaluated on structured verbal and linguistic interactions, our architecture successfully detects MCI with a classification accuracy of 0.733 +- 0.13 and an F1-score of 0.716 +- 0.14, outperforming advanced baselines. The framework provides a mathematically rigorous, scalable, and robust foundation for home-based, real-time dementia screening.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20877
PDF
-
Probing Negation Understanding in Vision-Language Models With Object Attention
2026-10-14 19:41:18Beth Pearson, Michael Wray, Martha LewisAbstract
Negation is a fundamental aspect of language understanding and it is important vision-language models (VLMs) are able to process it effectively. We investigate how VLMs process negation by analysing attention patterns of different models. We draw comparisons between human gaze and attention in models when processing affirmative vs negated statements. We compare object level model attention to human eye-fixations in both binary and multary settings, evaluating with both synthetic and realistic datasets. We find models generally do not follow human-like attention patterns when resolving negation, suggesting they rely on mechanisms outside of attention. While layerwise analysis shows that PaliGemma2-3B shifts object attention to the alternative under negation, our headwise ablations suggest this is a byproduct rather than a negation processing mechanism. Additionally, we find that LLaVA-1.5-7B displays a spatial bias towards left-side objects. Our evaluation of models on downstream tasks shows that all models have a discrepancy between affirmative and negated questions highlighting the need to better understand the underlying processing of negation to close this performance gap. Code available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.20852
PDF
-
Temporal Specification Guidance for Text-to-Video Generation
2026-10-14 19:09:33Avishree Khare, Hideki Okamoto, Bardh Hoxha, Georgios Fainekos, Rajeev AlurAbstract
Text-to-video models often generate the visual concepts requested in natural language prompts but fail to satisfy their intended ordering, timing, and persistence requirements. We propose representing these requirements as specifications in a formal Temporal Logic, and introduce Temporal Specification Guidance (TSG), a training-free inference-time framework that turns temporal specifications into executable objectives for guiding video generation during sampling. TSG scores traces of visual concept evidence using differentiable quantitative semantics and uses the score's gradients to construct a specification-guided denoising direction. We instantiate this framework with video-text attention traces, efficient differentiable scores for Metric Temporal Logic, and gradient-derived attention biases. This instantiation requires neither an external concept classifier nor backpropagation through the video transformer. Its attention-steered forward pass yields a denoising direction that augments classifier-free guidance (CFG). On a benchmark of 220 prompts spanning 11 categories of temporal constraints, including ordered object entry and state transitions, this instantiation improves specification satisfaction accuracy over CFG by 9.5-20 percentage points across five models, with minimal memory overhead. These results demonstrate the potential of temporal specifications and differentiable satisfaction objectives as a flexible interface for controlling video generation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20822
PDF
-
Verdict: Encode Once, Decide Many
Abstract
Automating organizational work means answering many small typed questions about one piece of state (does this email need a reply, how urgent is it, who owns it) with probabilities good enough to act on without a human. The open typed-decision model we compare against, like a standard cross-encoder, answers each question with a separate pass over the input. We present Verdict-MQ, an open encoder that reads a case once and answers all of its questions: given identical retained state tokens after truncation (verified on every reported case), a structured attention mask makes the state encoding independent of the questions, keeps questions from influencing each other, and makes option order irrelevant by construction. On the typed-decisions benchmark, a 395M Verdict-MQ matches the soft gold labels significantly better than the released 421M Laya specialist on all five fresh seeds (KL 0.080 vs 0.117, Brier 0.045 vs 0.061), an advantage that survives per-system temperature scaling; its higher accuracy (0.801 vs 0.767) lies past the benchmark's own saturation point, so we read that margin as closer agreement with the benchmark's labelling teacher. On RACE, BoolQ and MultiRC, against a matched fine-tuned baseline it wins on MultiRC and loses on RACE and on BoolQ, where its training is unstable; we measure zero-shot references ourselves, including a closed general system (Jev) that beats every fine-tuned encoder. A controlled comparison measures what the shared state costs: giving each question its own question-aware copy of the state gains 0.9-3.5 points on RACE and MultiRC at 1.4-2.5x the per-case latency, and nothing measurable on typed decisions, with the invariance guarantees unchanged.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20778
PDF
-
Optimal Hydra: Model Heterogeneous Federated Learning for LLMs using Data-Informed Hydra-Nets
2026-10-14 18:08:34Gwen Legate, Rahaf Aljundi, Eugene BelilovskyAbstract
Large language models (LLMs) trained on vast amounts of data have demonstrated impressive performance over a range of tasks, but future performance gains will require increasing amounts of data and compute. Federated Learning (FL) facilitates training on decentralized private data, but client resource limitations can make training impractical, particularly as model sizes continue to expand. These resource constraints make the assumption of a shared server-client architecture impractical. Model heterogeneous FL offers a promising alternative by adapting model size to client capabilities; however, previous work is limited to the pre-training setting, and conclusions on sub-network selection do not necessarily hold for fine-tuning. Furthermore, prior approaches create client sub-networks without considering their unique data distributions, thereby overlooking the opportunity to tailor sub-network architecture to local data. Motivated by evidence that attention heads exhibit task dependency within layer defined feature spaces, we propose Optimal Hydra, a transformer based heterogeneous FL framework for fine-tuning that combines global layer selection with client specific attention head selection, aligned with local data distributions. Optimal Hydra makes explicit use of client data to produce client-specific sub-networks we call Hydra-nets. To our knowledge, this is the first model heterogeneous FL approach tailored to LLM fine-tuning. Experiments on three datasets show significant improvements over baseline methods while also reducing local resource demands.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20753
PDF
-
Causal Analysis of Robot Foundation Models Enables Efficiency Gains
2026-10-14 18:06:47Aiden Swann, Mathilde Papillon, Mozes Jacobs, Siddharth Boppana, Tal Haklay, Eric Bigelow, Andrew Lee, Conor Watts, Thomas Icard, Mac Schwager, Owen Lewis, Jack Merullo, Ekdeep Singh Lubana, Atticus Geiger, Thomas Fel, Matthew KowalAbstract
Vision-language-action (VLA) models are a promising approach to general purpose robot manipulation. While the internal computation of vision-language models, which compose the backbone of VLAs, has been studied extensively, less is known about how their perceptual and semantic representations are transformed into action in VLA models. We study this question mechanistically in three frontier VLA models ($\pi_{0.5}$, MolmoAct2, and MolmoBot) by analyzing the causal pathways through which the input image and the instruction shape actions. Across the three models, activation patching shows that causally relevant information influences the action expert (AE) through a narrow band of mid-to-late cross-attention layers, which we term the action-conditioning interface. The content of this information changes with token and layer within the model. Early-to-mid layer representations of the target noun token determine which object is selected, whereas mid-to-late layers carry the spatial information in text tokens that redirects the movement when patched. Where spatial information is stored in prompt tokens various based on the architecture. In MolmoAct2 the model forms task-conditioned representations at the text positions that follow the instruction, whereas $\pi_{0.5}$ additionally stores this information at image token positions. We show that these findings can be translated directly into improving model efficiency. We find that restricting MolmoBot's action expert conditioning to only 3 out of 36 layers reduces inference FLOPs by up to $33\%$ while maintaining comparable task success with the default model across 3,000 evaluation episodes ($46.7\%$ vs. $46.9\%$), and the same principle carries over to real-robot deployment of $\pi_{0.5}$.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20750
PDF
-
Can Learning to Answer Spatial Questions Improve Novel-View Generation?
2026-10-14 17:59:58Jahyeok Koo, Hyeonseo Yu, Seungho Jang, Jaewoo Jung, Jiwon Kang, Seokju Cho, Seungryong KimAbstract
Generating novel views from limited observations requires preserving both the appearance and spatial structure of a scene. Although recent approaches exploit fine-grained geometric cues, they do not explicitly model high-level semantic and spatial properties such as object identity, attributes, and inter-object relationships. In this work, we investigate whether 3D language understanding can provide complementary supervision for novel-view generation. We introduce Ask4NV, built on a pretrained unified multimodal model with a Mixture-of-Transformers (MoT) architecture. Our model incorporates camera geometry into both VAE and ViT representations and leverages understanding features during generation through joint self-attention with a modality-aware attention masking strategy. Experiments demonstrate strong image quality and geometric consistency in novel-view generation, alongside competitive 3D language understanding. Further analysis shows that joint training improves cross-view correspondence in VAE representations and similarity in a 3D-aware feature space.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20610
PDF
-
SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving
2026-10-14 17:59:56Shaoyuan Xie, Zihan Zhang, Jingxuan Wang, Jiashu Qu, Xiaoqing Liang, Lingdong Kong, Junchi Lu, Henrik I. Christensen, Qi Alfred ChenAbstract
Vision-Language-Action (VLA) models are gaining attention in autonomous driving, inheriting commonsense priors from large language models (LLMs) and widening the navigation interface from a fixed command set to free-form natural-language instructions. However, this open-ended interface exposes a new vulnerability surface: an inattentive driver may issue instructions that are unsafe given the current scene, and a malicious user may issue dangerous ones on purpose. We present the first benchmark suite that evaluates navigation following and conflict awareness together, along with adversarial robustness. Existing VLAs fall short on both: SimLingo completes no language-instructed lane change in closed loop, and all models degrade sharply under conflicting signals. We propose SafeDriveVLA, the first driving VLA that (1) explicitly predicts a three-level navigation-compliance mode to decouple conflict reasoning from action generation, and (2) uses navigation-conditioned world model dreaming to inform this prediction. Unlike prior world models that roll out expert or planned actions, ours dreams of strictly following the navigation signal, showing when following is safe. SafeDriveVLA reaches 88.16 Driving Score (DS) and 68.64% Success Rate (SR) on Bench2Drive, lifts closed-loop language following by 49.3% and reduces collisions under unsafe instructions from 60 to 38 compared to the best SimLingo variant. Video demonstrations are available on our project page.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20609
PDF
-
Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion
2026-10-14 17:59:52Huixuan Cheng, Bike Zhang, Shankar Sastry, Koushil SreenathAbstract
Recent advances in robot learning have driven rapid progress in humanoid control, enabling increasingly agile behaviors in the real world. However, humanoid control networks are typically obtained through conventional training procedures, while comparatively little attention has been given to directly modeling the distribution of their learned weights and generating functional networks in weight space. In this work, we investigate whether high-performing humanoid control networks can be generatively modeled directly through their weights. We adopt a latent diffusion framework that couples a linear autoencoder with a Diffusion Transformer. The autoencoder compresses actor network parameters into a compact latent space, and the Diffusion Transformer models the distribution of these latents. Trained on networks collected from reinforcement learning for the Unitree G1 robot, our approach generates deployable networks that perform various motions in the real world, including walking, dancing, and spin kick. Beyond generation, we further study the learned weight space to better understand the structure captured by these humanoid control networks. We identify a structural principle in the first-layer weights that consistently reflects the robot's kinematic prior. Notably, this kinematic prior emerges under both Reinforcement Learning (RL) and Imitation Learning (IL) training, suggesting that it is not specific to a particular optimization process but may instead reflect common knowledge associated with humanoid embodiment. Together, these results demonstrate the feasibility of generative modeling in humanoid network weight space, while providing initial insights into the information encoded by their learned weights. Additional results are available on our project website: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.20606
PDF
-
Latent Compaction with Test-Time Training
2026-10-14 17:57:01Yingfa Chen, Xingyu Shen, Chaojun Xiao, Xu HanAbstract
Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs during generation. Compressing these caches must preserve information for unseen questions. We introduce **Latent Compaction**, which learns a compact document-specific cache through test-time training. We initialize cache entries using reconstruction importance, teacher attention, and key-space diversity. Training first fits attention outputs and log-normalizers to those of the full cache, with reconstruction-selected entries held fixed. It then jointly optimizes the entire cache through end-to-end distillation from the full-cache model. Model weights remain frozen throughout, and training uses only generic prompts about the document, without access to downstream questions or answers. Experiments on RULER, LongBench, and Phonebook Lookup demonstrate substantial gains over budget-matched compression baselines. On a RULER, for example, our method achieves 96.7% accuracy at 32$\times$ compression, compared with 40.0% for KVZip+. The method also transfers across model scales and families, with retrieval gains on both Qwen3 and MiniCPM5. These results show that treating the KV cache as a neural network and optimizing it with test-time training can preserve retrieval accuracy under aggressive compression. The code is released at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.20582
PDF
