# Research Papers — 2026-09 · bundle page 61 of 266 > Full-text markdown of papers #3001–3050 of 13298 published in 2026-09, newest first. AI-readable mirrors of the full 13298-paper 2026-09 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-09/page/61 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/60.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/62.md) ### MTLiquid: Enabling Efficient Multi-Task Learning using Liquid Neural Networks for Lightweight Healthcare Monitoring Systems Authors: Rachmad Vidya Wicaksana Putra, Fahad Abdul Rauf, Muhammad Shafique Published: 2026-09-27 Categories: cs.LG, cs.NE arXiv: https://arxiv.org/abs/2609.33232v1 Continuous-time sensing and monitoring with timely and accurate decision-making are critical for many real-world applications. In healthcare monitoring systems, physiological signals are often available or sampled at irregular time intervals, hence requiring continuous-time processing to provide accurate prediction. Moreover, such systems often need to solve multiple detection/prediction tasks to provide a comprehensive patient review from different physiological aspects for more accurate decision-making. To solve this, continuous-time neural networks (CTNNs) can be employed. However, state-of-the-art works typically solve only one task at each network, thereby limiting their efficiency gains. To address this limitation, we propose MTLiquid, a novel methodology to enable efficient multi-task learning in continuous-time processing for healthcare monitoring systems through effective network design and training strategy. MTLiquid employs: (1) multiple input and output heads to accommodate different tasks, while sharing the same backbone network across tasks; as well as (2) an effective training strategy that leverages a loss-weighting technique to balance learning updates across different tasks and a proportional data presentation technique to address imbalanced dataset sizes. Experimental results for mortality prediction (P12) and sepsis early detection (P19) tasks for ICU patients show that, MTLiquid achieves strong performance (AUROC: 0.84 for P12 and 0.94 for P19) comparable to the state-of-the-art single-task learning in both continuous-time networks (AUROC: 0.84 for P12 and 0.95 for P19) and discrete-time networks (AUROC: 0.79-0.82 for P12 and 0.92-0.94 for P19), while incurring significantly smaller memory cost by 44%-94%. These results highlight the potential of our MTLiquid methodology to enable lightweight continuous-time healthcare monitoring systems for better decision-making. --- ### AevaScenes: An FMCW LiDAR Dataset and Benchmark for Long-Range Perception Authors: Gautham Narayan Narasimhan, Heethesh Vhavle, Kumar Bhargav Viswanatha, James Reuther, Deva Ramanan Published: 2026-09-27 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.33230v1 FMCW LiDAR measures per-point radial Doppler velocity alongside range, providing a motion cue unavailable in conventional time-of-flight sensors. Exploiting this signal at long range remains understudied. We present an FMCW LiDAR dataset of 575 sequences (57.5K frames) with over 8 million annotated 3D boxes across 16 detection classes and per-point labels across 24 semantic classes, captured by six commercial FMCW LiDAR sensors and six paired 4K cameras across eight Bay Area cities, including 237 nighttime sequences, with annotations extending to 400m. We define a benchmark with three tasks: 3D object detection, scene flow estimation, and semantic segmentation. Detection and scene flow are evaluated across three range bins to 400m, with a public evaluation server. We explore the impact of Doppler measurements on flagship recognition tasks, and find significant improvements up to 2X in detection AP of far-away vehicles and pedestrians, particularly in low-latency single-frame settings. We similarly find scene flow accuracy is significantly improved with Doppler measurements across all ranges. Our dataset and benchmark have been publicly released at https://scenes.aeva.com. --- ### Beyond Memory Construction: Rethinking Memory Access for LLM-based Conversational Agents Authors: Donghua Cai, Yongheng Deng, Yifei Wang, Zijun Shen, Ju Ren Published: 2026-09-27 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.33226v1 Memory is a core component of conversational agents, enabling coherent and context-aware behavior over long interactions. Recent approaches commonly rely on LLM-based memory construction, where raw interactions are rewritten into structured memory units and later retrieved via a RAG pipeline. While effective in controlled settings, we show that this paradigm breaks down in long-horizon, high-entropy conversations: memory construction becomes increasingly lossy and unstable as context length and information complexity grow, and incurs prohibitive cost due to repeated LLM invocation. To address these limitations, we propose Threader, a memory system that shifts the focus from memory construction to efficient, structure-aware access over raw interactions. Instead of rewriting interactions, Threader preserves them as first-class memory, organizes them into topic-coherent segments via lightweight incremental segmentation, and enables accurate retrieval through multi-view representation. At query time, it performs multi-signal retrieval that combines segment-level access with localized evidence matching, ensuring both completeness and coherence. Extensive experiments demonstrate that Threader consistently improves answer accuracy and evidence recall, while significantly reducing the memory construction overhead. --- ### RMB: Reward Model Boosting Mitigates Reward Hacking Authors: Jiabin Fan, Dezhi Ye, Yongchang Hao, Lili Mou Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33221v1 Reinforcement Learning from Human Feedback (RLHF) is a powerful technique for aligning large language models (LLMs) with human preference. However, it often suffers from the reward hacking issue, where policy optimization improves the proxy reward model while actually degrading performance with respect to the true human preference, due to the imperfection of the proxy. To address this, we propose Reward Model Boosting (RMB), a novel approach that enhances the robustness and reliability of the reward signal for RLHF. RMB first trains a set of reward models with a diversity-promoting regularizer. This encourages each model to learn complementary aspects of the reward landscape. Then, RMB learns a lightweight aggregator in the principle of boosting to aggregate the outputs of the diverse reward models into a more accurate and robust reward signal. Our extensive experiments demonstrate that RMB significantly improves reward accuracy on both in-distribution and out-of-distribution datasets, substantially mitigating the reward hacking issue and ultimately improving RLHF performance. --- ### When Do Models Admit They Are Wrong? Failure Disclosure Is Unstable Under Reinforcement Learning Authors: Steven Y. Feng, Noah D. Goodman, Michael C. Frank, Evan Hubinger, Paul C. Bogdan, Andrew Lampinen Published: 2026-09-27 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.33220v1 Outcome-based reinforcement learning can produce models with similar task performance but very different ways of communicating about their mistakes. We study failure disclosure: whether a model admits that an attempted solution failed rather than staying silent or presenting it as successful. Across repeated outcome-only GRPO training runs, failure disclosure varies far more than task accuracy. The pattern extends to a second reasoning task and stabilized PPO, persists at 7B, and also appears in an instruction-conditioned 32B setting. We also find that small floating-point and sampling differences during training can redirect reporting behavior even when the task objective and earlier training history are held fixed. Additional tests show that failure disclosure is not a single decision: Checking the answer, entering a report, and completing the admission can separate, and the weak point depends on the task and response format. Further, experiments with neutral controls show more broadly that behaviors left weakly constrained by training are especially likely to vary across runs, of which failure disclosure is an example. We can reduce variability in failure disclosure by discouraging the model from drifting from its starting policy on failed, well-formed responses. This makes reporting substantially more consistent, though its effect on task performance depends on the setting. Stable task accuracy therefore does not guarantee stable safety-relevant behavior: Researchers should measure these behaviors directly across runs and design training methods that keep them reliable. --- ### Scope-WM: Scoped Computation for Efficient Visual World Models Authors: Chunzheng Li, Zesheng Jia, Hongda Zhang, Jiaying Tang, Yuntian Wang, Siao Liu, Jin Wang Published: 2026-09-27 Categories: cs.CV, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2609.33218v1 Visual world models enable robotic planning by predicting future observations, but dense latent-state propagation and sample-intensive trajectory optimization incur high inference latency and peak memory usage, limiting real-time deployment on resource-constrained platforms. Existing sparse world-model acceleration methods either rely on unguided token sparsification, which may discard planning-relevant information and restrict achievable sparsity, or introduce heavy auxiliary modules and cumbersome multi-stage training pipelines. In this work, we present Scope-WM, an efficient visual world model that scopes computation to prediction-relevant latent regions and promising action sequences. Scope-WM distills prediction relevance into a lightweight action-conditioned selector and applies full dynamics prediction only to a compact subset of selected tokens. It updates the remaining tokens using a compact summary of foreground states and their changes, allowing the background to perceive foreground dynamics without costly token-to-token interactions. During planning, Scope-WM preserves and reuses high-quality action sequences discovered during the initial MPC search, focusing subsequent search under reduced rollout budgets. The resulting pipeline requires only a one-off selector distillation followed by a single joint training stage for the sparse world model. On the challenging Push-T task, Scope-WM reduces peak GPU memory usage and planning time to $18.1\%$ and $14.3\%$ of those of dense DINO-WM, respectively, corresponding to a $6.97\times$ planning speedup, while maintaining competitive task performance. Further evaluations across five diverse visual planning tasks demonstrate the general applicability of Scope-WM. Code is available at https://github.com/ChunZheng2022/Scope-WM. --- ### RepFlow: Reciprocal Supervision Improves Generation and Representation in Flow Models Authors: Weili Zeng, Feng Tian, Shengqi Liu, Yichao Yan Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33217v1 Generative models learn visual structure through denoising, yet their internal states are entangled with both noise level and network depth, making it difficult to obtain a stable visual representation from the generator itself. We introduce RepFlow, which learns such a representation from the generator's evolving computation and uses it to guide generation. Specifically, a separate timestep-free encoder is trained, through a timestep-conditioned predictor, to recover generator states across depths and noise levels from masked clean images. By excluding the reference noise and masked-out content from the encoder's input, we encourage the encoder to distill visual information that is recoverable from the visible image context and predictive of generator states. The representation learned from the generator's evolving states is then fed back to guide and improve the generator, whose updated states provide supervision for further representation learning, forming a reciprocal learning process. This reciprocal process improves multi-step generation and representation quality, as measured by frozen linear probing on ImageNet with latent-space SiT and pixel-space JiT, without an externally pretrained representation teacher. Across the two unconditional settings, FID decreases by $19.2$--$40.8\%$ relative to native training, while linear-probe accuracy improves by 6.80--10.04 percentage points over the best searched raw generator features. The learned representation also serves as a distributional metric for one-step JiT post-training, extending its role from instance-level alignment to distribution-level supervision. --- ### CoLMbo-SV: A Grounded Language Model for Explainable Speaker Verification Authors: Massa Baali, Sarthak Bisht, Ziyue Qiu, Joseph Konan, Rita Singh, Bhiksha Raj Published: 2026-09-27 Categories: cs.CL, cs.SD, eess.AS arXiv: https://arxiv.org/abs/2609.33212v1 Speaker verification systems achieve high accuracy but provide little account of the acoustic evidence behind their judgments. Making these systems inspectable requires exposing interpretable evidence while retaining the richer information on which their decisions depend. We present \textbf{CoLMbo-SV}, a speaker language model that combines strong speaker discrimination with structured, acoustically grounded comparison reports. By connecting a pretrained speaker encoder to a language model and supplying explicit acoustic measurements, CoLMbo-SV makes voice comparisons inspectable without restricting verification to the evidence verbalized in its reports. We additionally introduce \textbf{VoxReason}, paired recordings with measured acoustic properties and comparison reports filtered through numerical and qualitative checks, providing supervision for this combined capability. We also develop an evaluation framework that separates what acoustic information a speaker representation encodes, what influences the verification score, and what the generated report discusses. On VoxCeleb1-O, CoLMbo-SV achieves 0.99\% EER, reducing verification error by approximately 80\% relative to the strongest audio-language baseline fine-tuned on VoxReason, while attaining a numerical-grounding score of 0.82. Our analysis further demonstrates that acoustic correctness and decision relevance are distinct properties of an explanation, exposing a gap that numerical-grounding metrics miss. Together, these contributions substantially advance audio-language speaker verification, bring its accuracy toward that of dedicated speaker encoders while adding checkable acoustic reporting, and establish an empirical framework for connecting natural-language explanations to the decisions they explain. --- ### Background Gradients Shape Memorization in Flow Matching Authors: Xuanhua Yin, Boyu Wei, Shuyi Zhang, Shunqi Mao, Chuanzhi Xu, Weidong Cai Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33210v1 Repetition is closely associated with memorization in generative models, but how other training images affect the retention and copying of targets remains unclear. We study this question in class-conditioned flow matching, where images outside the target set form the background. At fixed target repetition and same-class background row count, replacing repeated same-class images with distinct images reduces the target extraction rate from 80.7% to 18.0%. To explain this effect, we develop a paired-trajectory framework that isolates target-induced parameter displacement and the background gradient response to it. This response has an exact path-integrated curvature representation, connecting background loss geometry to target learning. Reciprocal response transfer between repeated and distinct backgrounds changes target retention and copying in both directions, establishing the response's causal role. After target removal, the response correction parallel to the target-induced displacement preserves approximately 90% of the copying effects of full response transfer. Directly scaling the displacement also changes copying without further training. The post-removal copying effects of reciprocal transfer are reproduced across datasets and architectures. Together, these results identify the background gradient response as a mechanism through which same-class training data shape the retention of target learning and the reproduction of target images. --- ### Beyond Calibration: Do a Typed-Decision Model's Probabilities Obey the Probability Axioms? Authors: Keyi Li, Yihao He, Quanyi Li Published: 2026-09-27 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.33209v1 Typed-decision models such as TypeSafe's Jev answer a declared yes/no or multiple-choice question about a state with a probability instead of text, and their evaluations report accuracy and calibration. Neither requires that the probabilities a model gives to logically related questions fit together item by item, which is what a system that acts on those probabilities needs. We test this property, coherence, with a battery of logically linked questions that needs no labels. For 160 items from ChaosNLI and PubMedQA, each with three mutually exclusive labels, we ask whether the label is X, whether it is not X, whether it is one of the other two labels, and which label applies. On 480 negation pairs, Jev's probabilities for "the label is X" and "the label is not X" miss summing to one by 0.064 on average (95% CI 0.055 to 0.072). Qwen3.8-27B, run from its official BF16 weights, misses by 0.293 with first-token probabilities and by 0.122 with verbalized probabilities. The gap to the first-token readout persists on pairs where both systems give similar probabilities, without double-negation labels, and after averaging Jev's repeated calls. Jev is not coherent either: its violations are about five times its repeat noise, and it over-endorses statements about single labels, so that its three single-label probabilities sum to 1.14 on average. The two systems also fail differently. Qwen3.8-27B's first-token readout under-endorses the complement of a label whether or not the question contains "not", rejecting both a statement and its negation in 196 of 480 pairs, and it does not become more coherent where it is more confident, whereas Jev's violations concentrate where its answer is uncertain. Because the checks need no labels, they expose biases that appear only when question forms are compared, and inconsistencies within items. --- ### WorldAgent: Verification-Guided Agentic Physical World Construction Authors: Caoliwen Wang, Mengdi Wang, Yige Chen, Zejia Wu, Bowen Huang, Siyuan Chen, Guanxiong Chen, Lifu Wei, Heng Zhang, Qinghai Zhang, Yin Yang, Guandao Yang et al. Published: 2026-09-27 Categories: cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.33208v1 Constructing complex physical worlds from language requires coordinating extensive 3D environments, detailed structures and objects at different spatial scales, and interacting physical processes under both stated goals and implicit physical constraints. We present WorldAgent, an agentic framework for verification-guided physical world construction from a single natural-language prompt, without iterative user debugging. A world construction layer expands the prompt into a structured world specification and uses physical knowledge to build scenes and run numerical simulations. After every step, a verification layer inspects scene geometry and simulation states alongside rendered views. Failed checks guide automatic revisions to the specification and re-execution of the affected steps. Accepted worlds pass the required checks and remain editable for further inspection and resimulation. We introduce AgenticSimBench, on which WorldAgent achieves the best scores among the evaluated agent-based methods on five of seven metrics. In a 26-participant user study, it receives the highest mean ratings across all four criteria. --- ### MorphAtt: A Neuromorphic Accelerator for Efficient Multi-Head Attention Processing in Spiking Vision Transformers Authors: Rachmad Vidya Wicaksana Putra, Amirhesam Jafari Rad, Muhammad Shafique Published: 2026-09-27 Categories: cs.AR, cs.LG, cs.NE arXiv: https://arxiv.org/abs/2609.33207v1 Spiking Vision Transformers (SViTs) are developed as an energy-efficient alternative to conventional ViTs for computer vision tasks at the edge. However, huge parameter counts and complex multi-head self-attention (MHSA) operations make it challenging to achieve high energy efficiency in SViT inference, especially in tightly constrained applications. To maximize efficiency gains of SViT processing, we propose MorphAtt, a novel digital accelerator that expedites SViT inference through streamlined processing. Specifically, it processes MHSA operations using cascaded hardware modules: a Spiking Query-Key-Value generator (SpikeQKV), a low-complexity Spiking Multi-Head Self-Attention engine (SpikeAtten), and Reparameterization Convolution (RepConv) modules. To mitigate traffic congestion in on-chip memory accesses and data reuse, specialized inter-module buffers are integrated within the dataflow. Under synthesis using 32nm CMOS technology, MorphAtt achieves 792-1605 GOPS of throughput, while incurring ~39-55 mW of power consumption and 1.5 mm^2 of area, which lead to 20.3-29.1 TOPS/W of energy efficiency. These results also demonstrate that our MorphAtt offers better performance and efficiency trade-offs than state-of-the-art, thereby enabling highly energy-efficient vision-based AI systems at the edge. --- ### SMORE: Stability-Promoting Mesh-Agnostic Model Reduction for Time-Dependent PDEs Authors: Yangyuan Li, Weichao Li, Shaowu Pan Published: 2026-09-27 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.33205v1 High-fidelity simulations of time-dependent partial differential equations (PDEs) are computationally expensive, motivating data-driven reduced-order surrogates for many-query tasks such as uncertainty quantification, design optimization, data assimilation, and optimal control. However, existing surrogate models often exhibit poor temporal stability, which can lead to unstable rollouts and exploding gradients during backpropagation, especially in multistep long-horizon forecasting. To address this, we propose SMORE, a mesh-agnostic framework for model order reduction of time-dependent PDEs. Its latent dynamics are trained with Lyapunov-guided stability regularization, which promotes stable long-horizon rollouts. We provide theoretical guarantees under the stated structural assumptions. Beyond forecasting PDE evolution, the learned latent dynamics, which are interpretable and linear or linear-quadratic, could bring benefits for downstream tasks such as data assimilation and optimal control. Moreover, our framework is capable of predicting continuous PDE solution fields from sparse measurements of the initial condition. We evaluate SMORE on a range of problems, including wave propagation, the Navier-Stokes equations, and the shallow water equations. Our results show that it improves long-horizon rollout generalization and empirical robustness, and achieves competitive accuracy at comparable parameter budgets relative to competitive baselines including DINo, FNO, CNO, and Transolver. --- ### Turning Speech Language Models into Multilingual Listeners Authors: Tolúlopé Ògúnrèmí, Dan Jurafsky, Chris Manning, Ahmet Üstün, Martijn Bartelds Published: 2026-09-27 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.33204v1 Speech Language Models (SLMs) that understand spoken language questions support only a few high-resource languages, limiting access to millions of people worldwide. This gap stems from the scarcity of multilingual speech instruction-tuning datasets. We present MULTISPEECHQA, a large-scale, synthetically generated and human-verified dataset comprising 9200 hours of 10.8 million spoken question-answer pairs in 23 typologically diverse languages. Using MULTISPEECHQA, we also introduce MULTISPEECH-BENCH, a multi-task benchmark for evaluating SLM performance on 23 languages. We compare the performance of a cascading system to open-weight and closed SLMs on MULTISPEECH-BENCH and find that the cascading system outperforms open-weight SLMs but not all closed SLMs. We use MULTISPEECHQA to finetune Qwen 2.5-Omni, which improves its performance on our benchmark. Our findings show that high-quality synthetic datasets offer a cheap solution to improving the multilingual capabilities of SLMs. --- ### Structured Residual Connectivity Matters for Diffusion Transformers Authors: Yuhe Liu, Xinyin Ma, Gongfan Fang, Songhua Liu, Xinchao Wang Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33203v1 Diffusion Transformers (DiTs) have established themselves as a scalable backbone for high-fidelity image synthesis. However, unlike U-Net based diffusion models that rely on rigid, hand-crafted skip connections, DiTs predominantly use a uniform residual stream that integrates all preceding layers as a monolithic state. In this work, we rethink residual connections in diffusion transformers and propose to transform them from passive summation into an active retrieval mechanism optimized for image denoising. First, we conduct a systematic analysis of DiT's internal representation, revealing a latent preference for early-layer feature reuse and symmetric layer guidance. Motivated by this, we introduce a structured connectivity design that explicitly integrates local residual connections with long-range pathways. Instead of static skip connections or dense all-layer routing, our method enables each transformer block to selectively ``attend'' to critical earlier representations, dynamically retrieving spatial and semantic cues through direct, differentiable cross-depth paths. Experiments show that our adaptive connectivity leads to faster convergence, with up to $1.73\times$ fewer training iterations, and significant gains in FID and visual quality with less than $0.1\%$ additional parameters, further improving a strong REPA-XL/2 model from $5.9$ to $4.34$ FID without guidance and reaching $1.39$ FID with classifier-free guidance. Our findings suggest that adaptive cross-layer connectivity is a critical yet underexplored factor in diffusion transformers, and that incorporating structured information pathways provides a simple and effective direction for improving scalable generative models. --- ### ReAL: Accelerating Flow Matching through Segment Advancement with Shared Lookahead Authors: Xuanhua Yin, Chuanzhi Xu, Haoxian Zhou, Shunqi Mao, Weidong Cai Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33202v1 Flow-matching models generate high-quality images and videos, but repeated neural network evaluations make sampling expensive. Skipping evaluations reduces this cost by extending an available velocity estimate over a longer span. However, local velocity agreement alone does not determine a suitable span, and checking each candidate endpoint adds costly model calls. We introduce ReAL, a training-free sampler that selects how far to advance using one shared lookahead. Our key insight is that the discrepancy between uncorrected and lookahead-corrected endpoint proposals can be computed directly from the observed velocity mismatch and the candidate span beyond the lookahead. This relation provides a span-dependent selection criterion without additional endpoint evaluations. The same lookahead selects the longest passing candidate span, corrects the accepted update, and supplies its velocity as the next starting estimate. After initialization, each regular iteration requires only one fresh evaluation. ReAL uses the pretrained velocity output and original noise schedule, with no additional training or access to internal features. Experiments cover four image-generation backbones, video generation, and image editing. ReAL achieves 4.91x measured speedup on FLUX.1-dev while retaining 97.0% of dense mean ImageReward. On HunyuanVideo, it achieves a 5.49x speedup while maintaining a VBench score close to that of dense sampling. --- ### Teach Yourself Where to Look: On-Policy Attention Self-Distillation for Reasoning Authors: Safaeid Hossain Arib, Rabeya Akter, Ismam Nur Swapnil, Md. Faiyaz Abdullah Sayeedi, Tasnim Mohiuddin, Md Mofijul Islam Published: 2026-09-27 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.33200v2 On-policy self-distillation trains reasoning models on their own trajectories using dense token distribution guidance from a privileged teacher with access to a verified solution. This supervision transfers what the teacher predicts without directly transferring where it attends within the preceding context. We introduce On-Policy Attention Self-Distillation (OPASD), which complements token-level supervision with solution-conditioned attention distillation. Because the privileged teacher can attend to verified solution tokens unavailable to the student, OPASD projects teacher attention onto student-visible positions and renormalizes the resulting distribution before alignment. Across three model sizes and four competition-level mathematics benchmarks, OPASD consistently outperforms token-only OPSD, improving average accuracy by 4.98 to 8.40 percentage points. OPASD also avoids the response-length inflation and performance degradation observed with token-only distillation, reducing generated rollout tokens by 73.9% and estimated model compute by 72.6% while training 1.53x faster. These results show that solution-conditioned attention provides a complementary supervision signal that makes on-policy self-distillation more accurate, stable, and compute-efficient. --- ### TAO-DA: Towards Autonomous Operation--A Dual-Arm Vision-Language-Action Model for Coordinated Manipulation Authors: Yongsheng Zhao, Han Gao, Baoping Cheng, Jingyao Tang, Dian Zhou, Deng Liang, Ji Ge, Xuanzhang Wen, Lei Zhao, Ye Wang Published: 2026-09-27 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.33197v1 Vision-Language-Action (VLA) models provide a unified framework for grounding high-level semantic information into low-level robot actions, enabling scalable robotic manipulation across diverse tasks. However, existing VLA models lack explicit mechanisms to disentangle the states and intents of the two arms, leading to unintended cross-arm interference that degrades task execution success. To address this issue, we propose a symmetric Dual-Arm Expert (DAE) architecture built upon a shared Vision-Language Model (VLM) backbone with decoupled, arm-specific expert towers. Expert selection is carried out through a two-stage dual-arm intent routing scheme, in which experts are routed either by explicit language instructions in the first stage or by implicit visual semantics in the second stage. Moreover, we introduce a lightweight task progress prediction module that leverages cross-attention between the pre-chunk temporal features and semantic representations of proprioceptive and visual observations to accurately estimate frame-wise task completion progress. This module facilitates task progress synchronization to support coordinated scheduling for collaborative multi-robot tasks. Experimental results demonstrate the effectiveness of our model in dual-arm intent routing and the disentanglement of cross-arm interference, and further provide preliminary evidence of emergent skill generalization from single- to dual-arm tasks (as well as the reverse), together with cross-arm motion-domain skill transfer. --- ### Are Benchmarks Reliable? Toward Structural Diagnosis via Sample-Level Capability Boundaries Authors: Haiquan Hu, Yuzhu Liang, Weicheng Tang, Yanzeng Li, Yao Shi, Tian Wang Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33196v1 Evaluating large language models (LLMs) relies heavily on benchmark scores, yet aggregate metrics can obscure whether benchmark samples reliably support model comparison. We introduce \textbf{BSDProbe}, a sample-level framework for \emph{benchmark structural diagnosis} that estimates capability boundaries from repeated-response trajectories along ordered model axes. BSDProbe summarizes samples by boundary position, boundary width, boundary-signal validity, and order consistency, then aggregates them into benchmark-level structural profiles. Experiments on six benchmarks show that benchmark reliability is axis-conditioned and heterogeneous: GSM8K and MATH exhibit the most stable measurement structures, MMLU and TriviaQA are relatively stable but heterogeneous, while GPQA and PopQA show stronger axis-conditioned risks. These profiles remain consistent across Qwen3, Qwen2.5, and cross-model axes. BSDProbe further selects compact high-value subsets whose model discriminability reaches up to $8.58\times$ that of the full benchmark. These results suggest that reliable benchmark use requires examining sample-level capability boundaries beyond leaderboard scores. --- ### Orthogonal Witness Control for Muon Optimization via Sigmoid Spectral Reshaping Authors: Dat Phi Van, Ngo Vu Minh, Tuc Nguyen, Thin Nguyen, Ngoc-Thanh Dinh, Trung Le Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33194v1 Matrix-valued optimizers such as Muon exploit the spectral structure of neural network updates through Newton--Schulz orthogonalization, but their near-flattening of the singular spectrum discards relative magnitude information across gradient modes. We introduce \emph{Soren} (\textbf{S}pectral \textbf{O}rthogonal \textbf{Re}shapi\textbf{n}g), a matrix-valued optimizer that preserves the singular subspaces of the gradient while applying a bounded, monotone sigmoid transformation to its singular values. This smoothly compresses dominant modes without fully flattening the spectrum. We interpret Soren as a positive-definite preconditioned gradient method and establish convergence guarantees under relative smoothness and metric Polyak--Łojasiewicz geometry. To avoid explicit singular value decomposition, we further develop a finite-depth Soft Newton--Schulz (SNS) polynomial realization of the sigmoid spectral map and characterize how its spectral approximation affects the induced convergence geometry. Experiments across LLM pre-training, supervised fine-tuning, and direct preference optimization demonstrate the effectiveness and robustness of Soren against established optimizers. --- ### Apparent Compression, Real Stability: The Intrinsic Dimension of Learning a Quantum Wavefunction Authors: Lu Wei, Yufeng Wang, Chenfeng Cao, Haibin Ling Published: 2026-09-27 Categories: cs.LG, quant-ph arXiv: https://arxiv.org/abs/2609.33193v1 How many directions in weight space does training need? The intrinsic dimension answers this with the smallest number of random directions in which training still reaches a target accuracy, and small values have motivated parameter-efficient methods such as LoRA. We measure it for variational Monte Carlo (VMC), which trains a neural network to represent the ground state of a quantum many-body system. VMC is a demanding test, because the network generates its own training samples and every gradient is noisy, and a revealing one, because the exact answer is known and every run can be scored. We train only a small latent vector that a frozen random map turns into the network's weights, with no change to the standard natural-gradient optimizer. We find that a small dimension can be misleading, while the stability it brings is real. On a magnet with a hard sign pattern, a network that cannot represent signs reaches its best energy in 8 of 28,642 directions, but only because no such network can go lower; once signs are learnable, neither the signs nor the magnitudes are cheap. The dimension rises across a quantum phase transition, so it tracks how difficult a state is at far less compute than fitting a scaling law, yet it never falls below a floor set by the random subspace itself, even where the ground state is nearly trivial. Training in the subspace, in contrast, never diverged in our experiments, whereas full-parameter training with the same settings did, and a control with matched solvers attributes the difference to the reduced dimension. --- ### AG-CoT: Verified Algorithmic Traces for LLM Program Synthesis on Clifford Circuits Authors: Lu Wei, Yufeng Wang, Chenfeng Cao, Lu Pang, Haibin Ling Published: 2026-09-27 Categories: cs.LG, cs.CL, quant-ph arXiv: https://arxiv.org/abs/2609.33192v1 Scientific code generation can produce executable programs that fail to compute the intended scientific object. We study this problem in language-model synthesis of Clifford circuits, which prepare the stabilizer states used in quantum error correction and admit exact classical verification. In our target-conditioned framework, each target is given as compact signed stabilizer generators, and an exact verifier checks the generated OpenQASM circuits. We supervise models with Aaronson-Gottesman chain-of-thought (AG-CoT) traces checked by the verifier, and continue training on model generations that the verifier accepts. Across two independently trained model families (3B and 7B), AG-CoT supervision multiplies greedy-decode state-equivalence accuracy by four to six times over circuit-only baselines, and verifier-filtered continuation training adds a further consistent gain atop both. A complementary 32B study shows that supervised models achieve near-perfect syntax and Clifford validity while the strongest direct model reaches 6.14% state equivalence per target, rising to over 10% under verifier-guided selection with multiple candidates. These results show that algorithmic trace supervision gives a large, statistically significant gain in both model families and that verifier-filtered continuation adds a further repeated gain. The persistent gap between Clifford validity and state equivalence confirms that exact verification is necessary: a circuit can be syntactically and physically valid yet prepare the wrong quantum state. --- ### FocusDrive: Reasoning with Visual Focus for Autonomous Driving Authors: Zhiyuan Liu, Zehong Ke, Yuanxin Tian, Hao Cheng, Jinhao Li, Yining Xing, Yanbo Jiang, Zhenhua Xu, Wenhao Yu, Jianqiang Wang Published: 2026-09-27 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.33190v1 Driving decisions depend on both where to focus and how to act on what is seen. Effective driving reasoning must establish which objects matter, where they are, and how they inform the intended action. Text-based rationales can describe a driving response while leaving its correspondence to specific visual evidence implicit. Visual focus provides a concrete starting point for this connection by identifying what matters in the scene and where it is. We propose FocusDrive, a structured multimodal reasoning framework that organizes end-to-end planning around explicit visual focus. It pairs descriptions of decision-relevant objects with image-patch references, bringing explicit visual focus into the reasoning that generates driving plans and trajectories. We first assess this focus representation through driver gaze prediction, then investigate its role in planning reasoning using driving-focus annotations within existing NAVSIM training scenes. Experiments on W3DA and NAVSIM demonstrate competitive gaze prediction and end-to-end planning performance, with FocusDrive improving over text-based chain-of-thought. These results support visual focus as an effective link between scene understanding and driving action. --- ### When an Evaluation Rule Writes Training Labels: Measuring Human-Reference Forgiveness in NAVSIM Authors: Jiaxuan Guo, Jingxin Yang, Jiaqi Ye, Youran Sun, Shuo Xin, Kejia Zhang, Haizhao Yang Published: 2026-09-27 Categories: cs.LG, cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.33189v1 When the human reference scores zero on a metric, the released GTRS-Dense label generator for NAVSIM marks every candidate trajectory in the scene as passing it. NAVSIM's authors introduced this human-reference forgiveness to avoid penalizing contextually justified maneuvers when scoring one trajectory, and warned that it could overlook important failures. In label generation it sets a whole column of 16,384 candidate targets to passing. To measure the consequences for supervision, we re-run the generator with the overwrite disabled and compare the pre-overwrite targets with the released labels on all 103,288 navtrain scenes. The rule erases a candidate distinction that the training loss reads on 11,237 of them (10.8793%). Firing usually changes most of a column: lane keeping carries 9,982 of the 13,042 forgiven loss columns, and its median forgiven column had 14,391 of 16,384 candidates failing before the overwrite. On held-out navtest scenes forgiven on lane keeping, the released lane-keeping head's median AUC against the pre-overwrite outcome is 0.7095; on unforgiven scenes matched on failing-candidate count it is 0.9807. For the Hydra-MDP checkpoint released with GTRS, whose configuration takes the same label file, the two values are 0.6627 and 0.9761. Continuing the released GTRS-Dense checkpoint for 300 optimizer steps with three paired seeds, we observe the forgiven-scene AUC 0.1086-0.1251 higher with pre-overwrite than with published targets, and a narrower gap between matched groups, still above zero. Scoring with forgiveness disabled, we observe lane keeping higher by 2.478-3.524 points on navtest scenes forgiven on any of five loss metrics, with lower adjacent-frame plan consistency. Both changes are larger there than on the rest. EPDMS, scored the same way, does not separate the two target sets. --- ### Offline Policy Evaluation via Mixed Bellman Residuals and Adaptive Critic Representations Authors: Amitakshar Biswas, Yuhan Li, Ruoqing Zhu Published: 2026-09-27 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2609.33186v1 Evaluating a target policy using data generated by a different behavior policy remains a fundamental challenge in reinforcement learning. While most existing work relies on the standard one-step Bellman residual, we consider a convex combination of one-step and two-step residuals with a fixed mixing weight. In ideal settings, this mixed Bellman formulation can provide a natural bias--variance trade-off between approximation error under a restricted value-function class and the increased variance arising from multi-step importance weighting. To solve this mixed residual optimization, we adopt a minimax formulation involving a critic function. Unlike standard approaches that rely on a fixed functional class, we construct a data-dependent critic representation using predicted future feature directions which effectively induces a kernel adapted to the underlying transition dynamics. This allows the critic to focus on directions that are most relevant for the estimated Bellman error. To control overfitting, we use sample splitting to construct the critic and estimate the value function on separate data subsets. Simulation studies and MetaWorld tasks illustrate the effect of the mixing parameter and show that intermediate residual combinations can improve value estimation in challenging settings. --- ### Identifying Temporal Features within Transcoders for Time Sensitive Factual Recall Authors: Sanjay Govindan, Yang Song, Maurice Pagnucco Published: 2026-09-27 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.33183v1 Large Language Models (LLMs) suffer from temporal misalignment, often due to the contradictory nature of their training corpora. While current mitigation strategies rely on computationally expensive fine-tuning or context-heavy retrieval augmented generation (RAG), the internal mechanisms governing time-sensitive recall remain under-explored. Unlike prior studies that identify temporal components such as attention heads and MLP layers, we provide the first feature-level map of temporal recall by isolating individual MLP features via transcoder circuit tracing. We identify three node categories (common temporal, common to the year, and chrono-semantic) which interact to generate a temporal filter during factual recall. By analysing Gemma 2 2B, LLaMA 3.2 1B, and Qwen3-4B, we show that these features do not follow a simple linear pipeline but represent time through a parallel and mixed syntactic-semantic interplay across layers. We additionally discover a class of higher-layer temporal components invisible to existing EAP-IG methods, establishing transcoders as a more complete lens for temporal interpretability in time-sensitive factual recall. These findings present MLP components for potential targeted interventions in time-sensitive factual recall --- ### Unlocking Latent Personalization in LLMs Authors: Wei Chen, Guanghui Zhu, Zhongliang Cai, Yihua Huang Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33182v1 Large language models (LLMs) are increasingly expected to adapt to individual users, yet effective personalization remains challenging when only limited user-specific samples are available. In this work, we take an alternative perspective: pretrained LLMs may already possess latent capacity for personalization, and a few user samples may therefore suffice to guide the model toward user-aligned behavior with minimal user-specific adaptation. From this perspective, we propose LatentPersonal, a framework that formulates personalization as navigation in a shared latent adaptation space. LatentPersonal infers a compact latent representation from a few user samples to guide user-specific model adaptation, regularized with a variational information bottleneck to encourage compact preference representations. We instantiate LatentPersonal with LoRA, leveraging its low-rank parameterization as a natural low-dimensional adaptation space for personalization. By simply inserting a user-specific guidance vector between the shared low-rank factors, the model can navigate toward personalized adaptations through lightweight inference of this compact representation, without updating the shared LoRA parameters. Experiments across multiple personalization datasets demonstrate that LatentPersonal substantially reduces user-specific adaptation overhead while achieving effective personalization from only a few user-specific interactions, with particularly strong performance in the one-shot regime. --- ### SeOPD: Self-Evolving LLMs via Online Policy Distillation from Self-Generated Chain-of-Thought Authors: Xiaoshu Chen, Xiangyu Wong, Sihang Zhou, Ke Liang, Xinwang Liu Published: 2026-09-27 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.33181v1 Recent advances in online policy self-distillation (OPSD) have demonstrated that large language models (LLMs) can improve their capabilities by leveraging external privileged information (PI), such as manual annotations or feedback from external environments. However, obtaining accurate annotations and constructing sophisticated environments often require substantial human effort and computation, limiting the scalability of OPSD. While a few recent studies have explored self-improvement without external PI, the resulting gains remain limited. In this work, we explore whether LLMs can achieve comparable self-improvement without external PI. Our key observation is that a single LLM can support multiple reasoning modes, such as deep-thinking and non-thinking modes, with deep thinking generating additional information during reasoning. Based on this observation, we propose Self-Evolving Online Policy Distillation (SeOPD), which enables LLMs to distill and internalize information generated by their own chain of thought (CoT). Specifically, it (1) generates CoT with the deep-thinking mode, (2) produces responses with the non-thinking mode, and (3) uses the generated CoT as PI to provide token-level supervision for the non-thinking response, allowing new information inferred during reasoning to guide the non-thinking mode and be internalized into the shared model parameters, thereby improving both non-thinking and deep-thinking capabilities. Extensive experiments across LLMs and tasks demonstrate the effectiveness of SeOPD. --- ### Which Self-Improvements Should We Trust? Reliable Self-Improvement When Agents Reuse Their Benchmarks Authors: Xiaojing Sun, Yuhan Zeng, Zihua She, Xiao Wang Published: 2026-09-27 Categories: stat.ML, cs.AI, cs.LG, stat.ME arXiv: https://arxiv.org/abs/2609.33180v2 As recursive self-improvement (RSI) rapidly advances, reliable evaluation becomes critical for guiding adaptive search. RSI typically relies on finite evaluation resources, such as fixed benchmarks, to determine which modifications are retained and what is proposed next. However, when these finite resources are repeatedly reused, new candidates are proposed based on feedback from the same evaluation set, so the search trajectory can adaptively overfit and empirical improvement may not reflect genuine population improvement on the underlying task distribution. Some existing methods account for multiple comparisons but assume that candidates are chosen independently of the evaluation set, and therefore do not control this adaptive dependence. To address this, we propose REUSE (Risk-controlled Evaluation Under Sequential Evolution), a certified evaluation and promotion framework that allows a fixed evaluation set to support repeated adaptive decisions while providing statistical guarantees. For a user-specified error level $α$, with probability at least $1-α$, every promoted modification is a genuine population improvement on the underlying task distribution. REUSE achieves this by strictly limiting the evaluation feedback returned to the search process and accounting for possible promotion histories within the error budget. We develop detailed statistical theory for RSI evaluation in this setting, including simultaneous error control, valid lower bounds on cumulative improvement, and a characterization of the fundamental limits of adaptive evaluation reuse. In live self-improvement experiments, REUSE commits substantially fewer false promotions than evaluation frameworks from current RSI systems and error-controlled baselines, reducing the proportion of false promotions from up to 20.7% to 0%, while achieving final true population performance comparable to the best baselines. --- ### Scoring the Wrong Question: Readout Failures in Constrained-Option Evaluation Authors: Jiaxuan Guo, Kejia Zhang, Shuo Xin, Jingxin Yang, Youran Sun, Haizhao Yang Published: 2026-09-27 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.33179v1 Constrained-option scoring reads a model's probabilities for a fixed set of permitted answers, so it returns a score even when the model is about to write something else. We study a prompt that quotes a multiple-choice item, ending in the item's own answer instruction, and asks for a forecast of whether a reader model given only a short note will answer it correctly. The model can begin answering the quoted item: on three Qwen3 checkpoints an answer letter is the most probable token on nearly every passage, and the forecast ranks the reader's correctness indistinguishably from chance. Beyond prior work's argmax check and prefill repair, we contribute a label-free diagnostic of the declared options' support, their first-token mass in context; a deletion test tracing the Qwen3 answer-letter start to the quoted instruction; and rendering and tokenisation checks for dropped separators and coinciding first tokens. Prefilling an answer stem lifts the median option mass from below to above one half on all nine checkpoints scored with this prompt and, with options and renormalisation unchanged, raises the Qwen3-averaged AUC from 0.4929 to 0.6152. lm-polygraph's default P(True) estimator shows a related failure: it reads True where Qwen3 without thinking favours an answer letter (MMLU) or the prompt's (A)/(B) label (TriviaQA), and its Qwen3-averaged AUROC is indistinguishable from chance on MMLU and below chance on TriviaQA. Scoring and renormalising the (A)/(B) labels after an answer stem raises it to 0.632 and 0.868, renormalising True against False in place reaches 0.660 and 0.860, and on TriviaQA both exceed all fourteen default single-answer library estimators. The original forecast is renormalised too, yet stays indistinguishable from chance: option mass flags low support in both settings without labels, and only checking the intended target shows which score still ranks. --- ### Can Protein-Derived Knowledge Improve Pathology Foundation Models? Authors: Di Zhang, Zhangpeng Gong, Jiashuai Liu, Zhi Zeng, Jiusong Ge, Chunze Yang, Xitong Ling, Kai Yi, Kai He, Weimiao Yu, Mireia Crispin-Ortuzar, Chen Li et al. Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33178v1 Molecularly guided pathology foundation models (PFMs) exploit transcriptomic or proteomic information to enrich whole-slide image (WSI) representations, yet effectively leveraging large standalone molecular corpora remains challenging. First, existing molecular foundation models encode protein sequences or single-cell states, not the patient-level bulk expression profiles paired with WSIs. Second, because cross-modal supervision is restricted to paired WSI-omics samples, knowledge from standalone molecular corpora reaches the pathology encoder only indirectly, creating a paired-support bottleneck. To address these challenges, we propose a three-stage framework that decouples proteomic knowledge acquisition from cross-modal transfer, yielding ProSlide, a slide-level hierarchical pathology foundation model. First, to close the modality gap, we pretrain a Proteomic Foundation Encoder (PFE) on 12,695 sample-level bulk protein profiles using virtual profile generation and expression-space multi-view pretraining. Second, we pretrain ProSlide, a patch-region-slide encoder, to predict protein expression from paired WSI-protein samples. Third, to relax the paired-support bottleneck, we introduce Prot2Path, a cross-modal relational distillation objective. For each paired sample, it aligns the similarity distributions of the WSI and its protein profile over a shared, frozen bank of PFE-encoded paired and standalone profiles. We evaluate ProSlide on 12 downstream tasks across breast, lung, and renal cancers. Despite being pretrained with only 2,229 WSIs and 12,695 sample-level protein profiles, ProSlide achieves the highest mean accuracy and AUC within each cancer group. --- ### DeltaWAM: Change-Centric Visual Foresight via Delta Tokens for an Efficient World-Action Model Authors: Tianyun Jiang, Wenrui Bao, Bingxin Xu, Yu Tian, Yuzhang Shang Published: 2026-09-27 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.33177v1 World-Action Models (WAMs) offer visual foresight for robotic manipulation, but pixel-space models repeatedly reconstruct entire future scenes, incurring high computational cost and spatio-temporal redundancy. In physical manipulation, consecutive frames often share most of their visual context; the changes between them are what an action policy needs to anticipate. We introduce DeltaWAM, a change-centric WAM that makes a compact delta token the unit of future prediction. Each token is a single vector encoding changes between consecutive dense DINO feature maps. DeltaWAM builds on DeltaWorld, a latent world model pretrained on large-scale videos, to autoregressively predict one delta token per future frame. A flow-matching action expert then conditions on the predicted transitions and current DINO features, which serve as spatial anchors, to generate action chunks. Trained for 256 GPU hours on two H100 GPUs, DeltaWAM has 0.725B parameters and achieves 92.8% average success on LIBERO. It also shows robust generalization under procedural perturbations on LIBERO-Pro. Inference takes 142.1 ms per action chunk with 3.86 GB peak memory. --- ### ABO-Med: Accelerated Bilevel Optimization for Few-Shot Medical Image Classification Authors: Ruoxuan Shi, Sheng Yang, Zhengxing Su, Xiaoyang Hou, Yating Liu Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33176v1 In recent years, bilevel optimization has been widely used in a variety of machine learning tasks. However, prior bilevel optimization algorithms generally require the computation of second-order information, which limits their practical scalability. Only recently has a first-order paradigm for bilevel optimization been established, attaining near-optimal theoretical guarantees for solving bilevel optimization problems. In this paper, we propose ABO-Med, a scalable instantiation of this paradigm for few-shot learning, by incorporating it into the model-agnostic meta-learning (MAML) framework and tailoring it to medical image classification. We also introduce Medical Adaptive RandomAugment (MedRAug), a modality-aware augmentation strategy designed for medical images. Theoretically, ABO-Med establishes the optimality of MAML-type meta-learning approaches. Empirically, ABO-Med outperforms prior baselines on several public medical datasets, with gains of 1.99% to 18.76%, while MedRAug further improves the average accuracy by 2.20% to 6.34%. Additional cross-domain experiments, augmentation ablation studies, backbone ablation studies, and training efficiency analysis further validate the effectiveness and efficiency of the proposed method. --- ### Dynamic Manipulation with World-Action Models via Counterfactual Planning Authors: Sunwoo Park, Wonbin Lee, Seonghyun Jin, Youngmin Kim, Jangho Park, Jong Chul Ye Published: 2026-09-27 Categories: cs.RO, cs.AI, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.33172v1 World-Action models (WAMs) trained on static demonstrations often fail to manipulate moving targets even when they possess the required manipulation skills. We attribute this failure to target-response collapse: as execution advances, the policy becomes increasingly biased toward the learned continuation of its ongoing behavior and less responsive to target relocation. To bridge the gap between what the model has learned and what it can generate from the current context, we formulate dynamic manipulation as counterfactual planning by decoupling the context used for plan generation from the physical state used for execution. Our framework, Dynamic Predictive Planning (DPP), first uses the WAM's predictive rollout to estimate when an interaction is expected to occur, and combines this timing estimate with observed target motion to predict the target's future interaction position. DPP then constructs a counterfactual observation that places this predicted target position in a familiar robot context, allowing the model to invoke an existing manipulation skill rather than generate a recovery behavior from an unfamiliar robot-target configuration. The resulting plan is connected to the robot's actual state during execution. DPP enables real-time dynamic manipulation on a single consumer GPU without additional training on dynamic data. Experiments in simulation and on a real robot demonstrate consistent improvements across diverse target motions, with simulation performance surpassing all evaluated baselines, including methods additionally trained on dynamic data. Project page: https://methoder00.github.io/DPP/ --- ### Perturb-and-Solve: Efficient Learned-Operator Conditioning for Latent Diffusion Inverse Problems Authors: Abduragim Shtanchaev, Arip Asadulaev, Luiza Labazanova, Aidar Alimbayev, Karim Salta, Eric Moulines Published: 2026-09-27 Categories: cs.LG, cs.CV arXiv: https://arxiv.org/abs/2609.33171v1 Latent diffusion models serve as powerful priors for solving inverse problems in image restoration, such as deblurring, inpainting, and super-resolution. Current methods have a trade-off between generality and efficiency. Solvers that are restricted to a fixed set of degradation operators are fast and efficient. Methods that support arbitrary degradation operators are slow and require gradients through the diffusion network. To break this bottleneck, we introduce PASEO (Perturb-And-Solve for Efficient Operator conditioning), a method that uses a small (1M parameters) learned network to degrade diffusion model predictions in latent space. PASEO supports learned degradation operators without back-propagating through the diffusion network. We efficiently sample reconstructions from an approximate posterior by combining the diffusion model's prediction with the observed image. We do this by adding noise and solving linear equations based on a local linear approximation of the learned network, without building or inverting large covariance matrices. Across super-resolution, deblurring, and inpainting on FFHQ and COCO, PASEO achieves strong perceptual quality while running up to 9x faster and using up to 34% less peak memory than the tested baselines, with the same or fewer model evaluations. --- ### When Does Backpropagating Through Policy Memory Matter? Physical Credit, Optimizer Updates, and Observability Authors: Xingjian Li, Yi Han, Jianhua Z. Huang Published: 2026-09-27 Categories: cs.LG, cs.RO arXiv: https://arxiv.org/abs/2609.33169v1 Policies with memory can learn along two backward paths: through the physical states their actions produce and through the representations they store. Transformer-XL and truncated backpropagation through time cut the second path at stored history while keeping its values. We ask when this cut matters. Holding the forward computation fixed and varying only derivative edges, we measure parameter gradients, the updates the optimizer applies, and continued training in a Transformer vessel-trajectory model and a quadrotor tracking policy. In the vessel model, detaching the key-value cache shrank the gradient to about a tenth of its norm, with little rotation, when gradients flowed through all earlier physical states, but barely changed it under one-step physical credit. In this strongly clipped regime the optimizer, not the gradient, set how far updates differed: global-norm clipping removed most of the gradient difference between memory-cut graphs, whereas AdamW turned a 2% gradient difference between two placements of the cut into update differences of up to 31% at the step where the placement was switched. In a quadrotor trained from initialization with 0.20 m/s velocity noise, removing memory raised tracking error by 43% and cutting memory gradients raised it by 32%; at low noise the cut's mean cost exceeded the value of memory. Two-step truncation segments gave no measurable gain, although with hidden velocity a two-step window captured most of the value of memory; eight-step segments removed half to three quarters of the cost. Switching the cut on only for the last fifth of training understated its cost about threefold at 0.20-0.30 m/s, but not at low noise or with hidden velocity. These results suggest measuring the cost of a memory cut by training with it from initialization, and comparing backward graphs by the updates the optimizer applies rather than by raw gradients. --- ### FloodDiffusion 2: Efficient and Path Controllable Streaming Motion Generation Authors: Yiyi Cai, Yuhan Wu, Kunhang Li, Tu Fangyuan, Xiangyue Zhang, Qiaoge Li, Zhixiang Wang, Kaipeng Zhang, Haiyang Liu Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33167v1 We present FloodDiffusion 2 (FD2), an efficient and controllable framework that builds upon FloodDiffusion (FD1), a state-of-the-art streaming motion generation model. While FD1 produces plausible motion, it suffers from low efficiency and limited controllability, as its attention design requires repeated computation over the entire history, and it lacks precise trajectory control for real-world applications. To address these limitations and improve generation quality, FD2 introduces three advances. First, Partial Attention makes finalized history representations independent of the active window, enabling KV-cached inference and shared-history packing for efficient training. Second, we establish a necessary-and-sufficient Bregman criterion for regression losses to preserve diffusion's conditional-mean velocity field. This criterion guides an FK-induced quadratic loss that incorporates motion geometry without online FK evaluation. Third, FD2 introduces precise path conditioning to control the character's root trajectory while preserving natural body motion. Experiments show that FD2 reduces training computation by 4.6$\times$ and accelerates denoising by 11.29$\times$, reaching 2.303 ms per update on long sequences. Alongside these efficiency gains, FD2 improves motion quality over FD1 and achieves state-of-the-art FID scores among streaming methods, with 0.048 on SEED and 0.053 on HumanML3D. --- ### Downstream-Aware Context Selection for Online In-Context Reinforcement Learning Authors: Ruihan A. Li, Shangtong Zhang, Rohan Chandra Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33166v1 In-context reinforcement learning (ICRL) enables large language model agents to adapt to new environments using their interaction history without updating model parameters. However, repeatedly conditioning on growing histories can lead to substantial token cost. We propose a bounded-history context-management framework that predicts the task-dependent downstream effect of removing historical interactions to guide history selection and determine a decision-dependent context budget. Formally, our framework uses the full rolling history as a reference. The predictor evaluates removal effects, defines a deletion ordering, and applies a shared selection criterion to determine how much history to retain at each decision. We evaluate the method in closed-loop SUMO driving under held-out in-distribution, unseen-domain, and unseen-route settings, and in ScienceWorld under a continual ICRL protocol. Relative to a baseline using the full context, our method reduces total token usage by 25.7%, 25.8%, and 23.2% across the three driving settings while maintaining comparable closed-loop driving performance. In ScienceWorld, it reduces total token usage by 52.1% compared to full context and uses 30.2% and 37.8% fewer tokens than the Recent and Similarity baselines, respectively, while maintaining performance. --- ### Beyond Tasks: A Vision for Reproducing an Animal-like Behavioral Substrate Using Modern Robot Learning Techniques Authors: Samiyuru Menik, Hemadri Jayalath Published: 2026-09-27 Categories: cs.RO, cs.AI, cs.HC arXiv: https://arxiv.org/abs/2609.33165v1 Recent advances in robot learning have produced increasingly capable embodied agents. Yet comparatively less attention has been given to a more basic form of competence that animals exhibit continuously: the ability to remain situated, responsive, and behaviorally coherent as physical, environmental, and social demands change over time. We propose the ethological behavioral substrate as a conceptual lens for studying this form of competence in artificial agents. Rather than treating these behaviors that animals exhibit as a set of isolated skills, we argue that their continual coordination under competing demands constitutes an important and underexplored target for modern robot learning. We further propose robotic animal companions as a useful research setting for studying sustained interaction and adaptation in human-centered environments. Such systems provide an opportunity to investigate how social behavior, memory, and continual learning develop over long periods of interaction. This perspective motivates further investigation of how such persistent behavioral competence may complement higher-level capabilities in embodied agents. --- ### Notes on Generative Modeling for Feedback Control and Planning Authors: Karthik Elamvazhuthi Published: 2026-09-27 Categories: math.OC, cs.LG, eess.SY arXiv: https://arxiv.org/abs/2609.33164v1 In these notes, we view control as a dynamically constrained sampling problem on the state-space of a control system. With this viewpoint, we extend methods from generative modeling, such as flow matching, normalizing flows and denoising diffusions to control problems. Concepts such as controllability, optimal control and trajectory planning play an important role in guiding the extension and understanding well-posedeness of the corresponding algorithms, with application to steering systems to target states or distributions and sampling from reachable sets. The notes are intended as an accessible introduction for readers with a background in control theory and robotics. --- ### FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs Authors: David Restrepo, Chenwei Wu, Luis Filipe Nakayama, Miguel L. Martins, Stergios Christodoulidis, Maria Vakalopoulou, Enzo Ferrante Published: 2026-09-27 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.33158v1 Progress in AI-based retinal image analysis has advanced with foundation models, yet evaluating their reliability remains challenging. Performance reported on a single dataset does not capture how models behave under dataset shift, across clinical definitions, or for different patient subgroups. This limitation is particularly critical in medical imaging analysis, where robustness, calibration, and fairness are essential for safe deployment. We introduce FOCUS (Foundation Ophthalmic Cross-Dataset Understanding under Shift), a cross-dataset benchmark for evaluating retinal fundus models that considers vision-only encoder models (VM), vision-language dual-encoder models (VLM), and multimodal large language models (MLLM). FOCUS harmonizes binary diabetic retinopathy, referable diabetic retinopathy, and glaucomatous optic neuropathy tasks across ten public datasets spanning diverse geographies, acquisition conditions, and label protocols. The benchmark evaluates models through a unified analysis layer that measures ranking performance, calibration, subgroup disparities, and image-quality robustness. We present a large-scale evaluation covering 532 base configurations and 228 MLLM configurations adapted through supervised fine-tuning with low-rank adaptation (LoRA). Results show that no model family consistently dominates across tasks and datasets: general VM encoders achieve the strongest average ranking performance, medical MLLMs are competitive but variable, and dual encoder VLMs benefit substantially from lightweight adaptation. Fine-tuning improves in-domain performance but exhibits heterogeneous transfer to external datasets, particularly in calibration. These findings demonstrate that retinal model evaluation is inherently multidimensional. FOCUS provides a practical framework and public benchmark to assess generalization, reliability, and robustness beyond single-dataset leaderboards --- ### Where Do Test-Time Scaling and Training Fall Short in Individual Stance Prediction? Authors: Yuyang Zhao, Xuan Liu, HaoYang Shangm Haojian Jin Published: 2026-09-27 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.33155v1 Test-time scaling and post-training have improved LLM performance in coding and mathematical reasoning, but their effectiveness for individual stance prediction remains unclear. We study this question by predicting a person's stance in a new discussion from their history. We evaluate widely used test-time scaling strategies and post-training methods, such as supervised fine-tuning and reinforcement learning, and identify four failure modes across generation, selection, and learning: (1) incorrect consensus, where repeated samples agree on the wrong stance; (2) selection failure, where generation covers the observed stance but selection misses it; (3) response overfitting, where supervised fine-tuning improves imitation but harms prediction; and (4) early plateau, where reinforcement learning shows modest initial gains followed by limited further improvement. We expose these failures using STANCE-BENCH, which contains 2499 prediction tasks from 500 Hacker News users. Guided by this analysis, we explore a simple approach that combines direct scores for all candidate stances with explicit assessments of support from the individual's history. On the 781-task test set, this approach achieves 21.83 discussion-specific Macro F1 with Qwen3-8B, compared with 19.27 for direct scoring. Our results motivate evaluating candidate generation, final selection, and person-specific evidence use separately. --- ### Sharp Critical Minimax Laws and No-Learning Thresholds in Continuous-Time Adaptive Control Authors: Chen Jia Published: 2026-09-27 Categories: math.OC, cs.LG arXiv: https://arxiv.org/abs/2609.33154v1 We study episodic continuous-time control with an unknown vector control gain, scalar state, quadratic action cost, and smooth convex terminal cost. In the scalar Gaussian experiment, let $Δ(H)$ denote the minimax improvement over zero control and set $δ=\sqrt2H^2-1$. We prove the critical law $$ Δ(H)\asymp δ^4\sqrt{\log(1/δ)} \qquad (δ\downarrow0), $$ with a matching lower and upper bound. The lower bound follows from a uniform deficit--energy inequality valid for fully adaptive controls with unbounded amplitudes, while a moving soft-threshold feedback attains the rate. For local parameters $θ=N^{-1/4}h$, $|h|\le H$, we show that the normalized minimax regret over $N$ episodes is within $O(N^{-1/2})$ of a fixed-horizon Gaussian sequential control problem, without an additional dimension factor. The terminal task enters the limit only through $c_g=\operatorname{Var}(g(Z))$. The Gaussian problem exhibits an exact no-learning phase transition: $$ C_d^T(H)=TH^2/2 \iff H^4T\le d^2/2, $$ yielding the asymptotic task boundary $c_gH^4=d^2/2$. --- ### What Does a Skill Actually Do? Estimands and Evaluation Validity for Tool and Skill Use in LLM Agents: A Critical Review Authors: Shuyang Zhang Published: 2026-09-27 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.33153v1 Large language model (LLM) agents increasingly draw on external tools and reusable skills selected at run time from libraries that hold thousands of entries. Reports that a retriever, router, or skill library "improves" an agent may refer to retrieval recall, the success change from enabling a library, a paired contrast restricted to triggered tasks, or a gain under an approximate budget constraint. This critical review asks what each design compares and under which assumptions. Building on estimand-based approaches to agent evaluation, we describe tool and skill designs along six axes: treatment contrast, target population, outcome, budget constraint, summary measure, and identification assumptions. Thirteen core empirical studies anchor the evidence synthesis, supplemented by related methodological work and design-level reading of the wider literature. Our contribution is to make explicit distinctions that some source authors already acknowledge through a decomposition of trigger-conditioned pairing, analytic counterexamples, and comparisons across studies. Pairing on the task does not by itself identify an invocation effect; paired gain and regression counts measure protocol-specific discordance rather than the share of tasks whose expected outcomes worsen; and total effects of module deployment answer a different question from budget-constrained efficiency. We compare curated skill provision with retriever replacement, triggered subsets with all-task outcomes, and observed cost reductions with budget-constrained comparisons. A reporting checklist and worked examples connect these distinctions to information that studies can report. The review runs no new experiments; empirical results come from the cited studies, and numerical toy examples are analytical illustrations. --- ### Convergence of Practical Muon Authors: Haonan Wang, Yu Wu, Minghui Liwang, Xinlei Yi, Yiguang Hong Published: 2026-09-27 Categories: cs.LG, math.OC arXiv: https://arxiv.org/abs/2609.33152v1 Muon is emerging as a promising alternative to AdamW for large-scale neural network training, yet theoretical understanding of its practical implementation remains incomplete, as existing analyses often simplify or omit two key components: (i) practical Newton--Schulz iterations with empirically tuned polynomial coefficients $(3.4445,-4.7750,2.0315)$; and (ii) decoupled weight decay for regularization. In this paper, we provide an optimization interpretation and establish convergence for practical Muon, jointly accounting for both components. Specifically, we interpret practical Muon as right-preconditioned optimization of the original loss with a dynamic weighted $\ell_2$ regularizer that vanishes as stationarity is approached, so that the optimization target remains the original objective. We then establish, to our best knowledge, the first convergence guarantee for practical Muon in the stochastic nonconvex setting, with an $\mathcal{O}(T^{-1/4})$ convergence rate in terms of the expected Frobenius norm of the gradient, improving the dimension dependence of the best known AdamW's convergence rate by a factor of $\sqrt{d}$, where $T$ is the iteration horizon and $d$ is the parameter dimension. Experiments further support the theoretical convergence results. --- ### Generalization Dynamics of LM Pre-training Authors: Jiaxin Wen, Zhengxuan Wu, Dawn Song, Lijie Chen Published: 2026-09-27 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.33150v1 People typically assume that LMs stably mature from pattern-matching parrots to generalizable intelligence during pre-training. We build a toy eval suite and show this mental model is wrong: throughout pre-training, LMs frequently and suddenly hop between parrot-like and intelligence-like computations. We call this mode-hopping. Across our suite, LMs suddenly latch onto memorized or in-context patterns instead of in-context learning, use System 1 instead of System 2 thinking, pick up what sounds true instead of what is true, fail at multi-hop persona QA, out-of-context reasoning, and emergent misalignment -- then just as suddenly revert and generalize. Mode-hopping is not explained by standard optimization dynamics: it is locally stable and cannot be fixed by checkpoint averaging. We instead think of it as a capacity allocation problem: in a capacity-bounded model, generalizable circuits must compete with the shallow ones learned early in training, and the data in each pre-training window may decide which circuits win. Our suite provides a new efficient lens on generalization. We demonstrate two concrete applications: (i) select intermediate pre-training checkpoints that strongly generalize reasoning and alignment, better than the final pre- or mid-training checkpoints, and (ii) select pre-training data that controls and stabilizes generalization dynamics. --- ### Not Too Hard, Not Too Easy: Learning from Intermediate States for LLM Structured Reasoning Authors: Hongbo Chen, Guohua Lu, Ting Dang, Hong Jia Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33149v1 A common principle of effective learning is to practice material that is neither already mastered nor too difficult to permit progress. We ask how to apply this principle to structured reasoning tasks such as Sudoku and maze solving. In these tasks, a model can repeatedly revise an incomplete or incorrect candidate solution until it satisfies the problem's constraints. The intermediate candidate solutions along this trajectory provide natural training examples: some are already solved, some cannot yet be repaired by the model, and others lie at its current frontier of achievable progress. We therefore investigate whether pretrained language models can learn to revise such states and whether training on states at this frontier improves reasoning more broadly. To achieve this, we couple a pretrained language-model backbone with a recurrent updater that repeatedly revises an explicit solution state, using the same parameters at every update step. We further introduce Frontier-Oriented Curation Using Self-trajectories (FOCUS), which selects training states from trajectories generated by the current model. FOCUS measures how much the model improves each state within a fixed number of recurrent updates and prioritizes states from which it can make substantial progress. With Qwen3-1.7B, FOCUS achieves 64.4% exact solve accuracy on Sudoku-Extreme and 91.1% on Maze-Hard, with similar gains observed across five Qwen and Llama backbones spanning 1.7B to 8B parameters. We further observe zero-shot transfer in the adapted LLM to mathematical reasoning and code execution, even when the recurrent updater is disabled and no downstream fine-tuning is performed. --- ### DroneWAM: Efficient World Action Model for Drone Visual Navigation Authors: Liang Yao, Fan Liu, Hongbo Lu, Wei Xu, Jianyu Jiang, Yijun Shen, Chuanyi Zhang, Pai Peng Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33148v1 World-action models give visual navigation agents a way to anticipate how candidate actions will change future observations and to act from the predicted consequences. For drones, this capability must operate under tight accuracy and efficiency constraints. We present DroneWAM, an efficient world-action model for drone visual navigation. DroneWAM adopts a JEPA-based architecture to model future states directly in representation space, avoiding the cost of explicit future image generation. A pretrained Resampler further compresses dense encoder features into fewer latent tokens, reducing the computation repeated at each imagined step. We also introduce adaptive rollout, where a preference-trained Gate adaptively allocates prediction depth according to the current scene. To support learning under richer aerial motion, we construct DroneNav-6D, a simulated visual navigation dataset with synchronized RGB observations, 6-DoF flight trajectories, control commands, and randomized wind disturbances. On DroneNav-6D, DroneWAM achieves the best trajectory accuracy among the compared methods. Adaptive rollout further reduces the average prediction depth from 8 to 4.58 while improving trajectory accuracy, demonstrating that predictive computation can be allocated more effectively across scenes. \href{https://github.com/1e12Leon/DroneWAM}{Codes and data} will be released. --- ### CFLoRA: Federated Fine-tuning of LLMs with Complementary Factors for Error-free Aggregation Authors: Yanan Ma, Qiyuan Chen, Zihan Fang, Xianhao Chen, Yuguang Fang Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33147v1 Federated low-rank adaptation (LoRA) enables collaborative fine-tuning of large language models without centralizing private client data. Its factorized update, however, creates a structural mismatch in federated averaging: averaging the two LoRA factors separately does not equal averaging their products. Existing exact methods resolve this issue mainly by freezing an entire factor or alternating factors across rounds, but none can update factors simultaneously without aggregation errors or expanding communication ranks. To address this fundamental problem, we present \texttt{CFLoRA}, a federated LoRA scheme that partitions latent LoRA channels into two complementary sets in every communication round. By ensuring that columns and rows are complementary across factors, we eliminate bilinear terms in matrix multiplications, making federated aggregation exact. Crucially, our framework also supports clients with heterogeneous rank budgets. Convergence analysis validates \texttt{CFLoRA} achieves $\mathcal{O}(1/\sqrt{T})$ convergence rate of the \textit{original} LoRA objective in homogeneous-rank cases. Extensive experiments with RoBERTa on the GLUE benchmark and with LLaMA-3.2-3B-Instruct on commonsense reasoning tasks demonstrate that \texttt{CFLoRA} achieves superior performance and training efficiency compared to state-of-the-art federated LoRA baselines. --- ### LiteEvo: Automated, Cost-Efficient Harness Evolution for Generalization to Unseen Tasks Authors: Euntae Choi, Sumin Song, Sungjoo Yoo Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33146v1 An LLM agent is defined by two things: the weights inside its model and the harness of components assembled around it. Harnesses are still handcrafted, and HarnessX, which evolves them automatically, starts each benchmark from a handcrafted harness, reports gains on the tasks it evolved on, and budgets 100 to 175 million meta-agent tokens per benchmark. We propose LiteEvo, a lightweight harness-evolution algorithm whose tool-free meta-agents mine agent trajectories for reusable components, curate them into a versioned library, and compose each round's harness from it, starting every benchmark from the same neutral harness and never naming the benchmark. Evolving on the graded tasks of five agentic benchmarks with a frozen Qwen3.5-9B, LiteEvo lifts pass@2 by 10.5 to 67.7pp and reaches comparable or higher pass@2 than a reproduction of HarnessX (71.0 against 67.3 on average) at 13.0 lower mean API cost. Harnesses evolved on train tasks keep their gains on unseen test tasks of four benchmarks, and LiteEvo also lifts Claude Code with Sonnet 4.6 by 1.2 to 71.4pp. ---