# Research Papers — 2026-09 · bundle page 57 of 266 > Full-text markdown of papers #2801–2850 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/57 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/56.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/58.md) ### Pretraining Transformers with Quantized Softmax in Attention Authors: Shangzhen Zhu, Muyan Hu, Tomasz Kozlowski Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33591v1 Low-precision Transformer systems increasingly quantize attention matrix multiplications, while softmax often remains at higher precision. During pretraining, an approximate softmax changes the gradients that train the model as well as its forward computation. We study this interaction with K-interval attention, which approximates the exponential using K+1 grid values. We vary per-row grid calibration, interpolation versus hard rounding, and the placement of a straight-through surrogate relative to normalization. We derive the corresponding backward rules, including calibration derivatives, and compare these choices in pretraining experiments matched on model, data, and optimizer. Detaching the row extrema leaves the forward computation unchanged but produces a delayed increase in validation loss. With hard rounding at K=4, min-max calibration and a pre-normalization surrogate incur a large loss gap; changing either choice substantially reduces it. At 124M parameters and 2.5B training tokens, fixed-window calibration with a post-normalization surrogate yields a validation loss gap of +0.019 nats relative to softmax at K=4, and with a pre-normalization surrogate yields +0.004 nats at K=16. --- ### TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs Authors: Zihan Lin, Xiaohan Wang, Jie Cao, Jiajun Chai, Wei Lin, Guojun Yin, Ran He Published: 2026-09-27 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.33589v1 Efficient exploration often remains a central bottleneck in reinforcement learning with verifiable rewards (RLVR). Although temperature control and test-time scaling strategies can increase rollout diversity of large language models (LLMs), they either expand the sample budget at rollout time or leave the benefit of exploration unquantified. To this end, we propose Temperature-Grouped Reinforcement Learning (TGRL), which turns temperature-induced diversity into an explicit training signal. For each prompt, TGRL partitions its rollout group into low- and high-temperature subsets, estimates exploration gain through their reward contrast, and allocates this group-level signal as token-level credit using Jensen--Shannon (JS) divergence between the corresponding temperature-scaled next-token distributions induced by the same logits. Notably, TGRL reaches equivalent accuracy up to 36% faster than strong RLVR baselines without expanding the rollout budget. Across 11 benchmarks from diverse domains, TGRL broadly improves over strong RLVR baselines: it improves the six-benchmark math average by 1.6% at 32B, raises CodeForces rating by 196.7 points and LiveCodeBench Pass@16 by 4.4%, and improves ALFWorld/WebShop success rates by 6.3%/4.9%. Comprehensive ablations and wall-clock analysis confirm the efficacy of all proposed components. Code is available at https://github.com/1229095296/TGRL/tree/main. --- ### Approximating Softmax in Pretrained LLMs: Model Sensitivity and Kernel Acceleration Authors: Shangzhen Zhu, Muyan Hu, Tomasz Kozlowski Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33586v1 On NVIDIA Blackwell B200, tensor-core throughput outpaces special-function exponential throughput by more than two orders of magnitude, exposing exponential evaluation in fused attention kernels. A pretrained Transformer, however, may not need it evaluated accurately at every element. We characterize what a pretrained model does need by approximating softmax at inference in ten frozen decoder-only models (0.5B-72B). The number of positions the softmax map assigns probability to and within-row resolution can be cut substantially, yet uniform weighting of the same positions is damaging. Where a fixed resolution budget is placed matters as much as its size, with resolution near the row maximum consistently favored. Perturbations matched on scalar distortion produce model-dependent responses of opposite sign. These findings motivate Rowmax-PoT, a coarse logarithmic weight representation anchored at each row maximum, and Rowmax-H15, its hardware specialization in FlashAttention-4. On B200, the patched FP8 attention forward is 12.4% faster at causal 8K and 25.8% faster at non-causal 8K in host-side call-latency measurements; board energy per forward falls by 8.4% at causal 16K. Measured separately on the BF16 kernel path at 2K, Rowmax-H15 increases perplexity by 0.091-0.492% across five models from three families. --- ### IVT-Guard: All-in-One Reasoning Model for AI-Generated Content Detection Authors: Hongwei Niu, Yunpeng Luo, Hanjun Li, Ziyin Zhou, Jianghang Lin, Ke Yan, Shouhong Ding, Shengchuan Zhang, Liujuan Cao Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33585v1 The rapid proliferation of highly realistic AI-Generated Content (AIGC) necessitates robust and interpretable detection mechanisms. However, existing detectors are predominantly confined to single modalities and provide binary outputs without reasoning. While Multimodal Large Language Models (MLLMs) present a promising solution, their development is constrained by the scarcity of multimodal reasoning data and the reasoning-detection optimization dilemma, where explicit reasoning supervision can compromise detection accuracy. To this end, we introduce IVT-Set, a comprehensive dataset comprising over 152K diverse image, video, and text samples equipped with multi-granularity Chain-of-Thought (CoT) reasoning trajectories. Based on it, we propose IVT-Guard, a pioneering framework for unified and interpretable AIGC detection across image, video, and text modalities. Furthermore, to overcome the aforementioned optimization dilemma, we design a novel three-stage training paradigm: Artifact-Aware Pre-training, Artifact-to-Evidence Supervised Fine-Tuning via artifact-aware injection, and Evidence-Verdict Consistency Group Relative Policy Optimization. Extensive experiments demonstrate that IVT-Guard achieves state-of-the-art detection performance across in-domain, out-of-domain, and cross-dataset settings while delivering faithful reasoning. Code and data will be released. --- ### Fill2SR: Repurposing Inpainting Diffusion Transformers for Real-World Super-Resolution Authors: Xingfu Yi, Xiaoxue Yu Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33582v1 Recent real-world image super-resolution (SR) methods often adapt text-to-image (T2I) backbones with ControlNet-style branches or spatial conditioning tokens, which increases memory and computes with resolution and often constrains training to a fixed scale. We propose Fill2SR, which repurposes a masked-inpainting Diffusion Transformer for SR without extra spatial branches. Our Inpainting-Interface Evidence Adapter (IIEA) writes the low-quality (LQ) observation into the native masked-image slot under a full-image mask, turning inpainting into a reverse-degradation conditional rectified flow trained with LoRA-only tuning. We further introduce RCDT, an offline pipeline that distills degradation descriptors from unpaired real images and transfers them onto clean targets using frozen open-source models. Fill2SR supports mixed-resolution training up to QHD and yields stable performance across $512/1024/2048$ outputs. On synthetic benchmarks, our base model with IIEA achieves the best LPIPS on DIV2K and LSDIR; adding RCDT trades a small LPIPS drop for consistently stronger no-reference quality on RealLQ250 and RealPhoto60. Fill2SR remains memory-predictable, running $1536^2$ inference on a single 32GB GPU and extending to multi-megapixel outputs via tiled restoration. --- ### ForeFly: A Dual-Horizon World Action Model for Aerial Vision-Language Navigation Authors: Kunhui Wang, Xintong Zhang, Junyu Gao, Changsheng Xu Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33581v1 Aerial Vision-Language Navigation (AVLN) requires UAVs to maintain reliable instruction following over long trajectories in complex 3D environments. However, existing AVLN approaches are predominantly reactive or limited to single-horizon prediction, overlooking complementary future cues across different temporal horizons. To address this limitation, we propose ForeFly, a dual-horizon latent world action model that predicts both a proximal future for local continuity and an adaptive route-critical future for long-range guidance. Horizon-specific foresight queries are primed with recent and route-critical visual memories, providing history-aware context for future prediction. To exploit their distinct roles in action generation, we introduce Foresight-Guided Action Refinement (FGAR), which asymmetrically exploits proximal foresight for local action enhancement and route-critical foresight for feature-wise correction and route-level guidance. Experiments on the TravelUAV and UAV-ON benchmarks show that ForeFly consistently outperforms strong baselines across seen and unseen settings, validating the effectiveness of dual-horizon foresight and FGAR learning. The code is available at: https://github.com/kunhuiW/ForeFly --- ### Compressing Value Predictions for Learning-Augmented Metrical Task Systems Authors: Sizhe Li, Yecheng Li, Kun He Published: 2026-09-27 Categories: cs.LG, cs.AI, cs.GT arXiv: https://arxiv.org/abs/2609.33580v1 Learning-augmented algorithms for metrical task systems (MTS) can exploit predictions of canonical dual values, but existing formulations typically require a prediction for every state. We study whether these predictions can be compressed to a small set of representative states while retaining their algorithmic value. We introduce landmark-compressed value predictions, in which the predictor reports predicted dual values only at $m$ landmarks and the remaining values are reconstructed by a Lipschitz extension. Our algorithm achieves additive excess cost $O(T\,r(L) + \sum_t δ_t)$, where $r(L)$ is the covering radius of the landmarks and $δ_t$ measures prediction error up to additive shifts; local and value-dependent bounds refine this guarantee. For sparse landmark sets on unit-spaced finite lines, we prove a matching $Ω(T r_m)$ lower bound for every randomized algorithm using fixed landmarks, even with advance access to their entire exact absolute-value table. The prediction interface also matters: on two states with one landmark, exact absolute values permit horizon-independent excess, whereas exact relative values force worst-case expected excess linear in $T$. We give PAC guarantees for learning compressed prediction tables, with efficient empirical-risk minimization for fixed landmarks. Our results connect metric coverage, prediction interfaces, and learning guarantees for compressed predictions in online MTS. --- ### OpenFC: Learning Verification Policies towards Open-Search Fact Checking Authors: Xinming Wang, Kaixiang Qiu, Yansong Lin, Chunji Lv, Yi Chen, Boran Wang, Hong-Ming Yang, Xu-Yao Zhang Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33579v1 Open-search fact checking is not merely retrieval followed by classification, but a sequential decision problem in which every query, source visit, and stopping decision reshapes the evidence available for verification. Yet existing systems often distribute these decisions across predefined pipelines or separately prompted modules rather than learning them as a unified task-specific policy. We introduce \textbf{OpenFC}, a unified verification-policy training framework that post-trains Qwen3-8B as a compact next-action controller over reasoning, evidence acquisition, and stopping. OpenFC learns this policy in two stages. \textbf{Stepwise-Calibrated Cold Start (SCCS)} uses a strong training-time supervisor to review post-initial reasoning, tool-use, and stopping proposals before execution, producing reliable trajectories for supervised fine-tuning without access to gold verdicts. \textbf{Verification-Aware Reinforcement Learning (VA-RL)} then improves the cold-start policy on unresolved claims through budget-aware tool rewards, label-aware advantage reweighting, and localized response masking. Across six fact-checking benchmarks, OpenFC achieves 70.39\% average accuracy and 63.30\% macro-F1, the highest overall averages among the evaluated methods. Stage-wise ablations further show that SCCS and VA-RL provide complementary gains, supporting the design of the two-stage training framework. These results position OpenFC as a strong and effective framework for open-search fact-checking. We will open-source our code and release the model checkpoints to support reproducibility. --- ### HiLoRe: What to Store, Compress, or Recompute for Efficient GRPO Training Authors: Xinrui Chen, Mengyang Li, Ou Wu, Ji Zhang Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33570v1 Group-relative policy optimization (GRPO) makes learner-side activations a major memory-computation bottleneck: gradient checkpointing reduces activation memory through recomputation, but fixed schedules can leave roughly 18 GB unused on a 48-GB GPU despite substantial recomputation overhead. Existing activation-management methods set state fidelity from execution cost, tensor properties, or generic compression sensitivity, without explicitly incorporating GRPO's analytic update structure into state-fidelity allocation. We formalize this dependence as policy-update exposure, linking the current GRPO loss coefficients to state-level approximation sensitivity. These coefficients are available before backward without an additional backward pass. We introduce HiLoRe, which allocates graph-attributed recovery units among high-precision storage, low-precision compression, and deterministic recomputation using measured recovery utility and update-conditioned approximation risk. It combines high-precision storage and deterministic recomputation with low-precision recovery under a calibrated risk budget. Across five model-task settings with 2K responses and memory < 1.10 times GC's per-GPU actor-update peak, HiLoRe's actor-update throughput gains reach 13.5% over GC and 7.9% over the fastest evaluated baseline, with paired mean downstream-score differences below 0.6 percentage points. --- ### Terminal-Register Certification for Finite-Measurement Learning of Multiscale Quantum States Authors: Bhvain Makwana, Kashyap Patel, Manjunath Joshi, Jaideep Mulherkar Published: 2026-09-27 Categories: quant-ph, cs.LG arXiv: https://arxiv.org/abs/2609.33567v1 Structured quantum-state learning not only depends on an expressive ansatz but also on an operational certificate that stays meaningful with finite measurements and imperfect implementation. We study pure one dimensional states learning by an inverse binary multiscale entanglement renormalization ansatz (MERA). In the learning procedure, the qubits removed during coarse graining are controlled coherently and measured together at the terminal register. We confirm that an ideal sequential and terminal measurement schedule delivers the same complete bit string distribution under matched causal operations, while normalized postselection can amplify perturbations inversely with prefix acceptance. A noise aware theorem introduces an individual calibrated total variation implementation budget to the finite shot certificate. The protocol is estimated on an open boundary transverse field Ising ground state. A frozen 8-qubit schedule using $560$ million simulated training measurements per run achieves fidelity above $0.99$ in all $60$ held-out runs, with a mean fidelity of $0.996886$. 1080 circuit-noise cells and 6480 confidence-coverage rows are covered by fixed-circuit robustness validation without a locked soundness violation. We then address architectural fairness at $n=16$ using three new studies. In a 120-run exact-gradient multistart diagnostic, MERA has higher fidelity in 58/60 paired restarts and lower long-range error in 60/60, although no run met the prespecified stationarity criterion. Finally, a causal cone-complete, parameter matched local circuit achieves $2.62\times$ greater aggregate gate exposure yet loses all 30 paired comparisons in fidelity, long-range error, energy, and entropy. --- ### Correct then Forecast: Observer State-Space Models for Time Series Forecasting Authors: Alexis-Raja Brachet, Guillaume Clavier--Frémond, Abdelhakim Ziani, Pierre-Yves Richard, Céline Hudelot Published: 2026-09-27 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.33566v1 Time series forecasting requires extrapolating the dynamics of an observed process beyond the last available measurement. Yet recurrent forecasting models typically treat observations as inputs that directly control their latent dynamics. It leads to a regime change when these observations become unavailable at prediction time. Following a state-estimation perspective, we introduce Observer State-Space Models (OSSMs), a class of recurrent models that interprets the observed input time series as measurements of an underlying autonomous dynamical system. OSSMs explicitly separate latent-state propagation from measurement assimilation: a single transition governs the dynamics across both context and forecasting intervals, while available observations correct the estimated state through an observer. This formulation naturally exposes classical control-theoretic properties, including observability and convergence of the state estimation error. We further show that conventional and recent SSMs can be recovered as particular instances of our OSSM framework, thereby providing a unified interpretation of their recurrent dynamics and revealing modeling inconsistencies. We perform experiments across several benchmarks showing that OSSM achieves substantial improvements while maintaining the same parameter count and training setup as the corresponding SSM baseline. These results support a simple principle for recurrent forecasting: observations should correct the estimated latent state, rather than control the dynamics used to propagate it. --- ### Dr. Free: You Don't Need Difficulty Rewards for Self-Evolving Search Agents Authors: Zhipeng Qian, Zihan Liang, Yufei Ma, Jie Ma, Ben Chen, Huangyu Dai, Lingtao Mao, Xinyu Sun, Tong zhao, Xuxin Zhang, Qingpeng Cai, Peng Jiang et al. Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33565v1 A central limitation of current data-free self-evolution methods for training search agents is their reliance on difficulty-based proposer rewards. These methods reward a proposer for generating questions that challenge a co-evolving solver, using solver difficulty as a proxy for question quality. Yet difficulty alone is insufficient to distinguish questions that require cross-passage evidence from those that are answerable via simpler shortcuts. In addition, measuring difficulty demands repeated solver rollouts for every candidate question, leading to substantial computational costs. In this paper, we introduce \methodname, the first self-evolving search framework that eliminates difficulty-based proposer rewards and directly optimizes for evidence necessity relative to shortcut contexts. Dr. Free samples relational chains from a knowledge graph and pairs them with aligned passages, giving question generation an explicit multi-hop structure. A generated question receives a positive information-gain reward only when the likelihood of the target answer under the complete evidence passages exceeds the maximum likelihood under all evaluated shortcut contexts. Because this signal is computed from teacher-forced likelihoods, it removes the need for pass-rate estimation and reduces proposer training time by over $7\times$. Experiments on seven open-domain QA benchmarks show that Dr. Free outperforms prior data-free search agents and the supervised baseline, with large improvements on multi-hop QA benchmarks. --- ### MA-JEPA: Joint-Embedding World Models for Multi-Agent Reinforcement Learning Authors: Brandon Gary Kaplowitz, Osaze James Obahor, Christian Schroeder de Witt Published: 2026-09-27 Categories: cs.LG, cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.33563v1 World models improve sample efficiency by training policies on imagined trajectories, but their usefulness depends on learning representations that capture the information needed for future control. We study whether self-supervised joint-embedding prediction (JEPA) can provide this learning signal for multi-agent reinforcement learning. We introduce MA-JEPA, a stochastic world model that replaces observation reconstruction with prediction of target representations, enabling model-based multi-agent reinforcement learning with centralized training and decentralized execution. A categorical latent state and a causal Transformer are trained with posterior and action-conditioned dynamics prediction objectives and are then used for actor-critic learning from latent imagination. A training-only joint predictor conditions on all agents' local states and actions to predict each agent's next local observation embedding. These predictions are passed through the same local posterior used during real interaction with a centralized critic that is used only for value learning, with execution remaining decentralized. Our experiments show that this architecture performs strongly on SMAC, matching or exceeding the strongest reported comparator mean win rate on four of eight evaluated maps. --- ### OOD Generalization as a Bifurcation Problem Authors: Nguyen-Thanh-Luong Doan, Quang-Vu Nguyen, Tang-Phu-Quy Le, Cong-Phap Huynh Published: 2026-09-27 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.33562v1 Systematic out-of-distribution (OOD) generation remains a critical bottleneck for continuous-time generative models. While standard joint classifier-free guidance (CFG) routinely fails to synthesize unobserved concept combinations, exact decomposed scoring generalizes robustly at the cost of severe computational overhead. In this work, we reveal that compositional binding is not a uniform process but a highly localized phase transition. We identify the semantic bifurcation window - the precise temporal interval where joint and decomposed vector fields meaningfully diverge. Exploiting this dynamic, we propose surgical guidance, a hybrid sampling strategy that restricts exact multi-pass scoring strictly to this critical window. On an OOD bi-digit MNIST testbed, surgical guidance achieves state-of-the-art compositional fidelity at a fraction of the inference cost, yielding a +5.3% absolute improvement in pairwise accuracy over the joint baseline by intervening during just the first 15% of the diffusion trajectory. Furthermore, our empirical analysis uncovers a fundamental topological divide: diffusion models (SDEs) force conceptual resolution immediately at peak noise, whereas Conditional Flow Matching (ODEs) delays structural binding until intermediate features emerge, establishing a new temporal framework for accelerating large-scale generative decoding. --- ### GraphSelect for Budgeted Representation Selection in Multimodal Graph Inference Authors: Xu Wang, Xunkai Li, Yinlin Zhu, Rong-Hua Li Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33561v1 Multimodal graph predictors combine text, images, and relations to classify connected entities. How much of this input is needed to preserve their predictions? We study budgeted representation selection, which chooses a subset of candidate text and image vectors under a separate capacity for each modality. Predictions from the complete candidate input define the classes to preserve. The challenge is that a representation's contribution depends on the other selected inputs, while graph propagation extends its effects across nodes. Our empirical study shows that candidate rankings change with the selected input, while predicted probabilities remain informative after the class stops changing. Updating scores improves selection, and exchanging inputs can improve a subset whose capacity is already filled. These findings lead to GraphSelect, which starts from individual candidate gains and refines the subset through jointly evaluated exchanges. It screens promising removals and additions, accepts an exchange when it reduces the prediction loss, and updates the scores. Experiments on six graphs show higher mean objective recovery than six attribution and explanation methods adapted to the selection task. Across nine trained architectures on two graphs, retaining 20% of the candidate representations per modality gives a mean accuracy drop of 0.10 percentage points relative to full candidate input, preserving classification performance with substantially fewer text and image representations. --- ### Fine Until Fine-Tuned: Repeated Solutions Make Reasoning Fragile Authors: Ely Sheikh Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33559v1 Recipes such as s1 and LIMO teach a model to reason with little data by showing it the same thousand or fewer worked solutions many times over. Judged when that training ends, the repetition looks harmless. But reasoning models are often trained again, and we find that repetition leaves their reasoning fragile to that next stage, even when the stage has nothing to do with reasoning. We fine-tuned Qwen3.5-9B-Base on its own correct solutions to competition math problems, either drilling a few hundred of them about eight times each or showing many more once; with the same amount of training, both solve about 95% of held-out problems. A single pass of ordinary instruction tuning leaves the once-trained model where it was, while the drilled one falls to 86.0%, and harsher later stages take it to 59.3% or below. A third model that visited the drilled problems just as often, with a new solution at every visit, was unharmed, so the damage comes from seeing the same texts again rather than from having few problems. The break recurs with a stronger model's traces, in further training runs and on other models and tasks. It is also cheap to undo: the reasoning is suppressed rather than erased, and five updates of reasoning training bring almost all of it back, as does brief training on the reasoning format with almost no mathematics. Fresh solutions prevented the damage, and so did replaying 6.25% of the original solutions in a gentler later stage, so our claim concerns later training without such replay. Sharpening alone does not explain the break, since a model sharpened three-quarters as much without repetition was unharmed. On a skill the base model could not perform within a token budget, repetition mainly cost learning. --- ### PGL-3D: Towards Progressive Geometric Learning for 3D Visual Query Localization Authors: Liang Peng, Shizhuo Mu, Bohan Tan, Wenyuan Wang, Chen Zhao, Xingping Dong, Heng Fan, Libo Zhang, Bo Du Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33558v2 3D Visual Query Localization (3DVQL) retrieves the latest contiguous occurrence of a queried object in an RGB--point-cloud sequence and predicts a 9-DoF cuboid for every response frame. The query is captured independently of the search sequence, so its annotated pose may differ from how the object appears in the search frames. The benchmark baseline predicts cuboids after feature modeling, leaving their geometry unused for subsequent feature refinement. We investigate whether complete intermediate cuboids can improve query and proposal representations before final decoding. We introduce Progressive Geometric Learning for 3DVQL (PGL-3D), a predict--select--refine--re-predict framework that uses intermediate cuboids to guide the aggregation of search evidence and update query and proposal representations. A shared head first predicts a complete cuboid for every proposal. Query--Tube--Memory (QTM) then selects reference observations by combining proposal association, cuboid quality, frame response, and target absence, since association confidence alone establishes neither target presence nor geometric accuracy. The center, size, and orientation of each selected cuboid define soft pooling weights over query-conditioned proposal features. The pooled memory updates the query and proposal representations, and the head re-predicts from the updated features. A training-only objective, ST-D9O, supervises cuboid geometry at every stage by adding boundary, signed-distance, and soft-overlap terms to parameter regression. PGL-3D achieves a mean stAP of $0.270 \pm 0.004$ on 3DVQL, compared with $0.044$ reported for LaF. Ablations support the benefits of geometry-guided feature updates, while stage-wise analyses show improved cuboid accuracy. Replacing the geometry objective in our PROT3D reproduction with ST-D9O improves mAO on GSOT3D from $21.63\%$ to $25.78\%$. Our code and models will be released. --- ### Sparsity by Default: The Theory and Practice of ARD in Gaussian Process Regression for Variable Selection Authors: Jia Cai Published: 2026-09-27 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2609.33550v1 Automatic relevance determination (ARD) is the standard device for input selection in Gaussian process (GP) regression. By giving the covariance kernel a separate lengthscale for every input and learning those lengthscales by maximizing the marginal likelihood, ARD lets the data decide which coordinates matter: irrelevant inputs receive very large lengthscales and are effectively switched off. We trace this mechanism to the Bayesian Occam's razor embodied in the marginal likelihood, derive the gradient through which it prunes inputs, and emphasize that ARD delivers effective rather than exact sparsity. We review the algorithms used in practice and the rules that turn lengthscales into selections, and we survey the asymptotic theory, distinguishing the fixed-domain identifiability obstruction on the lengthscales from the high-dimensional selection-consistency guarantees recently established for hierarchical GP priors, and noting what remains open for plain ARD. We compare ARD with spike-and-slab priors, sparse axis-aligned and global-local shrinkage priors including the Bayesian lasso and horseshoe, penalized likelihood kriging, sensitivity and projection criteria, and additive kernels. We argue that ARD endures because of its seamless integration with kernel learning, universal software support, and low cost, and we close with its limitations and remedies. --- ### Calibrated Derivative-Process Sensitivity for Gaussian-Process Variable Selection Authors: Jia Cai Published: 2026-09-27 Categories: stat.ML, cs.LG, math.NA arXiv: https://arxiv.org/abs/2609.33549v1 Automatic relevance determination (ARD), the default tool for variable selection in Gaussian-process (GP) regression, ranks inputs by inverse lengthscales -- which measure how fast a function varies, not how much an input contributes to prediction -- and offers no calibrated rule for deciding which inputs to keep. The prediction-centred alternative, the derivative sensitivity $ν_j = \mathbb{E}((\partial f/\partial x_j)^2)$, is available in closed form from a fitted GP, but turning it into a selection rule is harder than it looks: at a null input the estimator is a degenerate quadratic form, so Wald and Bernstein-von Mises cutoffs are anti-conservative, and the natural residual bootstrap is mis-scaled. We show that a studentized multiplier bootstrap of the GP derivative process repairs both, prove its validity through an invariance principle for quadratic forms, and obtain asymptotic family-wise and false-discovery-rate control across inputs. Over 100 replications the rule controls FDR wherever inputs are truly null, while uncalibrated derivative rankings breach the target by up to 2x and a Bernstein-von Mises cutoff by 2.2x; at matched FDR it loses no power; it holds under a Matérn kernel and input correlation up to 0.99; on real data with planted and authentic null inputs it admits 5-12x fewer spurious inputs; it costs 5-18% of the GP fit; and a block-averaged variant retains validity at cost linear in $n$. --- ### TerMeZO: Ternary Sparse Zeroth-Order Optimization for Fine-tuning BitNet Models at the Edge Authors: Houssem Sifaou, Prabodh Katti, Bipin Rajendran, Osvaldo Simeone Published: 2026-09-27 Categories: cs.LG, cs.AI, eess.SP arXiv: https://arxiv.org/abs/2609.33548v1 Fine-tuning anguage models (LLMs) with first-order optimizers requires a memory several times larger than that required for inference. Memory-efficient zeroth-order optimization (MeZO) sidesteps this cost by estimating gradients from forward passes only. However, for BitNet architectures, a family of LLMs with ternary {-1,0,1\} weights and 8-bit activations, fine-tuning requires updating full-precision latent weights, and thus the memory footprint of MeZO no longer matches that of inference. A promising solution is to finetune only a subset of the latent weights, but existing sparse zeroth-order (ZO) methods either ignore the ternary structure or require first-order gradient information to build a sparse mask, which is at odds with the purpose of ZO fine-tuning. We propose TerMeZO, a sparse MeZO scheme that exploits the geometry of the ternary quantizer itself to identify the latent weights that are more likely to change values during fine-tuning, at no additional data or memory cost. Our convergence analysis shows that TerMeZO can converge faster than full-parameter MeZO, owing to its optimized reduction of the fine-tuning effective dimension. We run extensive experiments on BitNet models ranging from 1B to 3B parameters, spanning classification, instruction-following, and mathematical reasoning tasks. TerMeZO matches or exceeds the performance of full-parameter MeZO while substantially reducing the fine-tuning memory footprint. --- ### OSCC: Certified Observation-Safe Coupling Optimization for Gradient-Noise Control in Imperfect-Information Learning Authors: Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Rui Chen, Daren Zha, Jun Xiao Published: 2026-09-27 Categories: cs.AI, cs.GT, cs.LG arXiv: https://arxiv.org/abs/2609.33543v1 Coupled rollouts can reduce the noise of counterfactual action comparisons, but two issues prevent standard common-random-number constructions from serving as a general learning primitive in imperfect-information environments. First, an invalid coupling may expose hidden state, synchronize endogenous policy randomness, or misalign chance events after counterfactual histories diverge. Second, in multi-action policy optimization, lower return-contrast variance is not by itself the relevant objective: the optimizer depends on the return covariance matrix after projection through the local policy-gradient geometry. We introduce observation-safe counterfactual coupling (OSCC), a framework that defines an admissible class through marginal preservation, information-state safety, branch-local policy randomness, semantic event alignment, and trace-before-oracle replay. We derive a gradient-aware coupling criterion showing that, for marginal-preserving couplings, policy-gradient noise changes are determined by policy-Jacobian-weighted off-diagonal return covariance. This motivates OSCC-Select, a calibration-only selector that chooses among independent, root-only, continuation-only, and fully coupled rollouts using separate safety and gain certificates. Its gain target combines projected gradient noise with measured physical sampling cost and falls back to independent sampling whenever a simultaneous lower confidence bound does not certify improvement. On 100,000 fixed-root Leduc comparisons, the fully coupled CP-GRPO instantiation reduces return-contrast variance from 41.1158 to 18.1441, a 55.87% reduction, while preserving the declared branch marginals. With three actions, OSCC-Select chooses continuation coupling and attains gradient-noise trace 0.0783 versus 0.0917 for return-variance selection. Increasing calibration from 64 to 2,048 groups raises certification from 0.327 to 0.995. --- ### Reasoning on the Simplex: Geometric Fixed-Point Models Authors: Talgat Daulbaev, Ilya Glazkov, Maxim Rakhuba, Ivan Oseledets Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33540v1 Looped reasoners spend test-time compute by iterating a weight-tied map, but a small residual does not mean the state is a fixed point when that map lives in unconstrained latent space. We propose Geometric Fixed-Point Reasoning (GFPR), in which the iterated state is the prediction itself: a field of categorical beliefs on a product of simplices, whose argmax is the answer at every step. Because the state is a belief, task structure can be imposed through compact convex relaxations, either as structured readouts or directly in the recurrent state; in the latter case the update remains a continuous self-map, so a fixed point exists for any parameters. At about 7M parameters, GFPR reaches 95.1% exact match on Sudoku-Extreme, 92.0% on Maze-Hard, and 100% sequence accuracy on S_5 length 128, above the published FPRM numbers at the same scale. The same update also trains a 201M language model on FineWeb-Edu in which each site is a distribution over the vocabulary; with 24 Picard steps it is above GPT-2 small on four zero-shot multiple-choice tasks and above GPT-2 medium on ARC-Easy. --- ### Jev Matches 7B Language Models for Speech-Neuroprosthesis Rescoring Authors: Gabriele Cinà Published: 2026-09-27 Categories: cs.CL, eess.AS arXiv: https://arxiv.org/abs/2609.33538v1 A speech neuroprosthesis decodes attempted speech from brain activity and ends by rescoring the decoder's candidate sentences with a language model of several billion parameters, the only component that needs a GPU. Replacing that model with a cheaper one is hard: general language models asked to pick one sentence from a list answer from where a label sits in the list rather than from the sentence itself. We pose rescoring as a single typed decision, one call that returns a probability for every candidate, served by Jev, a hosted model trained for calibrated decisions, and combine it with the decoder's own score. On 978 held-out sentences from a participant with ALS, where the published decoder alone reaches 8.1% word error, Jev reaches 7.5% against 7.8% for both OPT-6.7b and Qwen2.5-7B; with the decoder's weight re-tuned, 6.9% against 7.2% and 7.4%. Jev is ahead in all four comparisons and at most 0.2 points behind at the 95% bound. It costs 0.07 USD per thousand sentences and needs no GPU; a dedicated GPU running a 7B model is cheaper per sentence only above 43% utilisation, far beyond what one user generates. End-to-end latency over the internet is 262 ms, of which 62 ms is spent at the provider, the same order as a 7B model on a local GPU (27 ms) but not faster. --- ### Does Execution Require Target KV Fidelity? A Mixed-Fidelity KV Runtime for LLM Serving Authors: Jiantong Jiang, Yue Yang, Peiyu Yang, Feng Liu Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33536v1 Large language model (LLM) serving is increasingly constrained by the GPU memory consumed by key-value (KV) caches. Existing compression, eviction, and offloading techniques alleviate this pressure, but serving runtimes typically treat only the configured target KV representation as execution-ready. Under memory pressure, this target-only contract can turn KV shortage into request stalls and preemptions. We present ElasticKV, a mixed-fidelity KV runtime built on the observation that target fidelity need not gate execution. ElasticKV introduces a compact intermediate KV state, making fidelity a runtime-managed execution property. To realize this state in a paged serving runtime, ElasticKV combines (i) a pair-structured layout that turns fidelity reduction into reusable GPU capacity, (ii) a dual-mode attention backend that directly consumes the compact state while preserving the native target-only path, and (iii) pressure-aware fidelity management that adapts KV fidelity to memory pressure. Our extensive evaluation across diverse workloads, model families and scales, and GPU platforms demonstrates the effectiveness and generality of ElasticKV. Under high concurrency, ElasticKV achieves 3.8-4.0$\times$ lower time-to-first-token (TTFT) and 9.1$\times$ lower P90 TTFT than vLLM while preserving generation quality. --- ### ManiEdit: Sequential Unstructured Knowledge Editing for Language Models from a Manifold Perspective Authors: Rui Liu, Chenheng Zhang, Haoxuan Li, Zhouchen Lin Published: 2026-09-27 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.33534v1 Large language models (LLMs) inevitably generate some incorrect or outdated content, necessitating efficient and precise mechanisms for continual knowledge updates. However, existing model editing methods struggle to sequentially edit unstructured long-form knowledge, suffering from severe edit forgetting and degradation of general capabilities. To address these challenges, we reframe knowledge editing from a manifold perspective, viewing it as a localized displacement of an edit sub-manifold within the global knowledge manifold. Under this formulation, the problem can be decomposed into two key questions: (i) how to identify representative edit points that effectively anchor the edit sub-manifold, and (ii) how to preserve the remaining manifold structure during the sub-manifold displacement process. Based on this perspective, we propose ManiEdit, a novel manifold-aware autoregressive editing framework consisting of two core components. Pivot Localization addresses the mediocre-point dilemma by identifying high-leverage pivots to anchor the edit sub-manifold. Manifold-Aware Preservation preserves different knowledge types through an energy-weighted penalty combined with recursive null-space alignment. Experiments on two base LLMs and four unstructured editing benchmarks demonstrate that ManiEdit achieves state-of-the-art performance, outperforming the strongest baseline by up to +27.81 BERTScore and +8.50 ROUGE-L, while maintaining near-original general capabilities across six representative downstream tasks. Our code is available at: https://github.com/Areyliu/ManiEdit --- ### Neural Scaling Laws of Transformer Operator Network Authors: Haoran Yan, Zhongjie Shi, Yuanzhe Xi, Peng Chen, Wenjing Liao Published: 2026-09-27 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2609.33533v1 Transformers have emerged as powerful architectures for learning solution operators of physical systems. Empirically the prediction error has been observed to decrease when the data size and model size increase, suggesting neural scaling behavior. Yet a theoretical understanding of such scaling laws for transformer-based operator learning remains limited. In this work, we develop a theoretical framework for characterizing the approximation and generalization errors of transformer-based operator learning. Our analysis builds on a local-to-global approximation principle that is naturally aligned with the softmax attention mechanism and yields discretization-invariant output functions. On approximation theory, we derive a universal approximation error of transformer-based operator learning for Hölder-regular operators. On generalization theory, we establish a power scaling law between the prediction error and the training data size. The rate of convergence represented by the scaling exponent explicitly reflects the dimensions of the input and output domains, the regularity of the underlying functions and operators, and crucially, the intrinsic dimension of the input function class. By exploiting this intrinsic low-dimensional structure, our analysis yields a power-law generalization rate for operator learning, in contrast to the logarithmic-type power-law rates appearing in existing analyses of operator learning with feedforward neural networks. Numerical experiments validate the predicted power-law scaling and confirm that the convergence rate varies systematically with the intrinsic dimension of the input function class. --- ### E-CONAN (Entailment, CONtradition And Neutral) Diagnostics Dataset Investigating Linguistic Phenomena in Arabic Natural Language Understanding Authors: Khloud AL Jallad, Nada Ghneim, Ghaida Rebdawi Published: 2026-09-27 Categories: cs.CL, cs.AI, cs.HC, cs.LG, cs.LO arXiv: https://arxiv.org/abs/2609.33530v1 Natural Language Understanding (NLU) plays a crucial role in various applications, yet its performance suffers from weaknesses in handling the complexities of human languages, ranging from lexical ambiguity to high-level reasoning difficulties. Analyzing errors across diverse linguistic phenomena is crucial for NLU improvement, as it will help humans get insights to comprehensively assess models' limitations and capabilities, so optimizing models' generalization. Notably, several benchmarks contain diagnostics datasets designed for investigation and fine-grained error analysis. When highlighting the gaps in the state-of-the-art, we noted that there is no naming convention for macro and micro categories or even a standard set of linguistic phenomena that should be covered. To overcome this gap, we propose an initial hierarchy for Cross-Lingual NLU error analysis. Moreover, we propose a methodology to create an NLI hierarchical framework and applied a case study on Arabic NLU. Moreover, this paper introduces E-CONAN diagnostics dataset, a freely available dataset manually-annotated with coarse-grained and fine-grained categories based on our proposed Arabic hierarchy. E-CONAN dataset helps NLU designers better understand their models by doing error analysis and in-depth investigation. We used E-CONAN to investigate the performance of 9 pretrained language models and 5 LLMs. Results indicate that LLMs outperform pretrained models in world knowledge and commonsense reasoning macro-category, and underperform pretrained models in syntactic macro-category. Moreover, the hardest phenomena for all models is Reasoning, and the easiest phenomena for all pretrained models is Syntactic, and the easiest for LLMs is Lexico-Syntactic. --- ### Discovering Symmetries in Neural Network Parameter Spaces Authors: Bo Zhao, Nima Dehmamy, Robin Walters, Rose Yu Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33527v1 Parameter space symmetries are important for understanding neural networks' loss landscape, training dynamics, and generalization. However, systematically identifying these symmetries remains a challenge. In this paper, we formalize data-dependent parameter symmetries and characterize loss invariance and the group-action axioms through infinitesimal conditions, which provide objectives for jointly learning group generators and nonlinear action maps. Our framework systematically uncovers parameter symmetries, including previously unknown ones. To study larger networks, we establish conditions under which subnetwork symmetries extend to the full model. The same construction gives an explicit family of finite-batch symmetries, providing both analytical examples and a foundation for discovery through small subnetworks. Using the infinitesimal characterization and subnetwork construction, we implement a framework for automated discovery of parameter symmetries, and successfully uncovered symmetries in various architectures, including pretrained transformer models. --- ### EverMine: Dissecting the Self-Evolution of Research Capabilities in Long-Horizon Alpha Research Authors: Siyuan Li, Jiangfeng Zhang, Rui Yao, Weihua Qiu, Mingyang Xu, Zixuan Yuan Published: 2026-09-27 Categories: cs.AI, q-fin.ST arXiv: https://arxiv.org/abs/2609.33524v1 Self-evolving agents aim to turn research feedback into reusable skills, tools, and research rules. Whether these accumulated capabilities continue to improve later research requires controlled evaluation. Long-horizon alpha discovery provides a state-dependent setting: once a new factor enters the portfolio, the predictive information already covered changes, so the value of the same candidate or experience may change over time. We introduce EverMine, an empirical framework for studying self-evolving research capabilities in long-horizon alpha discovery. EverMine decomposes the research state into history (Hist), the current factor portfolio (Frontier), and reusable capabilities (Cap). Under matched resource limits, we compare complete runs with fixed or evolving Cap, and replace Cap while holding Hist and Frontier fixed to estimate the conditional value of accumulated capabilities. We also combine full trajectories with historical-state replay to examine how experience-based decisions affect candidate selection and portfolio outcomes. Across 18 long-horizon trajectories, end-to-end comparisons show no consistent gain from Cap evolution. Across 48 continuation branches from shared Hist and Frontier states, accumulated Cap also does not consistently outperform the initial Cap. Parameter tuning of existing factor structures can still improve the portfolio. In an exploratory replay of two screening batches from one Evolving trajectory, some screened-out candidates have positive marginal value at the original state, yet submitting all screened-out candidates sequentially slightly lowers final portfolio IC in both batches. These results show that candidate value depends on the evolving portfolio and submission order, and motivate evaluating self-evolving research capabilities through end-to-end outcomes, conditional capability value, and the consequences of experience-based decisions. --- ### In-Token Learning for High-Fidelity Image Restoration via Diffusion Transformers Authors: Xingfu Yi, Xiaoxue Yu Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33523v1 We present In-Token Learning, an image restoration framework that adapts a pretrained diffusion transformer using conditional rectified flow matching. Clean targets paired with degraded inputs supervise transport from Gaussian noise to restored images. Spatially aligned degraded-image tokens are fused with evolving latent tokens along the channel dimension, preserving the image-token count at a given resolution. Direct Low-Quality Guidance (DLG) combines frozen degraded-image embeddings with a fixed task prompt through the native conditioning pathway, without a trainable ControlNet-style branch or image captioning. We evaluate super-resolution and denoising on DIV2K, LSDIR, FFHQ, RealLQ250, and RealPhoto60, and automatic colorization on DIV2K and LSDIR. The tasks use separately trained checkpoints under the same framework. Results show competitive fidelity and perceptual quality under the evaluated protocols, with weaker generalization on RealLQ250. We report full-image QHD ($2560{\times}1440$) inference and a tiled $12$K restoration demonstration of Along the River During the Qingming Festival. Attention cost still increases with resolution. This technical report preserves the early broader study underlying Fill2SR, which subsequently developed the real-world super-resolution direction. --- ### AI-Driven Collaborative Assembly Line Inspection: System Integration and Deployment Challenges Authors: Asya Ünal, Amr Okasha, Ege Çırakman, Perin Ünal Published: 2026-09-27 Categories: cs.RO, cs.CV, cs.HC arXiv: https://arxiv.org/abs/2609.33522v1 Manual visual inspection on assembly lines is a persistent manufacturing bottleneck: operator fatigue over extended shifts lowers defect-detection rates. This paper presents the design, integration, and field deployment of an AI-assisted collaborative inspection cell at the Silverline kitchen-appliance factory, developed within the AI-PRISM project. The cell couples a Universal Robots UR 10e cobot carrying a machine-vision defect-detection pipeline with a Comau Racer-5 cobot for functional tests, coordinated through ROS 2 Humble on an Ubuntu 22.04 LTS server. Multi-modal data (Basler camera imagery, TIA microphone acoustics, and SPS electrical-safety measurements) are logged locally and visualised in real time with Grafana. We report the practical deployment challenges (close-proximity safety, AI robustness under glare and reflections, ROS 2 namespace collisions across two cobots, and operating-system and dependency issues) together with the engineering solutions adopted, and structure the integration through a four-level Human-Robot Interaction analysis. The deployed cell cuts per-unit quality-check time from 82 s to 61 s (about 25%), raises final-control resource efficiency from 0.75 to 0.88, reduces operator visual-inspection viewing time by 82%, and significantly lowers operator mental demand (p = 0.005, NASA-TLX). --- ### LLM4Trust: Exploring the Capabilities of Large Language Models for Trust Evaluation Authors: Jie Wang, Yanbo Sun, Zheng Yan, Jiahe Lan, Elisa Bertino Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33521v1 Trust evaluation plays a critical role in cybersecurity by supporting risk mitigation and decision-making. A variety of trust evaluation methods have been proposed, with learning-based approaches offering high accuracy and automation. However, they often require substantial ground truth, suffer from low training efficiency, lack support for basic trust properties, and provide limited explainability. Large Language Models (LLMs) offer a compelling alternative due to their strong zero-/few-shot reasoning abilities and broad knowledge. To this end, we propose LLM4Trust, the first benchmark framework that systematically explores the capabilities of LLMs for trust evaluation. We first construct diverse trust graphs to model five basic trust properties and design corresponding property understanding tasks. We then assess the ability of eight representative LLMs to understand these properties under nine prompt methods. Based on this exploration, we identify the most effective LLM-prompt combinations and apply them to five real-world datasets for validating LLMs' trust evaluation capability. During this process, we propose two strategies to extract key information from large-scale trust graphs, addressing the context window limitations of LLMs. Extensive experiments show that LLMs can effectively understand basic trust properties and have great potential for real-world trust evaluation, particularly under limited supervision. However, they remain vulnerable to attacks targeting trust graphs and demonstration examples used in few-shot prompting, and incur high inference costs. Accordingly, we propose a defense mechanism and batch inference to improve the robustness and efficiency of LLM-based trust evaluation. The source code of LLM4Trust is available at https://github.com/Jieerbobo/LLM4Trust --- ### Anatomy-Structured Hierarchical MIL for Weakly-Supervised Thoracic Disease Detection in Chest X-rays Authors: Jeongin Kim, Sohyun Ahn, Seo Young Kang, Jaeyi Sung, Soomin Kim, Sungho Cho, Rena Lee, Kwanchang Kim, Junhyug Noh Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33520v1 Weakly-supervised thoracic disease detection in chest X-rays (CXR) is challenging due to subtle appearances and complex anatomical overlap, motivating anatomy-aware modeling for improved localization. However, prior anatomy-aware methods typically rely on coarse region proxies or static spatial priors, which may restrict dynamic instance discovery and limit precise localization of small abnormalities. We propose Anatomy-Structured Hierarchical Multiple Instance Learning (ASH-MIL), a framework that introduces parallel anatomy-structured observation branches (cardiac, pulmonary, and agnostic) combined with hierarchical MIL aggregation. Anatomical priors are injected as soft spatial biases into decoder cross-attention, enabling anatomically grounded evidence maps without disease bounding-box supervision. Instance localization is derived directly from MIL-weighted cross-attention maps without bounding box supervision. Experiments on CXR8 and cross-domain MIMIC-CXR held-out sets demonstrate consistent improvements over prior weakly-supervised and anatomy-aware approaches, particularly under stricter localization criteria. Our code is available at https://github.com/jn-kim/ash-mil. --- ### SceneScaffold: Active Scene-State Construction for Unified 3D Scene Understanding Authors: Xiangqi Li, Libo Huang, Jiarui Zhao, Weilun Feng, Chuanguang Yang, Zhulin An, Yongjun Xu Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33518v1 Recent 3D large multimodal models (3D-LMMs) rely on a visual bottleneck to compress complex 3D scene evidence into a limited number of visual tokens compatible with large language models (LLMs). Current visual bottlenecks, however, often passively compress heterogeneous 3D evidence into a homogeneous object-centric token sequence, leaving the spatial organization of the scene under-represented. This under-representation forces the LLM to recover spatial relations from a flattened token sequence, leading to unstable reasoning in relation-intensive and spatially ambiguous scenes. To address this issue, we propose SceneScaffold, an active scene-state construction framework for unified 3D scene understanding. SceneScaffold reformulates the visual bottleneck from a passive feature compressor into an active scene organizer, constructing a role-aware spatial scaffold before language reasoning. Specifically, SceneScaffold organizes superpoint-level visual evidence into scene-state components with distinct structural roles: entity states preserve core object semantics, scene-frame states maintain spatial references via boundary and region anchors, relation states encode object-environment interaction cues, and a global summary provides compact context. Through this role-aware construction, SceneScaffold provides the LLM with a spatially organized scene representation before language reasoning. Experiments on unified 3D scene understanding tasks, including 3D visual grounding, question answering, and dense captioning, demonstrate the effectiveness of SceneScaffold, while diagnostic results further show its applicability to relation-intensive and spatially ambiguous cases. Code is available at https://github.com/lixiangqi707/SceneScaffold. --- ### PPG-LM: A Photoplethysmography-Language Model with Multi-Level Clinical Alignment Authors: Xiaoda Wang, Minxiao Wang, Maxwell A Xu, Patrick Langer, Kaiqiao Han, Defu Cao, Xiao Luo, Yuzhe Yang, Yan Liu, Xiao Hu, Yizhou Sun, Wei Wang et al. Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33516v1 Photoplethysmography (PPG) is widely recorded by clinical monitors and consumer wearables, providing a scalable source of continuous physiological information. These recordings offer an opportunity for physiological assessment at scale, but realizing this potential requires models to learn from both signal-derived physiological supervision and broader clinical context captured in electronic health records (EHRs). This involves aligning information spanning local observations, care events, and entire visits with PPG representations at corresponding temporal scales. However, existing PPG foundation models primarily rely on task-specific prediction heads, while the medical knowledge of large language models does not necessarily translate into waveform understanding. To bridge this gap, we introduce PPG-LM, the first PPG-language model family to learn physiological representations from both signal-derived supervision and broader clinical context captured in EHRs. To construct clinically grounded captions, we develop an automatic captioning pipeline that generates segment-, event-, and visit-level descriptions from signal measurements and structured EHR records. We then learn from these pairs through a two-stage framework that first establishes segment-language correspondence through contrastive learning and waveform-conditioned captioning, then extends alignment to events and visits through time-aware aggregation and temporal statement matching. Pretrained on approximately 73k hours of PPG, PPG-LM supports language-based recognition, cross-modal retrieval, and segment captioning. Experiments on MC-MED, MIMIC-III, and VitalDB show improved retrieval and caption factuality over language-model baselines and gains over PPG and time-series foundation models on multiple clinical prediction tasks. --- ### SLP-ProbHard: Probabilistic Hard-Constrained Learning via Structural Latent Parameterization Authors: Wondesen Teshome Bekele, Marco D'Oria Published: 2026-09-27 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.33515v1 Many probabilistic predictors must satisfy exact structure in every stochastic realization, yet common hard-constraint approaches form predictions in ambient coordinates and then correct or project them. We introduce SLP-ProbHard, a cross-family, representation-centered framework for probabilistic hard-constrained learning when explicit structural parameterizations are available. Its core object, a Structural Feasible Latent Parameterization (SFLP), combines a structural latent law $Z \sim P^Z_θ(\cdot\mid x)$ with a feasible map $Y=h_φ(x,Z)$ that satisfies the constraint for every latent realization. Together these components define the predictive law itself, including its support and boundary probabilities, rather than serving as a final feasibility wrapper. We study how feasible coordinates and maps affect stochastic dimension, dependence, calibration, expressiveness, and computation. Experiments use Gaussian latent laws and fixed geometry-derived maps across affine equalities, ordering and simplex constraints, nonlinear manifolds, and three structural representations of seven-basin hydrological flow-duration-curve (FDC) data. In an official-source affine comparison with ProbHardE2E/DPPL, both methods achieve zero practical constraint violations. SLP-ProbHard uses 8 instead of 11 stochastic coordinates and improves MSE/MAE, while DPPL yields better marginal CRPS and closer-to-nominal coverage; a paired test detects no Energy Score difference across ten seeds. Real-world affine and nonlinear FDC representations reduce 13 to 7 and 14 to 8 ambient versus computational coordinates, respectively. Exact feasibility alone thus does not determine a predictive law, motivating direct structural generation when meaningful feasible coordinates are available. --- ### Printability-Constrained Adversarial Decals for Near-Nadir Aerial Perception: Measured Ink Gamuts, Nested Realism Constraints, and a Physical-World Bound Authors: Sandesh Shrestha, K. T. Yasas Mahima, Asanka G. Perera Published: 2026-09-27 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33513v1 Adversarial patches for aerial perception are typically evaluated as digital composites, with printing left as an implementation detail. This study imposes three physical constraints during optimization rather than after it: the color range a particular printer can reproduce, the size of the flat panel a vehicle offers, and the loss of fine detail incurred when the patch is imaged from altitude. The principal comparison isolates the ink set. Two patches share all seventeen recorded optimization settings and differ only in the colors available to them. One is constrained to a uniform color cube; the other to a gamut measured by printing and scanning a 216-patch chart. Each was optimized at three seeds and evaluated against thirteen victim conditions, with every rate reported against a size-matched optimized control. The effect of the measured gamut is victim-dependent rather than uniform. Net attack success rises on three of six closed-set segmentation victims, and for these the seed ranges of the two ink sets are disjoint: $+0.120$ on DeepLabv3-R101 and $+0.041$ on SegFormer-B0. The color-cube patch is consistently stronger on the open-vocabulary segmenter and on two of four detectors, though no detector exceeds a net of $+0.026$ under either ink set. The natural explanation is that a printable palette is simply less chromatic and lower in frequency than a digital one. Eleven further patches test this account and it does not hold. Once cardinality is matched, a palette as chromatic as the cube attacks equally well. Cardinality itself shows no trend from three inks to thirty-two. Palettes matched on cardinality, lightness and chroma, and differing only in hue placement, span $0.035$ to $0.136$. A physical evaluation with printed decals did not detect transfer; it bounds the transferred rate at $0.133$, which does not exclude the simulated value of $0.121$. --- ### Source Anchoring for Physical Consistency in Flow Matching Models Authors: Giulia Romoli, Filippo Ruffini, Paolo Soda Published: 2026-09-27 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.33510v1 Deep generative models are used to solve partial differential equations and model distributions of physical system states, but ensuring that the generated samples satisfy the governing laws remains challenging. Projection-based flow-matching methods enforce physics by correcting the flow from an unconstrained noise distribution. These corrections shift the generated samples away from the distribution of target solutions, especially in high noise regions. To address this limitation, we propose Source Anchoring for Physical Consistency (SAPC), a Functional Flow Matching method that encodes the physical constraints into the source noise before generation begins. We evaluate SAPC on five systems governed by partial differential equations, covering six tasks with linear and non-linear dynamics, and compare results against five baselines and the unconstrained backbone. Anchoring the source reduces the need for large corrections that drive samples onto admissible but off-distribution states, and SAPC reproduces the target distributions most accurately on every evaluated task, while matching the constraint precision of the best projection-based baselines. Ablation experiments show that this gain arises from pairing source projection with a matched training objective that regresses toward the projected source. These results identify the source distribution as a key design choice for physically consistent generative modelling. --- ### What Happens During Autonomous Deep Research After the User Steps Away? Authors: Yimin Liu, Yijia Zhang, Yanmin Li, Tangwen Luo, Yuze Li, Ziling Yao, Zhi Yang Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33509v1 In autonomous deep research, a user provides a task and relevant background, then leaves the agent to conduct an extended investigation without further human intervention. We study how this initial user information is reflected in intermediate actions and how these actions relate to final recommendations. We introduce DRaligned, a counterfactual behavioral evaluation framework built on PDR-Bench. By varying one task-relevant user factor while keeping the remaining context fixed, we compare acquisition requests, working drafts, and final reports. Source-grounded extraction, blinded local judgments, and deterministic aggregation yield coarse directional measurements while leaving ambiguous cases unresolved. Our experiments show that strong user-specific delivery can emerge from a largely shared research process: agents investigate similar broad questions but allocate requests differently, and final recommendations distinguish user conditions more clearly than explicit requests do. Reports can also integrate user factors that were not jointly visible during acquisition. In readable draft-to-report comparisons, recommendations often retain their coarse user-specific direction despite substantial rewriting. Final directional differences recur across tested agent models, execution harnesses, and evaluator models, even as execution paths vary. These findings describe how initial user information shapes autonomous research and clarify the relationship between the process an agent follows and the recommendations it delivers. --- ### When Evidence Changes the Subject: Subject-Typed Claim Licensing for Learned Routing Authors: Jian Chen, Zixuan Yuan Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33505v1 Modern learned systems increasingly combine learned components with search, repair, or external solvers. Benchmarks often measure the resulting end-to-end system, while scientific claims may concern only one component, creating an attribution problem: evidence can fail to support the requested component-level claim while still supporting a positive conclusion about the larger system. Existing evidence-to-claim methods primarily calibrate claim strength. We argue that composite systems require a second dimension: scientific subject. We address this problem with subject-typed claim licensing, which separates weaker conclusions about the requested subject from positive but non-substitutive credit about another subject. We instantiate this idea in SCOPE-Routing for preference-conditioned multigraph routing. Non-authors reproducibly apply the declared semantics; held-out review yields fewer reference-relative upward deviations than unstructured review, while the difference from a strong evidence checklist remains unresolved; and a controlled routing study shows that score-optimal and claim-eligible methods can differ while valid hybrid-system credit is preserved. These results motivate treating claim strength and scientific subject as distinct dimensions of evidence-based evaluation. --- ### RelaxKV: Recomputation Guided by the Query with Sparse Context Attention for Efficient KV Cache Reuse Authors: Ruoling Qi, Yirui Liu, Xuaner Wu, Yuxin Jin, Jian Chen, Jiayu Qin, Yin Chen, Jiawei Shao Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33503v1 Cross-request KV caching reduces the prefill cost of Retrieval-Augmented Generation (RAG), but conventional prefix caching severely limits cache reuse across requests. Position-Independent Caching (PIC) removes this constraint by reusing independent chunks, but their KV states miss cross-chunk interactions. Existing methods selectively recompute token states to recover these missing interactions, but primarily allocate the recomputation budget to selecting which states to recompute, while fixing the recomputation context to the full causal prefix. We introduce RelaxKV, which formulates selective cache repair as a joint allocation problem over repair targets and recomputation context. Guided by the user query, RelaxKV identifies layer-specific repair targets and restricts their recomputation to a query-relevant context, reducing attention computation. Across four decoder models, RelaxKV at a 15% anchor ratio improves aggregate LongBench performance over ProphetKV on all models. On Qwen3-14B, RelaxKV provides a stronger quality-TTFT trade-off than ProphetKV across a 5%-30% anchor-ratio sweep, and achieves the best selective results on RULER-MV and LV-Eval at 16K and 32K context lengths. Controlled ablations further demonstrate the importance of recomputation context selection. --- ### Pulseflow: PPG Counterfactual Generation Via Latent Transport Authors: Hung Manh Pham, Dong Ma, Bin Zhu, Pan Zhou Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33501v1 Photoplethysmography (PPG) has become an important modality for continuous cardiovascular monitoring, including atrial fibrillation (AF) detection. However, labeled AF recordings remain limited in many clinical settings, making model adaptation difficult when only limited target data are available. Generative modeling offers a natural way to alleviate this scarcity by synthesizing additional AF signals. Existing approaches, however, mainly generate samples that match the target condition without explicitly modeling how an observed source recording should be transformed, making it difficult to leverage abundant source recordings from a specific population or cohort for targeted augmentation. We introduce PulseFlow, a source-conditioned counterfactual generation framework that combines conditional representation learning with invertible latent transport to edit cardiac rhythm while retaining information from the source. Experiments across two clinical cohorts demonstrate effective rhythm transformation, measurable source correspondence, and improved AF classification under limited labels. --- ### The cost of useful natural gradient updates Authors: Subhransu S. Bhattacharjee, Dylan Campbell, Rahul Shome Published: 2026-09-27 Categories: cs.LG, cs.IT, math.OC arXiv: https://arxiv.org/abs/2609.33499v1 What information is needed to turn a natural-gradient direction into a useful finite update? Under a population Kullback-Leibler (KL) budget, we call a step useful if it is feasible and loses at most a fraction $\varepsilon$ of the best feasible gain along the direction. We construct a four-state exponential family whose laws share their initial gradient, scalar Fisher information and natural gradient, yet two laws have disjoint useful-step sets. With these quantities supplied exactly and the law otherwise known only through draws, the family's worst-case sample complexity is $Θ(\log(1/δ)/(p\varepsilon^2))$ for small $\varepsilon$, where $p$ scales rare-state probabilities and $δ$ is the failure probability. The budget is fixed and the optimal gain stays bounded away from zero, so the step length, not the direction, carries this cost. For succinctly described event-tilt models, returning a useful step is NP-hard even with the exact natural gradient and efficient exact sampling. Recovering the unit natural gradient to constant error is also NP-hard even in a two-parameter logistic family with Fisher condition number at most 3. We also give matching sample bounds for event tilts, sample bounds for damped Fisher solves and a population-KL certificate for affine classifiers. In frozen-feature classifier heads, stopping at a sampled KL boundary succeeds in about half of the trials, and a 10% KL margin raises joint success above 93% at a KL budget of 0.01. Thus, knowing where to move is not enough: how far to move can carry an update's entire cost. --- ### Hamiltonian JEPA: Action-Conditioned World Models with an Inherited Control State Authors: Tamim Zoabi, Ameen Ali, Lior Wolf Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33497v1 Planning from pixels needs more than a latent space that is stable and predictable. The state the planner scores must also be organized by how actions move the system. Joint-embedding predictive architectures (JEPAs) avoid pixel reconstruction by predicting future representations, but existing action-conditioned JEPAs ask one embedding to serve both perception and control. We introduce H-JEPA, which separates the two. A wide perceptual code is regularized toward a well-scaled isotropic geometry with a Bures-Wasserstein prior, and a fixed orthonormal slice of that code is the control state, which inherits the code's covariance without any objective of its own. The state evolves under phase-conditioned dissipative port-Hamiltonian dynamics whose input port has orthonormal columns. Port-inverse consistency (PIC) reads the executed action back through the transpose of that port. We show that this readout is exactly the rollout error projected onto the port directions, so PIC is a parameter-free reweighting of prediction error and not an auxiliary action decoder. Untying the readout from the port breaks this identity and loses half of the gain. H-JEPA matches or exceeds reconstruction-free baselines, including the action-decoding Delta-JEPA, on four pixel-based control benchmarks after at most $10$ training epochs, and its largest gain is on OGB-Cube ($91.9$ against $79.3$ percent). Ablations on PushT and OGB-Cube separate the contributions of the structured predictor, PIC, the prediction horizon, the state rank, and the anti-collapse prior. --- ### Chameleon: Dynamic Format Adapter for Efficient Diffusion Authors: Arnab Sanyal, Sandeep Chinchali Published: 2026-09-27 Categories: cs.LG, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.33496v1 Post-training quantization (PTQ) is the standard way to run modern diffusion models on memory-constrained accelerators, yet every existing diffusion PTQ scheme fixes the $\mathit{number\ format}$ in advance and only tunes the scale, zero point, or per-layer bit-width. At a fixed bit-width the best format depends on the distribution being encoded, and that distribution differs across weight channels, across layers, and along the diffusion timestep, where activation distributions slide from heavy-tailed and noise-dominated to tightly clustered and structured. We propose Chameleon, a PTQ framework that holds the bit-width fixed and treats the format itself as a discrete variable, chosen per weight channel and per (layer, timestep bucket) activation tensor. Activation formats come from {INT8, FP8 E4M3, FP8 E5M2, MXFP8, MXINT8}, selected ahead of time from two cheap statistics (empirical kurtosis and the closed-form diffusion SNR) and stored in a lookup table; weight formats come from {INT8, MXINT8} at 8 bits or {INT4, NF4, FP4 E2M1, MXINT4, MXFP4} at 4 bits, selected offline by reconstruction error. An architectural fork adapts the same selection layer to multi-step UNets, single-step distilled models, and Diffusion Transformers. Across SDXL, SDXL-Turbo, and PixArt-$α$ on COCO-2014, Chameleon achieves the best FID in all six backbone $\times$ bit-width settings, with CLIP within 0.24 of the FP16 reference and the best of all quantized methods at $W_{4}A_{8}$. --- ### LLMs Trust Their Own: Identity-Dependent Conformity in Multi-Agent Systems Authors: Liron Soffer, Ravid Shwartz-Ziv, Chen Shani Published: 2026-09-27 Categories: cs.CL, cs.MA arXiv: https://arxiv.org/abs/2609.33495v1 Large language models (LLMs) are increasingly deployed in multi-agent settings, where agents observe and influence one another, making social influence a key dimension of AI behavior and safety. We investigate whether LLMs' responses depend on the social identity of other agents, beyond the effect of their consensus. We construct judgment tasks with a single correct answer, and place models in a multi-agent setting where they receive incorrect answers from other agents whose social identities (AI or human, model family, or an arbitrary minimal group) are either shared with or distinct from their own. Across 12 open-weights models and nine tasks, we find a bidirectional effect of group identity on conformity to incorrect answers: in-group consensus increases conformity (in-group favoritism), whereas out-group consensus decreases it (out-group divergence). Unlike humans, for whom one ally breaking the consensus sharply reduces conformity, models are unmoved by an ally from the majority's group. Worse, a correct ally from the opposing group intensifies this bidirectional effect. Chain-of-Thought reasoning suppresses most of these effects, yet an in-group ally still reduces conformity to an incorrect out-group majority. Labeling peers as safety-aligned shifts overall conformity but leaves in-group favoritism and out-group divergence intact. These results show that group identity shapes how LLMs aggregate information across agents, independently of its correctness, and identify a manipulation surface for multi-agent AI systems. --- ### Federated Multi-Modal Human Activity Recognition using Multi-Agent Reinforcement Learning Authors: Debasmita Dey, Tanmay Sen, Himel Mallick Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33492v1 Human Activity Recognition (HAR) from heterogeneous wearable sensors is fundamental to the Internet of Health Things (IoHT), supporting rehabilitation, elderly care, and smart healthcare. Existing multimodal fusion methods often assign fixed equal weights to sensor streams, overlooking differences in modality importance, acquisition cost, and sensor quality, which can vary due to movement, incorrect placement, or temporary blockage. We propose an adaptive and cost-aware multimodal HAR framework based on multi-agent reinforcement learning for centralized HAR and extend it to federated learning as FedMHAR. In the centralized setting, multimodal fusion is formulated as a cooperative Multi-Agent Reinforcement Learning (MARL) problem, where each sensing modality is assigned a PPO-based agent that learns per-sample fusion weights, enabling the model to emphasize informative modalities while down-weighting costly sensors when cheaper alternatives provide sufficient information. In the federated setting, we introduce BiFL-PPO, a bidirectional federated optimization strategy in which a server-side PPO policy learns client-specific trust weights and feeds them back to adapt local learning rates and proximal regularization. Unlike round-level optimization, BiFL-PPO uses dense batch-level rewards for more frequent feedback and stable training under heterogeneous client data. Evaluation on the MEx Rehabilitation and UTD Multimodal Human Action datasets shows that the centralized framework achieves 87.30% and 94.98% accuracy, respectively, outperforming conventional fusion methods and state-of-the-art HAR models. FedMHAR achieves 79.74% and 77.49% in the federated setting, consistently surpassing FedAvg, FedProx, FedBN, FedNova, and AdaFedProx, while providing more stable performance and reducing sensor acquisition cost. --- ### Domain-Adapted Diffusion Models for Conditional Independence Testing Authors: Yanfeng Yang, Junda Zhao, Yijie Gao, Jiaqi Yang, Xinyu Shi, Ziqi Chen, Shunyu Zhao, Shuai Li, Wei Huang, Eshant English, Kenji Fukumizu Published: 2026-09-27 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2609.33490v1 Conditional independence (CI) is a fundamental concept in statistics and machine learning. Recent advances in conditional generative modeling provide flexible tools for generative-model-based CI tests, which rely on an estimated conditional distribution to generate randomized samples. However, errors in estimating this distribution accumulate in existing Type I error bounds, and consistency of the generative estimator alone does not guarantee asymptotic Type I error control. To address this limitation, we formulate conditional generative modeling as a domain adaptation problem and leverage auxiliary data from multiple source domains to improve estimation in the target CI testing domain. We propose Domain-Adapted Diffusion (DA-Diff), a multi-source domain adaptation framework for conditional diffusion models based on weighted empirical risk minimization over both target and source domains. We establish the convergence rate of DA-Diff and show how transferable source data can improve target-domain estimation through an increased effective sample size while controlling transfer bias. Building on DA-Diff, we further propose Domain-Adapted Conditional Independence Testing (DA-CIT) and show that its Type I error satisfies $P(p \leq α) \leq α+ o(1)$. Experiments demonstrate that DA-Diff improved conditional generation quality compared with transfer-learning diffusion baselines, while DA-CIT provides strong Type I error control and competitive power. --- ### Predicting Block-Coordinate Performance via Cross-Curvature Authors: Shengkun Zhu, Jinshan Zeng, Zhiqiang Kou, Yongxin Tong, Yang Liu Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33489v1 Simultaneous and sequential block updates are two basic optimization strategies used across machine learning, such as neural-network training, federated learning, and low-rank adaptation. Choosing between them is difficult because their relative advantage depends on both the objective geometry and the number of iterations. We develop a unified theory for comparing Jacobi (JC), Gauss--Seidel (GS), and partially sequential deterministic block-gradient updates. Our analysis expresses the one-step loss difference through cross-block curvature, with an $O(η^3)$ remainder, where $η$ is the learning rate. We derive a signed loss comparison after $K$ iterations with $O(Kη^3)$ error under regularity conditions and $ηK\le T$ for fixed $T$, identifying the better method when the predicted difference exceeds this error. We evaluate these formulas along observed training trajectories across different machine learning settings. Over 500 iterations, our theory correctly identifies the lower-loss method in 98.0\% of iterations for the neural network, 83.4\% for federated learning, and 97.6\% for LoRA. Applying the loss recursion at each step using the measured parameter difference raises these rates to 100.0\%, 93.2\%, and 99.6\%, respectively. --- ### What masking geometry works best for EEG foundation models? Authors: Pierre Guetschel, Bruno Aristimunha, Yassine El Ouahidi, Arnaud Delorme, Thomas Moreau, Michael Tangermann Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33487v1 EEG foundation models hold promise for scalable brain-signal decoding across clinical and cognitive neuroscience applications, yet their pre-training pipelines remain poorly understood. Among design choices, the masking strategy is particularly critical: it determines what the network must predict and from which context. Yet it has never been ablated in isolation, as each new model bundles a new masking strategy with a new backbone and objective. In this paper, we formalize the design choices for spatio-temporal masking strategies and train various models with a single pipeline under varying masking configurations across two SSL frameworks (MAE and JEPA). We then systematically evaluate the resulting 58 pre-trained models on the 12 datasets of OpenEEGBench under a linear probe. Both frameworks agree on an optimal masking configuration and on shared failure modes. Outside these, performance is robust: 11 MAE and 9 JEPA configurations are statistically indistinguishable from the best. We further identify a novel JEPA-specific failure mode, tagged bias-inflation collapse, invisible to standard detectors. With a well-chosen mask, our pipeline reaches REVE-level downstream performance at a fraction of REVE's pre-training compute. ---