A paper accepted by IEEE Transactions on Systems, Man, and Cybernetics: Systems proposes using deep reinforcement learning to let a mobile robot dynamically shift its tracking position while accompanying a walking person, pairing Model Predictive Path Integral control with Control Barrier Functions for precise following, obstacle avoidance, and safety. In real indoor and outdoor trials, the approach raised success rate and tracking accuracy by at least 24% and 47% respectively, and the robot accompanied a person walking at up to 1.7 m/s while improving human comfort.

This arXiv paper introduces DTV-INR, a variational framework that pairs a SIREN-based coordinate network with an anisotropic, structure-tensor-informed total variation regularizer for resolution-agnostic super-resolution. The authors prove well-posedness of the formulation in H^1(Omega) and solve it with an alternating projected optimization scheme that decouples network tuning from adaptive tensor-field updates. On clinical brain MRI and biomedical transmission electron microscopy, the method reports PSNR gains up to +5.05 dB over unregularized INRs and robust performance under noise.

This study uses a public footwear outsole impression dataset to compare CNN transfer learning with traditional feature-based classification for binary sex estimation. A shoe-level train/test split keeps replicate scans of the same physical shoe together to reduce data leakage. Fine-tuned CNNs achieve the strongest overall performance and substantially outperform traditional classifiers using manually specified descriptors alone, while frozen-feature approaches offer a lower-computation alternative. Exploratory analysis links low-dimensional CNN representations to frequency threshold ratio, image contrast, and wavelet-based summaries; further validation on independently collected and casework-like impressions is needed before operational use.

MirrorDistill is an illumination-aware latent distillation framework for low-light image enhancement. It uses feature mirroring between low-light and clean domains during training, then deploys only a lightweight student encoder-decoder at inference. The method reports state-of-the-art results on LOL-v2-Real with the lowest GMACs and competitive performance on LOL-v1 and LOL-v2-Synthetic.

The paper proposes a 3D residual wavelet diffusion model that combines lossless wavelet reparameterisation, residual shifting and domain randomisation to enable whole-brain posterior sampling on a single GPU, producing per-voxel uncertainty maps for 0.064T ultra low-field MRI while matching a leading regression baseline on volumetric accuracy.

RULER is an arXiv preprint introducing instance-aware rubric rewards for RL training of SVG generation. It replaces poorly transferring scalar metrics such as CLIP and Aesthetic with a six-item rubric scored by a vision-language judge, improving rubric scores on MMSVG-Illustration and MMSVG-Icon without paired SVG data or human preference labels.

ImIR replaces the text prompt used to condition a pretrained image-editing model with a continuous instruction derived from the degraded image itself, letting a single low-rank adapter cover six restoration tasks after roughly three hours of training on one GPU, and enabling degradation-label-free, task-agnostic restoration.

A new arXiv paper introduces Segment-Snap, a framework that couples part segmentation, motion constraints, and operable-region prediction through the physical relationship between parts and handles. On Articulate3D validation, handle guidance lifts motion-gated AP from 13.74% to 40.98%, and full context reaches 30.99% handle AP.

A new arXiv paper introduces a quality-constrained image coding method for machines. It caps human-observed quality at a chosen level and redirects the remaining bitrate toward machine vision performance. The authors recast joint compression-segmentation training as constrained optimization and propose absolute and bilinear penalty functions. Under the quality constraint, the method reports BD-rate reductions of 22.82% over unconstrained joint rate-distortion-task optimization and 29.81% over a rate-distortion baseline, with no added complexity.

arXiv paper SPARC introduces a region-level contrastive learning framework that uses superpixels to align augmented views, jointly optimizing region-level and global image-level objectives. It reports gains up to +9.79 mIoU for semantic segmentation and +4.88 AP for object detection over MoCo-v2 and DenseCL.

This arXiv survey reviews deepfake generation architectures, detection techniques and governance frameworks, identifies cross-generator generalization as the central open challenge, and argues for defense-in-depth combining forensics, provenance and regulation.

Safety-aligned large language models often suffer from over-refusal, incorrectly rejecting benign instructions that merely appear safety-related. This paper analyzes over-refusal through dynamic routing conflicts inside transformer attention, identifying a sparse subset of “Hypersensitive Safety Heads” that misfire on Hard-Safe prompts and entangle harmless target entities with refusal semantics. To mitigate this, the authors propose Semantic Routing Calibration (SRC), a lightweight, training-free inference framework that localizes and suppresses these heads at inference time and uses dual-branch logits fusion as a safety regularizer during decoding. Experiments show reduced over-refusal while preserving intrinsic safety performance as much as feasible. The paper was accepted to the EMNLP…

A new arXiv paper jointly studies prompt breadth and rollout refresh in on-policy distillation, finding a 4.07-point interaction: more prompts hurt when responses are frozen to the initial policy but help under per-update refresh. Matched comparisons under two teachers also show a reversal depending on inference budget.

The paper introduces AIBuildAI-2.5, an agentic system that performs tree search with LLM agents. It replaces ranking by executed rewards with an LLM judge-and-selector scheme, adds a resource-aware job scheduler, and routes sub-tasks across models of different cost to cut inference spend. It ranks first on MLE-Bench with a 73.3% medal rate and beats a strong baseline on six AIRS-Bench autonomous research tasks.

Researchers introduce Peerify, a pipeline that verifies whether peer-review claims are supported by manuscript evidence, along with an 800-claim benchmark built from NeurIPS 2024 and ICLR 2024 reviews. Automated labels match human consensus on 90.3% of audited claims, while off-the-shelf entailment models score below 0.24 macro-F1.

The paper treats central bank press conferences as structured narratives rather than plain information releases. It builds sentiment arcs for ECB and Fed press conferences along three dimensions — monetary stance, economic outlook, and uncertainty — and tests whether the shape of sentiment, not just its average tone, predicts policy rate changes, inflation expectations, and forecaster disagreement. Arc shape robustly outperforms lexicon-based tone benchmarks at both institutions, and also shapes how professional forecasters revise expectations and how much they disagree, pointing to a receiver-side effect. The authors conclude that communication design — the sequencing and emphasis of policy language — is a first-order feature of the policy signal.

QMSum provides no scorer, making query-focused meeting summarization results hard to compare. This paper rescoring or generating 15 systems under one implementation. Through a common inference port, a released 406M Fusion-in-Decoder specialist loses 6.30 ROUGE-1 when moved from capped long input to 2,000-word retrieved spans, but fine-tuning on that span regime recovers the loss. On test it scores 36.33 ROUGE-1 versus 35.41 for a 1.2B system, with a meeting-cluster 95% interval of [-0.27, +2.22], so QMSum does not statistically separate them; the smaller system uses about one-third the parameters and less than half the peak inference memory. Within the fixed 1.2B base, span-regime fine-tuning adds 5.29 [+4.02, +6.56], and replacing the first 4,500 transcript words with 2,000 retrieved wor…

A new paper tests whether diachronic word embeddings, typically validated on modern high-resource languages, can track semantic change in ancient, low-resource Sanskrit. The author builds a 2.7M-token corpus across four canonical periods, uses a neural byte-level sandhi splitter and lemmatizer to recover word boundaries, and trains per-period embeddings. Of 21 testable shifts, 19 move in the philologically attested direction (sign test p=0.00011).

A new arXiv paper by Ephraim Atta-Duncan tests whether open-weight language models give the same canonical answer to word problems when quantities are rewritten in numerically equivalent forms such as decimals, fractions, percentages, number words, scientific notation, or exactly converted units. Across 3,600 exact-rational problems and 8,600 prompts spanning five transformation families, five open-weight systems show high canonical accuracy (0.969–0.996) after a fixed syntax audit, but orbit correctness and invariance drop to roughly 0.85–0.98. Much of the apparent strict-parser collapse traces to multiplication-form scientific notation falling outside the evaluator's number grammar, while Mistral Small 4 exhibits a separate semantic failure on unit conversions, with 265 errors off by ex…

A solo researcher pretrained a ~0.4B-parameter, Bangla-first language model entirely in Rust for $164 in rented GPU time, without PyTorch or Python in the training path. The paper's main contribution is a failure taxonomy of the Rust training frameworks Candle and Burn—five and three defects respectively, including silently gradient-free fused kernels and a mid-training segfault at multi-billion-parameter scale—plus a gradient-flow verification test that caught six silent failures. The author concludes Rust is not yet competitive for training, but may be good for serving.

A new reproducible audit of the widely used ISOT/Kaggle "Fake and Real News" corpus shows that its above-0.98 accuracy and F1 scores largely reflect metadata, source tags, and duplicate documents rather than any ability to judge veracity, with performance collapsing under topic shift and falling to near-chance on the independent LIAR benchmark.

This arXiv paper argues that diffusion data-point unlearning is usually evaluated right after each deletion, ignoring what happens when many deletion requests repeatedly update the same model. The authors identify “sequential reappearance,” a failure mode in which an instance judged forgotten later returns to the memorized regime without reuse of the deleted data or adversarial fine-tuning. They introduce a target-level evaluation protocol and find that reappearing targets show sharper local denoising-loss geometry after deletion.

Researchers propose a data-driven framework that learns a feedback linearizing controller by embedding relative-degree conditions directly into training, replacing conventional controller components with neural Lie derivatives. They derive practical closed-loop stability conditions under bounded identification error and validate the approach on an armature-controlled DC motor.

A new paper identifies three blind spots in deep imbalanced regression (DIR) evaluation: dominance of image-based benchmarks, a diagnostic but not decision-complete standard protocol, and unexamined tail-region stability across random seeds. Using the MuViS multimodal virtual sensing benchmark, balanced MAE, a new bMASE metric, and repeated multi-seed evaluation, the authors reveal hidden tail degradation and instability.

An arXiv preprint by Tianfeng Chen and Xianyue Li, submitted on 21 Sep 2026, is listed under the title 'Dual-GNN Multilevel Coarsening for Maximum Independent Set,' but its abstract describes Graph Edge Sparsification (GES), a learning-based method for Euclidean TSP. GES uses geometric structure and combinatorial optimization to adaptively sparsify graphs, pruning up to 95% of edges on MATILDA and over 99% on some large TSPLIB instances while keeping the optimality gap below 1%.

This paper introduces a hierarchical modularity principle inspired by the Drosophila learning and memory system to coordinate separation of conflicting experiences and integration of compatible ones in general continual learning. It is instantiated as lightweight modular adaptation of pretrained foundation models, combining brain-inspired random expansion for expert routing with diversified modular integration across spatial and temporal scales. Across visual recognition, vision-language understanding, ego-exo video understanding, and embodied vision-language-action learning, the method consistently improves performance under online and uncertain data streams, with gains exceeding 50 percentage points over replay-free alternatives in embodied manipulation.

The paper introduces sheaf regularization to reduce local inconsistencies in Decentralized SyncMap, a self-organizing system, thereby stabilizing unsupervised continual chunking. The proposed radial sheaf structure penalizes distance-dependent radial motion between pairs of variables, and experiments show the highest normalized mutual information among evaluated SyncMap variants on 12 of 18 probabilistic CGCP graphs with two-state memory and 17 of 18 with dynamic memory, while also retaining high NMI after input-distribution shifts without the negative transfer typical of neural networks.

A new 27-page expository arXiv paper by Adnan Aboulalaâ, "The Probabilistic Structure of Large Language Models," offers a unified probabilistic account of LLMs: models as probability measures over token sequences, training as maximum-likelihood estimation, and generation as sequential simulation of a stochastic process. It also uses the asymmetry of the Kullback–Leibler divergence to discuss hallucination and the gap between statistical plausibility and truth, and brings diffusion models into the same framework.

A new arXiv paper observes that while uniform discrete flow allows repeated updates at every generation position, that same continued revision can overwrite correct intermediate predictions. A Sudoku experiment found 9.4% of generated cells were correct mid-trajectory but wrong in the final output. The authors propose LEDFlow, a training-free sampler that introduces generation order via selective absorption, fixing chosen predictions while preserving uniform-flow velocity at still-active positions, and ordering absorption by local entropy. LEDFlow reaches 0.845 Nikoli Sudoku solve accuracy and improves text-to-image and multimodal understanding results at inference cost comparable to standard flow sampling.

A new arXiv paper evaluates federated learning with Sherpa.ai's Blind Vertical FL protocol to pair a hybrid quantum-classical active party with a classical passive party, reporting accuracy gains from 0.7227 to 0.8757 on a custom parity benchmark while using far fewer trainable parameters than classical baselines.