The paper introduces probe-driven test-time reinforcement learning (TTRL) for code generation, using probe inputs derived from problem statements to evaluate candidate programs and define a Probe Consensus Reward (PCR). To address PCR’s unreliability and prevent reward hacking, the authors propose Entropy-Regularized Rank-Masked Policy Optimization (ERPO), which applies rank masking and an entropy ceiling to produce conservative policy updates. Experiments on coding benchmarks show that ERPO significantly improves pass@1 and pass@k metrics in both in-domain adaptation and zero-shot transfer scenarios.
By Jiacheng Xu, Feng Chen, Xiuneng Xu, Bo An
The paper introduces NIO Bench, a benchmarking framework that profiles storage I/O for six machine learning model types, using Python hooks and Linux strace to capture detailed access patterns. Experiments on a Ceph-backed Kubernetes cluster show that I/O is dominated by data preparation, model loading, and checkpointing, with training becoming compute-bound once data is staged. The study finds a power‑law distribution of file usage and identifies cache‑miss read tail latency as the main storage bottleneck, recommending aggressive prefetching, page‑cache pinning, and bursty write handling for ML‑optimized storage.
By Jonathan W. Morris, Ionut Mistreanu, Connor Louie
ZetaDial introduces a post‑sampling, per‑protein secant controller that adjusts the net charge of protein binders generated by fixed‑backbone ProteinMPNN, enabling inference‑time tuning of developability‑relevant charge. On benchmark datasets, the secant loop reduced mean absolute charge error compared to fixed‑slope and global‑bias approaches, achieving significant improvements on RCSB complexes and comparable performance on Cas13 monomers. The method also highlights trade‑offs between charge calibration and foldability, with exploratory tests on PD‑L1, IL‑7R‑alpha, and RBD designs illustrating its impact on design success rates and interface properties.
By Mohammed Sameer Syed, Tamara Dinneen
The paper introduces a contrastive learning approach for anomaly detection in multi-illumination and multi-focus display images. It builds on Multiresolution Knowledge Distillation (MKD) and proposes Multiresolution Contrastive Distillation (MCD), which eliminates the need for explicit positive/negative pairs by adjusting distances between teacher and student features. A blending module aggregates multi-channel data into a three‑channel input, and the method achieves superior AUROC and accuracy on the MMdAD dataset compared to state‑of‑the‑art baselines.
By Jihyun Lee, Hangil Park, Yongmin Seo, Taewon Min, Joodong Yun, Jaewon Kim, Tae-Kyun Kim
The paper introduces GP-Refiner, a plug‑and‑play framework that uses Gaussian Process Regression to correct feature predictions in Diffusion Transformers. By observing that residuals between cached and full‑compute features follow a local zero‑mean Gaussian distribution, the method treats cached features as noisy observations of the true trajectory, enabling online correction without needing explicit labels. Experiments show that integrating GP‑Refiner with existing acceleration techniques, such as TaylorSeer, cuts computational load by 19.3% while improving image quality metrics (PSNR +0.9 dB, LPIPS 0.46→0.29).
By Zhirong Shen, Rui Huang, Chang Zou, Shikang Zheng, Jiacheng Liu, Peiliang Cai, Zhengyi Shi, Yaosong Du, Liang Feng, Xiaobing Tu, Jinkui Ren, Xiantao Zhang, Linfeng Zhang
FALCON‑S is a modular, high‑fidelity simulator designed for fixed‑wing aerial robots operating near the ground. It models full 6DoF rigid‑body physics, semi‑empirical ground‑effect aerodynamics, actuator dynamics, sensor noise, and environmental disturbances, and supports both CPU and GPU backends via Torch and NVIDIA Warp for large‑scale reinforcement learning and optimal control. The framework offers a unified interface for various controllers, including RL and optical control algorithms, and allows cross‑validation with X‑Plane and JSBSim for engineering integration and visual fidelity.
By Matteo El Hariry, Pedro Lima, Andrej Orsula, Matthieu Geist, Miguel Olivares-Mendez
SolarBench is an open global benchmark for image-based solar nowcasting that consolidates over six million sky and satellite images from 11 sites across a decade, paired with irradiance, PV output, and atmospheric data. The benchmark includes a toolbox for reproducible data access, processing, model development, and evaluation. Using SolarBench, the authors benchmark representative models, uncover a gap between average forecasting accuracy and the capture of rapid solar fluctuations, quantify predictability across cloud regimes, and demonstrate data‑efficient adaptation to new PV systems.
By Yuhao Nie, Stephen Campbell, Quentin Paletta, Liwenbo Zhang, Tao Jing, Samer Chaaraoui, Jonathan Giezendanner, Andea Scott, Tao Sun, Cong Feng, Max Aragon, Jacques Camier, Adam Jensen, Florian Kotthoff, Yuexing Yang, Yang Ming, Mengying Li, Stefanie Meilinger, Yupeng Wu, Adam Brandt, Sherrie Wang
The paper introduces Circuit Reasoning Score (CRS), a data‑selection signal for reinforcement learning with verifiable rewards that uses attention‑head activity from a frozen base model to gauge reasoning engagement. CRS is computed in a single forward pass without reward labels or rollouts, and it shows that selecting problems with the lowest reasoning‑circuit engagement can outperform random selection on several medium‑difficulty benchmarks. However, the benefit depends on domain, model scale, and reward conditions, indicating that data selection in this setting is regime‑dependent rather than a fixed ranking of problem quality.
By Zhuofan Chen, Ziqian Jiao, Yikai Cui, Zhixin Cai, Jun Bai, Wenge Rong
The study evaluates audio provenance attribution systems, showing that high clean‑benchmark accuracy does not translate to robustness after codec compression. Using a prospectively registered protocol, the authors measured closed‑set attribution performance on two corpora after single‑stage codec transport, finding significant degradation—up to 70.3 Macro‑F1 points for WavLM‑Base+ and 61.0 for W2V2‑BERT 2.0—depending on codec settings and representation. The results demonstrate that clean accuracy alone cannot guarantee deployment robustness across different codecs and representations.
By Gang Shi (Independent Researcher)
MamMA is a pedestrian trajectory prediction algorithm that leverages LiDAR-generated occupancy maps and egocentric vision sensor data. It partitions the occupancy map into patches to extract obstacle features and incorporates pedestrian awareness states, which influence perception and speed. Using a Mamba-based model, MamMA predicts future trajectories and outperforms state‑of‑the‑art methods on multiple benchmark datasets.
By Juncen Long, Xiaofeng Jin, Gianluca Bardaro, Simone Mentasti, Matteo Matteucci
The paper introduces Retrieval-Augmented Decoding (RAD), a decoding-time method that improves the truthfulness of large language models without retraining. RAD uses a small reference set of up to ten annotated examples to build a grounding space of context embeddings and next-token logits, which it retrieves and aggregates during inference to shape the model’s output. Experiments on four open-ended generation benchmarks and four different LLMs show that RAD consistently outperforms strong baselines and generalizes well across tasks.
By Manh Nguyen, Sunil Gupta, Hung Le
The paper introduces a semi‑supervised framework that learns to predict nuclear magnetic resonance (NMR) chemical shifts from millions of literature‑extracted spectra without explicit atom‑level assignments. By treating the prediction as a permutation‑invariant set supervision problem, the authors show that optimal bipartite matching can be reduced to a sorting‑based loss, enabling stable large‑scale training. The resulting models outperform state‑of‑the‑art methods, generalize better to diverse molecules, and for the first time capture systematic solvent effects across common NMR solvents.
By Yongqi Jin, Yecheng Wang, Jun-jie Wang, Rong Zhu, Guolin Ke, Weinan E
The paper introduces ControlG, a control‑theoretic framework for coordinating multi‑objective graph self‑supervised learning. It treats objective coordination as a temporal allocation problem, estimating each objective’s difficulty and antagonism, planning budgets with a Pareto‑aware log‑hypervolume planner, and scheduling updates via a PID controller. Experiments on nine datasets show that ControlG consistently outperforms state‑of‑the‑art baselines and provides an auditable schedule revealing which objectives drive learning.
By Karish Grover, Theodore Vasiloudis, Han Xie, Sixing Lu, Xiang Song, Christos Faloutsos
FuseDiff is an end‑to‑end diffusion model designed for dual‑target structure‑based drug design, jointly generating a ligand graph and two pocket‑specific binding poses conditioned on both target pockets. It employs a message‑passing backbone with Dual‑target Local Context Fusion (DLCF) to fuse ligand atom contexts from both pockets, preserving symmetry while enabling expressive joint modeling. The model enforces topological consistency across the two poses and allows target‑specific geometric adaptation, achieving state‑of‑the‑art docking performance and enabling systematic assessment of dual‑target pose quality before docking‑based pose search.
By Jianliang Wu, Anjie Qiao, Zhen Wang, Zhewei Wei, Sheng Chen
The paper investigates how hard‑ReLU training behaves when perturbations have a finite radius. It shows that the usual infinitesimal sensitivities are insufficient to predict the response at a chosen radius, and it characterizes the intermediate regime where the perturbation radius scales with the gradient‑descent step. The authors derive crossing indices, a uniform endpoint expansion for separated transverse events, and provide explicit remainder terms in contractive affine regions to certify finite candidate comparisons, supported by experiments on nonlinear networks.
By Xiaoyang Li, Runni Zhou, Xinghao Yan
The paper introduces Reward Ensemble under Confidence (REC), a probabilistic reward learning framework for preference-based reinforcement learning that models per‑timestep reward uncertainty using an ensemble of distributional reward models. REC incorporates uncertainty into the preference loss and uses model disagreement to drive exploration, achieving 88.4% of shaped‑reward performance on acrobatic quadrotor control versus 55.2% with standard Preference PPO. The authors train policies in simulation and transfer them zero‑shot to real quadrotors, demonstrating complex acrobatic maneuvers learned solely from human preference feedback, and validate REC on a continuous‑control benchmark.
By Colin Merk, Ismail Geles, Jiaxu Xing, Angel Romero, Giorgia Ramponi, Davide Scaramuzza
RePro is a web‑recycling technique that trains a small language model (as little as 1 B parameters) with reinforcement learning to produce high‑quality, faithful rephrasings of pretraining data. The method uses one quality reward and three faithfulness rewards to preserve core semantics and structure while converting organic data into better training examples. Experiments show that RePro boosts downstream accuracy by 3.7–14.5 % over organic‑only baselines and improves data efficiency 2–3×, outperforming prior prompting‑based recycling approaches.
By Zichun Yu, Chenyan Xiong
The paper introduces TokCode, a token encoding framework that enhances robustness in generative semantic communication by restructuring redundancy in the semantic domain. TokCode leverages a lightweight adapter to transform a large language model into a token encoder, avoiding the need for a dedicated deep model. A channel-quality-aware distillation method (CADET) trains the adapter across diverse erasure rates, producing a reconfigurable low‑rank adapter that enables efficient reinforcement learning and achieves significant improvements in image similarity over existing receiver‑side recovery benchmarks.
By Jingzhi Hu, Ouya Wang, Geoffrey Ye Li
The paper introduces a web‑data curation recipe for pretraining medical encoders, addressing the scarcity of large, diverse corpora in dense‑terminology domains like medicine. It proposes two complementary techniques: medical‑term density filtering to select documents rich in medical terminology, and signal‑amplifying rephrasing that uses an LLM to rewrite documents into denser variants with broader entity contexts. Applied to French medical NLP, the recipe produces the FineMed corpus and the DoctoBERT encoder family, achieving state‑of‑the‑art results on the DrBenchmark public benchmark and a proprietary clinical NER task.
By Bofeng Huang, Jacques Sun, Diane Bouchacourt, Nicolas Barascud, Fajwel Fogel
The paper investigates whether self-consensus—stopping a reasoning model when its partial trajectory’s answers agree—can safely reduce inference cost. A large sweep of 3,520 consensus rules on two models and three benchmarks failed to meet predefined safety criteria, while a boundary‑confidence control (DEER) succeeded. The study shows that agreement signals that an answer persists under a fixed probing procedure, not that reasoning has finished, leading to premature stops and missed corrections even when token savings are significant.
By Yunxiang Mo, Donghao Zhao, Hejia Geng