On-Policy Distillation (OPD) suffers from position bias where later tokens provide poor supervision. Importance-Weighted OPD (IW-OPD) assigns dynamic weights based on distribution discrepancy, prioritizing early tokens and suppressing late ones. IW-OPD converges faster and achieves up to 6.9 point performance gains on AIME-2025 compared to standard OPD.
ROVER enables reward-free pretraining by maximizing occupancy coverage in state space, using a learned world model to estimate occupancy without density or entropy estimation. It introduces a virtual sink state to balance exploration of known and unknown regions, achieving more uniform coverage and better downstream performance in tabular and pixel-based navigation tasks.
TeaNet, a task-enhanced augmentation network, reconstructs randomly masked spectra to generate augmented samples that preserve original spectral features while introducing domain-specific variations. This approach enables deep neural networks to identify discriminant wavenumbers more effectively, outperforming CNNs by 17% in challenging synthetic scenarios and offering improved interpretability in few-shot learning tasks.
TASER introduces a framework that dynamically expands and routes atomic skills for continual learning across highly heterogeneous tasks. It reduces catastrophic forgetting and improves plasticity by ensuring semantic distinctness and efficient capacity allocation through skill detection and routing mechanisms. Evaluated on HeteroCLBench, a benchmark with 19 diverse tasks across 9 cognitive dimensions, TASER outperforms existing baselines.
A user reports that Qwen3.6-27B runs significantly less intelligently in vLLM than in llama.cpp, exhibiting issues like ignoring messages, hallucinating tool calls, and failing to recognize prior conversation context. Despite proper configuration and prompt templates, the model appears to lose coherence and misinterprets its own tool usage, with errors occurring consistently rather than sporadically.
MedTS-TTT introduces a test-time training framework for medical time series classification. Built on CLSA-TTT and a Gated Convolutional Backbone, it enables rapid, single-step adaptation without iterative optimization. On four public datasets, it achieves 11 top-1 rankings out of 12 evaluations across nine baselines and three metrics.
KaLM-Reranker-V1 is a fast but not late-interaction reranker that decouples query and passage computation while maintaining strong relevance modeling through cross-attention. It achieves state-of-the-art performance on BEIR, outperforms industrial models like Qwen3-Reranker, and shows excellent results on MIRACL and LMEB, with the 0.27B Nano model remaining competitive against 7-12B models.
A Kolmogorov--Smirnov test on reservoir computer output weights detects regime changes in nonlinear systems. The method distinguishes visually identical attractors, resolves parameter drifts seven times smaller than deep-learning baselines, and identifies ventricular flutter in ECG recordings.
Sea-Scan uses machine learning to detect and localize dark vessels from unlabeled data. It achieves a 97.8% detection rate with only a 1.98% false-trigger rate, using weak supervision from imperfect AIS labels.
DataClaw0 introduces an agentic paradigm for actively refining raw multimodal data to align with user and downstream intents. It uses a two-stage pipeline grounded in factual anchors to generate a large-scale dataset across five domains and combines supervised fine-tuning with GRPO to achieve strong alignment with complex refinement tasks. Evaluated on video generation, VQA, and GUI navigation, DataClaw0 produces high-information-density tailored data, enabling efficient model adaptation with minimal training data.
A study finds that Transformer-based trajectory prediction models degrade significantly with noisy object state data. Accuracy drops by 1.3x under mild noise and up to 3.9x under realistic high noise conditions, highlighting the models' sensitivity and the need for noisier, real-world training data and mitigation strategies.
A new framework uses public infrastructure and OpenStreetMap data to reconstruct urban power grid topology from transmission to building-level connections. It successfully maps the grid for 7,330 buildings in Oslo's Alna district, enabling detailed power system analysis such as flow optimization and resilience studies.
SOHET introduces a hierarchical transformer architecture with event-type-specific tabular encoders and self-supervised pre-training. It outperforms existing methods by 5.8% on Booking.com's fraud detection task and achieves state-of-the-art results on 6 out of 8 EBES benchmark tasks.
Graph-of-Differences (GoD) introduces anatomy-structured difference alignment for medical image re-identification. It represents images as anatomy graphs, computes differences over matched anatomical regions, and anchors retrieval signals to homologous structures. GoD improves Rank-1 accuracy by 7.1 pp on fundus and 3.1 pp on CXR, with better generalization in zero-shot settings.
VLA-FAIL introduces a lightweight, failure detection framework for vision-language-action models that uses last-layer Mahalanobis distance and action chunk consistency without requiring failure data or expensive action sampling. The framework combines these detectors to achieve reliable, early failure detection across diverse tasks, outperforming baseline methods in both accuracy and efficiency.
CAT-Translate introduces a family of small, open-source models (0.8B to 7B parameters) specialized for Japanese-English bidirectional translation. Using synthetic parallel corpora and a two-stage fine-tuning approach with Multi-Objective GRPO, the models outperform multilingual models on real-world benchmarks across business, legal, medical, financial, and patent domains.
ADualVUOT introduces a dual-encoder VAE with Continuous Normalizing Flows to improve latent representation flexibility in medical image segmentation. It uses Gaussian-Gromov-Wasserstein distance for domain alignment and adversarial augmentation to boost robustness, outperforming prior optimal transport-based methods on medical imaging benchmarks.
The LDT-FRL framework introduces a privacy-preserving defense system for IoMT devices, combining temporal attention, lightweight digital twins, and federated reinforcement learning. It achieves 99.66% and 99.95% accuracy on CICDDoS 2019 and TON-IoT benchmarks, with perfect F1 on the MITM class, converging 81% faster than prior methods and offering interpretable defense decisions via SHAP and Grad-CAM.
Fast-TurboQuant introduces a multiplier-free projection method using a structured fast Johnson-Lindenstrauss transform. It replaces dense random rotation matrices with Rademacher phase inversion and fast Walsh-Hadamard transform, reducing arithmetic to only additions and improving Recall@10 with lower mean squared error.
A new post-training method uses multi-metric perceptual rewards to optimize speech enhancement models. It directly applies non-differentiable metrics like DNSMOS, WER, and UTMOS as rewards via Group Sequence Policy Optimization, achieving state-of-the-art results on DNS2020. Human evaluation confirms that combining multiple metrics outperforms single-metric approaches, reducing reward hacking.