Deep learning models for breast MRI tumor classification suffer near-chance external accuracy (0.5048–0.5265) when dataset origin is perfectly correlated with label, but a class-wise dataset mixing method using Duke Brea…
A new study from arXiv (2607.18695v1) finds that descriptors generated by large language models (LLMs) for zero-shot classification carry little visual evidence, collapsing ImageNet accuracy from 59.5% to 15.5% when the …
Researchers propose SAGE (ScAffolded Generative models for Explanation), a neuro-symbolic framework combining language models with cognitive models for pragmatic reasoning, achieving high accuracy across three case studi…
A study using fine-tuned large language models (LLMs) to analyze nearly 3,000 de-identified UK police incident logs found that mental ill health indicators appear in approximately one in five incidents, with lower preval…
Researchers introduced Search-on-Graph-R1 (SoGR1), an 8B-parameter model that internalizes knowledge graph navigation through supervised fine-tuning and reinforcement learning, outperforming all frozen frontier-LLM syste…
A new study from arXiv (2607.18566v1) finds that narrative framing of a task explains 5-31 times more behavioral variance in large language model agents than assigned personas, based on 1,890 sessions across 3 models and…
A new study from researchers using NVIDIA PhysicsNeMo finds that Monte Carlo Dropout with concrete dropout achieves well-calibrated uncertainty estimates at a fraction of the computational cost of Deep Ensembles for AI-d…
A new study introduces LatentMT, the first systematic application of latent-reasoning looped language models (LoopLMs) to machine translation, achieving performance comparable to models three to five times larger using a…
Researchers have introduced Fusion Embedding, a family of models that unify text, image, video, and audio into a single embedding space, enabling cross-modal retrieval without paired audio-visual training data. The first…
A rationale-guided knowledge distillation framework for cross-lingual stance detection outperforms competitive baselines on multilingual benchmarks, according to a new arXiv preprint (2607.18693v1). The method uses Chain…
Researchers introduced E-SpecFormer (Edge Spectrum monitoring Transformer), an edge-efficient model for end-to-end automatic modulation and covert channel recognition. The Nano variant achieves 86.5% average accuracy on …
Researchers at urbanAIthi have developed a preference-conditioned multi-objective reinforcement learning controller for transit signal priority (TSP) that allows runtime tuning of the trade-off between bus delay and over…
BearingNAS, a Hardware-Aware Neural Network Architecture Search framework, enables in-sensor intelligent fault diagnosis for bearings using only a laptop CPU, achieving 99.50% diagnostic accuracy on the LSM6DSO16IS Intel…
A new method called Soft-Prompt Privileged Context (SPPC) for on-policy self-distillation in large language models uses a teacher that differs from the student only by a learnable soft prompt, avoiding post-hoc rationali…
A new method called domain-conditional position offsets reduces the cold-start penalty in autoregressive language models by adding a single learned vector to the embedding activation at the first sequence positions while…
A new study from arXiv finds that cross-lingual factual inconsistency in large language models can be mitigated at inference time, with persona prompting outperforming more invasive methods like Contrastive Activation Ad…
Researchers propose RLAES, a unified LLM framework that jointly optimizes automated essay scoring and feedback generation through reinforcement learning, achieving the best scoring performance among LLM-based methods on …
Researchers at arXiv identified a failure mode called repetitive copying in long-context large language models, where models extensively copy input text instead of reasoning. They propose GEAR (Grounding Evidence-Aware R…
A study from arXiv finds that structured JSON extraction achieves 96% feasibility accuracy in two-agent LLM relays for travel planning, while narrative summarization drops feasibility to 48% despite producing the smalles…
ExecuTorch's experimental MLX delegate, released in May 2026, enables PyTorch models to run on Apple Silicon GPUs. Testing Qwen3-0.6B on an M1 Max, MLX INT4 achieved 188.9 tokens/s, 4.52x faster than PyTorch MPS BF16 (41…