This study benchmarks 25 open-source language models for local deployment on edge hardware, evaluating inference efficiency, general knowledge, and teaching effectiveness. Results show pronounced trade-offs: throughput and energy efficiency vary by over an order of magnitude, MMLU accuracy ranges from near-random to 57.2%, and teaching effectiveness does not correlate monotonically with either metric. Granite4 Tiny Hybrid (7B) achieves a strong overall balance, and a three-tier local inference architecture is proposed for the Robot Study Companion.

Compositional diffusion planning for long-horizon trajectories suffers from mode-averaging when local plan distributions are multimodal. The proposed training-free guidance method, RCD, uses self-reconstruction error and overlap consistency to steer sampling toward high-density, globally coherent plans, outperforming existing methods on OGBench tasks.

RoboULM is a human-in-the-loop methodology using LLMs to systematically identify and analyze uncertainties in self-adaptive robots at design time. It includes a detailed uncertainty taxonomy and was evaluated positively by 16 practitioners across four industrial use cases.

This paper introduces a diffusion-model-driven synthetic data augmentation pipeline to address long-tail class imbalance in medical image classification. The approach uses an inpainting diffusion model with an Out-of-Distribution post-selection mechanism to generate diverse, realistic synthetic samples. Evaluated on ISIC2019, tail classes see over 28% improvement.

Sentinel2Cap is a human-annotated multimodal captioning dataset containing Sentinel-1 SAR and Sentinel-2 multispectral image patches. Zero-shot evaluation with Qwen3-VL-8B-Instruct shows RGB images achieve highest performance, SAR images remain challenging, and modality-specific prompts improve results. The dataset is publicly available.

CAFe-DINO leverages DINOv3 backbone to achieve state-of-the-art open-vocabulary semantic segmentation on remote sensing imagery without fine-tuning on RS data.

A new framework called Fire-Centered Evaluation Region (FCER) focuses uncertainty quantification on critical fire zones, comparing ensemble and distilled student models on wildfire spread prediction. The student model achieves comparable calibration and complementary uncertainty rankings in boundary-sensitive areas, pointing toward more operationally relevant evaluation.

In computational pathology, nuclear instance segmentation is fundamental but evaluation pipelines have been overlooked. This paper identifies four critical issues and proposes NucEval, a unified framework that handles vague regions, score normalization, overlapping instances, and border uncertainty for more robust evaluation.

This paper proposes GPU-optimized data augmentations that improve cross-modality spine segmentation by 155% Dice on unseen domains with minimal in-domain accuracy loss, and releases an open-source toolbox.

Medical experts often manually segment images for diagnostic statistics and discard annotations. This paper trains segmentation models using only retained summary statistics (e.g., area) plus a few weakly supervised pixels. A novel loss function combines reconstruction quality, statistic matching, and weak signal overlap, showing effectiveness on standard, ultrasound, and CT data.

Researchers developed a composite method combining Segment Anything Model (SAM) with a domain-specific tool for automated organoid image segmentation, achieving accuracy close to human inter-observer variability.

This paper proposes VANGUARD, a framework that unifies anomaly classification, spatial grounding, and chain-of-thought reasoning in a single VLM, achieving 94% ROC-AUC and 84% F1 on UCF-Crime while providing interpretable explanations and precise localization.

A study reveals that memorization in Stable Diffusion is primarily caused by padding embeddings in the CLIP text encoder, which inadvertently duplicate the end-of-text embedding, leading the model to over-rely on it. Two simple inference-time mitigation strategies are proposed.

S^2tory introduces a narratology-grounded framework for movie script summarization that leverages character development trajectories to identify plot nuclei, achieving state-of-the-art semantic fidelity at 3.5x compression on MovieSum and strong generalization on BookSum.

Sparse Memory Finetuning (SMF) addresses catastrophic forgetting by adding key-value memory layers and updating only the most heavily read memory rows per batch. Experiments on Qwen-2.5-0.5B-Instruct using MedMCQA show SMF improves performance by 2.5 percentage points while keeping forgetting probes within 1 point of the base model, whereas LoRA and full finetuning achieve larger gains but with clear drift. Two row-selection rules (KL-divergence and TF-IDF) are also compared.

This research investigates whether measuring the deviation of LLM hidden states from a reference set of answerable queries can serve as a pre-generation signal for detecting unanswerable inputs. Across three instruction-tuned models and three prompt types, geometry primarily encodes task form. For math prompts, unanswerable inputs deviate consistently from the answerable centroid (ROC-AUC 0.78-0.84), outperforming a simple refusal baseline and capturing cases where models do not explicitly refuse. No reliable signal emerges for factual prompts, and code prompts show large effect sizes with high variance. Layer-wise analysis reveals the signal arises in early layers and attenuates toward output.

This paper explores the challenges of extracting Key Performance Indicators (KPIs) from unstructured earnings call transcripts. The authors evaluate SEC-trained models, introduce three new benchmarks (SECB, ECB, ECB-A), and find that encoder-based models struggle with domain shift. They propose an LLM-based system achieving 79.7% precision in human evaluation, providing a baseline for tracking emergent KPIs.

MedStruct-S benchmark is proposed to evaluate semi-structured information extraction under unknown keys and OCR noise, containing 3,582 annotated real-world clinical report pages. Experiments show encoder-only models excel at non-null-value key-conditioned QA, while fine-tuned decoder-only models are strongest overall.

This paper enhances the StockEmotions dataset with semantically structured opinion graphs to capture the 'why' behind investor sentiment. Using a declarative LLM pipeline on 10,000 StockTwits comments and Graph Neural Networks, the authors show that opinion semantics improve emotion classification performance.

Open-source and commercial ASR systems perform poorly on niche-domain Indic speech recognition tasks such as digit strings, currency amounts, addresses, and code-mixed phrases. This paper proposes a TTS-STT flywheel that generates ~22,000 synthetic entity-dense utterances, achieving an entity hit rate (EHR) of 0.98 on a Telugu test set, compared to 0.027 for the best open-source model and 0.16 for a commercial system. Code, data, and dictionaries are open-sourced.

LLMs have internal components that correctly process negation but often give wrong answers due to late-layer attention shortcuts. Ablating those modules improves accuracy. The models use both suppressive and constructive mechanisms, with the constructive being more prominent.

A new study finds that reasoning models slightly outperform non-reasoning models (2-11% accuracy gain) but still fail to challenge 26-42% of false presuppositions, and are susceptible to presupposition strength.

This study evaluates same-model self-verification as a confidence signal for selective prediction, finding task- and model-dependent results that preclude it as a general-purpose uncertainty estimator.

PRISM-CTG is a novel self-supervised foundation model that leverages large-scale unlabelled cardiotocography (CTG) recordings to learn transferable domain-level representations. It uses a multi-view self-supervised framework jointly optimizing three complementary pretext objectives: random-projected guided masked signal reconstruction, clinical variable prediction, and feature classification. The model outperforms baselines on 7 downstream CTG tasks and demonstrates strong generalization on external datasets, marking the first large-scale foundation model for CTG.

This paper proposes ASDAgent, a strategy-aware framework that addresses data scarcity and strategy consistency issues in AI-assisted early intensive behavioral intervention for autism. It includes a DoctorAgent with O-T-A-C reasoning loop and a ChildAgent with probabilistic behavior modeling. Experiments show generated dialogues closely match therapist strategy distribution (KL divergence 0.083), achieve nearly 80% strategic consistency with human experts, and distills clinical knowledge into small language models.

Fine-tuning guard models on purely benign data can cause complete loss of safety alignment, not through adversarial manipulation but standard domain specialization. This failure is demonstrated across three safety classifiers—LlamaGuard, WildGuard, and Granite Guardian—and originates in the destruction of latent safety geometry. Granite Guardian undergoes complete collapse with refusal rate dropping from 85% to 0%, CKA falling to zero, and 100% of outputs ambiguous. The proposed Fisher-Weighted Safety Subspace Regularization (FW-SSR) mitigates this, recovering 75% refusal on Granite Guardian and reducing WildGuard's Attack Success Rate to 3.6%.

Reinforcement learning (RL) is a key post-training tool for improving LLM reasoning, but rollout design is often underreported. This survey proposes the Generate-Filter-Control-Replay (GFCR) lifecycle taxonomy that decomposes rollout pipelines into four modular stages, complemented by criteria of reliability, coverage, and cost sensitivity. It synthesizes methods across RL paradigms and presents case studies in math, code, multimodal reasoning, and agentic skills, along with a diagnostic index for common rollout pathologies.

This paper proposes an agentic AI-based network optimization framework integrating mixture of experts (MoE) with large language models (LLMs) for future 6G networks. The LLM acts as a semantic gate to reason over operator objectives and dynamically compose optimization agents. Applied to a joint communication and computing network, the framework achieves near-optimal performance and outperforms individual experts.

This study examines the impact of systematic verification errors in Reinforcement Learning with Verifiable Rewards (RLVR). Controlled experiments on arithmetic tasks show that systematic false negatives behave like random noise, while systematic false positives can cause sub-optimal plateaus or performance collapse. The results challenge prior assumptions that errors only slow training, emphasizing the need to understand verifier quality beyond sample-level error rates.

This paper defines the "energy field" (row-centered attention logit) and reveals two classes of invariants: mechanism-level (per-row zero-sum, rank bound, spectral signatures) and model-level (delocalized variance, key incoherence). These findings provide theoretical insights and practical tools such as low-dimensional subspace constraints and per-head training monitors.