Benchmarks and studies reveal persistent gaps in LLM reasoning and self-description, while domain-specific models advance in chemistry, medicine and physical systems.
Safety fine-tuning is not neutral: it narrows the strategic space of LLMs in real debates. For on-premise deployments, control over weights, pipelines, and TCO becomes the real variable. AI-Radar's analysis focuses on the trade-off between ready-to-u...
A benchmark on political dialogues shows that LLMs prioritise logical defences and fail to exploit ethotic counterattacks in ad hominem exchanges. The study attributes this limit to safety fine-tuning, which narrows the strategic action space. For te...
A team built a 10,000-headline supervision set and classified 902,111 deduplicated headlines from 25 French outlets between 2022 and 2025. The framework separates rhetorical salience from editorial selection. Default thresholds inflate salience estim...
A 0.4M-parameter multimodal model combines SMILES transformers and GATv2 graphs via cross-attention, reaching AUC-ROC 0.987 and F1 0.906 on ClinTox. The ChemBERTa-2-backed variant posts a mean AUC-ROC of 0.750 on Tox21, comparable to larger baselines...
A study on thousands of tutoring-session transcripts shows that using disagreement among multiple LLMs to target expert attention shortens codebook revision from months to days. Among three methods, Rationale Labeling reaches 64.9% accuracy against 5...
Hungarian startup DiffuseDrive generates synthetic data to train physical AI systems in defence, mining, aerospace and autonomous vehicles. Its platform identifies missing edge cases in datasets, produces them with control over object position and si...
A study on eight open-source instruct models up to 9B parameters shows that the chat template acts as a switch: when present it raises disclaimers like "I'm just an AI" and lowers experiential expressions, and vice versa. Inside three models' activat...
A population-supervised framework links 2D red-cell images to latent biophysical quantities, aggregating them into mean corpuscular volume, red-cell distribution width, and mean corpuscular haemoglobin. It combines shared local inference, a structure...
A token-matched study shows didactic and clinical data shape medical large language models asymmetrically: clinical data improves clinic-oriented tasks while remaining competitive on knowledge-intensive ones, whereas didactic data mainly boosts theor...
AstroForge will give its next Autonomy-1 space probe a small transformer-based model responsible for onboard decisions. Local inference becomes unavoidable without connectivity and under severe physical constraints. The shift from deterministic softw...
PsyAgentBench reruns five classic psychological experiments on open-weight LLMs under named, blind, canonical, and counterfactual conditions. The results show that effects like Asch conformity depend on the prompt and safety filter, not on a stable b...
Microsoft Research publishes in Nature a framework combining a Transformer model and a graph neural network. Learned ranking reduces unconstrained proposals and outperforms previous models in tests with chemists. Code and weights are open source.
Bonsai's own documents show Ternary Bonsai 2 27B scoring 52.8 on Terminal-Bench 2.1 and 60.8 on SWE-bench Verified, roughly three quarters of the full-precision model used as comparison. Yet the competitor name appears wrong, and the 98.2% retained-c...
