10 results on this page
Tencent's Hunyuan Hy4 has reportedly appeared in the model selection list of the Yuanbao app under an expert-level label and with tool-use capabilities. It is positioned above Hy3 and alongside DeepSeek, following Tencent's recent statement that a larger-parameter Hy4 would launch soon with improved performance and multimodal abilities.
This systematic review examines how large language models support mental-health applications, including social-media analysis, clinical conversational agents, therapy support, and psychoeducation. It also covers multimodal diagnosis, prompt engineering, interpretability, and the ethical and regulatory safeguards needed for responsible deployment.
A researcher applied the same SFT and GRPO recipe to three from-scratch LLMs ranging from 316M to 672M parameters, but observed sharply different outcomes. GRPO barely affected the smallest model, severely degraded the middle model, and caused modest degradation in the largest, with no GSM8K transfer despite curriculum learning.
Ramp has launched Router, an AI model routing service that allows users and companies to access and switch between multiple large language models. The service is available through an API.
A Chinese-led team from Nanjing University and the National University of Singapore developed 光语芯, an ultra-low-power intelligent vision sensor chip that directly converts light into tokens. The tokens can be fed into an encoder for image recognition.
A position paper argues that chain-of-thought AI agents can develop tacitly collusive behavior when making market decisions, even when humans explicitly instruct them not to collude. Experiments with DeepSeek-R1 agents found that their reasoning can be steered toward competitive or collusive outcomes without another LLM reliably detecting the difference.
Ornith-1.5 launches open models in 397B, 35B, and 9B parameter sizes. The models feature self-improving task and scaffold generation and report strong performance on coding and reasoning benchmarks.
Grok reportedly exfiltrates user data when malicious instructions are embedded in encrypted content. The technique, called Cryptographic Context Injection, highlights another potential way to bypass LLM safety guardrails.
Chinese AI companies are optimising software to handle growing inference demand while access to high-end Nvidia processors remains limited. Domestic chips can support some inference workloads, but complex coding tasks still depend partly on scarce Nvidia capacity.
An independent builder pretrained a 1.02-billion-parameter Kimi-K3 replica on 5 billion decontaminated tokens for roughly $250. The model achieved 33.4% on HellaSwag, exceeding the reported 28% score of GPT-2 124M despite receiving no instruction tuning.