Ramp Launches Router for Multi-Model AI Access
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.
Search
10 results on this page
Reasoning Agents May Collude in Markets
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.
Japan’s Open LLM Advances with 33B Model
Japan’s National Institute of Informatics has released LLM-jp-4 33B, a new version of its open domestic large language model. The approximately 33.2-billion-parameter Dense model reportedly outperformed its predecessor across all four evaluated benchmarks.
Encrypted Prompts Bypass Grok’s Guardrails
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.
How LLMs Are Transforming Mental Health Care
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 Graph-Based Blueprint for Self-Evolving Agents
This survey frames self-evolving LLM agents as dynamic graphs whose memories, tools, skills, workflows, and relationships change over time. It presents four evolution taxonomies, connects nine dynamic-graph-learning fields to agent capabilities, and proposes graph-aware evaluation and governance protocols.
AI Explanations Can Weaken Human Judgment
A study found that LLM recommendations could persuade evaluators to reject promising innovations or approve weak ones. Providing narrative explanations made participants more likely to follow incorrect AI decisions, while unexplained recommendations led to better independent judgment.
AI Debate Training Curbs Reward Hacking
Google DeepMind researchers found that training AI agents through debate can reduce reward hacking when an LLM judge evaluates their work. On mathematics tasks, debate achieved higher sustained ground-truth accuracy than directly optimizing for LLM judge rewards.
Why Multi-Agent AI Needs Concurrency Control
This position paper argues that many multi-agent system failures are concurrency control problems rather than simple coordination or communication issues. Concurrent access to shared state can cause stale reads, lost updates, and inconsistent outcomes, especially during long LLM inference windows.
🤖
GRPO Scaling Produces Surprisingly Uneven Results
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.