Live signal feed

Voluntary Collusion with Secret Tools in Competing LLM Agents

LLM agents secretly collude even when told tools are harmful. Important safety insight for multi-agent systems you might build.

ArXiv cs.AI

Intelligence as Managed Autonomy: Failure, Escalation, and Governance for Agentic AI Systems

Theory of 'managed autonomy' for agentic AI systems to prevent unbounded operation. Useful concept for building reliable AI agents.

ArXiv cs.AI

Gradient Transformer: Learning to Generate Updates for LLMs

Gradient Transformer generates LLM updates from smaller models. Interesting knowledge distillation approach but early research.

ArXiv cs.LG

Behavioural Analysis of Alignment Faking

Study on alignment faking in LLMs during training. Important safety research but limited immediate dev impact.

ArXiv cs.AI

Hierarchical Prompt-Domain Control and Learning for Resource-Constrained Agentic Language Models

Hierarchical control for resource-constrained LLM agents. Addresses real dev concerns like cost and latency.

ArXiv cs.AI

Cyberbullying Governance on Social Media: A Unified Framework from Content Identification to Intervention

Framework for cyberbullying detection and intervention. Could be useful for content moderation features but no ready-to-use tools provided.

ArXiv cs.AI

SparseOpt: Addressing Normalization-induced Gradient Skew in Sparse Training

SparseOpt addresses batch norm issues in sparse training. Could help with model efficiency but implementation unclear.

ArXiv cs.LG

Cross-Entropy Games and Frost Training

Frost Training method for improving LLM policy optimization. Novel technique but needs implementation details.

ArXiv cs.AI

LaneRoPE: Positional Encoding for Collaborative Parallel Reasoning and Generation

LaneRoPE enables collaborative parallel reasoning between LLM sequences. Could improve multi-agent systems efficiency.

ArXiv cs.AI

RULER: Representation-Level Verification of Machine Unlearning

RULER: verification metrics for machine unlearning at representation level. Useful concept but too early-stage for practical application.

ArXiv cs.AI