Fetching from the wire…
Public story · 2026-03-02 · source-backed
Training reasoning models to follow instructions in their reasoning traces (not just final answers) improves privacy by up to 51.9 percentage points. Uses a dual-adapter generation strategy. Critical for builders deploying reasoning models that handle sensitive data — reasoning traces can leak private info. arXiv 2602.24210
Each link below shares sources, entities, or timing with this story.
Reasoning models engage in "performative" CoT — the model's final answer is decodable from activations far earlier than visible CoT suggests. Activation probing enables up to 80% token reduction on MMLU. Critical for safety monitoring: visible reasoning may not reflect actual...
Detects economically motivated deviations — model substitution (provider swaps in cheaper model), quantization abuse, and token overbilling — without model internals access. Uses verifiable computation on a small fraction of requests. Critical for anyone relying on third-party...
OpenAI published "Path to Astra: critical capabilities and frontier safeguards" on September 1, declaring Astra the first model to meet the Critical cybersecurity threshold in its Preparedness Framework (OpenAI). Critical, in their own definition, means the model can find and...
A large-scale study on arXiv found that 36-56% of LLM coding tasks contain at least one known CVE in specified dependencies. Not in the generated code itself. In the packages the model tells you to install. The numbers get worse. 62-75% of those CVEs are rated Critical or High...
Uses confidence metrics and steering vectors to dynamically guide reasoning depth — reduces output redundancy while improving accuracy across 9 benchmarks, 4 model sizes, no training cost. Drop-in for o1-style reasoning models. arXiv
Expected-cost constraints fail under tail risk. Distributional safety addresses the blind spot. Critical for deploying safety-critical agents. arXiv:2603.10938
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.