AI / ML

Cloudflare has introduced Clef, a new open-weight decision model designed to optimize routing and infrastructure choices in real time. Alongside the model, they have released a dedicated platform for reinforcement learning (RL) fine-tuning to improve these decision-making capabilities. This release aims to provide developers with transparent, customizable tools for managing complex network and compute decisions.

• Clef is an open-weight model focused on decision-making rather than generative text.
• New RL fine-tuning platform allows practitioners to customize model behavior.
• Enables real-time optimization for routing and infrastructure resource allocation.
• Open weights facilitate transparency and integration into existing Cloudflare stacks.

This study evaluates nine NER systems across classical, encoder, and generative paradigms for on-device deployment, ranging from 13M to 8B parameters. It moves beyond standard leaderboard metrics by jointly assessing accuracy, latency, and output validity without requiring human-annotated gold labels. The authors utilize a silver-grounded corpus derived from RSS-News to enable scalable evaluation of model reliability and confidence.

• Evaluates spaCy, GLiNER, and local LLMs (Qwen3, DeepSeek-R1) for on-device NER.
• Prioritizes latency and output validity alongside accuracy for deployability.
• Uses silver-grounded annotation to scale evaluation without human labeling.
• Covers parameter sizes from 13M to 8B across three distinct model paradigms.

Researchers audited the GooseReason-0.7M dataset to check if models learn synthetic distractor patterns instead of reasoning. A classifier using only surface statistics achieved an AUROC of 0.562, proving the asymmetry between real correct answers and synthetic wrong ones is detectable. Code-based distractors were particularly easy to identify, performing below chance, suggesting models might exploit these artifacts rather than learning genuine reasoning.

• Synthetic distractors in RLVR corpora are statistically distinguishable from human prose using surface features.
• Models risk learning provenance shortcuts instead of reasoning if data asymmetry is not addressed.
• Code distractors are easier to flag than natural language ones, indicating domain-specific artifact risks.
• Data pipelines should audit for synthetic artifacts to ensure models learn genuine capabilities.