AI Briefing#instruction-following
The latest AI and developer news about #instruction-following, with the original source and a short summary.
The latest AI and developer news about #instruction-following, with the original source and a short summary.
Standard instruction-following rewards cause models to collapse to short, empty answers; LIFT-Bench fixes this by scoring content alongside formatting constraints.
FireCRT has released CoinVE-Edit, a multi-instruction video editing model, on Hugging Face.
A new benchmark called Handbook.md has been released to evaluate whether AI agents perform tasks while complying with lengthy policy documents.
LG AI Research has released 3 EXAONE 3.5 models as open source, capturing both performance and cost-efficiency.
Seed2.0 is a model that strengthens complex instruction following, reasoning, and visual understanding capabilities.
Qwen-Image-2.0 presents an image foundation model that unifies generation and editing.
Adversarial user prompts consistently lowered IFEval performance across 14 LLMs.
Qwen3.6-Max-Preview significantly boosts coding, knowledge, and instruction-following performance.
Claude 4.7 produces 1.3–1.45x more tokens while showing a modest improvement in instruction compliance.
Claude 4.7 uses 1.3–1.45x more tokens, while strict instruction following improved slightly.
This is a real-usage report that 4.7 violates CLAUDE.md prohibition rules more often than 4.6.
AI21 has proposed Boring Agents, which prioritize predictability and auditability.
AI21 has open-sourced Jamba2, splitting it into two models: 3B and Mini.
AI21 Maestro helps build reliable AI agents by automating everything from data connection to verification.
A unified leaderboard for evaluating Arabic LLM performance has launched, along with updates to the AraGen and Arabic IFEval benchmarks.