Web agents learn on the fly, while LLMs learn to compress without losing their reasoning.

In today’s roundup: new approaches to adapting and compressing AI models, as well as tools for evaluating agents and systematic reviews. At the end, a discussion of whether Astra has reached AGI-level capabilities.

🔥 Hot:

🔹 Researchers propose budget-friendly online adaptation for web agents — The approach is designed for lightweight local models that need to learn after deployment. 🔹 A new LLM compression method protects critical reasoning chains — Reasoning-Aware Compression abandons uniform quantization of all model components to reduce energy consumption without unnecessarily compromising quality.

➡️ Useful materials:

🔹 SCAFFOLD turns a web agent’s accumulated skills into a hierarchical library — The agent does not start every task from scratch but reuses procedural knowledge across changing interfaces. 🔹 SciLitBench tests LLMs at every stage of a systematic literature review — The benchmark covers everything from title and full-text screening to structured data extraction. 🔹 AutoFyn trains long-term agents without changing model weights — The system stores verified results in a persistent state and brings them into new sessions through explicit interfaces. 🔹 CriticGen turns LLM response evaluation into concrete improvement recommendations — The method takes the generation process itself into account and provides more detailed feedback than conventional general evaluations. 🔹 ARC-Bench highlights the problem of action ranking in frozen world models — The study tests the assumption that the proximity of a future state to the goal in latent space actually indicates the best action.

➡️ Discussions and case studies:

🔹 Habr users discussed whether Astra can be considered AGI-level — The author interviewed the model and examines what lies behind claims of achieving artificial general intelligence.

📝 If you would like to add other news and materials to the list, write in the comments.