DeepSearch-World: A Framework That Lets Web Agents "Self-Evolve" in a Reproducible Environment

Featured on Hugging Face Daily Papers, "DeepSearch-World" presents a verifiable environment in which web search and page browsing can be performed reproducibly. On top of this environment, the research team built a self-distillation framework called "DeepSearch-Evolve," which strengthens agents by repeatedly iterating through trajectory generation, filtering, data mixing, and fine-tuning — without relying on a powerful external teacher model.

The main reported metrics are as follows:

BrowseComp: 31.2%

GAIA: 61.5%

HotpotQA: 93.4%

Notably, the team constructed a pool of 420K multi-hop QA tasks and supports agent cognitive behaviors such as progress verification, grounded reflection, and failure recovery. The researchers plan to release the environment, training data, validation sets, models, and code.

Xiaomi-Robotics-1: "Data Volume" Beats Model Size — 100,000 Hours of Motion Learning

In an article on Xiaomi's robot learning reported by The Decoder, "Xiaomi-Robotics-1" was found to have been trained on more than 100,000 hours of motion data. What is particularly noteworthy is that the data was collected not by robots, but by humans using handheld grippers equipped with cameras.

The report states that adding data contributes more to performance gains than scaling up the model, and that the approach has not yet hit a ceiling. That said, absolute success rates are still noted as low, so it seems premature to conclude that we have entered a regime "solved by data scale."

DeepSWE: A New Benchmark for Evaluating AI Coding Agents

Covered by i-programmer.info, "DeepSWE" is drawing attention as a benchmark for evaluating AI coding agents. While existing evaluations tend to focus narrowly on task completion rates and patch adherence, DeepSWE aims to provide consistent comparisons on real-world SWE tasks.

A debate over whether it is "the best evaluation benchmark" has also sprung up on Hacker News. Whether it becomes the future standard for agent evaluation will likely depend on how well it integrates with existing toolchains and how reliably its results can be reproduced.

Trump Administration Pushes Open-Source AI Regulation Again, With Kimi K3 as Backdrop

According to an Axios report (via Reddit's LocalLLaMA), the Trump administration — aided in part by lobbying from major US AI companies — is reinvigorating moves toward regulating open-source AI models. Tom's Hardware also reports that, in the wake of Kimi K3's release, there is a renewed push to ban Chinese AI models on cybersecurity grounds.

On the other hand, some point out that openly downloadable weights make a US ban difficult to enforce in practice. As adoption widens, tightening regulation seems more likely to underscore the limits of policy than of the technology itself.

Local LLM + Pi.dev Fixed a PC Just by Saying "This Computer Is Slow"

A user in Reddit's LocalLLaMA community reported an interesting experience. After connecting to their home Qwen 27B (RTX 3090) via Pi.dev, they simply typed "This computer is really slow, can you please fix it" on a friend's laptop that had become extremely sluggish. Within about five minutes, the Windows Search Indexer storage I/O issue was resolved, delivering an overall 5–10x speedup.

The task itself was simple enough to handle manually, but the UX of completing everything through natural language alone has earned praise. It demonstrates that even a local LLM, combined with the right harness (such as Pi.dev or Hermes), can deliver a practical agent experience.

---