Thu 17 Sept

Previous Mixture‑of‑Agents pipelines treated routing and fine‑tuning as separate stages, limiting adaptation. CERA‑MoA’s predictive familiarity estimator avoids costly full rollouts by inspecting hidden activations. The cumulative‑threshold routing keeps the active agent pool small, enabling continual specialization after deployment. This design aims to improve overall system efficiency and responsiveness to evolving capabilities.

Current large language models freeze weights after training and rely on prompts to inject live information, which is discarded after each request. The proposed system treats user‑supplied facts and corrections as weight updates, allowing continual learning without retraining. This could reduce dependence on massive pretraining corpora and personalize behavior, though challenges remain around stability, latency, and security of on‑the‑fly weight generation.

Earlier compression methods focused solely on redundancy in final‑score distributions, ignoring rich process information. DualViewEval extracts six complementary trajectory signals and learns an exact‑size miniset that predicts full‑benchmark results. Tested on five agent benchmarks against five baselines, it consistently outperformed prior approaches, enabling faster evaluation of costly agent systems.

kmscon, in development for over a decade, offers better Unicode handling, configurable fonts, and easier debugging than the traditional console. The change affects all Fedora spins that rely on the console during boot and rescue, and may influence other distributions to adopt a similar userspace approach.

The updates address numerous CVEs in components such as OpenSSL, nginx, and Python. Administrators are urged to apply the patches promptly; the bulletin provides per‑distribution package lists and links to upstream advisories for detailed vulnerability information.

Highlights include new server‑side encryption APIs, a controversial PostgreSQL patch set fixing replication bugs, and performance improvements that reduce kernel compilation times. The edition also features a lessons‑learned article from a recent DPL and assorted announcements from the open‑source community.

Solano discussed challenges of encrypting headless systems without keyboard access and reviewed solutions such as TPM‑based sealing, network‑bound disk encryption, and initramfs key retrieval. He emphasized trade‑offs between security guarantees and operational complexity for production environments.

The contest is a Division 1 event, targeting top‑rated participants. It follows the standard Codeforces format with problems of increasing difficulty. The UTC start time helps global competitors plan their schedule. No registration is required beyond a Codeforces account.

AccelPact keeps device memory quiescent after a communication failure, allowing training to resume without writing checkpoints to disk. It integrates with frameworks such as PyTorch FSDP by rebinding internal communication handles. By avoiding asynchronous snapshotting, it eliminates storage back‑pressure and reduces overall training time. The approach is transparent to the training code, requiring only runtime support.

Existing Byzantine‑tolerant samplers lose effectiveness as the fraction of malicious nodes grows. AUPE lets honest participants share sampling information, counteracting attempts to inflate Byzantine representation in local views. The protocol maintains low overhead while providing stronger guarantees in permissionless blockchain settings. It is the first to combine collaboration with Byzantine fault tolerance in peer sampling.

DAT captures logical schema definitions together with data‑distribution strategies (e.g., sharding, replication) and uses guided query mutation to provoke diverse execution paths. By focusing on distribution‑specific optimizations, it finds bugs that centralized test generators miss. The approach improves test executability and bug‑detection rates across multiple DDBMS implementations.

Given a reference implementation, DiverseFT prompts LLMs to generate alternative code that passes the original test suite while differing at the binary level. The generated variants reduce common‑mode failures caused by shared bugs. The system can be integrated into CI pipelines to continuously refresh replica diversity without manual coding effort.

Netkit inserts eBPF programs that directly redirect packets between network namespaces, eliminating unnecessary buffering in the Linux backlog queue. It preserves full compatibility with the standard Linux network stack, so existing containerized applications run unchanged. The implementation resides in the kernel and has been shown to reduce latency for same‑host container traffic.

Developers declare ordered expansions of parent items into child records; the RayOrch compiler validates matching gather operations. At runtime, the system records each child's parent and ordinal position, enabling GPUs to batch children across many parents while respecting ordering constraints. This lineage awareness simplifies downstream processing of large, heterogeneous training datasets.