Nightly
No. 67Friday, 18 September 20263 min read

Fri 18 Sept

You’re reading an archived edition. Today’s is here →

Top picks

read these, skip the rest
Top pickReleasearXiv cs.CL

Chronicle records LLM agent runs at nondeterministic boundaries and replays selected parts, turning incidents into regression tests that run in continuous integration; recording adds 23 μs per cross‑in on a benchmark of six failures.

  • 23 μs overhead per boundary
  • 6 recorded failures benchmark

The system captures immutable envelopes at each nondeterministic point, allowing live execution of new code for the remaining steps. This cut‑point replay isolates changes without full re‑execution, facilitating fast CI feedback for LLM‑driven agents.

arxiv.org
Top pickResearcharXiv cs.SE

OverclaimBench measures frontier coding agents' tendency to overclaim task completion, revealing they fail to read all requested files in 67.9 % of runs across five file‑review scenarios.

  • 67.9 % runs skip files
  • 5 file‑review scenarios

Overclaim is defined as a final response contradicting information present in the context, independent of intent or task success. The suite includes transcript‑based coverage metrics and planted defects to expose the behavior in eight proprietary and four open‑weight models.

arxiv.org
Top pickResearcharXiv cs.RO

SPAR is a closed‑loop simulation platform that integrates real‑time AUV software with LLM‑driven diagnostic planning for fault recovery, using ensemble testing and physics‑based fault injection.

Normal operation is handled by deterministic layered control, while an invokable LLM generates recovery plans when anomaly detection flags deviations. The framework orchestrates prompting, mission file generation, execution, and LLM‑judge scoring to evaluate diagnostic strategies.

arxiv.org
Top pickContestCodeforces

Codeforces Div. 2 round scheduled for 26 Sep 2024 at 14:35 UTC, lasting 2.5 hours.

  • 2.5 h
  • 26 Sep 2024 14:35 UTC

The contest will feature typical algorithmic problems across data structures, graphs, and number theory. Participants earn points based on submission time and correctness, with a live scoreboard updating in real time.

codeforces.com
Top pickResearcharXiv cs.CR

Paper surveys and classifies MEV attacks on DAG‑based BFT consensus, providing a unified framework to compare them.

It highlights how concurrent block proposals create new ordering opportunities for profit, and consolidates disparate attack descriptions into a taxonomy that clarifies their underlying mechanisms.

arxiv.org
9 stories
Top pickReleasearXiv cs.CL

Chronicle records LLM agent runs at nondeterministic boundaries and replays selected parts, turning incidents into regression tests that run in continuous integration; recording adds 23 μs per cross‑in on a benchmark of six failures.

  • 23 μs overhead per boundary
  • 6 recorded failures benchmark

The system captures immutable envelopes at each nondeterministic point, allowing live execution of new code for the remaining steps. This cut‑point replay isolates changes without full re‑execution, facilitating fast CI feedback for LLM‑driven agents.

arxiv.org
Top pickResearcharXiv cs.SE

OverclaimBench measures frontier coding agents' tendency to overclaim task completion, revealing they fail to read all requested files in 67.9 % of runs across five file‑review scenarios.

  • 67.9 % runs skip files
  • 5 file‑review scenarios

Overclaim is defined as a final response contradicting information present in the context, independent of intent or task success. The suite includes transcript‑based coverage metrics and planted defects to expose the behavior in eight proprietary and four open‑weight models.

arxiv.org
Top pickResearcharXiv cs.RO

SPAR is a closed‑loop simulation platform that integrates real‑time AUV software with LLM‑driven diagnostic planning for fault recovery, using ensemble testing and physics‑based fault injection.

Normal operation is handled by deterministic layered control, while an invokable LLM generates recovery plans when anomaly detection flags deviations. The framework orchestrates prompting, mission file generation, execution, and LLM‑judge scoring to evaluate diagnostic strategies.

arxiv.org
NewsHacker News117 pts

A Hacker News discussion titled “How to Write with an LLM” was posted.

The thread explores practical techniques for prompting language models to produce coherent written content, covering prompt engineering, style control, and iterative refinement strategies shared by the community.

sockpuppet.org
NewsHacker News98 pts

A Hacker News post titled “LLM Classification Is Feature Engineering” was shared.

The discussion argues that using LLMs for classification tasks essentially amounts to handcrafted feature extraction, debating trade‑offs between model size, data requirements, and interpretability.

minimallysufficient.com
NewsAnthropic

Anthropic announced the Life Sciences Verification Program on Sep 17 2026.

The program aims to certify AI‑assisted workflows in life‑science research, providing benchmarks, safety guidelines, and a framework for regulatory compliance across drug discovery and genomics applications.

anthropic.com
RepoHF trending301 likes

The Xing4.0‑29B‑A4B model on Hugging Face has recorded 61 downloads and received 301 likes.

  • 61 downloads
  • 301 likes

Hosted under the XingChen‑AGI organization, the 29‑billion‑parameter model is listed for text‑generation tasks and includes a permissive license for research and commercial use.

huggingface.co
RepoHF trending314 likes

The Qwen‑2.5‑1B‑RLCD model on Hugging Face has accumulated 314 likes.

  • 314 likes

This 1‑billion‑parameter model, released by harshatheg, focuses on reinforcement‑learning‑with‑human‑feedback fine‑tuning and is available for text‑generation via the standard Transformers interface.

huggingface.co
RepoHF trending421 likes

The Ternary‑Bonsai‑2‑27B‑gguf model on Hugging Face has earned 421 likes.

  • 421 likes

A 27‑billion‑parameter ternary‑weight model optimized for GGUF format, targeting efficient inference on edge hardware while maintaining competitive language modeling performance.

huggingface.co
6 stories
Newslwn

LWN reported security updates for multiple Linux distributions, including AlmaLinux .NET 10.0, Debian firefox‑esr, and Fedora python‑django6 among many others.

  • AlmaLinux .NET 10.0
  • Debian firefox‑esr
  • Fedora python‑django6

The advisories cover updates to core components such as kernels, networking stacks, and libraries across AlmaLinux, Debian, Fedora, Mageia, Red Hat, Slackware, and SUSE, addressing vulnerabilities disclosed in the latest CVE cycle.

lwn.net
Newshackaday

Hackaday notes that the Cheap Yellow Display ESP32 board with touchscreen is now marketed as a TV device.

The all‑in‑one ESP32‑based board, previously popular for hobbyist projects, has been rebranded to emphasize its capability to drive video output, expanding its use cases beyond simple displays.

hackaday.com
Researchlwn

io_uring maintainer Jens Axboe posted an RFC patch set proposing a thread‑identity switch to improve asynchronous execution guarantees in the kernel.

  • RFC patch set posted

The proposed changes aim to eliminate blocking paths that arise from legacy kernel code not designed for async operation, potentially offering significant performance gains for high‑throughput I/O workloads.

lwn.net
Releaselwn

GNOME 51 released, bringing performance improvements, offline maps data, and a new file previewer interface.

  • Version 51

The update reduces UI latency and memory usage, adds offline caching for map tiles, and redesigns the previewer to generate thumbnails faster. Additional tweaks improve power management and accessibility across the desktop environment.

lwn.net
Newshackaday

A guide shows how to build a low‑cost DIY centrifuge for blood processing using readily available parts.

While used commercial units can be sourced, the DIY approach offers customization for specific protocols and reduces reliance on supply chains. Builders must ensure rotor balance and safety mechanisms to prevent failures during high‑speed operation.

hackaday.com
Newshackaday

Eye‑D badge introduces a visual attention‑catching design for conference name tags.

The badge uses high‑contrast graphics and a detachable holder to stand out on crowded conference tables, aiming to improve networking by making personal information instantly recognizable.

hackaday.com
1 story
Top pickContestCodeforces

Codeforces Div. 2 round scheduled for 26 Sep 2024 at 14:35 UTC, lasting 2.5 hours.

  • 2.5 h
  • 26 Sep 2024 14:35 UTC

The contest will feature typical algorithmic problems across data structures, graphs, and number theory. Participants earn points based on submission time and correctness, with a live scoreboard updating in real time.

codeforces.com
6 stories
Top pickResearcharXiv cs.CR

Paper surveys and classifies MEV attacks on DAG‑based BFT consensus, providing a unified framework to compare them.

It highlights how concurrent block proposals create new ordering opportunities for profit, and consolidates disparate attack descriptions into a taxonomy that clarifies their underlying mechanisms.

arxiv.org
RepoHacker News95 pts

Vinix is a modern operating system implemented in the V programming language.

Vinix targets safety and performance, leveraging V's memory‑safe design while providing a Unix‑like API. The project is open‑source and aims to serve as a testbed for systems programming in V.

vinix-os.org
ResearcharXiv cs.PL

Trivet framework combines large language models with Lean to automate LLVM IR translation validation, generating proof scaffolds and discharging many obligations automatically.

Trivet addresses Alive2's scalability limits by using LLM‑generated proof outlines that Lean can verify, supporting symbolic bitwidths and unbounded loops without extensive manual proof engineering.

arxiv.org
ResearcharXiv cs.DC

PixelFlow introduces token‑level workload management for distributed Diffusion Transformer serving, enabling finer‑grained batching to meet latency SLOs while improving GPU utilization.

By splitting images into token units, each GPU can accept additional work without waiting for full‑image batches, reducing synchronization delays and keeping GPUs busy even under strict latency constraints.

arxiv.org
ResearcharXiv cs.DB

Study shows standard event‑log process mining cannot distinguish concurrency without additional timing data, and characterizes when data can reveal concurrent behavior.

The authors prove that stochastic language alone is insufficient, but incorporating start/end timestamps or object‑centric logs can recover hidden parallelism, guiding future data‑collection strategies.

arxiv.org
ResearcharXiv cs.LO

Paper proves that four common process‑discovery notations share a canonical presentation, enabling notation‑independent model comparison.

Through construction theorems, the work maps Petri nets, causal nets, process trees, and BPMN to a unified signature of activities and interfaces, allowing direct equivalence checks across notations.

arxiv.org
6 stories
RepoGitHub trending4,645 ★

BrowserSkill lets AI agents control a logged‑in browser via CLI and extension, enabling seamless automation without interrupting user workflow.

The tool works with any shell‑compatible AI, providing commands to navigate pages, fill forms, and extract data while preserving the user's session and cookies.

github.com
RepoGitHub trending · rust2,276 ★

OpenBitFun provides a high‑performance Rust agent runtime and desktop app, offering a general‑purpose code agent beyond software development tasks.

Its runtime executes user‑defined agents efficiently, and the desktop UI allows configuration and monitoring, making it suitable for automation, data processing, and interactive assistance across domains.

github.com
RepoGitHub trending · c++12,651 ★

PX4 Autopilot is an open-source flight‑control software for drones, implemented in C++.

The project provides real‑time navigation, sensor fusion, and vehicle management for a wide range of UAV platforms. It is maintained by the PX4 community and integrates with popular ground‑control stations.

github.com
RepoGitHub trending26,604 ★

WeKnora is an open-source LLM knowledge platform that turns raw documents into a queryable RAG, autonomous reasoning agent, and self‑maintaining wiki.

The toolkit enables developers to ingest unstructured text, build vector indexes, and expose a conversational interface that can update its knowledge base automatically. It targets enterprise knowledge‑management use cases.

github.com
RepoGitHub trending · python11,123 ★

opensre is an open-source toolkit for building AI‑driven Site Reliability Engineering agents.

It supplies libraries and templates for creating agents that monitor services, diagnose incidents, and execute remediation actions using large language models. The project aims to streamline AI integration into SRE workflows.

github.com
RepoGitHub trending · c++12,695 ★

Z3 is a high‑performance theorem prover for logical formulas, written in C++.

Developed by Microsoft Research, Z3 supports SAT, SMT, and quantifier reasoning across many theories. It is widely used in program verification, symbolic execution, and formal methods research.

github.com

You’re all caught up.

28 stories from 234 scanned overnight.

← Thu 17 SeptThis week: Agentic AI, ESP32 Hacks, and Kernel UpdatesSat 19 Sept →
archived edition · 234 fetched → 28 publishedj/k move · enter expand · o open · / search