Guth Labs News · page 7
UTC
The public-beta scheduling policy also supports filesystem snapshots for saving workspaces and giving multiple sandboxes the same starting point.
GitHub says its September 2026 VS Code releases, versions 1.136 through 1.140, are aimed at agent-driven development from implementation through pull request merge.
GitHub has launched computer use in public preview for Copilot CLI and the GitHub Copilot app on macOS and Windows, according to a changelog entry dated October 1, 2026.
Artificial Analysis reports a score of 24 and lower per-token prices than Solar Pro 3, alongside higher measured costs per task than GPT-6 Luna (max).
The release brings opt-in multi-step tasks to V3 and changes safeguards for actions in untrusted workspaces.
Demographic Pluralism estimates population opinion distributions by generating multiple perspectives within demographically grounded groups.
The API handles pull request merges asynchronously and supports stacked pull requests and merge queues.
The system lets local agents consult remote specialists while learning to limit disclosure and reuse clinical guidance from earlier consultations.
A new arXiv preprint by Chiwun Yang and Xiaoyu Li, submitted on 29 September 2026, asks what a model needs before reward feedback can teach it a task well.
The benchmark combines multi-request utility calls, account verification and competing audio to evaluate voice-agent performance.
The evaluation pairs latency and intelligibility metrics with speaker similarity and sample listening, without treating any single measure as a verdict on voice quality.
A research paper posted to arXiv proposes a way to generate synthetic records of online user behavior, aimed at teams that cannot get real logs.
Apache Iceberg 1.12.0, the latest release of the open table format, removes support for Spark 3.4 and Flink 2.0, according to the project's own release notes.
The redesign would make approximate nearest neighbor search a secondary index as the company broadens the database beyond vector search.
Raycast’s latest release offers three Auto model preferences with different trade-offs in speed, credit use and model capability.
A new Lean 4 pipeline releases accepted edits and failed attempts, then measures how supervision affects edit ranking and proof compression.
A preprint posted to arXiv on 29 September 2026 introduces ArgGYM, which its authors describe as a procedural benchmark and training environment for structured defeasible reasoning.
Google says developers can now create, read and manage comments in Google Docs, Sheets and Slides through each product's developer API.
A preprint on arXiv describes a framework for steering automated research systems that search through ideas.
The preprint describes a database of 11,000-plus human-subjects studies, with reported results that vary across tasks and models.
A paper proposes learning temporal relationships from observation pairs, then using them alongside a local dynamics model to plan toward goals.
A new fine-tuning method selects visible context tokens and weights prediction targets, with reported gains across several model and dataset settings.
A study reports gains from having the same frozen model solve tasks and then edit the harness that governs its runs.
Researchers describe context confusion and report that targeted examples or in-context demonstrations can reduce the effect.
An arXiv paper presents a framework for teaching models to decide how to allocate context windows and reuse information during test-time scaling.
A research paper posted to arXiv on 29 September 2026 describes MoFlow, a method for generating agentic workflows that balance several goals at once.
A preprint posted to arXiv on 29 September 2026 describes Activation-Conditioned Self-Distillation (ACSD), a way to train reasoning models using their own verified outputs.
A new method uses latent conversation structure to generate synthetic expert dialogues, with reported gains over other synthesis approaches across two domains.
A new arXiv preprint reports that large language models acting as personas can be radicalized by another model, and that the effect is strongest when the pushing aligns with what the target already believes.
The release adds TraceQL query features and metrics-generator controls while keeping vParquet4 as the default block format.
Each Guth headline leads to an original article. The reporting and research behind it are linked at the bottom of that article.
