Nov 1, 2026 · 02:18 UTC
The point
DecompGym benchmarks coding agents on designing decomposition algorithms. Agents write solvers across 81 instances from nine problem families, and each solver must run in 60 seconds, pass a method audit, and land within 1 percent of the optimal objective.
Why it matters
For builders, output checks alone miss bad methods, so agent feedback needs to name the exact defect.
Key details
- Evaluated 12 GPT, Claude, and Gemini models in coding agent tools
- Reflection turned 66 percent of retries into successes, versus 39 percent for regeneration
Nov 1, 2026 · 00:20 UTC
The point
A setup guide turns OpenClaw into a LaTeX assistant in a Prismer container with TeX Live. You describe a section, it generates source, compiles to PDF, previews it inline, and fixes errors from the log.
Why it matters
Writing and compiling stay in one flow, with no local TeX install. Evidence: setup guide, from the Good Bot case list at goodbot.guide.
Key details
- Compilers: pdflatex, xelatex, and lualatex
- Templates: article, IEEE, beamer, and Chinese article, plus BibTeX and BibLaTeX support
- Server: Prismer on port 8080, started with Docker
Oct 31, 2026 · 04:18 UTC
The point
Researchers present ExLabel, a self-evolving LLM assistant that judges task clarity and test coverage in SWE-bench style datasets. It uses an evolving experience bank to guide assessments of new instances.
Why it matters
Builders can filter training and evaluation data before trusting benchmark scores.
Key details
- New SWE-TQ benchmark spans eight languages, cleaned from human annotations in SWE-bench and Multi-SWE-bench
- Tested with three backbone LLMs, ExLabel achieved the best F1 across all settings
- A model fine-tuned on data filtered by ExLabel resolved 66.2% of SWE-bench Verified issues
Oct 31, 2026 · 02:19 UTC
The point
Researchers present DualGauge, the first fully automated framework for jointly evaluating correctness and security of specification-only code generation. DualGauge-Bench has 307 coding tasks, each with functional and security tests from the same specification.
Why it matters
Builders cannot trust functional tests alone to show code is secure and correct.
Key details
- Evaluated 10 representative LLMs across Python, C++, and JavaScript
- Even the strongest model remained below 15% joint security-functionality success in every language
- Codex, OpenHands, and Claude Code were tested, and two failed to outperform a comparable direct model
Oct 30, 2026 · 06:19 UTC
The point
Researchers present an agentic workflow that ports Fortran kernels to portable Kokkos C++. Specialized LLM agents translate, validate, compile, test, debug and optimize without human intervention.
Why it matters
It offers HPC teams a way to modernize legacy code for GPUs without manual porting.
Key details
- Evaluated over five runs per configuration on an NVIDIA A100, including a production climate model kernel
- GPT-5 and o4-mini-high completed it most reliably, beating Fortran baselines for a few US dollars
- Open weights models handled simpler kernels but struggled with parallel reductions and hierarchical parallelism
Oct 30, 2026 · 04:18 UTC
The point
Researchers present Benchmark Radar, a living database and search engine for AI benchmarks. It covers LLM evaluation, agentic and tool use, coding, reasoning, safety, and domain specific evaluations.
Why it matters
Builders can find evaluations and check score evidence in one catalog instead of hunting sources one by one.
Key details
- Catalog: 1,283 source records from 4 catalogs, plus 12,916 numeric observations on 790 records
- Daily discovery uses 37 sources: 13 direct connectors and 24 first party feeds
- Ships a dashboard with leaderboard and trend views, plus a CLI for offline queries
Oct 30, 2026 · 00:50 UTC
The point
A Claude Cowork user points the agent at more than 10 PDFs on their own machine and asks for specific data points. The agent scans the files and puts what it finds into one simple Markdown table.
Why it matters
Local files in, one checkable table out: a small pattern for document grunt work. Evidence: user-reported, from the Good Bot case list at goodbot.guide, not independently verified.
Key details
- Agent: Claude Cowork, scanning local files
- Input: more than 10 local PDFs
- Output: one simple Markdown table of the requested data points