Coding research is tightening around tests, runtime visibility, and precise targeting

Work on coding systems centers on executable checks as the training signal. ZeroCoder improves a code model without reference solutions or human tests by co-evolving code and test generation, then scoring both through execution. On Qwen2.5-Coder-7B-Instruct, the fully label-free setup improves code generation by 14.5%, and DyB4 raises that gain to 21.6%; across three model families and six benchmarks, average gains reach 18.8% for code generation and 62.7% for test generation. A smaller paper on Test-Oriented Programming pushes the same idea into workflow design: developers review generated tests, then let the model write production code. Its Onion prototype completed a small BibTeX CLI task in all 10 runs, with no direct production-code edits, though the evidence stays narrow because the evaluation is a single small application.

Several items on this day argue that agent quality depends on what the runtime records, exposes, and constrains. The externalization review gives the broad claim: memory, reusable skills, protocols, and the harness around them are now central to reliable agents, even though the paper is conceptual rather than empirical. Tokalator makes that argument concrete for coding assistants by showing token budgets inside the IDE, ranking open tabs for relevance, and reporting an illustrative 21.2% context reduction with a tab scorer that runs in under 5 ms for 30 or more tabs. A separate security tooling post applies the same visibility idea to agent behavior after the fact. Its coverage viewer reconstructs which files and lines an agent actually read, and shows large differences across models and reasoning budgets, such as about 8.3k to 17.7k median uniquely covered lines for GPT-5.4 settings versus about 30k to 32k for Opus 4.6 on an OpenSSH audit task.

Security-oriented code analysis papers are putting pressure on a common assumption: more context will help. In a 509-case ReposVul study across C, C++, and Python, code-only prompts beat caller-augmented and callee-augmented prompts for every tested model. Gemini 3 Flash reached 0.9853 accuracy and 0.9926 F1 in the code-only setting, while extra interprocedural context often cut accuracy and nearly doubled token cost. The practical response in the corpus is to improve targeting before generation. GALA tackles multimodal bug repair by aligning screenshot structure with repository files and then with functions, so the model edits code that matches the visual evidence. The excerpt supports the localization claim and lists prior SWE-bench Multimodal baselines, but it does not include GALA's own final resolved rate, so the strongest grounded takeaway is about better bug localization, not a verified repair percentage.