LLM coding research is getting more serious about control surfaces

Software engineering work on LLMs focused on narrowing the problem surface before generation. SWEzze learns a structured code-context compressor for issue resolution and reports about 6x compression, a 51.8% to 71.3% token cut, and a 5.0% to 9.2% gain on SWE-bench Verified. A separate repair study asks what happens after localization is already strong. Giving Agentless, KGCompass, and ExpeRepair oracle file and line spans lifts results, but native success still stays below 50%, and the best fixed added-context probe solves only six extra instances beyond the three-system Solved@10 union. The combined message is practical: better retrieval and smaller prompts help, but prompt packing and patch synthesis still cap repair quality.

Security attention centered on the boundary between code and model calls, plus the operational safety of agent tooling. The NL/PL paper treats LLM calls as a dataflow boundary and labels how much input survives into outputs such as SQL, JSON, or code. On its benchmark, the taxonomy-backed pipeline reaches F1 0.923 for taint propagation and cuts backward slices by a mean of 15% when placeholders do not propagate. On the tooling side, greywall proposes a deny-by-default sandbox for coding agents, with filesystem, network, and syscall controls instead of full user permissions. In applied review practice, a Claude Code workflow is useful when the human reviewer keeps narrow tasks and verifies the model's claims; the case study also records a concrete factual error that had to be corrected. The thread here is simple: model-enabled coding now needs program analysis and runtime containment around it.

Generated tests are back in the loop for code generation, but with less trust placed in any single test. BACE keeps populations of candidate programs and generated tests, then updates belief in both with a Bayesian noise model. On LiveCodeBench v6, it beats CodeSIM by 3.8% with GPT-5-Mini and 5.0% with Qwen2.5-Coder-7B. The idea fits the day's broader pattern: evaluation signals are treated as noisy evidence that must be filtered, weighted, or compressed before they can guide generation well.