Coding progress is coming from stricter intermediate checks

Repo-level coding work is getting stricter about what counts as understanding. LogicLoc targets cases where the prompt gives no file names or identifier hints, then converts the request into Datalog over extracted program facts. The paper’s own example asks for functions with more than 15 parameters and returns two exact matches in Astropy. REA-Coder makes the same point at generation time. It checks whether the model understood the requirement before and after writing code, then rewrites the task when the understanding is off. Across 20 model-benchmark settings, it beats eight baselines, with especially large gains on CodeContests and CodeContests-raw.

Control is becoming an explicit layer in coding agents. Zoro turns project rule files into active checks tied to plan steps. An agent has to submit proof that a rule was followed, and for testable rules it must also provide unit tests before it can continue. The reported result is a 57% increase in rule following over standard vibe coding across 36 sessions. OpenAI’s Codex deployment points to the same operational pattern in a larger setting: agents act on a connected store of lineage, ownership, documentation, dashboards, permissions, and production code, then expose their assumptions and citations when they investigate incidents or prepare fixes.

Retrieval for code is widening beyond text. CodeMMR places natural language, code, and images in one embedding space, then tests retrieval across web pages, charts, SVG, diagrams, and UML. On MMCoIR, the 2B model reaches 68.0 nDCG@10 and 65.4 Hit@1, ahead of VLM2Vec-v2 and GME baselines. The per-dataset spread matters: Web2Code, Chart2Code, DATIKZv3, and PlantUML are very strong, while MMSVG remains hard. That makes the current signal practical, not universal. Visual grounding is becoming useful for code search and code RAG, but the gain depends on the artifact type.

Verification work is putting more effort into checking the specification before spending effort on proof. LeetProof uses one stack inside Lean for property-based testing, automated verification conditions, and interactive proving. Its strongest grounded claim in the available excerpt is 97.4% semantic accuracy on VERINA for specification inference, plus evidence that about 10% of reference specifications in VERINA and CLEVER are defective. That matters because a weak or wrong spec can make later proof effort meaningless. The result fits the day’s broader emphasis on tighter checks before expensive search or repair.