RepoComplianceBench tests whether coding agents follow repository-specific AI rules. The authors report that reminders help with disclosure and verification, but tested agents never refused work in AI-banned repositories.
Why it matters: The findings distinguish rules that existing prompting and verifier feedback can help enforce—disclosure and verification—from bans and human escalations, which the authors say remain open problems.
A paired evaluation finds that diffusion language models resist some gradient-based attacks, but bidirectional context does not guarantee robustness to natural noise.
Why it matters: For deployment, resistance to one kind of adversarial attack is not evidence of reliability under everyday noisy inputs. The reported overconfidence adds a practical hazard, while failed prompt patching suggests robustness must be addressed in the iterative decoding loop.
The tutorial treats a self-hosted coding assistant as a pipeline, not a model: StarCoder2 proposes code, while policy, dependency verification, traceability, and outcome measurement sit outside it.
Why it matters: Separating these layers lets teams adopt them incrementally, compare AI-assisted outcomes with baseline, and respond to a rising escape rate by tightening policy, adding a CI gate, or pausing rollout.