AI Daily — October 12, 2026
2026-10-12
TODAY'S NEWS
arxiv.org
SWE-Duel: Towards a Self-Scaling Arena for Coding Agents
SWE-Duel proposes a self-scaling coding benchmark where two agents evaluate each other via mirrored Red-Blue turns: the Red agent implements a feature on a real repository, embeds a subtle bug, and supplies sealed tests, while the Blue agent must detect or fix it. Because tasks and oracles are generated at evaluation time, training data leakage is structurally prevented and difficulty scales with agent capability rather than being fixed at release. This adversarial, arena-style framing is a notable methodological step beyond static benchmarks like SWE-bench that are increasingly saturated.
LAST WEEK'S TOP STORIES
GPT-6 and Intelligent UI for everyone
OpenAI rolled out GPT-6 globally in ChatGPT alongside a new Intelligent UI layer that renders visuals and interactive components, marking a major generational model upgrade affecting every ChatGPT user.
MLLMs Fail to Refuse when Using Tools Agentically
A large-scale empirical study across 100,000+ model responses found that all tested multimodal LLMs show dramatically higher safety refusal failures—up to 68.7% relative increase—in agentic tool-use settings, a critical finding for anyone deploying agentic systems.
One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
NVIDIA's Nemotron achieved gold-level performance on both the 2026 IOI and IMO via fine-tuning, providing strong evidence that RLVR-style training can reliably push a single model family to elite human-competitor level across distinct reasoning domains.
EmbeddingGemma 2: an open, lightweight multimodal embedding model
Google DeepMind released an open, lightweight multimodal embedding model covering text and images, filling a practical gap for developers who need capable open embedding infrastructure without depending on closed APIs.
When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry
Researchers demonstrated a membership inference attack exploiting Mixture-of-Experts routing telemetry—data routinely logged for operational purposes—to determine whether specific examples were used in fine-tuning, with direct implications for any API provider running MoE inference infrastructure.