Adaptive models cut computation as robotics tests expose practical limits

LEV, developed by Lehigh University with the University of British Columbia and the Vector Institute, adjusts model depth to question difficulty without switching checkpoints. Its stopping mechanism predicts agreement with the model’s deepest answer using unlabelled requests. On-demand depth saves 15% to 24% of computation while losing under one point against the deepest readout. Mixed-depth batching could reduce real-world latency savings, and agreement-based stopping cannot recover answers that become incorrect at greater depth.

A SenseNova-led study tested GPT-6-Astra as a robot controller across 12 benchmarks. Strong simulation results depended on task demonstrations and extended execution budgets. Physical tests exposed weaknesses in precision, navigation and response time: plug insertion succeeded 0/11 times. Simulators that pause during deliberation can conceal delays that matter on hardware. Reviewing proposed actions in chunks improved results over open-loop control.

Anthropic and Theorem used Claude to prove that 239 fuzz entry points in 123 OSS-Fuzz libraries never crash for any input under 64 MB, uncovering bugs missed by fuzzing. The proofs assume single-threaded static RISC-V binaries on Linux with musl, limiting their applicability to common deployments. OpenAI also launched Patch the Planet to coordinate fixes as agents discover critical open-source vulnerabilities.

Mistral introduced speech models and a LiveKit-based voice agent SDK integrating transcription, tool calls and speech output. OpenAI announced frontier models and Codex workflows for Google Cloud’s Model Garden, but left model availability, pricing, regions and data-residency terms unspecified.