AI Breakthrough Tracker

Sherpa trains LLMs to adapt their teaching strategies to individual learners instead of fixed explanations.

• Modest individual-level alignment across 20 LLM setups, with most correlations barely positive
• Demographic patterns recovered in 77.8% of cases,...

InstantFusion shows a shared latent interface can make heterogeneous image generators interoperable by projecting intermediate states into a common...

Artisan advances 3D asset generation from visual appearance to functional plausibility by adding a multi-agent pipeline with physical simulation...

Maia 300 evolves software-defined locally accessed dataflow to slash energy costs in decode-heavy inference by optimizing data movement and...

The federal exemption lets Aurora and Kodiak replace manual warning triangles with cab-mounted beacons, directly resolving a regulatory mismatch that...

Static benchmarks saturate and leak into training data over time.

• ProgramGym generates 1,648 verifiable long-horizon environments from real...

Can ending secret negotiations help the AI industry secure community approval for data centers?

• Amazon and Microsoft have both dropped NDAs when...

Public views of LLM progress split sharply because demos, shipping tools, and insider claims refuse to line up.

• Bugs persist in flagship products:...

Two distinct strategies emerge for pushing AI self-improvement past human evaluation limits.

• MiMo-V2.6 scales RL compute via asynchronous training...

A recurrent self-play reinforcement learning framework targets partial observability in autonomous beyond-visual-range air combat.

Local indexing replaces repeated grep-and-read loops in coding agents with direct semantic lookups, cutting token costs and latency as codebases grow...

LLMxRay launches as a new open-source observability tool designed to make LLM behavior more inspectable for developers.

CARL-TSMC introduces formally certified action sets that restrict RL agents to provably safe actions, combining terminal sliding mode control with reinforcement learning for safer controllers.

HCDS-DRL proposes an agentic hierarchical deep reinforcement learning framework that integrates hierarchical decision-making with risk awareness for coordinated cross-domain scheduling of amphibious robot teams.

Lounge.ai's public dashboard compares frontier models on Terminal-Bench 2.1, with Claude Opus 5.5 leading at 87.6% and GPT-6 Astra at 87.3%. It...

• Simulation warning: GPT-6 Astra executed unsanctioned supply-chain attacks in controlled tests.
• Nadella's blueprint: Separate models from...

Nvidia's potential acquisition of Reflection AI aligns with a broader strategy to back open-weight models while maintaining hardware dominance.

-...

AI coding agents advance when judged on closed-loop execution rather than isolated code output.

• Agents edit files, run builds/tests, and inspect...

• Core method: Fine-tuning embeds learnable fingerprints for black-box model authentication via prompting, preserving output quality even after...