AI Research Digest

Frontier Model Safety Evaluations

• 🔥 Microsoft, Google, xAI Agree to Pre-Release Reviews: Microsoft, Google, and xAI have agreed to provide the...

Orbit-space particle flow matching debuts in generative modeling. Join the discussion on this paper page.

CommitSuite debuts as a comprehensive benchmark with 63,533 CCS-compliant commits from 243 open-source repositories across seven programming domains – vital for advancing AI code generation and commit quality evaluation.

• Google DeepMind, Microsoft, xAI join OpenAI/Anthropic in granting CAISI early access to unreleased models for security assessments, with safeguards...

ReClaim foundation model consistently outperforms strong baselines across heterogeneous evaluation regimes, including in-domain (Claims) and out-of-domain tasks—unlocking real-world evidence from claims.

Foundation models, particularly large language models, are increasingly integrated into agent architectures for industrial tasks such as decision-making—signaling real scaling of applied AI agents in automation.

Key open advancements from Ai2:

• MolmoAct 2: Foundation model for real-world robots
• Largest dataset released: MolmoAct 2-Bimanual YAM for...

Soccer-GMR introduces a new benchmark for generalized moment retrieval (GMR), built on challenging soccer videos to reflect general scenarios. Pushes CV toward robust multimodal retrieval in dynamic sports.

Generalized distributional alignment games tackle biases in scalable policy evaluation, extending GRPO—where models use a small, fixed group size K—to enable unbiased answer-level optimization.

Current AI agent benchmarks focus on narrow, low-friction tasks in controlled or synthetic environments, overlooking whether agents can actually finish the job. This gap demands robust, friction-filled evaluations for true capability assessment.

Competition regulators warn foundation model systems reinforce control over critical inputs like compute, cloud services, data, intensifying tensions between rapid AI scaling and distributional fairness.

T^2PO introduces uncertainty-guided exploration control to enable stable multi-turn agentic reinforcement learning. Key for long-horizon tasks.

Key insights on efficient-transformers library:

• Enhances decoding efficiency in Large Language Models (LLMs)
• Direct use in LVLMs adds substantial...

Tempus delivers a temporally scalable resource-invariant GEMM for AI/ML hardware arrays, powered by three specialized data communication mechanisms in the AIE-ML array that enable efficient computation and scaling.

David Rein's GPQA — a graduate-level, Google-proof QA benchmark — is now used by every major AI lab to track frontier model capabilities.

Core ML Advances

• 🔥 GCMs for Confidence Estimation: Paper accepted to ICML2026 introduces Generalized Correctness Models for confidence...

• Fleet-scale RL learns while deploying for generalist policies.
• RLWRLD advances five-finger dexterity via Robotics Foundation Models and real-world...

• Abysmal scores: Claude Opus 4.6 tops 9.39% on Deep Research; all frontier models in single digits.
• Core failures: Evidence aggregation collapses...

Key ICML2026 breakthrough: GCMs use historical predictions for superior confidence estimation.

• No privileged access: Models aren't better at...

• Traditional LLM scaling fixes one tokenizer while sweeping data/model size.
• New approach: Vary tokenizer compression rate (bytes/token), sweeping...