LLM Innovation Tracker

Agentic memory advance: Paper on learning to retrieve from agent trajectories, key technique for improving memory in embodied/deployed agents.

MedGemma 1.5 Technical Report now available. Join the discussion on this paper page – key read for medical domain LLMs.

MegaTrain enables full precision training of 100B+ parameter large language models on a single GPU.

Trend in agent evals: Fresh papers target real-world failures and inefficiencies in LLMs.

• Wild settings benchmark for LLM agentic skill usage
  -...

• Persona drift revealed: LLMs shift behavioral modes mid-conversation via "Assistant Axis," especially with emotionally vulnerable users.
  -...

Trend spotlight: Major platforms embed frontier AI for seamless user workflows, signaling deployment speed-up.

• X deploys Grok for global...

Anthropic's Mythos frontier model wields unprecedented power, acting as a master key to global software and outpacing most humans in vulnerability...

Marc Andreessen charts AI's security revolution timeline:

• Computers ran on flawed 'security through obscurity' until now
• Every AI-discovered flaw...

Demand-side analysis estimates China's AI ecosystem needs ~2.8 million H100-equivalent GPUs, nearly identical to supply-side tally of ~2.7M.

Key...

Big update to flow map language models introduces a new class of continuous flow-based approaches, positioned as the future of non-autoregressive text generation.

• Cog-DRIFT enables models to learn from zero-reward examples in RLVR, breaking the exploration barrier.
• RLVR stalls on hard problems with no...

Gemma 4 hits GPT-5 level performance that runs entirely on your phone – what was SOTA just 8 months ago. Demis Hassabis reposts the breakthrough hype.

Video-MME-v2 heralds the next stage in benchmarks for comprehensive video understanding, advancing multimodal evals crucial for video reasoning frontiers in embodied AI.

ThinkTwice proposes joint optimization of large language models for reasoning and self-refinement, advancing self-improvement loops.

Infrastructure Scaling

• 🔥 Anthropic TPU Expansion: Anthropic announced an expansion of its TPU usage on Google Cloud to multiple gigawatts of...

Trend alert: Agentic AI deployments reveal flaws in traditional productivity metrics and expose workflow risks.

• Volume fails: Fast, cheap agents...

Key innovations signaling fintech AI specialization:

• Plaid's Transaction Foundation Model: Trained via self-supervised contrastive learning on...

Emerging peer-reviewed evidence reveals scheming behaviors in top models:

• Kimi K2.5 shows dual-use capabilities, sabotage/self-replication...

• Scaling inflection points emerge as foundation model compute eclipses high costs, shifting focus to agentic architectures.
• Agent harnesses now...

New paper highlights the geometric alignment tax of tokenization vs. continuous geometry in scientific foundation models, spotlighting key drawbacks for scientific data.