AI Innovation Radar

UB-SMoE resolves client heterogeneity-induced load imbalance in federated fine-tuning via Dynamic Modulated Routing for expert rebalancing and...

• 100M-param LBM pre-trains self-supervised on 100k+ raw EEGs via masked spectral prediction to learn brain signal patterns without labels
• Achieves...

Traditional benchmarks measure model smarts on exams, but GAIA checks if agents can actually complete work.

• Core gap exposed: Humans hit 92% on...

AdaJEPA turns frozen world models into adaptive ones by closing the loop between planning and learning. After each MPC step, the model executes its...

GPT-5.6 Sol scored 88.8% on TerminalBench 2.1, nearly ten points ahead of Claude Opus 4.8, with the Ultra variant reaching 91.9% via parallel...

Method Advances

• 🔥 Program-as-Weights (PAW): Researchers from University of Waterloo, Cornell, and Harvard introduced PAW, where a...

The article investigates whether LLMs reason equally well across 43 languages.

PAW reframes LLMs as compilers that generate reusable 23MB adapters instead of answering queries repeatedly.

• 600M interpreter with PAW adapter...

OpenAI previewed its GPT-5.6 family of three vision-language models (Sol, Terra, Luna) with tiered pricing and performance, currently restricted to...

Large-scale evaluations reveal simpler ML often matches expensive tabular foundation models for routine clinical predictions.

• TabPFN beat classic...

• Meta's upcoming Muse Spark targets GPT-5.5 parity on SWE-Bench Pro coding tasks while boosting agent performance
• OpenAI's GPT-5.6 Sol claims SOTA...

A developer reports switching completely to open models, using GLM-5.2 daily in Claude Code via Hugging Face Inference Providers and hf-claude. Open models are becoming easier to plug directly into real developer workflows.

New Agent Evaluation Benchmarks

• 🔥 PACE: Introduces atomic proxy evaluations for agentic capabilities with 4% MAE as stated in the change...

Two recent advances highlight efficient inference without heavy retraining or custom hardware:

• FPGA ViT attention: BRAM-free 16-segment PWL...

High exam scores hide critical failures in clinical LLMs. Models hitting 92% on licensing tests plummet to 44.8% on real EHR benchmarks like BRIDGE,...

New benchmarks target real agent weaknesses instead of final scores.

• EvoPolicyGym tests iterative policy edits in RL environments, showing GPT-5.5...

Diffusion language models are shifting from experimental releases to practical tools, showing clear speed and flexibility edges over autoregressive...

Meta's Watermelon reportedly matches GPT-5.5 on undisclosed benchmarks, marking its next frontier push after Muse Spark while Zuckerberg admits slower-than-expected AI progress.

WorldDirector decouples semantic motion from pixel rendering via LLM-orchestrated 3D trajectories, delivering strict physical consistency and...

LLM Table Reasoning and Post-Training

• 🔥 When LLMs Read Tables Carelessly: Presents systematic evaluation of data referencing errors in LLMs and...