Bleeding Edge AI

Four papers reveal converging advances tackling LLM training from multiple angles.

• Implicit Curriculum shows skill acquisition follows a stable,...

May 22 trending papers spotlight frontier work in agents, long-context, and multimodal models.

• π-Bench introduces evaluation for proactive personal...

UC Berkeley's MOSAIC microscope generates petabytes of 5D live-cell data to train an LVLM that acts as a sherpa for interpreting complex biological dynamics—essentially a ChatGPT for biology.

Current LLM agents score 87-95% on control decisions when given explicit options, yet collapse to 54-62% in open-ended scenarios. This gap shows why outcome-only leaderboards fail to capture safety, recovery, and reasoning gaps in autonomous agents.

AGIBOT's BFM-2 delivers a two-stage locomotion foundation model that enables autonomous, stable motion interpolation and closed-loop task execution from any state. This marks a concrete step toward fluid, adaptive robot control in embodied AI.

Attention and Optimization Advances

• 🔥 Gated DeltaNet-2: Decouples erase and write operations in linear attention mechanisms.
  -...

• CODA rewrites transformer blocks as GEMM-epilogue programs targeting kernel-level hardware utilization.
• Multi-Stream LLMs parallelizes prompts,...

DelTA introduces discriminative credit assignment at the token level for reinforcement learning from verifiable rewards, directly tackling a core challenge in fine-tuning LLMs on math and code tasks.

Four parallel advances are accelerating truly autonomous agents.

• Agentic interpretability now targets models readable by other agents, not humans,...

Different optimizers induce distinct spectral scaling laws in the FFN representation geometry of language models, even under identical architectures....

Two distinct routes to cheaper attention emerge from recent work:

• Rapid conversion: Full attention models can be transferred to sparse variants in...

Chain-of-Thought decomposition reduces error according to a power law in the number of classes—a general result that applies to any learning or inference mechanism.

Reasoning Self-Play Advances

• 🔥 PopuLoRA: Introduces co-evolving LLM populations for reasoning self-play, presented as an arXiv paper with 48...

New work shows scaling laws deliver predictable gains but efficient pretraining methods now push performance further without proportional compute...

Two fresh approaches tackle general agent pretraining from complementary angles.

• DiLA introduces a dual-pathway architecture that disentangles...

Three complementary angles reveal how to assemble reliable autonomous research agents:

• Regulatory frameworks introduce a three-tier hierarchy that...

Lightweight post-training methods are gaining traction for scaling LLM reasoning without heavy compute.

• Minimal RLVR needs only rank-1 updates for...

Recent work reveals rapid progress in audio reasoning via refined encoders and tokenization strategies that better align with LLM backbones.

• Stable...

Multimodal evaluators now score image-to-text outputs by feeding the source image and query directly into an MLLM judge, bypassing text-only intermediaries for more accurate large-scale assessment.

Velox learns unified native 4D geometry and appearance representations, powering video-to-4D, 3D tracking, cloth simulation and related tasks in this CVPR 2026 paper.