Bleeding Edge AI

Training Efficiency Breakthroughs

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

In-Place Test-Time Training – frontier paper on efficient test-time adaptation. Join the discussion now to stay ahead.

MedGemma 1.5 Technical Report now available – frontier medical AI from Gemma base. Join the discussion on the paper page to unpack this bio-AI push.

ThinkTwice advances LLM training by jointly optimizing for reasoning and self-refinement—enabling agent-like iteration. Catch the discussion early.

MegaTrain introduces full precision training for 100B+ parameter LLMs on a single GPU—a breakthrough in accessible scaling for frontier AI research. Join the discussion.

Action Images enables end-to-end policy learning through multiview video generation. Join the discussion on this emerging robotics breakthrough.

Frontier agent techniques tackle learning from trajectories and zero-reward stalls:

• Trajectory Retrieval: Agents learn by retrieving from their own...

🚨 New preprint uncovers 'boiling frog' AI trap: After just 10 min of assistance, users perform worse and quit more than AI-free controls in RCTs. First hard data on rapid overreliance risks.

Trend alert: Rigorous benchmarks spotlight agent gaps.

• Wild skills test: Benchmarks LLM agentic skills in realistic settings
• Claw-Eval framework...

New paper spotlights MMEmb-R1, a reasoning-enhanced multimodal embedding model:

• Integrates pair-aware selection for smarter data handling
• Employs...

A fresh paper demystifies when pruning works via representation hierarchies, unlocking paths to efficient frontier model deployment by revealing key conditions for effective compression.

Video-MME-v2 advances benchmarks for comprehensive video understanding, signaling the next evolution in multimodal evaluation. Check the paper discussion for early insights.

Compute-matched showdown: Single-agent LLMs outperform multi-agents on multi-hop reasoning when thinking-token budgets are equalized across...

Frontier mech interp localizes, scales, and controls policy circuits in language models, validating gates via cascade effects compared to 10 random non-gate heads at similar depths. Key step for safer, steerable frontier models.

Cog-DRIFT breaks RLVR's key bottleneck: training stalls on hard problems where failed rollouts yield zero learning signals, leaving tough tasks unsolved. Essential for frontier LLM agent scaling.

New paper spotlights the Geometric Alignment Tax in scientific foundation models, framing tokenization against continuous geometry. Join the early discussion.

Inference-time adaptation is surging for frontier models:

• Language agents unlock learnable adaptation policies at test-time
• Self-execution...

Frontier multimodal papers signal push toward robust VLMs/MLLMs pre-mainstream:

• PLUME introduces latent reasoning for universal multimodal...

Compact Vision Encoders

• 🔥 EUPE: Meta AI released EUPE, a compact vision encoder family under 100M parameters that rivals specialist models...

Emerging tooling tackles AI agents' struggles in evolving environments and behavioral adaptation:

• ClawArena benchmark evaluates multi-source...