AI Frontiers Digest

Agent Scaling Laws and Benchmarks

• 🔥 ByteDance EdgeBench: ByteDance Seed released EdgeBench with 134 real-world tasks (51 open) that track AI...

Diffusion language models are moving medical foundation models beyond autoregressive limits, matching AR performance on VQA benchmarks while...

ByteDance's EdgeBench reveals agents follow a log-sigmoid scaling law (R²=0.998) over 12-72 hour tasks, offering a post-deployment alternative as...

Two distinct strategies advance text-to-image generation without overlap.

• Training-free acceleration: MrFlow stages low-to-high resolution sampling,...

DuoMem's dual-space distillation transfers procedural skills from large teachers to compact models via teacher-generated memories and lightweight LoRA...

New agent benchmarks reveal a shift toward realistic, long-horizon evaluation across separate capability axes.

• EvoPolicyGym measures iterative...

WorldDirector achieves persistent dynamic object memory by decoupling LLM-orchestrated 3D motion trajectories from visual rendering, preserving exact entity identities even after long absences and supporting unrestricted viewpoint control.

This meta-article curates nine recent papers across generative models, agents, and systems.

• Program-as-Weights compiles natural-language specs into...

HOLA pairs a compressive delta-rule state with a small exact KV cache that stores only high-residual tokens, restoring needle recall without...

Anthropic is in early talks with Samsung to build its own AI chips, aiming to reduce dependence on external suppliers like Google and Amazon amid...

AI agents are moving beyond coding assistance toward full automation of empirical workflows.

• Article 1 shows agents handling data analysis and IPO...

Test-time compute budgets can dramatically alter AI agent evaluation results, yet most benchmarks reduce performance to a single score that conceals...

Vision-Language Model Advances

• 🔥 From Seeing to Thinking: Decoupling perception and reasoning in VLMs via staged training with specialized data...

Yann LeCun spotlighted the v2 JEPA-WM paper's acceptance to TMLR, complete with reproducibility certification. This milestone validates the world model's approach and strengthens its standing in self-supervised learning research.

Two recent papers target different stages of the text-to-image pipeline to cut compute while boosting quality.

• Lens delivers competitive results...

Two recent papers tackle core VLM bottlenecks through complementary strategies.

• SWIM aligns vision-language representations for fine-grained video...

DeepMind's AlphaProof Nexus uses evolutionary algorithms and Lean to autonomously generate formal proofs, solving nine open Erdős problems—including...

Two developments signal maturing support for AI agents in research:

• Paradigma tackles output chaos in auto-research by treating DAGs as the core...

A new theoretical framework models LLMs as noisy channels per the Shannon-Hartley theorem, mapping parameters to bandwidth and tokens to signal power....

Standard scalar RL post-training produces low-entropy responses that hinder inference-time search. VPO replaces the GRPO estimator with vector-valued...